Calibrated uncertainty for molecular property prediction using ensembles of message passing neural networks

Abstract Data-driven methods based on machine learning have the potential to accelerate computational analysis of atomic structures. In this context, reliable uncertainty estimates are important for assessing confidence in predictions and enabling decision making. However, models can produce badly calibrated it is therefore crucial detect handle carefully. work we extend a message passing neural network designed specifically predicting properties molecules materials with probabilistic predictive distribution. The method presented paper differs from previous by considering both aleatoric epistemic unified framework, recalibrating distribution unseen data. Through computer experiments, show that our approach results accurate molecular formation energies well out training data two public benchmark datasets, QM9 PC9. proposed provides general framework evaluating ensemble able estimates.