Deterministic global optimization for sample-efficient molecular design with generative machine learning

Generative machine learning (ML) models such as variational autoencoders (VAEs) learn continuous molecular latent spaces that can facilitate the exploration of novel molecules and materials. However, such latent spaces are typically high-dimensional, making targeted molecular optimization challenging. We therefore propose deterministic global optimization of molecular property prediction models in the form of artificial neural networks (ANNs) trained on VAEs' latent spaces. By using ANNs with ReLU activations, we formulate molecular design as a mixed-integer linear program (MILP) guaranteeing optimal molecular properties, as predicted by the ANN. Our results show superiority of the identified molecules with global optimal predicted property values compared to those found with frequently-used optimization strategies such as Bayesian optimization. Our approach thus enables finding the most promising molecules/materials according to the ANN predictions for subsequent investigation in simulations/experiments, thereby increasing the sample efficiency of ML-guided molecular design.