Automated Learning of GNN Ensembles for Predicting Redox Potentials with Uncertainty

Accurate prediction of redox potentials of iron (Fe) complexes, in tandem with uncertainty quantification, is essential to advance technologies related to electro-deposition and energy storage by enabling reliable modeling, guiding experimental design, and improving the efficiency of material discovery. Since experimental measurements and first-principles simulations are time-consuming, machine learning (ML) based on existing high-fidelity data is a fast alternative. In this contribution, we present a general ML framework for automated learning of Graph Neural Network (GNN) ensembles for predicting redox potentials of Fe complexes and quantifying the corresponding uncertainty. Using a recently reported high-fidelity dataset of transition metal complexes by Janet et al. [ACS Cent. Sci. 2020, 6, 4, 513–524], we leverage the DeepHyper hyperparameter optimization framework to optimize the hyperparameters of a user-defined GNN model to obtain a large set of high performing models in an automated way. This set of candidate models is then used to generate ensembles. We demonstrate that the proposed ensemble approach achieves superior performance in comparison to a single best-performing GNN model and also over a varied set of ML model classes applied to a standard benchmark. Further, the proposed framework offers insights into the confidence of model predictions through epistemic uncertainty measures. In fact, we demonstrate that through effective uncertainty quantification, our framework can be used to identify molecular structures and attributes that significantly affect model performance, particularly in challenging datasets with complex ligand environments. This general framework opens the door to a scalable and accurate high-throughput screening of transition metal complexes, facilitating the rapid discovery of new materials for energy and catalysis applications.