Accelerating Quantum Mechanical Simulations Using Physics-Based Machine Learning Potentials

Quantum mechanical modeling such as density functional theory (DFT) have allowed chemists to computationally explore the properties of materials. However, these methods still require immense amounts of computational power and time. To save resources and time, we developed an offline active learning method for physics-based machine learning potentials. By training a model through active learning, the total number of first principle calculations required for a simulation can be curtailed. Moreover, a Morse potential based delta- machine learning approach prevents the generation of unpysical structure configurations, allowing for a more stable training than traditional pure machine learning approaches.