The prediction frontier is the set of problems for which predictive models still evade us. For most of the past century the prediction frontier has moved at a hard-earned but glacial pace, with big breakthroughs occurring slowly through improvements in atomistic modeling, molecular dynamics, numerical methods, and the scaleup in computational resources. But in the past few years, the prediction frontier has moved too quickly to easily define. Protein structures? We can often predict those now. Sequence to function mappings of polymers? We can often predict those, too. For reaction conditions, reaction networks, molecular properties, synthetic pathways, large-scale optimization, and many other areas—Prediction methods are daily improving.
In this presentation, I’ll discuss three prediction problems that are at the frontier of being predictable under useful circumstances. In all three cases, I will argue that we had no realistic pathway to prediction prior to recent developments in machine learning. The common thread for all three is that the maturation of machine learning approaches—combined with the decades of physics-based methods preceding it—puts the field within reach of “solving” each of these problems under useful constraints.
