Historically, chemistry was based on generalizations from a relatively small number of chemical examples. The quantitative models were simple, in part because there were not enough training data to determine many parameters, and also very few test data. Larger data sets have become available, either by extracting data from many journal papers or patents, and or from high-throughput experimental and computational chemistry methods. This suggests it may be fruitful to switch to more flexible models to achieve better coverage of chemical space and also more accuracy in quantitative predictions. Here we present a variety of models members of my group have constructed using Machine Learning methods, to predict a variety of different chemical quantities (e.g. thermochemical parameters, solvation parameters, spectra, reaction barriers, transition state geometries, regioselectivity). We examine the models with a critical eye, with attention to quantitative errors in the model predictions and errors in the uncertainty estimates. We also discuss technical and practical limitations of the approaches employed. One promising approach for handling the paucity of accurate experimental data is to train a model on a large low-accuracy data set (e.g. from quantum chemistry calculations on many molecules or reactions), and then calibrate the model using the much smaller set of available experimental data, a type of Transfer Learning. Some of the many remaining challenges are outlined.

Challenges in quantitatively predicting chemistry with ML methods

ACS Fall 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.