Kris Reyes, University at Buffalo
Decision-Making Under Uncertainty for Multi-Stage Experimental and Computational Pipelines
Written by Don Monroe
Reyes described a goal to “meaningfully apply machine learning to materials science,” for example using nested materials/device codesign or exploiting autonomous experimentation. In the current work, he analyzed the common situation of sequential filtering to choose top contenders from a large initial library of candidates, such as materials or drug molecules.
Frequently such screening will need to weigh various desirable properties. Reyes considered a different scheme, in which the various stages measure the same kind of property but become more accurate (and expensive) as the field is winnowed down. The screening may include both computational evaluations and experimental assays, ending with some kind of “gold standard.”
Results for simulations of different screening protocols were not always intuitive. Some of the most important decisions were how to allocate the tests between different stages.