Across Dimensions and Scales: How Imaging and Machine Learning Will Help Design Tomorrow’s Energy Conversion Devices
MARIANA I. BERTONI
Arizona State University
Energy is the single most important factor that impacts the prosperity of any society, underpinning advances on which all depend. To supply the more than 7 billion people on this planet at the level of energy that the developed world is accustomed to, 60 terawatts would need to be generated—the equivalent of 900 million barrels of oil per day. Where could this astonishing amount of energy come from? The term “terawatt challenge” refers to the endeavor to produce energy at the level needed in an economically, socially, and technologically sustainable way.
When one searches for potential sources of energy at the terawatt (TW) scale it is striking to find that the biggest resources and most technically exploitable options are those that barely make up 10 percent of the energy mix today—solar, wind, and geothermal. If the TW challenge is solved, though, the world energy breakdown will look quite different by 2050.
It is not unreasonable to think that renewables can handle heavy loads. Reports have proposed that 100 percent of the world’s energy needs by 2030 (11.5 TW) can be fully achieved with an energy mix of roughly 50 percent wind, 40 percent solar (concentrated solar power and photovoltaic [PV]), 4 percent hydroelectric, 4 percent geothermal, and 1 percent tidal turbines (Jacobson et al. 2015).
No renewable energy source is as abundant as the Sun, and in recent years its potential has been capitalized to the point that solar has moved from a niche source toward a mainstream electricity generation source with grid parity. With 15 gigawatts (GWDC) installed, solar was the number one source of new US capacity in 2016—an unprecedented 39 percent. That momentum carried into 2017, as solar accounted for 30 percent of all new electric capacity installed in
the first quarter. Global PV shipments reached an astonishing 75 GW in 2016, arguably making the solar industry the largest optoelectronic sector in the world, worth $110 billion/year (Perea 2017).
The aspirational goal set by the US Department of Energy Sunshot initiative to meet $1/watt by 2020 initially seemed unrealistic and even comical to many in the industry (US DOE 2010). But 6 years later and 3 years ahead of schedule, module prices have dropped to $0.99/watt for fixed-tilt utility-size installations (Perea 2017).
For 18 states the levelized cost of solar energy is below gross electricity bill savings in the first year of a solar PV system’s life. This means grid parity under business-as-usual conditions is a reality, and 32 more states are expected to follow suit by 2020 (Munsell 2016). If PV reaches grid parity it will be an important milestone, but it is just the beginning. High market penetration will require that PV system costs drop to offset the additional costs of storage and transmission so that solar generation can be distributed to meet electricity demand both cost effectively and more broadly in time and space (Kurtz et al. 2016).
The rapid pace of change brings its own challenges and opportunities. For example, there are concerns about the maximum penetration possible with PV power because of its impact on utility demand, lowering its value as PV penetration increases and requiring further cost reductions. In addition, the important metrics of photovoltaics for sustainable energy are expanding to include factors previously not analyzed, such as the impact of capital expenditures on realizing sustained high growth rates (Haegel et al. 2017; Powell et al. 2015).
Technological barriers to PV have in some ways increased. Cost reductions from economies of scale are plateauing, the cost of photovoltaics is a moving target, and efficiency from single-junction technologies is approaching its technological limits, hampering the ability to use efficiency boosts as the lever to overcome previous barriers.
In this context, like Moore’s Law, an underlying law based on fundamental physics can help make a specific, quantitative prediction about innovation as a function of time. For semiconductors, the technical parameter has been transistor density; for photovoltaics the analogue is energy produced per unit volume.
Figure 1 puts a lot of this discussion in perspective. It shows that computing and photovoltaics have seen significant and steady cost reductions during the past 20 years by “packing more in a smaller volume,” while oil and natural gas have remained relatively constant despite shorter-term price fluctuations. It also depicts how competitive today’s solar energy prices are (light blue inset data in $/kWh) compared to other electricity sources, that unlike solar benefit from federal and state subsidies. The achievement of silicon module prices below $1/W—and projections of $0.50/W—has fundamentally changed solar R&D.
Slim margins have pushed some companies into bankruptcy and challenged the annual profitability of others. Cost, intermittency, and dispatchability have been major challenges in the pursuit of utility-scale solar energy generation. More
recently, materials’ degradation studies and long-term system performance R&D have become crucial for the bankability of projects. However, the standard business model of the solar industry, with each company eager to outcompete the next in price, has made the industry very risk averse when it comes to implementation of innovation.
What are the next steps? Movement toward an “electric-powered world” and increased demands for clean and efficient electricity (e.g., electric vehicles, portable electronics, rural electrification) raises new challenges.
The first of these challenges concerns portability: the availability of lightweight and flexible modules necessary for implementation in everyday life. Additional challenges involve the achievement of high power in small areas and the use of sustainable materials for device manufacturing.
As with other consumer applications, solar margins will improve and engineering hurdles associated with aesthetics, customization, and functionality will become standard R&D considerations. An analogy is the introduction of Ford’s Model T car: photovoltaics has demonstrated its affordability, impact, and potential, and now a whole new technology is taking off.
The path for improved PV is to make cells thinner and more efficient. The industry is maturing, costs are becoming dominated by those of materials, and
expensive process changes are yielding very small incremental benefits. Fundamental scientific breakthroughs are necessary to propel this energy source to next-generation levels.
Higher-power cells can be achieved by stacking cells with different band-gaps to efficiently capture a wider portion of the solar spectrum. The efficiency limits rise from 33 percent for a single-junction cell to 43 percent for two junctions under no concentration, 49 percent for three junctions, and 66 percent for all greater numbers of junctions. This approach is not novel; multijunction cells are well known in space applications, where very high quality single crystals are epitaxially grown and cells are engineered to withstand radiation and high levels of illumination (Takamoto et al. 2005). The automotive analogy would be a limited-edition Ferrari.
Although these modules are very expensive, epitaxial liftoff techniques enabling substrate reuse have demonstrated a path to lower costs. The future of solar lies in merging the ubiquitousness of the Model T solar cells with the performance of the limited-edition Ferrari cells.
The first thing to realize is the necessity of relying on the mass-production low-cost manufacturing lines of the Model T cells, which most likely means a silicon cell will be the bottom cell and high-quality single-crystal films will not be available for the multilayer stack. Instead faster deposition methods, like evaporation or sputtering, and defect-engineered top films will have to be used to achieve the desired electrical and optical properties (Bobela et al. 2016). This is crucial to the success of next-generation solar absorbers; engineered defect–tolerant materials are the best way to enable ultra-low-cost manufacturing technology for high-efficiency devices. A top-cell bandgap of 1.7 electron volt (eV) and an efficiency comparable to standard silicon cells today (20 percent) can enable 32 percent efficient tandems (Yu et al. 2016).
The task seems daunting, especially when one considers that the performance of a full device is usually governed by the concentration and distribution of nanoscale inhomogeneities and defects throughout the solar cell.
How can discovery and defect engineering be accelerated to facilitate a high-power, portable, and economic solar industry? The paradigms for materials discovery have to be redefined, especially for systems with complex functionalities, to move beyond serendipitous discoveries, Edisonian approaches, and the classical synthesis-characterization-theory methods. The answer lies in highly correlative imaging methods under operating conditions combined with big data analytics.
Understanding the fundamental relationships between composition and structure properties on a nanopixel basis, under real operating conditions and in situ (with both controlled and ambient temperature) is necessary to unravel the causes and effects of certain defects, including their impact on performance.
Current imaging techniques do not merely provide a picture of the system under study, they actually contain compositional, structural, and functional infor-
mation. The correlation of multiple 2D or 3D mapping modalities on a pixel-to-pixel basis and the multiple dimensions of these maps, based on time, temperature, and ambient conditions, creates a big data challenge.
In situ and operando measurement techniques combined with nanoscale resolution have proven invaluable in multiple fields of study. I argue that correlative hard X-ray microscopy (HXM) with <100 nm resolution can radically change the approach for optimizing solar absorbers, interfaces, and full devices in solar cell research.
Unlike other fields of microscopy, HXM has excellent penetration through layers and entire devices, yielding 3D imaging of buried structures. It can easily penetrate gases and fluids, enabling studies at pressure and under process conditions. It also enables quantitative studies of sample composition with trace element sensitivity in structured materials and devices. Chemical state information of individual atomic species can be obtained using X-ray spectroscopic techniques. X-rays do not interact with external fields and are therefore useful for studies in electric or magnetic fields (Stuckelberger et al. 2017).
As acquisition speeds and resolution increase, providing more density of data points, and the functionality of measurements adds more dimensions to be analyzed, the handling, management, and analysis of datasets become more and more complicated. Operando measurements as well as in situ studies pose a new challenge: Finding correlations in the 3D+ datasets that result from many of these measurements is not straightforward, and the possibility of missing connections, relationships, and trends is cause for concern.
Machine learning techniques, including principal component and cluster analyses, have been widely used in fields plagued with tremendous amounts of data (Hastie et al. 2013). A key benefit of these approaches is the ability to identify trends in highly dimensional data, a task that is otherwise difficult and sometimes even impossible.
The first step toward full information recovery from high-resolution multifunctional imaging data is the adoption of big data analytics (Kalinin et al. 2015; Rajan 2012, 2015; Runkler 2016). This requires implementation of dimensionality reduction, clustering techniques, and statistical unsupervised learning (Hastie et al. 2013). Unsupervised image analysis tools targeted to high-performance computing platforms can analyze high-resolution scanning and electron microscopy data in 2D in real time (Belianinov et al. 2015). Advances in high-resolution experimental imaging and high-performance computing (Dongarra et al. 2011) will undoubtedly propel materials discovery and ultimately “materials by design.”
However, just because a statistical correlation is observed does not imply an accurate understanding of the underlying physics in the multimodal imaging. Transitioning from big data to “deep data” is the next step. All the structure-property relationships at the nanoscale retrieved from big data can be examined with real physical models, allowing for verification and improvements in predictive modeling (Kalinin et al. 2015). This step makes it possible to close the loop and
propose design guidelines to develop or process a material with desired properties and functionalities.
Materials informatics is ready to lay the foundation for a new paradigm in materials discovery, especially for complex functional systems like solar cells. It could very well end up being data that ultimately push the cost of solar power to the levels of subsidized fossil fuel.
Belianinov A, Vasudevan R, Strelcov E, Steed C, Yang SM, Tselev A, Jesse S, Biegalski M, Shipman G, Symons C, and 3 others. 2015. Big data and deep data in scanning and electron microscopies: Deriving functionality from multidimensional data sets. Advanced Structural and Chemical Imaging 1:6.
Bobela DC, Gedvilas L, Woodhouse M, Horowitz KAW, Basore PA. 2016. Economic competitiveness of III-V on silicon tandem one-sun photovoltaic solar modules in favorable future scenarios. Progress in Photovoltaics: Research and Applications 25(1):41–48.
Dongarra J, Beckman P, Moore T, Aerts P, Aloisio G, André J-C, Barkai D, Berthou J-Y, Boku T, Braunschweig B, and 55 others. 2011. The International Exascale Software Project roadmap. International Journal of High Performance Computing Applications 25:3–60.
Haegel NM, Margolis R, Buonassisi T, Feldman D, Froitzheim A, Garabedian R, Green M, Glunz S, Henning HM, Holder B, and 11 others. 2017. Terawatt-scale photovoltaics: Trajectories and challenges. Science 356:141–143.
Hastie T, Tibshirani R, Friedman J. 2013. The Elements of Statistical Learning. Berlin: Springer Science & Business Media.
Jacobson MZ, Delucchi MA, Bazouin G, Bauer ZAF, Heavey CC, Fisher E, Morris SB, Piekutowski DJY, Vencilla TA, Yeskoo TW. 2015. 100% clean and renewable wind, water, and sunlight (WWS) all-sector energy roadmaps for the 50 United States. Energy and Environmental Science 8:2093–2117.
Jones-Albertus R, Feldman D, Fu R, Horowitz K, Woodhouse M. 2016. Technology advances needed for photovoltaics to achieve widespread grid price parity. Progress in Photovoltaics: Research and Applications 24:1272–1283.
Kalinin SV, Sumpter BG, Archibald RK. 2015. Big-deep-smart data in imaging for guiding materials design. Nature Materials 14:973–980.
Kurtz S, Atwater H, Rockett A, Buonassisi T, Honsberg C, Benner J. 2016. Solar research not finished. Nature Photonics 10(3):141–142.
Munsell M. 2016. GTM Research: 20 US states at grid parity for residential solar. Greentech Media, February 10. https://www.greentechmedia.com/articles/read/gtm-research-20-us-states-at-grid-parity-for-residential-solar.
Perea A. 2017. US Solar Market Insight 2017 Q2. Washington: Solar Energy Industries Association.
Powell DM, Fu R, Horowitz K, Basore PA, Woodhouse M, Buonassisi T. 2015. The capital intensity of photovoltaics manufacturing: Barrier to scale and opportunity for innovation. Energy and Environmental Science 8(12):3395–3408.
Rajan K. 2012. Materials informatics. Materials Today 15:470.
Rajan K. 2015. Materials informatics: The materials ‘gene’ and big data. Annual Review of Materials Research 45:153–169.
Runkler TA. 2016. Data Analytics. Wiesbaden: Springer Fachmedien Wiesbaden.
Stuckelberger M, West B, Nietzold T, Lai B, Maser JM, Rose V, and Bertoni MI. 2017. Engineering solar cells based on correlative x-ray microscopy. Journal of Materials Research 32(10):1825–1854.
Takamoto T, Kaneiwa M, Imaizumi M, Yamaguchi M. 2005. InGaP/GaAs-based multijunction solar cells. Progress in Photovoltaics: Research and Applications 13:495–511.
US DOE. 2010. $1/W Photovoltaic Systems: White Paper to Explore a Grand Challenge for Electricity from Solar. https://www1.eere.energy.gov/solar/sunshot/pdfs/dpw_white_paper.pdf.
Yu ZJ, Leilaeioun M, Holman Z. 2016. Selecting tandem partners for silicon solar cells. Nature Energy 1:16137–16134.
This page intentionally left blank.