Investigating the Relationship of As-Constructed Asphalt Pavement Air Voids to Pavement Performance (2021)

Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

38 Analysis Dataset Count Type New Pavements Rehab. Pavements Total Perf. Curves 215 167 382 Ride Observations 1,765 1,634 3,399 Test Sections 199 170 369 Perf. Curves 199 172 371 Analysis The work plan for this 2018 study recognized that the objective could be difficult to achieve because the target variable, as-constructed AV, was not expected to be a dominant variable in pavement performance. The research team elected to apply three different analysis methods to the LTPP dataset with the understanding that each method had strengths and weaknesses. A different researcher led each analysis and independently incorporated the LTPP dataset variables using a unique set of variable abbreviations. To maintain the independence of each analysis, this report did not consolidate the abbreviations. Table 2-4 provides a comparison of the abbreviations for the numerous variables in Analysis Method 2 and 3. More details for each variable are provided in the discussion of each analysis. Table 2-4. Comparison of abbreviations used in Analysis Methods 2 and 3 Variable Analysis Method 2 Analysis Method 3 Air voids, surface layer AV_s AVt Air voids, all asphalt layers, weighted average AV_m AVw Layer thickness, asphalt overlay OvTh AC H OL Layer thickness, asphalt layer Th AC H AC Layer thickness, base and subbase Th BSB H base Surface asphalt layer gradation, % passing No.4 Sieve Gr 4 P no4 Surface asphalt layer gradation, % passing No.200 Sieve Gr 200 P no200 Surface asphalt layer gradation curve shape factor, k Gr k Surface asphalt layer gradation curve shape factor, lambda Gr lambda Surface asphalt layer binder content BC s Pbt All asphalt layers binder content BC m Pbw Asphalt layer stiffness, E* for rutting E* RUT E 130,10 Asphalt layer stiffness, E* for fatigue and roughness/ride E* FtC, E* IRI E 70, 10 Asphalt layer stiffness, E* for thermal cracking E* ThC E 14, 0.5 Base and subbase layer stiffness Mr BSB E base Subgrade stiffness Mr SG E sg Total annual traffic ESAL TR Cumulative traffic CumESAL Cumulative traffic, log scale lCumESAL Annual average precipitation Prec P

39 Annual average temperature Temp T Freezing index FI FI Pavement performance, time AGE t Pavement performance, rutting RUT Pavement performance, fatigue cracking FTC Pavement performance, thermal cracking TRC Pavement performance, roughness or ride IRI Analysis Method 1: Common Data Subgroups Analysis Method 1 subdivided the study LTPP dataset into smaller subgroups of LTPP sections with common characteristics such that each group could be examined for the influence of as-constructed AV under the premise that pavements in that group would perform in a similar manner and as-constructed AV would be a distinguishing factor between sections. This analysis involved a number of steps to process the LTPP dataset, assemble the common subgroups of data, and examine the influence of as-constructed AV within each subgroup. The major steps of this analysis were: 1. Create performance curves for each LTPP section; 2. Establish a single value performance criteria for each performance type; 3. Establish subgroup parameter ranges; 4. Examine the subgroup categories for sufficient LTPP sections; 5. Assemble subgroup datasets; and 6. Examine each subgroup and establish the trend. Creating Performance Curves A performance curve was generated for each performance characteristic for each LTPP section identified in this study. As each curve was created, there were conditions related to insufficient data and questionable data that needed to be addressed. Not all LTPP sections had sufficient data to create a performance curve. This situation was often a result of the pavement failing early such that there was sufficient rutting data in the first six years but insufficient time to generate fatigue cracking. Another situation regarding insufficient data were pavements that did not have measured cracking over the monitoring period. This situation was particularly common with transverse cracking. Questionable data was noted in a few sections. The common issues identified were special surface mixtures (porous surface mix), very low as-constructed AV, and abnormally large deviations between measured performance values. LTPP sections with porous surface mix were removed from the analysis because their material properties (high AV) and the inability to observe cracking made them poor candidates for the study. Air void values are computed values that rely on good measurements of material volumetric properties. Although the LTPP database took steps to filter out errant data through a quality control process, a number of the LTPP sections pulled for this study had as-constructed air void values below 1% and were omitted from the study because these values are not realistic. Analysis Method-1 used the as-constructed AV values generated for the other two analysis methods. The third type of questionable data involved variation between measured performance data. While performance data regularly exhibits variation due to equipment and conditions, some cases noted a single value that was out-of-line with the trend over time. These values were removed without losing the entire section. It should be noted that some sections had performance data that varied over the entire time period. These sections were kept in the dataset.

40 Establish a Single Value Performance Criteria for Each Performance Type Once the performance curves were generated, the second step was to determine one or more values related to the performance curve that would distinguish between the levels of performance between LTPP sections. For rutting performance, the primary interest was rutting related to the stiffness of the surface asphalt layer. This performance characteristic is evident in the first two to four years of service under traffic. The research team used two types of performance criteria to quantify rutting: measured rutting after four, six, and eight years, and pavement performance time to reach 0.3 inches of rutting. For wheel path cracking, the primary interest was measured fatigue cracking, which is commonly observed on the pavement surface later in the pavement structure’s design life. After scanning the wheel path cracking performance curves, the researchers elected to quantify differences between LTPP sections as the percentage of cracking after 10 years of traffic. For transverse cracking, the performance is related to the ability of the asphalt mixture (specifically the binder properties in the surface layer) to relax when the pavement experiences tensile strain related to climate temperature changes. The potential for these transverse cracks increases as the surface binder stiffens with age. For this study, the selected performance criteria for thermal cracking was the increment of time when transverse cracking is first measured and the length of cracking three years after the transverse cracking appears. From these two values, the researchers computed a composite value named “thermal factor” during the analysis as an attempt to quantify transverse cracking as a single value. The thermal factor was the length of transverse cracking divided by the increment of time to crack initiation. When the distribution of the thermal factors was examined, it was determined that it did not provide a good distinguishing value and was dropped from further consideration. The last performance characteristic, ride, is a commonly measured and evaluated value, but is generally a lagging performance indicator that reflects other performance distress. For this study, ride was quantified as the measured ride value after 10 years of traffic and as the increment of time it took the ride value to increase by 25 from the post-construction initial measured ride value. This approach is similar to quantifying the wheel path cracking as quantity and time (rate of change). Establish Subgroup Parameter Ranges To establish common subgroups for Analysis Method 1, the research team examined the distribution of the climate, traffic, and pavement structure for the selected new construction and rehabilitation LTPP sections. The research team applied the four LTPP climate zones as an accepted norm for designating both temperature and precipitation. Unfortunately, the distribution of the LTPP sections between the LTPP climate zones placed 75% of the sections in wet climate zones as shown in Figure 2-13. The research team elected to adjust the climate zones using the 25-year average of annual temperature and precipitation. From the distribution of those annual values shown in Figures 2-14 and 2-15, an average annual temperature of 56oF and average annual total precipitation of 37 inches were applied to divide the sections into four zones with more uniform distribution shown in Figure 2-13. For traffic, the research team elected to use 1000 kESALs as the break-point between low and high traffic based on the distribution shown in Figure 2-16. This created a 58% to 42% split between the number of LTPP sections with high and low traffic. After further examination, it was noted that the traffic split within new construction and rehab did not reflect the combined proportion. The low:high traffic split was 73:27 percent for new construction, and the split was 41:59 for rehabilitation. Although the splits were not as even as planned, the 1000 kESAL five-year traffic value was determined to be a practical value for users to apply. For pavement structure, the team examined the distribution of the SN values for the LTPP sections shown in Figure 2-17 and selected two break-points, SN=4.4 and 6.4. The clear breaks in the distribution are 3.8 and 7.0, which divide the sections into 12:74:14

41 percent, but the team elected to narrow the center distribution to 53% to have more sections in the upper and lower subgroups. Figure 2-13. Distribution of LTPP sections by LTPP and study climate zones. Figure 2-14. Distribution of LTPP section by average annual temperature.

42 Figure 2-15. Distribution of LTPP section by average annual precipitation. Figure 2-16. Distribution of LTPP sections by traffic for new and rehab pavements.

43 Figure 2-17. Distribution of LTPP sections by pavement structure for new and rehab pavements. Examine the Subgroup Categories for Sufficient LTPP Sections Using the subgroup ranges established in Step 3, the LTPP sections were divided into their respective subgroups to examine how many sections were included in each subgroup. Table 2-5 is a summary of the LTPP section distribution divided into new construction and rehabilitation. The concept of Analysis Method 1 requires a sufficient number of LTPP sections in each subgroup to observe trends. Subgroups with less than six sections were considered insufficient to generate reliable trends. As expected, subgroups with impractical combinations of traffic and pavement thickness, such as high traffic and thin pavement structure, had few or no sections. 27 of the 48 subgroups moved to the next step in the analysis and the other 21 subgroups were deemed to have insufficient LTPP sections for further analysis. As noted in Step 3, the dominant range of pavement structure was SN=3.8 to 7.0 and the subgroup break-points were set at 4.4 and 6.4 knowing that 51% of the LTPP sections were bound between 4.4 and 6.4. Although it was possible to further narrow the thickness range for medium pavement structure to place more sections in the thin and thick category, this would further diminish the difference for thin and thick structures. The research team elected to move forward with 27 subgroups for analysis. Table 2-5. Distribution of LTPP sections into subgroups for new and rehab pavements. Climate Zone Traffic Pavement Structure New Rehab Climate Zone Traffic Pavement Structure New Rehab 1 High Thin 4 7 3 High Thin 0 0 Medium 0 25 Medium 12 28 Thick 12 4 Thick 0 14 Low Thin 28 6 Low Thin 10 8 Medium 19 6 Medium 0 14 Thick 8 2 Thick 4 0 2 High Thin 0 2 4 High Thin 3 1 Medium 3 9 Medium 14 14 Thick 14 2 Thick 0 4 Low Thin 12 0 Low Thin 12 11 Medium 15 11 Medium 30 14 Thick 11 2 Thick 12 2

44 Assemble Subgroup Datasets Using the criteria developed in the previous steps, a master dataset was created in a spreadsheet that was sorted into separate sets for new construction and rehabilitation. Each row of the spreadsheet represented one LTPP section and included: (1) section identification; (2) subgroup climate, traffic, and pavement structure; (3) as-constructed AV; and (4) identified performance values for rutting, wheel path cracking, transverse cracking, and ride. The rows were sorted by climate, traffic, and pavement structure. This organization of the data allowed the researchers to quickly locate the subgroups for each performance trend analysis. Examine Each Subgroup and Establish the Trend A separate X-Y scatter plot was created for each subgroup and performance characteristic with the performance value on the y-axis and the as-constructed AV on the x-axis. Each plot represented a subgroup of LTPP sections with similar climate, traffic, and pavement structure, but constructed to each section’s as- constructed AV. Each plot was examined for a trend in performance for the range of as-constructed AV. Trends were classified as meeting expectation (improved performance with AV between 4 to 8% and poorer performance above 8% voids and below 4% voids), no influence (changes in AV had no influence on performance), or contradicting expectation (reduced performance as AV decreased to 4%).The complete set of trends is provided in the User Guide. Qualifying the expectation of each set of LTPP sections in a subgroup was a subjective exercise. While sorting the LTPP sections into common subgroups provided some level of control for Analysis Method 1, the dataset does not represent a controlled experiment with the specific objective of evaluating the influence of as-constructed AV. Some subgroups have very strong trends, while others are marginal. None of the sections were removed based on observed difference in the performance value from the collective subgroup. Rutting performance was expected to improve with as-constructed AV between 4 to 8% and show poorer performance above 8% voids and below 4% voids. Examples of each type of rutting trend are shown in Figure 2-18. The performance trend for 9 of the 13 subgroups in dry climate and 7 of the 13 subgroups in wet climate met the expectation. Only 3 of the 26 subgroups trends contradicted expectation and all three were new construction thin pavements with low traffic.

45 Performance meeting expectation No influence of air voids Performance contradicting expectation Figure 2-18. Examples of influence of as-constructed air voids on rutting performance. Wheel path cracking (fatigue) performance data displayed more variability in the subgroups than the rutting data. There are fewer subgroups, only 21, because a number of LTPP sections did not have ten years of performance data. Examples of each type of fatigue cracking trend are shown in Figure 2-19. Of the 21 subgroups, 12 (57%) had both full cracking and no cracking performance at the same as-constructed AV. The trends for 5 of the 10 subgroups in dry climate and 8 of the 11 subgroups in the wet climate met the expectation. The performance trend of 4 of the 21 subgroups contradicted expectation.

46 Performance meeting expectation No influence of air voids Performance contradicting expectation Figure 2-19. Examples of the influence of as-constructed air voids on fatigue performance. Transverse cracking (thermal) performance data was unique compared to rutting and fatigue. Over half (114 of 209) of the new construction sections and over one-third (76 of the 197) of the rehab sections did not have observed transverse cracking over their period of performance monitoring. The monitoring period for an individual section generally ranged from 5 to over 20 years as shown in Figure 2-20. The research team concluded that after some identified years of service, the lack of transverse cracking could be considered actual performance, not insufficient data. This would avoid a significant loss of LTPP sections for the analysis.

47 Figure 2-20. Years of performance monitoring for LTPP sections with no transverse cracking. The two performance measures for transverse cracking are the year that transverse cracking is first measured (initial year) and the amount of total transverse cracking three years after the initial year (total cracking). The analysis examined the distribution of the initial year values for the LTPP sections with both a measured initial year and total cracking. As shown in Figure 2-21, the distribution for new construction sections displays a prominent peak at 12 years and decline to 20 years. The distribution for rehab sections show the majority of transverse cracks re-appear in the first six years and most sections have measured transverse cracking after 12 years. Using these distributions, the research team selected 12 years as the performance year for new pavements and 10 years for performance of rehab pavements. This reduced the number of sections classified as insufficient data from 114 to 71 for the new construction analysis and from 76 to 35 for the rehab analysis. Figure 2-21. Distribution of time to initial transverse crack. The reduction in LTPP sections due to insufficient data caused some subgroups to have fewer sections in the analysis and required two subgroups to be dropped. For example, the new construction subgroup for climate 1, low traffic and medium pavement structure changed from 19 to 3 available sections. This subgroup was left in the analysis but should have been deleted for having less than six sections. The next subgroup, climate 1, low traffic and thick pavement structure, was eliminated because all eight sections

48 were deemed as having insufficient data. The analysis of transverse (thermal) cracking was limited to only 24 subgroups. Both the year of initial crack and the amount of cracking three years after the initial crack were plotted for each subgroup of LTPP sections. The trends exhibited in each subgroup were often marginal or sometimes conflicting. Examples of transverse cracking trends are shown in Figure 2-22. The results for new construction resulted in only 5 of the 12 trends meeting expectation and 6 trends contradicting expectation. All 6 contradicting trends were in freeze climates where thermal cracking is a more critical distress. A trend meeting expectation should have a higher number of years before crack initiation and a lower amount of cracking as the as-constructed AV decrease. The results for rehab projects relate to reflective cracking from the underlying asphalt pavement and indicate that 6 of the 12 subgroups had trends that meet expectation and 6 trends contradicted expectation. These mixed results suggest that decreasing as-constructed AV does not improve thermal cracking performance. Performance meeting expectation Performance contradicting expectation Figure 2-22. Examples of influence of as-constructed air voids on thermal cracking performance. Pavement ride performance data displayed more variability in the subgroups than the rutting data. Examples of each type of ride performance trend are shown in Figure 2-23. The trends for 7 of the 13 subgroups in new construction but only 3 of the 12 subgroups in the rehab category met the expectation. The performance trend of 10 of the 25 total subgroups contradicted expectation. Since ride is a lagging performance indicator, it is likely that these results mostly reflect other performance trends. Table 2-6 shows how the ride performance trends compare to the wheel path cracking trends. Of the 21 paired subgroups, 9 had similar performance trends as shown in the green blocks and 6 had opposite trends as shown in the yellow blocks.

49 Performance meeting expectation No influence of air voids Performance contradicting expectation Figure 2-23. Examples of the influence of as-constructed air voids on ride performance.

50 Table 2-6. Comparing ride and wheel path cracking trends. Climate Zone Constr. Perform. LOW traf-THIN pvmt LOW- MED LOW- THK HIGH- THIN HIGH- MED HIGH- THK 1 New ride E E E ND ND N WP crk C E ND ND ND E 1 Rehab ride ND N ND N C ND WP crk ND E ND N C ND 2 New ride E C E ND ND ND WP crk N E E ND ND ND 2 Rehab ride ND E ND ND N ND WP crk ND N ND ND ND ND 3 New ride C ND ND ND E ND WP crk ND ND ND ND E ND 3 Rehab ride E C ND ND N E WP crk ND C ND ND E E 4 New ride E C C ND C ND WP crk E E E ND E ND 4 Rehab ride C C ND ND C ND WP crk E C ND ND N ND legend Similar performance trend No performance trend Opposite performance trend E = meets expectation, N = no influence, C = contradicts expectation, ND = no data for trend A summary of Analysis Method 1 is presented in Tables 2-7 to 2-10 with each table providing a summary of a specific performance measure. Each cell is color coded to show subgroups that met expectation, had no influence, contradicted expectation, or had insufficient data to establish a trend. Table 2-7. Summary of as-constructed air voids impact on rutting performance. RUTTING PERFORMANCE NEW CONSTRUCTION LOW TRAFFIC HIGH TRAFFIC DRY-FREEZE WET-FREEZE DRY-FREEZE WET-FREEZE THIN PAVEMENT MEDIUM PAVEMENT THICK PAVEMENT DRY-NO FRZ WET-NO FRZ DRY-NO FRZ WET-NO FRZ THIN PAVEMENT MEDIUM PAVEMENT THICK PAVEMENT REHAB CONSTRUCTION LOW TRAFFIC HIGH TRAFFIC DRY-FREEZE WET-FREEZE DRY-FREEZE WET-FREEZE THIN PAVEMENT MEDIUM PAVEMENT THICK PAVEMENT DRY-NO FRZ WET-NO FRZ DRY-NO FRZ WET-NO FRZ THIN PAVEMENT MEDIUM PAVEMENT THICK PAVEMENT legend Meets expectation No influence Contradicts expectation Insufficient data

51 Table 2-8. Summary of as-constructed air voids impact on fatigue cracking performance. FATIGUE CRACKING PERFORMANCE NEW CONSTRUCTION LOW TRAFFIC HIGH TRAFFIC DRY-FREEZE WET-FREEZE DRY-FREEZE WET-FREEZE THIN PAVEMENT MEDIUM PAVEMENT THICK PAVEMENT DRY-NO FRZ WET-NO FRZ DRY-NO FRZ WET-NO FRZ THIN PAVEMENT MEDIUM PAVEMENT THICK PAVEMENT REHAB CONSTRUCTION LOW TRAFFIC HIGH TRAFFIC DRY-FREEZE WET-FREEZE DRY-FREEZE WET-FREEZE THIN PAVEMENT MEDIUM PAVEMENT THICK PAVEMENT DRY-NO FRZ WET-NO FRZ DRY-NO FRZ WET-NO FRZ THIN PAVEMENT MEDIUM PAVEMENT THICK PAVEMENT legend Meets expectation No influence Contradicts expectation Insufficient data Table 2-9. Summary of as-constructed air voids impact on thermal cracking performance. THERMAL CRACKING PERFORMANCE NEW CONSTRUCTION LOW TRAFFIC HIGH TRAFFIC DRY-FREEZE WET-FREEZE DRY-FREEZE WET-FREEZE THIN PAVEMENT MEDIUM PAVEMENT THICK PAVEMENT DRY-NO FRZ WET-NO FRZ DRY-NO FRZ WET-NO FRZ THIN PAVEMENT MEDIUM PAVEMENT THICK PAVEMENT REHAB CONSTRUCTION LOW TRAFFIC HIGH TRAFFIC DRY-FREEZE WET-FREEZE DRY-FREEZE WET-FREEZE THIN PAVEMENT MEDIUM PAVEMENT THICK PAVEMENT DRY-NO FRZ WET-NO FRZ DRY-NO FRZ WET-NO FRZ THIN PAVEMENT MEDIUM PAVEMENT THICK PAVEMENT legend Meets expectation No influence Contradicts expectation Insufficient data

52 Table 2-10. Summary of as-constructed air voids impact on ride performance. RIDE PERFORMANCE NEW CONSTRUCTION LOW TRAFFIC HIGH TRAFFIC DRY-FREEZE WET-FREEZE DRY-FREEZE WET-FREEZE THIN PAVEMENT MEDIUM PAVEMENT THICK PAVEMENT DRY-NO FRZ WET-NO FRZ DRY-NO FRZ WET-NO FRZ THIN PAVEMENT MEDIUM PAVEMENT THICK PAVEMENT REHAB CONSTRUCTION LOW TRAFFIC HIGH TRAFFIC DRY-FREEZE WET-FREEZE DRY-FREEZE WET-FREEZE THIN PAVEMENT MEDIUM PAVEMENT THICK PAVEMENT DRY-NO FRZ WET-NO FRZ DRY-NO FRZ WET-NO FRZ THIN PAVEMENT MEDIUM PAVEMENT THICK PAVEMENT legend Meets expectation No influence Contradicts expectation Insufficient data Analysis Method 2: Regression Modeling This Chapter describes the method, results, and main findings from Analysis Method 2. The effect of as- constructed AV on pavement performance was assessed by quantifying and interpreting the coefficients of statistical models estimated for the different performance characteristic types: rutting, fatigue cracking, thermal cracking, and ride and for both new and rehabilitated pavements, separately. These models were developed considering various climate, traffic, pavement structure, and material properties related variables as shown in Table 2-11 to account for potential interactions on the effect of as-constructed AV on pavement performance. Table 2-11. List of variables used for Analysis Method 2 regression models. Variable Description Units AV_s as-constructed surface asphalt layer air voids %Gmm AV_m as-constructed mean asphalt air voids %Gmm ThAC total thickness of asphalt layers inches ThBSB total thickness of base and/or subbase layers inches OvThAC total thickness of asphalt overlay inches E*Rut surface asphalt dynamic modulus at 130°F and 10 Hz ksi E*FtC compound asphalt dynamic modulus at 70°F and 10 Hz ksi E*ThC surface asphalt dynamic modulus at 14°F and 0.5 Hz ksi E*IRI surface asphalt dynamic modulus at 70°F and 10 Hz ksi MrBSB compound resilient modulus of base and/or subbase layers ksi MrSG resilient modulus of subgrade ksi Gr#4 percent passing sieve #4 of aggregates in surface asphalt layer % Gr#200 percent passing sieve #200 of aggregates in surface asphalt layer % BCs surface asphalt layer binder content %mix wt

53 BCm mean asphalt binder content %mix wt Ageyr age of the pavement for year yr years Precyr total annual precipitation at year yr inches Tempyr annual mean temperature at year yr °F FIyr annual freezing index at year yr °F-days TRFyr annual total traffic at year yr kESALs CumESALyr cumulative traffic at year yr (since opening to traffic date for new pavements, or since assignment date for rehabilitated pavements) computed as 𝐶𝑢𝑚𝐸𝑆𝐴𝐿 = 𝑃𝑟𝑒𝑇𝑅𝐹 + ∑ (𝑇𝑅𝐹 ) kESALs lCumESALyr log-transformed (natural logarithm) cumulative traffic at year yr computed as lCumESALyr = logCumESALyr + 10 RUT Wheel path rutting inches FTC Fatigue cracking, reported as wheel path cracking for % of lane width % TRC Thermal cracking, reported as transverse cracking per lane-mile ft/ln-mi IRI Ride, reported as international roughness index in/mi The first section of this chapter describes aspects of the methodology used for developing the different pavement performance models as part of Analysis Method 2. Each of the remaining four sections describe the analysis of a performance characteristic, which includes a presentation of the characteristics of the analyses datasets, the specification of each model, and the result and main observations from the estimation of the models. Performance Model Development Methodology A total of eight performance regression models were developed, one for each combination of the four performance characteristic types (rutting, fatigue cracking, thermal cracking, and ride) and the two surface types (new and rehabilitated). Two additional models were developed to assess the effect of as-constructed AV on the increment of time until a transverse crack appears in the surface as part of the analysis of thermal cracking performance. All of these models were estimated using the four analysis datasets (one per performance type) produced from the LTPP dataset. The following list discusses different methodological steps applied when developing the statistical regression models to assess the effect of as-constructed AV on pavement performance. The five steps are in logical sequence, but the model development process involves an iterative back-and-forth effort. These steps were applied to each of the models that follow. 1. Model functional form: The functional form specified for each performance model was determined from evaluating a number of alternatives initially estimated by pooling data from both new and rehabilitated pavement sections. The alternative functional forms considered for each performance model were selected from the patterns observed in the performance data and reported in related studies. The functional form selected for each model was the one that balanced best model fit and lowest complexity, determined as the one with the lowest Akaike information criterion (AIC) value (Akaike, 1974). Among the well performing models, functional forms (either linear or non-linear) that can be linearized were preferred when specifying the model due to being easier to interpret and more computationally efficient for estimation. 2. Estimation technique: The performance model for each performance characteristic metric was estimated through quantile regression by estimating the conditional median. The estimation was performed using R “quantreg” package (Koenker, 2019). The method for estimation used by this package is a modified version of the Barrodale and Roberts algorithm described in Koenker and d’Orey (1987, 1994). This robust estimation technique was selected to reduce the effect of the outliers observed

54 in the data on the estimated model parameters without the loss of information resulting from removing data with outliers. As an example of the outliers present in the data, Figure 2-24 shows the boxplot of scaled variables for some of the variables considered for the ride models. The dots before and after the whiskers of each boxplot represent the outliers for each scaled variable. The red dots show the extreme outliers–i.e., greater than or less than 3 times the inter-quartile range (IQR) from the 1st and 3rd quartiles, and the black dots show the mild outliers–i.e., between 1.5 and 3 times the IQR from the 1st and 3rd quartiles. As shown in the plot, most variables present outliers, some of which include extreme outliers. The boxplots of scaled variables corresponding to the models for the other three performance characteristics show similar trends. Figure 2-24. Boxplots of scaled variables considered for ride performance models. 3. Pre-processing of variables: Two types of processing were conducted to the variables during development of models. The first consisted of proposing alternative transformations to the variables– such as logarithmic or power transformations–to either reduce their skewness, to increase the predictive power of the model, or both. As an example, the two boxplots at the bottom of Figure 2-24 show how the log transformed version of the cumulative ESALs variables, defined as lCumESAL=log(CumESAL+10), has a less skewed distribution, thus presenting less extreme values. Given the very different scales of magnitude among variables, the second type of pre-processing consisted of scaling all variables before estimating the model for better numerical stability and more direct comparison of effects. The scaling of each variable consisted of subtracting its mean and dividing by its standard deviation. 4. Selection of variables: The initial step in variable selection consisted of identifying combinations of variables that may cause issues in the estimation (such as pairs of variables presenting high correlation) or that may not be included together due to engineering reasons. The next step consisted of selecting the optimum combination of variables by running a stepwise algorithm. This algorithm increases the number of variables and interaction variables in each step and quantifies the model AIC. The algorithm stops when the addition of variables does not result in significant improvement in model fit–i.e., when the AIC stops increasing. This variable selection process was implemented using the R “stats” package. The combination of variables tried by the stepwise algorithm were restricted to those only including interaction variables involving AV or Age. Once the optimum set of variables was found, the final step

55 consisted of removing any variables that were not statistically significant–i.e., with a p-value lower than 0.10. 5. Interpretation of regression analysis: Each performance characteristic for both new and rehabilitated pavements identified for this study has a separate regression equation (eight models in total). The final results of each regression analysis are assembled in a table format for making observations relative to the influence of as-constructed AV and other variables in the regression equation (Devore, 2015). For each regression summary table that follows, the first column identifies the regression model intercept in the first row and divides the remaining rows into individual variables, the variables interacting with age, and the variables interacting with as-constructed AV. The second column lists the estimated coefficient value for each variable in the regression equation which, to some extent, can be compared because the values for each variable were scaled to minimize the influence of the different magnitudes of the input variable values. The third column presents the standard error associated with each variable’s estimated model coefficient and can be viewed as the amount of variation between the variable input values and the values generated by the estimated coefficients (a measure of the precision with which the coefficient was estimated). The fourth column presents the t value, which represents the relative effect (impact) of each variable on the model. The fifth column presents the p-value, which measures the statistical significance of the variable. In general, higher absolute values for the estimated coefficient and t value are associated with greater influence on the regressed performance value. Further, variables and interactions associated with time (age and ESALs) impact the progression of performance and the variables and interactions associated with material properties impact the magnitude of the performance. Rutting Performance Models This section describes the analysis datasets and reports summary statistics, estimation results, and main findings from the development of the rutting performance model for new and rehabilitated pavements. Rutting Performance Dataset. The plot in Figure 2-25 shows the median rutting over time of the available data for model development grouped by the as-constructed AV quintiles. Group Q1 contains test sections with 20% lowest AV_s values, while Q5 contains sections with the 20% highest AV_s values. The median rutting tended to increase with time and the observed rate of rutting was greater during the first few years, suggesting the use of a non-linear functional form for the model specification. The longer the age, the more unreliable the rutting trend (due to site conditions and possible unrecorded treatments applied to the surface), which explains the unexpected sudden change in rutting after 12 years of some curves. Taking this into consideration, only observations with Age ≤ 15 years were used for developing the rutting performance models. Figure 2-25. Median rutting versus age grouped by surface air voids quintiles.

56 The rutting performance data plot in Figure 2-25 provides an overview of general characteristics and trends in the input data subsets. However, the observed effects of AV_s grouping in the plot may be misleading due to confounding factors. A more accurate assessment of AV_s was obtained from interpreting the parameters in the regression model as they were estimated while accounting for several other significant factors. Rutting Performance Model Specification. A log-based math function, which creates early post- construction asphalt surface layer consolidation rutting, was used for developing the rutting performance models and is shown in Equation 9. The model parameters on AV_s are of special interest for the objective of this study. The parameters quantify the effect of as-constructed AV_s on rutting growth over time β%Gmms,Age and the effect on the magnitude of rutting β%Gmms and the interaction terms β%Gmms,j. 𝑅𝑢𝑡 = 𝛽 + ∑ (𝛽 𝑋 , ) + %𝐺𝑚𝑚 (𝛽% + ∑ (𝛽% , 𝑋 , )) + 𝑙𝑜𝑔(𝐴𝑔𝑒 ) (𝛽 + ∑ (𝛽 , 𝑋 , )) + 𝛽% , 𝑙𝑜𝑔(𝐴𝑔𝑒 ) %𝐺𝑚𝑚 (9) Where: 𝑅𝑢𝑡 = estimated rutting for year yr; Ageyr = age of pavement surface for year yr; %Gmms = as-constructed surface air voids; Xj,yr = model variable 𝑗 for year yr; and βj = model parameter on variable j. Summary Statistics of Rutting Dataset for New Pavements. Table 2-12 shows the summary statistics of the variables considered for developing the rutting performance model for new pavements. The minimum and maximum values presented in this table define the ranges under which the rutting model for new pavements was developed. For many of the variables, these minimum and maximum values represent extreme values well outside the norm. For example, the Th BSB maximum is 221.5 and the 90th percentile value is only 42.5. It is cautioned that inputting variable values into the regression model outside the 10th and 90th percentile values may lead to unreliable estimates.

57 Table 2-12. Summary statistics of data used for estimating rutting model for new pavements. Variable Units Mean Std Dev Min 10%ile 50%ile 90%ile Max RUT inches 0.21 0.13 0.04 0.08 0.20 0.35 0.91 AV_s %Gmm 7.11 2.93 1.64 3.45 6.95 10.82 21.82 AGE years 6.26 4.04 0.00 1.00 6.00 13.00 15.00 Th_AC in 5.64 1.57 1.40 4.00 5.20 7.40 13.80 Th_BSB in 20.03 17.15 0.00 7.70 15.90 42.50 221.50 E*_Rut ksi 75.79 41.41 42.61 45.58 64.22 135.28 327.77 Mr_BSB ksi 32.72 26.72 9.10 15.99 21.49 100.00 100.00 Mr_SG ksi 9.09 2.74 4.89 5.35 8.83 12.55 16.70 BC_s % 5.02 0.82 3.63 4.12 4.80 6.20 7.30 Gr_4 % 56.07 9.29 22.00 48.00 55.00 69.00 82.00 Gr_200 % 5.83 1.32 2.10 3.70 6.30 7.10 10.20 Gr_k - 5.27 1.76 2.50 3.63 5.10 7.18 15.33 Gr_lambda - 0.84 0.16 0.60 0.70 0.79 1.04 1.65 Temp °F 56.68 9.65 34.88 43.88 58.28 68.72 75.74 Prec in 31.72 17.71 3.76 9.11 32.30 54.64 117.99 FI °F-days 488.56 631.51 32.00 32.00 165.20 1,416.20 3,272.00 ESAL kESAL 262.76 267.72 0.00 2.19 183.85 578.26 2,553.54 CumESAL kESAL 1,520.47 1,845.82 0.00 12.03 819.13 3,948.95 15,338.44 Estimation of Rutting Model for New Pavements. Table 2-13 shows the estimated parameters of the final rutting performance model for new pavements. The median absolute error (MAE) of the model was 0.04 inches, and the approximate R2 was 0.31. Since the model was estimated using quantile regression to fit the median, R2 is not applicable and, therefore, the approximate R2 suggested by Koenker and Machado (1999) was applied. This approximate R2 is similar in concept to R2 but is not directly comparable. Figure 2-26 shows the comparison between the measured rutting pavement performance in the LTPP dataset and the predicted results from the regression model. The variables and interactions listed in Table 2-13 were the ones statistically significant–i.e., with a p- value lower than 0.10. Based on the statistical parameters for each variable, the AC surface layer properties (stiffness, gradation, and binder content) had the most influence on rutting performance along with climate and traffic. As-constructed AV was shown to have an effect as an independent variable. This effect was negative, which indicates that pavements with higher AV tended to have less rutting, keeping everything else fixed in the analysis dataset, which contradicts expectation. Since CumESAL is a function of the surface age and β%Gmms,CumESAL) was significant, the as-constructed AV had a statistically significant effect on the deterioration rate of rutting over time for new pavements. The parameters in Table 2-13 also show that the interactive effect of AV with AC surface layer thickness, gradation, and binder content had a statistically significant effect on the value of rutting, which may be a reflection of the influence of the independent variables. The mixed positive and negative signs on these six interactions imply that as-constructed AV did not influence performance. The variables in this model were scaled before regression model development (scaled back using the respective mean and standard deviation of the variable). The estimated parameters in Table 2-13 are shown in the units of the scaled variables–not the units of the listed variables. The model parameters in the original units are shown in the User Guide document.

58 Table 2-13. Estimated parameters of rutting model for new pavements. Variable Value Std. Error t value Pr(>|t|) (Intercept) 0.208 0.007 31.525 0.000 Th_AC -0.014 0.007 -2.181 0.029 E*_Rut -0.045 0.006 -7.794 0.000 Mr_BSB -0.005 0.002 -3.028 0.002 Gr_4 -0.043 0.006 -7.076 0.000 Gr_200 0.037 0.006 5.724 0.000 BC_s 0.052 0.004 11.939 0.000 Temp 0.096 0.004 26.604 0.000 Prec -0.040 0.004 -9.413 0.000 FI 0.058 0.004 12.974 0.000 cumESAL_log 0.060 0.005 13.106 0.000 logAge 0.015 0.002 9.039 0.000 AV_s -0.003 0.001 -3.746 0.000 logAge:Gr_200 0.004 0.001 4.267 0.000 logAge:Prec 0.011 0.001 9.538 0.000 logAge:FI 0.009 0.001 7.904 0.000 logAge:cumESAL_log 0.003 0.001 3.324 0.001 logAge:Mr_SG -0.007 0.001 -7.023 0.000 logAge:Th_AC -0.004 0.001 -3.807 0.000 logAge:Mr_BSB -0.002 0.001 -2.364 0.018 AV_s:Th_AC 0.003 0.001 2.900 0.004 AV_s:Gr_4 0.005 0.001 5.207 0.000 AV_s:Gr_200 -0.004 0.001 -3.973 0.000 AV_s:BC_s -0.005 0.001 -9.711 0.000 AV_s:Prec 0.004 0.001 6.301 0.000 AV_s:cumESAL_log -0.003 0.000 -6.206 0.000 Legend for color coding: Yellow Dominant variable in the model Green Air voids variable with positive value Grey Air voids variable with negative value Figure 2-26. Comparison of new pavement rutting performance between LTPP measurements and regression model predictions.

59 Summary Statistics of Rutting Dataset for Rehabilitated Pavements. Table 2-14 shows the summary statistics of the variables considered for developing the rutting performance models for rehabilitated pavements. The minimum and maximum values presented in this table define the ranges under which the rutting model for rehabilitated pavements was developed. For many of the variables, these minimum and maximum values represent extreme values well outside the norm. For example, the Th AC maximum is 21.9 and the 90th percentile value is only 13.4. It is cautioned that inputting variable values outside the 10th and 90th percentile values into the regression model may lead to unreliable estimates. Table 2-14. Summary statistics of data used for estimating rutting model for rehabilitated pavements. Variable Units Mean Std Dev Min 10%ile 50%ile 90%ile Max RUT inches 0.20 0.12 0.04 0.08 0.16 0.35 0.91 AV_s %Gmm 5.15 2.54 0.10 2.30 4.89 8.46 15.17 AGE years 6.83 4.23 0.00 1.00 7.00 13.00 15.00 OvTh_AC in 2.49 2.28 0.00 0.00 2.10 5.60 9.20 Th_AC in 8.87 2.88 2.50 5.40 8.30 13.40 21.90 Th_BSB in 15.86 9.22 0.00 7.30 13.60 27.00 64.00 E*_Rut ksi 108.63 69.85 16.91 45.64 87.66 159.18 384.13 Mr_BSB ksi 31.24 25.99 12.27 15.30 22.03 100.00 100.00 Mr_SG ksi 9.44 3.38 3.30 5.30 8.83 14.45 17.90 BC_s % 5.02 0.64 3.80 4.40 4.95 5.78 9.00 Gr_4 % 55.90 9.33 29.00 44.00 56.00 67.00 90.00 Gr_200 % 6.58 2.00 0.90 4.40 6.40 9.20 13.00 Gr_k - 4.98 1.70 2.18 3.17 4.96 6.48 15.28 Gr_lambda - 0.84 0.19 0.57 0.62 0.81 1.11 1.47 Temp °F 54.00 10.66 32.36 37.40 54.68 67.10 75.38 Prec in 37.21 16.86 3.52 12.86 39.00 59.19 83.36 FI °F-days 714.95 966.71 32.00 32.00 221.00 2,402.60 4,460.00 ESAL kESAL 451.21 421.61 17.00 89.50 303.82 932.28 3,384.13 CumESAL kESAL 2,878.15 2,982.87 43.55 315.42 1,881.39 6,833.94 20,232.25 Estimation of Rutting Model for Rehabilitated Pavements. Table 2-15 shows the estimated parameters of the final rutting performance model for rehabilitated pavements. The median absolute error (MAE) of the model was 0.04 inches, and the approximate R2 was 0.16. Since the model was estimated using quantile regression to fit the median, R2 is not applicable and, therefore, the approximate R2 suggested by Koenker and Machado (1999) was applied. This approximate R2 is similar in concept to R2 but is not directly comparable. Figure 2-27 shows the comparison between the measured rutting pavement performance in the LTPP dataset and the predicted results from the regression model. The variables and interactions listed in Table 2-15 were the ones statistically significant–i.e., with a p- value lower than 0.10. Based on the statistical parameters for each variable, the thickness of the AC overlay and pavement, stiffness of the base and subgrade, and P4 gradation and binder content of the AC material had the most influence on the rutting performance along with temperature, precipitation, traffic, and age. As-constructed AV was shown to have an effect as an independent variable. This effect was negative, which indicates that pavements with higher AV tended to have less rutting, keeping everything else fixed in the analysis dataset, which contradicts expectation. The effect of AV on the deterioration rate of rutting over time was not statistically significant for rehabilitated pavements. The parameters in Table 2-15 also show that the interactive effect of AV with AC pavement (thickness, P4 gradation, binder content) and climate

60 had a statistically significant effect on the value of rutting, which may be a reflection of the influence of the stronger AC pavement and climate independent variables. The dominance of negative signs on these six interactions implies that the influence of as-constructed AV on performance contradicts expectation. The variables in this model were scaled before regression model development (scaled back using the respective mean and standard deviation of the variable). The estimated parameters in Table 2-15 are shown in the units of the scaled variables–not the units of the listed variables. The model parameters in the original units are shown in the User Guide document. Table 2-15. Estimated parameters of rutting model for rehabilitated pavements. Variable Value Std. Error t value Pr(>|t|) (Intercept) 0.181 0.010 17.537 0.000 OvTh_AC -0.012 0.003 -4.646 0.000 Th_AC 0.024 0.007 3.386 0.001 Mr_BSB -0.007 0.004 -1.748 0.081 Mr_SG 0.018 0.003 6.523 0.000 Gr_4 -0.026 0.006 -4.433 0.000 BC_s 0.052 0.011 4.808 0.000 Temp 0.020 0.010 1.973 0.049 Prec -0.033 0.007 -4.655 0.000 cumESAL_log -0.039 0.008 -4.753 0.000 logAge 0.026 0.003 9.462 0.000 AV_s -0.006 0.002 -4.185 0.000 logAge:FI 0.010 0.002 4.338 0.000 logAge:cumESAL_log 0.009 0.003 2.970 0.003 logAge:Mr_BSB 0.003 0.002 1.723 0.085 logAge:Gr_4 -0.003 0.002 -1.932 0.054 AV_s:Th_AC -0.003 0.001 -2.205 0.028 AV_s:Gr_4 0.004 0.001 5.103 0.000 AV_s:BC_s -0.007 0.001 -5.434 0.000 AV_s:Temp -0.007 0.002 -3.030 0.002 AV_s:Prec 0.004 0.001 3.405 0.001 AV_s:FI -0.006 0.001 -6.374 0.000 Legend for color coding: Yellow Dominant variable in the model Green Air voids variable with positive value Grey Air voids variable with negative value

61 Figure 2-27. Comparison of rehabilitated pavement rutting performance between LTPP measurements and regression model predictions. Fatigue Cracking Performance Models This section describes the analysis datasets and reports summary statistics, estimation results, and main findings from the development of the fatigue cracking performance models for new and rehabilitated pavements. Fatigue Cracking Performance Dataset. Fatigue cracking was captured by the cracking percent metric defined in the 2016 HPMS Field Manual (FHWA, 2016) for asphalt pavements. This metric was referred to as wheel path cracking (WpC) for new pavements and as reflected fatigue cracking (RfC) for rehabilitated pavements in this study. The HPMS cracking percent for AC pavements, CP, is computed as the total area of the wheel paths with longitudinal cracks divided by the total area of the highway section, as shown in Equation 10. The value of 𝐶𝑃 ranges from zero (when no longitudinal surface cracking is visible within the left or right wheel path) to CPmax=6.5feet/W (when both wheel paths present longitudinal cracks along the entire pavement section length). Consequently, CPmax decreases as the lane width, W, increases and ranges between 81% and 50% for typical lane widths. For this analysis, the dependency of CP on the lane width complicates the direct comparison between sections. The dependent variable used for modeling the performance of fatigue percent consisted of a scaled version of CP, which is independent of the lane width and ranged from 0% to 100%. This scaled CP variable provided the methodological benefit of bounding the possible outcomes to feasible values. Once predictions are made for the scaled variable, they can be scaled back to the original CP for reporting purposes. The scaled CP variable, CP*, was defined as Equation 10. 𝐶𝑃∗ = 𝐶𝑃 = ( ) (10) Where: CP = HPMS cracking percent for AC pavements; CP* = scaled HPMS cracking percent for AC pavements; W = lane width; wpwidth = wheel-path width; L = pavement section length;

62 Lcrklwp = total length of left wheel path with detected cracking; and Lcrkrwp = total length of right wheel path with detected cracking. The plot in Figure 2-28 shows the median scaled cracking percent (CP*) over time of the available fatigue cracking data for model development grouped by the as-constructed AV quintiles. Group Q1 contains test sections with 20% lowest AV_m values, while Q5 groups sections with the 20% highest AV_m values. The median scaled cracking percent tended to increase with time, as expected, and the observed increments were greater for longer ages, suggesting the use of a non-linear functional form for the model specification. The longer the age and the lower the number of observations cause more unreliable performance trends (due to site conditions and possible unrecorded treatments applied to the surface), which is explained by the differences in the duration of performance data curves between the quintile groups in the later years of each curve. Taking this into consideration, only observations with Age≤20 years were used for developing the fatigue cracking performance models. Figure 2-28. Median scaled cracking percent versus age grouped by mean air voids quintiles. The fatigue cracking performance data plot in Figure 2-28 provides an overview of general characteristics and trends in the input data subsets. However, the observed effects of AV_m grouping in the plot may be misleading due to confounding factors. A more accurate assessment of AV_m was obtained from interpreting the parameters in the regression model as they were estimated while accounting for several other significant factors. Fatigue Cracking Performance Model Specification. An exponential-based math function which creates an increasing rate of cracking with time was used for developing the fatigue cracking performance models consisted of a Gompertz s-shape growth curve, as shown in Equation 11. The model parameters on AV_m are of special interest for the objectives of this study. The parameters quantify the effect of as- constructed AV on cracking percent growth over time (β_%Gmmm,Age), and the effect on the magnitude of cracking percent (β%Gmmm and the interaction terms β%Gmmm,j). 𝐶𝑃∗ = 𝑒𝑥𝑝(−𝑒𝑥𝑝(𝛽 + ∑ (𝛽 𝑋 , ) + %𝐺𝑚𝑚 (𝛽% + ∑ (𝛽% , 𝑋 , )) + 𝐴𝑔𝑒 (𝛽 + ∑ (𝛽 , 𝑋 , )) + 𝛽% , 𝐴𝑔𝑒 %𝐺𝑚𝑚 )) (11) Where: 𝐶𝑃∗ = estimated scaled cracking percent for year yr; Ageyr = age of pavement surface for year yr;

63 %Gmmm = as-constructed air voids, weighted average; Xj,yr = model variable j for year 𝑦𝑟; and βj = model parameter on variable j. Summary Statistics of Fatigue Cracking Dataset for New Pavements. Table 2-16 shows the summary statistics of the variables considered for developing the fatigue cracking performance model for new pavements. The minimum and maximum values presented in this table define the ranges under which the fatigue cracking model for new pavements was developed. For many of the variables, these minimum and maximum values represent extreme values well outside the norm. For example, the Th BSB maximum is 221.5 and the 90th percentile value is only 42.5. It is cautioned that inputting variable values outside the 10th and 90th percentile values into the regression model may lead to unreliable estimates. Table 2-16. Summary statistics of data used for estimating fatigue cracking model for new pavements. Variable Units Mean Std Dev Min 10%ile 50%ile 90%ile Max FTC % 19.31 29.99 0.00 0.00 1.84 73.77 101.07 AV_m %Gmm 6.76 2.42 1.85 3.45 6.76 10.03 15.19 AGE years 6.47 4.52 0.00 1.00 6.00 13.00 20.00 Th_AC in 5.69 1.57 1.40 4.00 5.65 7.40 10.50 Th_BSB in 20.23 17.52 0.00 7.90 15.90 42.50 221.50 E*_FtC ksi 1,064.39 343.66 412.40 638.37 1,065.42 1,458.37 2,480.68 Mr_BSB ksi 33.11 27.04 9.10 15.80 21.56 100.00 100.00 Mr_SG ksi 9.27 2.69 4.89 5.70 9.48 12.55 16.70 BC_m % 4.75 0.76 3.55 3.72 4.62 5.94 7.00 Gr_4 % 56.46 9.10 22.00 50.00 56.00 69.00 82.00 Gr_200 % 5.69 1.31 1.80 3.70 6.10 7.10 8.90 Gr_k - 5.25 1.75 2.50 3.63 5.10 7.18 15.33 Gr_lambda - 0.83 0.16 0.60 0.70 0.79 1.04 1.65 Temp °F 57.81 9.35 32.00 44.06 60.08 68.90 75.74 Prec in 32.12 17.38 3.76 8.43 33.93 56.09 75.72 FI °F-days 423.17 626.47 32.00 32.00 131.00 1,243.40 4,379.00 ESAL kESAL 244.89 257.62 0.00 0.93 178.93 543.91 1,952.54 CumESAL kESAL 1,440.63 1,793.18 0.00 7.12 798.66 3,617.80 12,448.26 Estimation of Fatigue Cracking Model for New Pavements. Table 2-17 shows the estimated parameters of the final fatigue cracking performance model for new pavements. The median absolute error (MAE) of the model was 1.74%, and the approximate R2 was 0.41. Since the model was estimated using quantile regression to fit the median, R2 is not applicable and, therefore, the approximate R2 suggested by Koenker and Machado (1999) was applied. This approximate R2 is similar in concept to R2 but is not directly comparable. Figure 2-29 shows the comparison between the measured fatigue pavement performance in the LTPP dataset and the predicted results from the regression model. The variables and interactions listed in Table 2-17 were the ones statistically significant–i.e., with a p- value lower than 0.10. Based on the statistical parameters for each variable, the thickness of the base and material properties (subgrade stiffness and AC P200) had the most influence on fatigue cracking along with traffic and pavement age. The parameters in Table 2-17 show that the effect of as-constructed AV on the wheel path cracking value was statistically significant for new pavements as an independent variable. However, the negative estimated value (-0.037) implies that increasing AV would decrease cracking which

64 contradicted performance expectations. Further, AV was significant as an interaction with the thickness of the base, stiffness of the subgrade, and AC layer P200 gradation, which may be a reflection of the influence of these pavement material properties as independent variables. Both interactions β%Gmmm,Age and β%Gmms,CumESAL were statistically significant and negative. Therefore, the effect of as-constructed AV on the deterioration rate of wheel path cracking over time was significant for new pavements. The negative sign on these parameters indicates that pavements with higher AV tended to deteriorate slower with longer age and higher traffic–keeping everything else fixed, which contradicts expectation. The variables in this model were scaled before regression model development (scaled back using the respective mean and standard deviation of the variable). The estimated parameters in Table 2-17 are shown in the units of the scaled variables–not the units of the listed variables. The model parameters in the original units are shown in the User Guide document. Table 2-17. Estimated parameters of fatigue cracking model for new pavements. Variable Value Std. Error t value Pr(>|t|) (Intercept) 2.926 0.105 27.740 0.000 Th_BSB 0.210 0.091 2.315 0.021 E*_FtC -0.050 0.018 -2.753 0.006 Mr_BSB 0.086 0.019 4.464 0.000 Mr_SG 0.394 0.064 6.198 0.000 Gr_200 -0.766 0.130 -5.877 0.000 cumESAL_log 0.443 0.074 5.992 0.000 Age -0.111 0.020 -5.425 0.000 AV_m -0.037 0.016 -2.368 0.018 Age:Th_BSB 0.070 0.006 11.017 0.000 Age:E*_FtC 0.052 0.004 11.801 0.000 Age:Gr_4 0.025 0.005 5.411 0.000 Age:Gr_200 0.028 0.007 4.291 0.000 Age:BC_m -0.024 0.005 -5.227 0.000 Age:Temp -0.064 0.006 -10.168 0.000 Age:Prec -0.026 0.005 -5.561 0.000 Age:cumESAL_log -0.055 0.005 -11.219 0.000 AV_m:Th_BSB -0.043 0.012 -3.452 0.001 AV_m:Mr_SG -0.052 0.010 -5.096 0.000 AV_m:Gr_200 0.106 0.020 5.383 0.000 AV_m:cumESAL_log -0.059 0.010 -5.678 0.000 Age:AV_m -0.014 0.003 -5.061 0.000 Legend for color coding: Yellow Dominant variable in the model Green Air voids variable with positive value Grey Air voids variable with negative value

65 Figure 2-29. Comparison of new pavement fatigue cracking performance between LTPP measurements and regression model predictions. Summary Statistics of Fatigue Cracking Dataset for Rehabilitated Pavements. Table 2-18 shows the summary statistics of the variables considered for developing the fatigue cracking performance model for rehabilitated pavements. The minimum and maximum values presented in this table define the ranges under which the fatigue cracking model for rehabilitated pavements was developed. For many of the variables, these minimum and maximum values represent extreme values well outside the norm. For example, the ESAL maximum is 3480 and the 90th percentile value is only 840. It is cautioned that inputting variable values outside the 10th and 90th percentile values into the regression model may lead to unreliable estimates. Table 2-18. Summary statistics of data used for estimating fatigue cracking model for rehabilitated pavements. Variable Units Mean Std Dev Min 10%ile 50%ile 90%ile Max FTC % 20.73 31.94 0.00 0.00 1.84 81.16 100.99 AV_m %Gmm 4.95 1.88 1.74 2.86 4.68 7.41 10.61 AGE years 7.69 4.63 0.00 2.00 7.00 14.00 20.00 OvTh_AC in 2.68 2.27 0.00 0.00 2.50 5.70 9.20 Th_AC in 8.95 2.93 2.50 5.50 8.30 13.40 21.90 Th_BSB in 16.41 9.40 0.00 7.90 15.00 28.90 64.00 E*_FtC ksi 1,109.79 399.27 225.05 662.90 1,047.58 1,622.62 2,513.61 Mr_BSB ksi 30.83 26.53 12.27 15.30 20.00 100.00 100.00 Mr_SG ksi 9.29 3.35 3.40 4.20 8.73 14.45 16.43 BC_m % 4.79 0.57 3.75 4.23 4.67 5.55 9.00 Gr_4 % 54.41 8.48 29.00 44.00 53.00 65.00 74.00 Gr_200 % 6.49 1.72 0.90 4.60 6.40 9.00 13.00 Gr_k - 5.12 1.42 2.48 3.80 5.02 6.49 15.28 Gr_lambda - 0.85 0.19 0.57 0.63 0.83 1.11 1.47 Temp °F 53.07 9.98 32.36 37.38 53.78 66.02 75.38 Prec in 37.97 16.40 1.51 16.13 39.47 57.66 110.27 FI °F-days 707.49 884.08 32.00 35.60 279.50 2,210.00 3,930.80

66 Variable Units Mean Std Dev Min 10%ile 50%ile 90%ile Max ESAL kESAL 403.65 372.82 12.83 103.42 281.96 840.00 3,480.00 CumESAL kESAL 2,869.20 2,773.16 15.23 429.52 1,963.13 6,340.44 22,366.37 Estimation of Fatigue Cracking Model for Rehabilitated Pavements. Table 2-19 shows the estimated parameters of the final fatigue cracking performance model for rehabilitated pavements. The median absolute error (MAE) of the model was 1.82%, and the approximate R2 was 0.35. Since the model was estimated using quantile regression to fit the median, R2 is not applicable and, therefore, the approximate R2 suggested by Koenker and Machado (1999) was applied. This approximate R2 is similar in concept to R2 but is not directly comparable. Figure 2-30 shows the comparison between the measured fatigue pavement performance in the LTPP dataset and the predicted results from the regression model. The variables and interactions listed in Table 2-19 were the ones statistically significant–i.e., with a p- value lower than 0.10. Based on the statistical parameters for each variable, the thickness of the AC overlay and base, AC overlay gradation, and FI had the most influence on reflected wheel path cracking performance on rehabilitated pavements. The effect of as-constructed AV on the deterioration rate of reflected fatigue cracking over time was significant for rehabilitated pavements based on the interaction of AV and age (β%Gmmm,Age). The negative sign on the interaction of AV and age indicates that pavements with higher AV tended to deteriorate slower with longer age–keeping everything else fixed. Further, the interactive effect of AV with pavement thickness, AC material properties (gradation and binder content), and temperature had a statistically significant effect on the magnitude of reflected fatigue cracking, which may be a reflection of the influence of those pavement variables as independent variables. The variables in this model were scaled before regression model development (scaled back using the respective mean and standard deviation of the variable). The estimated parameters in Table 2-19 are shown in the units of the scaled variables–not the units of the listed variables. The model parameters in the original units are shown in the User Guide document. Table 2-19. Estimated parameters of fatigue cracking model for rehabilitated pavements. Variable Value Std. Error t value Pr(>|t|) (Intercept) 2.766 0.065 42.737 0.000 OvTh_AC -0.242 0.053 -4.572 0.000 Th_BSB -0.678 0.161 -4.221 0.000 Gr_4 0.203 0.065 3.141 0.002 Gr_200 -0.288 0.071 -4.060 0.000 FI 0.283 0.062 4.556 0.000 Age -0.094 0.019 -4.968 0.000 Age:OvTh_AC 0.016 0.005 3.008 0.003 Age:Th_AC 0.044 0.005 9.302 0.000 Age:Th_BSB -0.032 0.014 -2.355 0.019 Age:E*_FtC -0.032 0.005 -6.368 0.000 Age:Gr_4 -0.026 0.006 -4.254 0.000 Age:Temp -0.075 0.012 -6.233 0.000 Age:FI -0.127 0.012 -10.467 0.000 Age:cumESAL_log -0.057 0.010 -5.431 0.000 OvTh_AC:AV_m 0.042 0.008 4.913 0.000 Th_BSB:AV_m 0.183 0.030 6.127 0.000 Gr_4:AV_m -0.041 0.011 -3.739 0.000

67 Variable Value Std. Error t value Pr(>|t|) Gr_200:AV_m 0.063 0.017 3.709 0.000 AV_m:BC_m -0.047 0.008 -5.689 0.000 AV_m:Temp 0.073 0.013 5.486 0.000 Age:AV_m -0.017 0.003 -5.433 0.000 Legend for color coding: Yellow Dominant variable in the model Green Air voids variable with positive value Grey Air voids variable with negative value Figure 2-30. Comparison of rehabilitated pavement fatigue cracking performance between LTPP measurements and regression model predictions. Thermal Cracking Performance Models This section describes the analysis datasets and reports summary statistics, estimation results, and main findings from the development of the thermal cracking performance models for new and rehabilitated pavements. Thermal cracking was captured as the total length of transverse cracks per mile. Of the 382 performance curves available for the analysis of thermal cracking, 184 had zero transverse cracks reported for the entire period of analysis. The remaining performance curves presented zero transverse cracks for an initial number of years before the first transverse crack was visible in the pavement surface. Given the large number of zero thermal cracking data, the effect of as-constructed AV on thermal cracking performance was assessed in two parts. The first part evaluated the number of years until crack initiation. The second part of the analysis consisted of estimating and interpreting the thermal cracking performance model, as conducted for the other performance characteristics, using only the thermal cracking data since crack initiation. For this second analysis, the amount of thermal cracking (ThC) was referred to as transverse cracking (TrC) for new pavements and as reflected thermal cracking (RtC) for rehabilitated pavements. Time to Thermal Cracking Initiation Analysis. The effect of as-constructed air surface voids on the increment of time until thermal cracking initiation was assessed by comparing the survival curves corresponding to each as-constructed AV quintile. These survival curves were estimated using R “survival”

68 package (Therneau and Grambsch, 2000). The plot in Figure 2-31 shows the Kaplan-Meier estimates of the probability of zero surface thermal cracking over time, where the range of each AV_s quintile used for the analysis is shown in the legend. The different survival curves show that thermal crack initiation occurred after 11 to 13 years, depending on the AV_s group, for half of the performance curves in the data. The difference in the number of years until thermal crack initiation did not vary significantly across the AV_s groups for most performance curves. These observations suggest that as-constructed surface AV have a limited effect on the increment of time until thermal cracking initiates. In addition, no clear, consistent patterns across the AV_s group were observed in the data. Figure 2-31. Probability of zero surface thermal cracking over time grouped by surface air voids quintiles. Thermal Cracking After Crack Initiation Performance Dataset. The plot in Figure 2-32 shows the median thermal cracking over time from the year since the first thermal cracking was visible in the surface (i.e., crack initiation) and grouped by the as-constructed AV quintiles. Group Q1 contains test sections with 20% lowest AV_s values, while Q5 groups sections with the 20% highest AV_s values. The median thermal cracking after crack initiation tended to increase with time and the observed increments were fairly constant or slightly greater for longer ages, as expected. The longer the age and the lower the number of observations, the more unreliable the cracking trend (due to site conditions and possible unrecorded treatments applied to the surface), which explains the unexpected sudden decrease in thermal cracking in the last years of four of the five curves. Taking this into consideration, only observations with Age≤10 years after crack initiation were used for developing the thermal cracking performance models.

69 Figure 2-32. Median thermal cracking after crack initiation versus age grouped by surface air voids quintiles. The thermal cracking performance data plot in Figure 2-32 provides an overview of general characteristics and trends in the input data subsets. However, the observed effects of AV_s grouping in the plot may be misleading due to confounding factors. A more accurate assessment of AV_s was obtained from interpreting the parameters in the regression model as they were estimated while accounting for several other significant factors. Thermal Cracking After Crack Initiation Performance Model Specification. An exponential-based math function which creates a rapid rate of increase in cracking was used for developing the performance models for thermal cracking after crack initiation is shown Equation 12. The model parameters on AV_s are of special interest for the objectives of this study. The parameters quantify the effect of as-constructed AV on crack growth over time (β%Gmm_s,Age), and the effect on the magnitude of cracking (β%Gmms and the interaction terms β%Gmms,j). 𝑇ℎ𝐶 = 𝑒𝑥𝑝(𝛽 + ∑ (𝛽 𝑋 , ) + %𝐺𝑚𝑚 (𝛽% + ∑ (𝛽% , 𝑋 , )) + 𝐴𝑔𝑒 (𝛽 + ∑ (𝛽 , 𝑋 , )) + 𝛽% , 𝐴𝑔𝑒 %𝐺𝑚𝑚 ) (12) Where: 𝑇ℎ𝐶 = estimated thermal cracking for year yr; Ageyr = years elapsed since crack initiation to year yr; %Gmms = as-constructed surface air voids; Xj,yr = model variable j for year yr; and βj = model parameter on variable j. Summary Statistics of Thermal Cracking Data After Crack Initiation for New Pavements. Table 2-20 shows the summary statistics of the variables considered for developing the thermal cracking performance model for new pavements. The minimum and maximum values presented in this table define the ranges under which the thermal cracking model for new pavements was developed. For many of the variables, these minimum and maximum values represent extreme values well outside the norm. For example, the AV_s maximum is 21.82 and the 90th percentile value is only 11.22. It is cautioned that inputting variable values outside the 10th and 90th percentile values into the regression model may lead to unreliable estimates.

70 Table 2-20. Summary statistics of data used for estimating thermal cracking after crack initiation model for new pavements. Variable Units Mean Std Dev Min 10%ile 50%ile 90%ile Max TRC ft/ln-mi 456.78 808.83 0.00 0.00 128.00 1,257.50 5,779.00 AV_s %Gmm 6.97 2.78 1.92 3.86 6.80 9.98 21.82 AGE years 2.66 2.72 0.00 0.00 2.00 7.00 10.00 Th_AC in 5.72 1.56 1.60 4.10 5.30 7.50 9.17 Th_BSB in 22.07 18.75 0.00 8.00 16.70 43.20 123.30 E*_ThC ksi 3,172.98 553.88 1,085.10 2,632.37 3,073.92 3,932.90 4,468.08 Mr_BSB ksi 26.76 19.75 11.50 17.10 20.40 33.35 100.00 Mr_SG ksi 9.90 2.57 4.89 5.70 10.38 12.55 16.70 BC_s % 4.78 0.80 3.63 3.63 4.53 6.10 7.00 Gr_4 % 55.62 7.19 22.00 50.00 55.00 66.00 72.00 Gr_200 % 5.48 1.56 1.80 3.70 5.80 7.10 9.70 Gr_k - 5.53 1.74 2.53 3.64 5.55 7.18 15.33 Gr_lambda - 0.81 0.14 0.60 0.71 0.79 0.90 1.49 Temp °F 56.86 9.74 35.96 43.61 60.35 68.36 74.66 Prec in 28.94 18.32 4.38 8.23 25.88 58.30 70.20 FI °F-days 561.43 785.18 32.00 32.00 147.20 1,963.40 3,221.60 ESAL kESAL 269.45 369.00 0.00 1.00 114.46 802.96 1,952.54 CumESAL kESAL 2,023.61 2,497.62 0.00 8.44 1,154.10 5,964.79 12,448.26 Estimation of Model for Thermal Cracking After Crack Initiation for New Pavements. Table 2-21 shows the estimated parameters of the final thermal cracking performance model for new pavements. The median absolute error (MAE) of the model was 164.5 ft/ln-mi, and the approximate R2 was 0.38. Since the model was estimated using quantile regression to fit the median, R2 is not applicable and, therefore, the approximate R2 suggested by Koenker and Machado (1999) was applied. This approximate R2 is similar in concept to R2 but is not directly comparable. Figure 2-33 shows the comparison between the measured thermal cracking pavement performance in the LTPP dataset and the predicted results from the regression model. The variables and interactions listed in Table 2-21 were the ones statistically significant–i.e., with a p- value lower than 0.10. Based on the statistical parameters for each variable, the thickness of the base and AC material properties (stiffness and gradation) had the most influence on thermal cracking performance along with temperature and pavement surface age. The parameters in Table 2-21 also show that the interactive effect of as-constructed AV with pavement thickness, AC P200 gradation, and temperature had a statistically significant effect on the value of thermal cracking, which may be a reflection of the influence of those independent variables. The effect of as-constructed AV on the deterioration rate of thermal cracking over time was not statistically significant for new pavements. The variables in this model were scaled before regression model development (scaled back using the respective mean and standard deviation of the variable). The estimated parameters in Table 2-21 are shown in the units of the scaled variables, not the units of the listed variables. The model parameters in the original units are shown in the User Guide document.

71 Table 2-21. Estimated parameters of thermal cracking model for new pavements. Variable Value Std. Error t value Pr(>|t|) (Intercept) 4.763 0.177 26.837 0.000 Th_BSB -1.337 0.473 -2.828 0.005 E*_TrC -0.408 0.141 -2.887 0.004 Gr_4 0.236 0.084 2.821 0.005 Gr_200 -0.689 0.270 -2.551 0.012 Temp -1.249 0.274 -4.550 0.000 Age 0.314 0.029 10.993 0.000 Age:E*_TrC 0.134 0.030 4.528 0.000 Age:BC_s -0.051 0.013 -3.777 0.000 AV_s:Th_AC -0.022 0.010 -2.175 0.031 Th_BSB:AV_s 0.183 0.071 2.574 0.011 Gr_200:AV_s 0.089 0.039 2.249 0.026 Temp:AV_s 0.173 0.045 3.869 0.000 Legend for color coding: Yellow Dominant variable in the model Green Air voids variable with positive value Grey Air voids variable with negative value Figure 2-33. Comparison of new pavement thermal cracking performance between LTPP measurements and regression model predictions. Summary Statistics of Thermal Cracking Data After Crack Initiation for Rehabilitated Pavements. Table 2-22 shows the summary statistics of the variables considered for developing the thermal cracking performance models for rehabilitated pavements. The minimum and maximum values presented in this table define the ranges under which the thermal cracking model for rehabilitated pavements was developed. For many of the variables, these minimum and maximum values represent extreme values well outside the norm. For example, the AV_s maximum is 15.17 and the 90th percentile value is only 7.80. It is cautioned that inputting variable values outside the 10th and 90th percentile values into the regression model may lead to unreliable estimates.

72 Table 2-22. Summary statistics of data used for estimating thermal cracking model for rehabilitated pavements. Variable Units Mean Std Dev Min 10%ile 50%ile 90%ile Max TRC ft/ln-mi 526.73 589.11 0.00 0.00 322.00 1,292.00 2,841.00 AV_s %Gmm 4.93 2.52 0.96 2.22 4.24 7.80 15.17 AGE years 4.50 3.15 0.00 0.00 4.00 9.00 10.00 OvTh_AC in 2.10 2.05 0.00 0.00 1.90 4.80 7.60 Th_AC in 8.40 2.62 2.50 5.40 8.20 11.40 21.90 Th_BSB in 16.81 8.94 0.00 8.40 15.60 30.00 51.40 E*_ThC ksi 3,167.28 591.22 227.55 2,499.23 3,274.16 3,854.78 4,443.50 Mr_BSB ksi 34.29 30.58 13.44 16.20 20.00 100.00 100.00 Mr_SG ksi 8.97 3.32 3.40 4.10 8.73 14.45 15.13 BC_s % 4.98 0.60 3.80 4.00 4.95 5.85 6.80 Gr_4 % 55.03 8.73 29.00 44.00 56.00 65.00 72.00 Gr_200 % 6.79 1.79 0.90 4.90 6.40 9.10 13.00 Gr_k - 5.12 1.71 2.67 3.71 4.88 6.81 15.28 Gr_lambda - 0.83 0.20 0.57 0.63 0.79 1.11 1.44 Temp °F 51.59 10.20 32.36 36.32 53.06 65.30 73.94 Prec in 33.77 16.97 1.51 13.77 34.90 54.09 110.27 FI °F-days 827.07 970.70 32.00 46.40 350.60 2,402.60 3,930.80 ESAL kESAL 392.42 346.77 25.81 96.61 262.00 722.47 2,866.65 CumESAL kESAL 2,924.48 2,626.14 46.46 544.84 2,050.15 6,271.94 16,607.85 Estimation of Model for Thermal Cracking After Crack Initiation for Rehabilitated Pavements. Table 2-23 shows the estimated parameters of the final thermal cracking performance model for rehabilitated pavements. The median absolute error (MAE) of the model was 158.37 ft/ln-mi, and the approximate R2 was 0.32. Since the model was estimated using quantile regression to fit the median, R2 is not applicable and, therefore, the approximate R2 suggested by Koenker and Machado (1999) was applied. This approximate R2 is similar in concept to R2 but is not directly comparable. Figure 2-34 shows the comparison between the measured thermal cracking pavement performance in the LTPP dataset and the predicted results from the regression model. The variables and interactions listed in Table 2-23 were the ones statistically significant–i.e., with a p- value lower than 0.10. Based on the statistical parameters for each variable, the pavement thickness (AC and base) and material properties (stiffness of base and AC P4) had the most influence on reflected thermal cracking along with precipitation, traffic, and surface age. The effect of as-constructed AV on the deterioration rate of reflected thermal cracking over time was not statistically significant for rehabilitated pavements. The parameters in Table 2-23 show that the interactive effect of as-constructed AV with the pavement thickness (AC and base) and material properties (stiffness of base and AC P4) had a statistically significant effect on the value of reflected thermal cracking, which may be a reflection of the influence of pavement thickness and material properties as independent variables. The negative sign on these four interactions implies that the influence of as-constructed AV contradicted the expectation. The variables in this model were scaled before regression model development (scaled back using the respective mean and standard deviation of the variable). The estimated parameters in Table 2-23 are shown in the units of the scaled variables–not the units of the listed variables. The model parameters in the original units are shown in the User Guide document.

73 Table 2-23. Estimated parameters of thermal cracking model for rehabilitated pavements. Variable Value Std. Error t value Pr(>|t|) (Intercept) 4.748 0.082 57.615 0.000 Th_AC 0.409 0.099 4.117 0.000 Th_BSB 0.599 0.120 4.990 0.000 Mr_BSB 0.321 0.102 3.137 0.002 Gr_4 0.366 0.072 5.058 0.000 Prec -0.162 0.025 -6.534 0.000 cumESAL_log -0.467 0.039 -12.043 0.000 Age 0.219 0.014 15.271 0.000 Th_AC:AV_s -0.088 0.016 -5.379 0.000 Th_BSB:AV_s -0.097 0.025 -3.802 0.000 Mr_BSB:AV_s -0.059 0.015 -3.862 0.000 Gr_4:AV_s -0.056 0.010 -5.432 0.000 Legend for color coding: Yellow Dominant variable in the model Green Air voids variable with positive value Grey Air voids variable with negative value Figure 2-34. Comparison of rehabilitated pavement thermal cracking performance between LTPP measurements and regression model predictions. Ride Performance Models This section describes the analysis datasets and reports summary statistics, estimation results, and main findings from the development of the ride performance model for new and rehabilitated pavements. Ride Performance Dataset. The plot in Figure 2-35 shows the median IRI over time of the available ride data for model development grouped by the as-constructed AV quintiles. Group Q1 contains test sections with 20% lowest AV_s values, while Q5 groups sections with the 20% highest AV_s values. The median IRI tended to increase over time, as expected, and the observed increments were greater for longer ages, suggesting the use of a non-linear functional form for the model specification. The longer the age and lower the number of observations, the more unreliable the IRI trend (due to site conditions and possible unrecorded treatments applied to the surface), which explains the unexpected sudden decrease in IRI at the

74 later years of four of the five curves. Taking this into consideration, only observations with Age≤20 years were used for developing the ride performance models. Figure 2-35. Median IRI versus age grouped by surface air voids quintiles. The ride performance data plot in Figure 2-35 provides an overview of general characteristics and trends in the input data subsets. However, the observed effects of AV_s grouping in the plot may be misleading due to confounding factors. A more accurate assessment of AV_s was obtained from interpreting the parameters in the regression model as they were estimated while accounting for several other significant factors. Ride Performance Model Specification. An exponential-based math function which creates a rate of distress increase similar to the cracking models was used for developing the ride performance models is shown Equation 13. The model parameters on AV_s are of special interest for the objectives of this study. The parameters quantify the effect of as-constructed AV on IRI growth over time (β%Gmms,Age) and the effect on the magnitude of IRI (β%Gmms) and the interaction terms β%Gmms,j). 𝐼𝑅𝐼 = 𝑒𝑥𝑝(𝛽 + ∑ (𝛽 𝑋 , ) + %𝐺𝑚𝑚 (𝛽% + ∑ (𝛽% , 𝑋 , )) + 𝐴𝑔𝑒 (𝛽 + ∑ (𝛽 , 𝑋 , )) + 𝛽% , 𝐴𝑔𝑒 %𝐺𝑚𝑚 ) (13) Where: 𝐼𝑅𝐼 = estimated IRI for year yr; Ageyr = age of pavement surface for year yr; %Gmms = as-constructed surface air voids; Xj,yr = model variable j for year yr; and βj = model parameter on variable j. Summary Statistics of Ride Dataset for New Pavements. Table 2-24 shows the summary statistics of the variables considered for developing the ride performance model for new pavements. The minimum and maximum values presented in this table define the ranges under which the ride model for new pavements was developed. For many of the variables, these minimum and maximum values represent extreme values well outside the norm. For example, the AV_s maximum is 21.82 and the 90th percentile value is only 11.22. It is cautioned that inputting variable values outside the 10th and 90th percentile values into the regression model may lead to unreliable estimates.

75 Table 2-24. Summary statistics of data used for estimating ride model for new pavements. Variable Units Mean Std Dev Min 10%ile 50%ile 90%ile Max IRI in/mile 68.46 26.14 33.39 43.62 63.68 99.30 272.89 AV_s %Gmm 7.65 2.90 1.92 4.20 7.27 11.22 21.82 AGE years 6.69 4.60 0.00 1.00 6.00 14.00 20.00 Th_AC in 5.69 1.60 1.40 3.90 5.70 7.40 10.50 Th_BSB in 19.40 16.49 0.00 7.80 15.80 37.00 221.50 E*_IRI ksi 982.23 375.58 412.40 631.03 895.47 1,644.06 2,480.68 Mr_BSB ksi 33.98 27.17 9.10 15.99 21.90 100.00 100.00 Mr_SG ksi 9.42 2.75 4.89 5.70 9.55 12.55 16.70 BC_s % 5.08 0.86 3.63 4.12 4.99 6.28 7.30 Gr_4 % 56.48 9.53 22.00 49.40 55.00 69.00 82.00 Gr_200 % 5.79 1.33 1.80 3.70 6.20 7.10 10.20 Gr_k - 5.22 1.74 2.50 3.63 5.10 7.18 15.33 Gr_lambda - 0.85 0.16 0.68 0.70 0.80 1.07 1.65 Temp °F 56.24 8.73 34.88 43.59 58.10 67.64 75.74 Prec in 33.46 17.49 3.47 8.03 35.46 57.96 75.72 FI °F-days 501.59 672.47 32.00 32.00 165.20 1,707.80 3,272.00 ESAL kESAL 237.23 251.87 0.00 1.74 168.98 534.94 1,952.54 CumESAL kESAL 1,431.45 1,712.46 0.00 10.70 827.81 3,763.62 11,607.04 Estimation of Ride Model for New Pavements. Table 2-25 shows the estimated parameters of the final ride performance model for new pavements. The median absolute error (MAE) of the model was 7.88 in/mile, and the approximate R2 was 0.24. Since the model was estimated using quantile regression to fit the median, R2 is not applicable and, therefore, the approximate R2 suggested by Koenker and Machado (1999) was applied. This approximate R2 is similar in concept to R2 but is not directly comparable. Figure 2-36 shows the comparison between the measured ride pavement performance in the LTPP dataset and the predicted results from the regression model. The variables and interactions listed in Table 2-25 were the ones statistically significant–i.e., with a p- value lower than 0.10. Based on the statistical parameters for each variable, the stiffness of the pavement structure (subgrade, base, and asphalt) had the most influence on ride performance along with precipitation and traffic. Since CumESAL is a function of the age and β%Gmms,CumESAL was statistically significant, the as-constructed AV had a statistically significant effect on the deterioration rate of IRI over time for new pavements. However, further analysis of the sensitivity of this interaction based on the small parameter value and t value showed that although it is statistically significant, it was most likely not significant in practical engineering terms. The parameters in Table 2-25 also show that the interactive effect of as- constructed AV with the stiffness of the pavement materials and climate had a statistically significant effect on the value of IRI, which may be a reflection of the influence of pavement material stiffness as independent variables. The plus sign on most of these interactions implies that the influence of as-constructed AV met expectations. The variables in this model were scaled before regression model development (scaled back using the respective mean and standard deviation of the variable). The estimated parameters in Table 2-25 are shown in the units of the scaled variables–not the units of the listed variables. The model parameters in the original units are shown in the User Guide document.

76 Table 2-25. Estimated parameters of ride model for new pavements. Variable Value Std. Error t value Pr(>|t|) (Intercept) 3.953 0.011 344.680 0.000 Mr_BSB -0.153 0.013 -11.598 0.000 Mr_SG -0.128 0.018 -7.226 0.000 Gr_4 0.161 0.018 9.011 0.000 Gr_200 -0.149 0.026 -5.660 0.000 BC_s -0.118 0.015 -8.109 0.000 Prec -0.111 0.016 -6.898 0.000 cumESAL_log -0.116 0.017 -6.634 0.000 Age 0.020 0.001 15.888 0.000 Age:Th_BSB -0.008 0.001 -7.310 0.000 Age:Mr_BSB 0.005 0.001 4.535 0.000 Age:Gr_200 -0.004 0.001 -2.627 0.009 Age:Temp -0.010 0.001 -7.919 0.000 Age:cumESAL_log 0.003 0.001 4.655 0.000 AV_s:E*_IRI -0.004 0.001 -8.460 0.000 Mr_BSB:AV_s 0.008 0.001 6.906 0.000 Mr_SG:AV_s 0.015 0.002 6.952 0.000 Gr_4:AV_s -0.022 0.002 -10.207 0.000 Gr_200:AV_s 0.026 0.003 8.765 0.000 BC_s:AV_s 0.025 0.002 15.777 0.000 AV_s:Temp 0.012 0.001 8.059 0.000 Prec:AV_s 0.015 0.002 7.583 0.000 cumESAL_log:AV_s 0.007 0.002 3.479 0.001 Legend for color coding: Yellow Dominant variable in the model Green Air voids variable with positive value Grey Air voids variable with negative value Figure 2-36. Comparison of new pavement ride performance between LTPP measurements and regression model predictions.

77 Summary Statistics of Ride Dataset for Rehabilitated Pavements. Table 2-26 shows the summary statistics of the variables considered for developing the ride performance model for rehabilitated pavements. The minimum and maximum values presented in this table define the ranges under which the ride model for rehabilitated pavements was developed. For many of the variables, these minimum and maximum values represent extreme values well outside the norm. For example, the Th_AC variable maximum is 21.9 and the 90th percentile value is only 12.6. It is cautioned that inputting variable values outside the 10th and 90th percentile values into the regression model may lead to unreliable estimates. Table 2-26. Summary statistics of data used for estimating ride model for rehabilitated pavements. Variable Units Mean Std Dev Min 10%ile 50%ile 90%ile Max IRI in/mile 70.20 24.87 32.11 47.85 64.00 98.45 294.08 AV_s %Gmm 4.99 2.48 0.89 2.30 4.48 8.42 15.17 AGE years 7.08 4.87 0.00 1.00 6.00 14.00 20.00 OvTh_AC in 2.85 2.23 0.00 0.00 2.80 5.73 9.20 Th_AC in 8.86 2.76 2.50 5.70 8.30 12.60 21.90 Th_BSB in 16.62 9.54 0.00 8.40 15.00 30.20 64.00 E*_IRI ksi 1,109.10 489.14 406.00 632.13 1,045.63 1,645.13 4,398.86 Mr_BSB ksi 30.03 25.85 12.27 15.30 20.00 100.00 100.00 Mr_SG ksi 9.05 3.26 3.30 4.10 8.42 14.45 17.90 BC_s % 5.03 0.64 3.80 4.40 4.97 5.78 9.00 Gr_4 % 54.58 8.54 29.00 44.00 55.00 65.00 74.00 Gr_200 % 6.52 1.67 0.90 4.60 6.40 9.00 13.00 Gr_k - 5.12 1.47 2.48 3.71 5.02 6.75 15.28 Gr_lambda - 0.84 0.18 0.57 0.63 0.82 1.10 1.47 Temp °F 52.03 9.53 33.08 37.40 53.06 65.48 74.30 Prec in 37.29 16.29 0.71 16.68 39.74 57.00 110.27 FI °F-days 767.26 938.30 32.00 41.00 334.40 2,334.56 4,134.20 ESAL kESAL 393.09 353.54 12.83 89.17 291.63 755.90 3,480.00 CumESAL kESAL 2,664.07 2,744.84 41.64 335.23 1,785.83 6,325.67 22,366.37 Estimation of Ride Model for Rehabilitated Pavements. Table 2-27 shows the estimated parameters of the final ride performance model for rehabilitated pavements. The median absolute error (MAE) of the model was 7.88 in/mile, and the approximate R2 was 0.21. Since the model was estimated using quantile regression to fit the median, R2 is not applicable and, therefore, the approximate R2 suggested by Koenker and Machado (1999) was applied. This approximate R2 is similar in concept to R2 but is not directly comparable. Figure 2-37 shows the comparison between the measured ride pavement performance in the LTPP dataset and the predicted results from the regression model. The variables and interactions listed Table 2-27 were the ones statistically significant–i.e., with a p-value lower than 0.10. Based on the statistical parameters for each variable, the thickness of the AC and base, stiffness of AC and subgrade, material properties P200, binder content, and climate variables, temperature and FI had the most influence on ride performance. The parameters in Table 2-27 show that the effect of as-constructed AV on the IRI value was statistically significant for rehabilitated pavements as an independent variable and the influence of as-constructed AV met expectations. Further, as-constructed AV was significant as an interaction with the stiffness of the materials and the AC layer gradation, which may be a reflection of the influence of these pavement material properties as independent variables. The effect of as-constructed AV on the deterioration rate of IRI over time was statistically significant for rehabilitated pavements. This effect was negative, which indicates that pavements with higher AV tended to deteriorate

78 slower–keeping everything else fixed. However, further analysis of the sensitivity of this variable and interactions based on the small parameter values and t values show that although they are statistically significant, they are most likely not significant in practical engineering terms. The variables in this model were scaled before regression model development (scaled back using the respective mean and standard deviation of the variable). The estimated parameters in Table 2-27 are shown in the units of the scaled variables–not the units of the listed variables. The model parameters in the original units are shown in the User Guide document. Table 2-27. Estimated parameters of ride model for rehabilitated pavements. Variable Value Std. Error t value Pr(>|t|) (Intercept) 4.063 0.025 160.038 0.000 Th_AC -0.065 0.007 -9.824 0.000 Th_BSB -0.047 0.008 -5.744 0.000 E*_IRI 0.097 0.020 4.883 0.000 Mr_SG 0.075 0.013 5.561 0.000 Gr_200 -0.089 0.015 -6.050 0.000 BC_s 0.058 0.009 6.701 0.000 Temp -0.075 0.017 -4.405 0.000 FI -0.043 0.013 -3.254 0.001 Age 0.015 0.003 5.112 0.000 AV_s 0.011 0.004 2.391 0.017 Age:OvTh_AC -0.003 0.001 -4.878 0.000 Age:Mr_BSB 0.005 0.001 6.732 0.000 Age:Gr_4 0.004 0.001 3.774 0.000 Age:Temp -0.005 0.001 -3.609 0.000 Age:Prec -0.002 0.001 -2.761 0.006 Age:cumESAL_log 0.005 0.001 3.359 0.001 AV_s:E*_IRI -0.013 0.004 -3.596 0.000 AV_s:Mr_SG -0.006 0.002 -2.716 0.007 AV_s:Gr_200 0.018 0.003 6.555 0.000 Legend for color coding: Yellow Dominant variable in the model Green Air voids variable with positive value Grey Air voids variable with negative value

79 Figure 2-37. Comparison of rehabilitated pavement ride performance between LTPP measurements and regression model predictions. Analysis Method 3: Artificial Neural Network Modeling Methodology for Development of Performance Models The ANN approach is a computer-based adaptive information processing technique that allows for establishing correlations between the input variables Xi and the output variables Yj through the inter- connected neurons (i.e., weight factors, wji). The input variables Xi and the output variables Yj are usually normalized using a minimum-maximum strategy to xi and yj respectively. Normalizing the variables allows a more equal weighting between input variables and improves the stability of the model. Normalizing the variables is accomplished by converting each value along a linear scale from -1 (the minimum value) to +1 (the maximum value). The correlations developed by the ANN models between the normalized input parameters xi and the normalized output variables yj are shown in Equation 14. 1 j n j ji i i by f w x =   +    =  (14) where f is a transfer function, normally a sigmoidal, Gaussian, or threshold functional form; wji are unknown weight factors; and bj is an unknown bias term. Specifically, the ANN model adjusts the weight factors wji and bias bj in Equation 14 based on a minimum error function. In pavement engineering, the ANN approach is usually used to develop prediction models based on a large number of data (typically over 1000 data points) collected from historical databases, experiments, and numerical analyses (Gu et al., 2018). In this study, LTPP pavement sections were carefully examined, and those with comprehensive information of traffic, climate, material, structure, and long-term performance were identified. The identified pavement sections were then randomly divided into the two clusters, namely training-validation cluster and test cluster, with a ratio of 9:1 in terms of the number of sections. The training-validation cluster was used to develop the ANN models and the test cluster was selected to examine the prediction accuracy of the models. Table 2-28 shows the number of pavement sections used for training-validation to develop each performance model. The newly constructed and rehabilitated pavement sections were separated for

80 analysis because they required some different model inputs. Note that each pavement section contained multiple data points (observations) for historical climate, traffic, and long-term performance. For newly constructed pavement sections, the training-validation dataset was further divided into a training subset and validation subset with a ratio of 7:3 in terms of the number of data points. For rehabilitated pavement sections, the ratio of training subset to validation subset was assigned as 8:2. For each run of ANN code, the training and validation data were randomly divided by the set ratio. The model is established by numerous runs of ANN code, which prevent from data skewing in the random selection. Table 2-28. Number of pavement sections and data sets development of ANN models. ANN Performance Model No. of Training-Validation Sections (Observations) No. of Test Sections (Observations) Rutting – New T=124, V=53 (T=885, V=379) 22 (131) Rutting – Rehab T=112, V=28 (T=690, V=172) 15 (82) Fatigue Cracking – New T=131, V=56 (T=846, V=363) 22 (112) Fatigue Cracking – Rehab T=117, V=29 (T=564, V=141) 17 (82) Transverse Cracking – New T=135, V=58 (T=774, V=332) 22 (141) Transverse Cracking – Rehab T=118, V=29 (T=569, V=142) 17 (79) IRI – New T=125, V=54 (T=872, V=142) 20 (142) IRI – Rehab T=122, V=31 (T=830, V=207) 16 (118) As shown in Figure 2-38, a four-layered ANN architecture consisting of one input layer, two hidden layers, and one output layer was constructed to explore the influence of as-constructed asphalt pavement AV on pavement performance. The input variables contained the traffic, climate, material, and structure characteristics, and the output variables included the pavement performance data (rut depth, percent wheel path cracking, length of transverse cracking, and roughness). Each ANN performance model had slightly different input parameters and number of neurons in hidden layers. Table 2-29 presents the input parameters and Table 2-30 identifies what input variables are required for each ANN model. Table 31 shows the ANN architecture for each performance model. For example, the model architecture 14-25-5-1 represents an ANN model that has one input layer with 14 variables, two hidden layers with 25 neurons in the first hidden layer and 5 neurons in the second hidden layer, and one output layer with one output variable.

81 Figure 2-38. Illustration of four-layered neural network architecture. Table 2-29. Input variables of ANN performance models. Variable Description Units t Pavement Age yr T Average Annual Temperature °F P Average Annual Precipitation in FI Average Annual Freezing Index °F-days TR Average Annual Traffic Volume kESAL hac Total Asphalt Layer Thickness in hOL Asphalt Overlay Thickness in Pbw Asphalt Content of Asphalt Layer, Weighted Average % mix wt Pbt Asphalt Content of Top Asphalt Lift % mix wt AVw As-Constructed Air Voids of Asphalt Layer, Weighted Average %Gmm AVt As-Constructed Air Voids of Top Asphalt Lift %Gmm PNo.4 Aggregate Passing Sieve No.4 % PNo.200 Aggregate Passing Sieve No. 200 % E130°F, 10Hz Dynamic Modulus of Asphalt Mixture at 130°F and 10Hz ksi E70°F, 10Hz Dynamic Modulus of Asphalt Mixture at 70°F and 10Hz ksi E14°F, 0.5Hz Dynamic Modulus of Asphalt Mixture at 14°F and 0.5Hz ksi hbase Base Layer Thickness in Ebase Composite Base Modulus ksi Esg Subgrade Modulus ksi

82 Table 2-30. Input variables required for each ANN performance model. ANN Models Input Variable Rutting Fatigue Cracking Thermal Cracking Ride New Rehab New Rehab New Rehab New Rehab t × × × × × × × × T × × × × × × × × P × × × × × × × × FI × × × × × × × × TR × × × × × × × × hac × × × × × × × × hOL × × × × Pbw × × Pbt × × × × × × AVw × × AVt × × × × × × PNo.4 × × × × × × × × PNo.200 × × × × × × × × E130°F, 10Hz × × E70°F, 10Hz × × × × E14°F, 0.5Hz × × hbase × × × × × × × × Ebase × × × × × × × × Esg × × × × × × × × Table 2-31. Architecture of ANN performance models. ANN Performance Models Model Architecture Rutting - New Construction 14-25-5-1 Rutting - Rehabilitation 15-15-5-1 Fatigue Cracking - New Construction 14-25-5-1 Fatigue Cracking - Rehabilitation 15-15-5-1 Thermal Cracking - New Construction 14-25-5-1 Thermal Cracking - Rehabilitation 15-20-3-1 Ride - New Construction 14-25-5-1 Ride - Rehabilitation 15-15-5-1 In this study, the transfer function uses a sigmoidal function form as shown in Equation 15. ( ) ( ) 1 1 exp ϕ = + −i i f I I (15) where iI is the input quantity, and ϕ is a positive scaling constant. The training algorithm used the Levenberg-Marquardt back propagation method to minimize the mean squared error (More, 1978). The gradient descent weight function was employed as a learning algorithm to adjust the weight factors wji (Amari, 1998). The ANN models were programmed using the Matlab software, which cannot be expressed as an equation in a report.

83 Figures 2-39 through 2-42 show the comparisons between the measured pavement performance in the LTPP dataset and the predicted results from the ANN models for training and validation. In general, the ANN model predictions are in good agreement with the measured pavement performance, indicating that the ANN models reasonably predict the various pavement performance after the training and validation process. a. New Construction Rutting Model b. Rehabilitation Rutting Model Figure 2-39. Comparisons of rutting performance between LTPP measurements and ANN model predictions.

84 a. New Construction Fatigue Cracking Model b. Rehabilitation Fatigue Cracking Model Figure 2-40. Comparisons of fatigue cracking performance between LTPP measurements and ANN model predictions. a. New Construction Transverse Cracking Model

85 b. Rehabilitation Transverse Cracking Model Figure 2-41. Comparisons of transverse cracking performance between LTPP measurements and ANN model predictions. a. New Construction IRI Model b. Rehabilitation IRI Model Figure 2-42. Comparisons of ride performance between LTPP measurements and ANN model predictions.

86 In this study, 10% of the pavement sections were randomly selected for the test set and the data of these pavement sections were not used for the development of ANN models. Figures 2-43 through 2-46 demonstrate the pavement performance prediction accuracy of the developed ANN models using the test data over the first ten years of measured performance. Although the ANN models show lower prediction accuracy for the test set than the training-validation set, they still exhibit better prediction accuracy compared to the regression models created in Analysis Method 2. The low prediction accuracy of the ANN models may be attributed to the relatively small amount of data representing the diverse parameters of the LTPP national pavement database and the inherent variation in the measurement of some input and output variables. Similar to the AASHTO PavementME national pavement design program, the models should improve with a narrower dataset. a. New Construction Rutting Model b. Rehabilitation Rutting Model Figure 2-43. Prediction accuracy of ANN rutting models using test data.

87 a. New Construction Fatigue Cracking Model b. Rehabilitation Fatigue Cracking Model Figure 2-44. Prediction accuracy of ANN fatigue cracking models using test data.

88 a. New Construction Transverse Cracking Model b. Rehabilitation Transverse Cracking Model Figure 2-45. Prediction accuracy of ANN transverse cracking models using test data. a. New Construction Ride Model

89 b. Rehabilitation Ride Model Figure 2-46. Prediction accuracy of ANN ride models using test data. Predicted Influence of As-constructed Air Voids Using ANN Models In the LTPP dataset, each section has a unique combination of input variables, which results in a unique performance curve. The research team attempted to investigate the influences of as-constructed AV on pavement performance by examining the predicted performance of each input section. The approach used the following steps: 1) assemble the input data of each LTPP pavement section in this study; 2) vary as- constructed AV input values from 3% to 9% for each section; 3) predict pavement performance from one to ten years using the developed ANN models; and 4) average the predicted pavement performance for each combination of air void and pavement age. Using this approach, the established relationship between as- constructed AV and pavement performance would be representative for most of the pavement sections. Figures 2-47a and 2-47b present the influence of as-constructed AV on rut depth of new and rehabilitated pavements, respectively. Compared to the newly constructed pavements, the rehabilitated pavements have a slightly lower rut depth on average. In general, lowering AV reduces the rut depth of asphalt pavements; however, this influence is not practically significant. At the 10th year, decreasing AV from 9% to 3% only reduces the rut depth of new pavements by 0.05 inch and that of rehabilitated pavements by 0.015 inch. Based on these predicted performance model outputs, the as-constructed AV may not be a significant factor affecting the rutting of asphalt pavements, in particular for those rehabilitated pavements. Rutting performance can also be stated as the duration of time to achieve the same level of rutting. In the case of new construction using a critical rut depth of 0.25 inches, pavements constructed at 9% as- constructed AV would experience 0.25-inch rut depth at five years and pavements at 3% as-constructed AV would experience 0.25 inches of rutting at eight years. Using these values, reducing as-constructed AV for new construction from 9% to 3% improved rutting performance life 60%. For rehabilitated pavements, basing performance on an increment of time was not reasonable because the amount of rutting never reached the 0.25-inch critical value. The 10-year performance window for rutting was adequate to sufficiently predict performance trends to satisfy the study objective. Both models correctly predicted approximately 0.15 inches of rutting immediately after construction due to surface compaction by traffic.

90 a. New Construction Rutting Model b. Rehabilitation Rutting Model Figure 2-47. ANN model predicted influence of air voids on rut depth. Figures 2-48a and 2-48b show the influence of as-constructed AV on fatigue cracking of new and rehabilitated pavements, respectively, using a 0-100% fatigue cracking scale. Overall, the reduction of as- constructed AV decreases the fatigue cracking of asphalt pavements. For new pavements after ten years, the fatigue cracking increases from 37% to 46% when as-constructed AV of asphalt increase from 7% to 9%. However, AV from 7% to 3% had negligible effect on the fatigue cracking of new pavements. For rehabilitated pavements, the decrease of AV was effective in improving their fatigue performance. Ten years after rehabilitation, decreasing AV from 9% to 3% results in the reduction of fatigue cracking from 45% to 20%. This corresponds to an approximate 4-5% increase in fatigue performance of asphalt pavements for a 1% decrease in as-constructed AV. Fatigue performance can also be stated as the duration of time to achieve the same level of cracking. In the case of new construction, pavements constructed at 9% as-constructed AV would experience 35% cracking at 7.5 years and pavements at 7% as-constructed AV would experience 35% cracking at 9 years. Using these values, increasing as-constructed AV from 7% to 9% reduced fatigue performance life by 17%. For rehabilitated pavements, basing performance on an increment of time was significant. At 20% reflected fatigue cracking, the years of performance for 9% as-constructed AV were only 4 years and for 3% as-

91 constructed AV the years of performance were 10 years, so the higher AV reduced reflected fatigue cracking life by 60%. The 10-year performance window for fatigue cracking could be considered too short, but the predicted performance trends are developed sufficiently to satisfy the study objective. a. New Construction Fatigue Cracking Model b. Rehabilitation Reflected Fatigue Cracking Model Figure 2-48. ANN model predicted influence of air voids on fatigue cracking. Figures 2-49a and 2-49b demonstrate the influence of as-constructed AV on transverse cracking performance of new and rehabilitated pavements, respectively. For the newly constructed pavements, the influence of as-constructed AV of asphalt on transverse cracking was negligible. Compared to the new pavements, the rehabilitated pavements had much more transverse cracking over the 10-year prediction period. For the rehabilitated pavements, the decrease of AV was effective in improving transverse cracking performance. Specifically, decreasing AV from 9% to 3% reduced transverse cracking by 110 ft/mile (approximately 10 cracks per mile) on average for a 10-year-old rehabilitated pavement. Transverse cracking performance can also be stated as the duration of time to achieve the same level of cracking. In the case of rehabilitated pavements, pavements constructed at 9% as-constructed AV would experience 300 ft/mi transverse cracking (approximately a 200-ft crack spacing) at 4.5 years and pavements

92 at 3% as-constructed AV would experience 300 ft/mi at 6.5 years. Using these values, decreasing as- constructed AV from 9% to 3% improved transverse cracking performance life by 45%. The 10-year performance window for transverse cracking could be considered too short, particularly for new construction. The predicted performance trend for rehabilitated pavements was developed sufficiently to satisfy the study objective. a. New Construction Transverse Cracking Model b. Rehabilitation Transverse Cracking Model Figure 2-49. ANN model predicted influence of air voids on transverse cracking. Figures 2-50a and 2-50b demonstrate the influence of as-constructed AV on ride performance of new and rehabilitated pavements, respectively. As presented, both the new and rehabilitated asphalt pavements exhibited the highest IRI when the as-constructed AV of asphalt surface layer was 9%. The roughness reached a minimum at the optimal air void content of 4-5% as shown in Figure 2-51. This finding was consistent with Choi et al. (2004) that the lowest roughness of asphalt pavement corresponds to the air void content of 5%.

93 Ride performance can also be stated as the duration of time to achieve the same level of roughness. However, the average quality of ride predicted for both new construction and rehabilitated pavements at all as-constructed AV are below an IRI value of 90 and are considered good performing pavements. It would be unreasonable to compare performance time with these predicted values. The 10-year performance window for ride could be considered too short, but the predicted performance trends are developed sufficiently to satisfy the study objective. Both models correctly predicted approximately 60 to 70 in/mi of ride immediately after construction. a. New Construction Ride Model b. Rehabilitation Ride Model Figure 2-50. ANN model predicted influence of air voids on ride performance.

94 a. New Construction Ride Model b. Rehabilitation Ride Model Figure 2-51. ANN model predicted influence of air voids on ride performance at the 10th year. As noted earlier, the low prediction accuracy of the ANN models may be attributed to the measurement variation of the input and output parameters. Figure 2-52 shows a set of rutting performance curves of three LTPP pavement sections. As illustrated, the measured rut depth values fluctuate with pavement age, which conflicts with the fact that rut depth cannot reduce over time when no rehabilitation is performed. However, as shown in Figures 2-47 through 2-50, the predicted performance indicators (i.e., rut depth, fatigue cracking, transverse cracking, and IRI) always increase with the pavement age. This indicates that the ANN model development process successfully reduced the noise induced by measurement variation.