The total amount of numerical information generated in a typical computer simulation is immense. Whether the model system under study is inorganic or biological, classical or intrinsically quantum mechanical, in thermal equilibrium or violently nonstat ionary, the processing of the data created is usually structured to address a small set of often conceptually orthodox questions that have been selected beforehand. Given the inappropriateness of trying to examine raw numerical information in tabular for m, this is entirely reasonable. Nevertheless, serendipitous discoveries that fall outside the preselected query set indeed do occasionally occur, and their consequences can produce significant advances.
The usual procedures for carrying out computer simulations probably are far from optimal in detecting initially unsuspected correlations and phenomena. This leads one to suspect that generic pattern-recognizing algorithms might be constructed that would operate in conjunction with Monte Carlo or molecular dynamics simulations and would effectively amplify the human scientist's limited capacity to pull novel correlations out of huge files of numerical data. Generally speaking, humans are best at recogni zing visual patterns--after all, the human visual apparatus has been honed to exquisite sensitivity by approximately a billion years of evolutionary trial. But in the scientific context this requires that just the right graphs or figures be presented to the investigator, and in one respect it presupposes something about the nature of the correlations to be examined. One hopes to transcend such human limitations, perhaps with a carefully crafted form of open-ended artificial intelligence that incorporate s pattern-detecting capacity free from our human biases and limitations. In particular, computers ought to be able to "see" in spaces of dimension higher than three and pick out significant patterns whose projection into two or three dimensions would obs cure the phenomena of interest.
This opportunity is not so much intended as an attempt to replace human researchers as it is to magnify their wisdom and insight. In this respect, generic pattern-detecting software could be a powerful device operating to maximize productivity. And it should not escape notice that such a tool is not restricted just to simulation data, but in principle could also apply to other classes of large chemical databases discussed elsewhere in this report.
NAS Home Page | NAP Home Page | Reading Room | MCTCC Home Page