Machine Learning Assisted Simulation and Experimentation for Accelerated Scientific Discovery

Addressing the global energy challenge requires breakthroughs in both clean energy generation and the optimization of manufacturing processes. The rapid advancement of machine learning (ML) has created new opportunities for accelerating scientific discovery, particularly in materials modeling and experimental workflows. ML has been integrated into molecular simulations to reduce computational costs and into experimental design to enhance process efficiency. However, significant challenges remain. In computational materials science, quantum mechanical methods such as density functional theory (DFT) are essential but computationally expensive, limiting their applicability to large-scale simulations. On the experimental side, optimizing process parameters is constrained by time and resource limitations, making trial-and-error approaches impractical.

In this thesis, we introduce Finetuna, a workflow that leverages pretrained ML models to accelerate atomistic simulations. We construct a computational dataset for perovskites containing neutral defects and systematically benchmark the performance of machine learning interatomic potentials. Additionally, we present a data-driven framework for process window optimization, where a constrained search strategy efficiently explores the parameter space while minimizing the number of required experimental trials. By integrating ML into both computational and experimental workflows, this work contributes to the broader effort of accelerating materials discovery and process optimization.