Improving Machine Learning Frameworks for Catalyst Simulations

The looming threat of climate change is driving an urgent need to redesign industrial processes worldwide with new and more efficient catalysts. Machine learning methods have the potential to massively accelerate the discovery of novel catalysts, which are extremely costly and therefore slow to discover by traditional methods. Quantum mechanical simulators like Density Functional Theory (DFT) allow computational simulations to be used to discover new catalysts at great computational expense, which prevents large scale DFT efforts from meeting this demand. Graph Neural Networks (GNNs) have emerged as promising machine learned surrogates for DFT, with the capacity to achieve high accuracy when learning from DFT generated data and improve upon computational costs by up to O(10⁵). However, GNNs can suffer from challenges with consistency, performance, and generalization, and require improvements to be sufficient for the task at hand. This dissertation explores the development of improved machine learning frameworks for catalyst discovery in the areas of trustability, efficiency, and adaptability. First I present Finetuna, an open source package for ac celerating catalyst simulations and efficiently acquiring new data with active learning and fine tuning. Second I demonstrate the successes and pitfalls of using Finetuna as a transfer learning strategy to adapt pre-trained GNNs to out-of-distribution structures. Third I benchmark and develop novel im provements to uncertainty quantification methods for GNNs to intelligently identify when to trust the model and when it must be adapted. Finally I ex plore the usefulness of GNNs for determining the potential energy Hessian, an extremely high value task which is related to, but not explicitly repre sented in, its training data set.