Abstract

Computer simulations in chemistry, materials- and nano-science generally rely on a trade-off between accuracy and computational speed. Quantum mechanical methods can come close to experimental values, but the computational cost of these methods grows rapidly with system size and complexity. Empirical methods, such as force-fields and coarse-grained models, can calculate properties of larger systems at reasonable timescales but tend to be limited to specific sets of systems. The nascent field of machine learning (ML) poses a different approach to this speed/accuracy trade-off by learning system properties through inference instead of direct calculation. The prediction error of a ML model tends to decrease systematically with the number of compounds used to train the model. Hence, given enough training data, a ML model can in principle reach arbitrary predictive accuracies. I will discuss some of the challenges that are encountered when ML is applied to predict properties throughout chemical space. These challenges include how to represent an atomic environment, which combination of representations and regressors is best suited for a given property, and how response operators can be used to efficiently learn response properties.