Robust Data Generation, Heuristics and Machine Learning for Materials Design
- Despite significant progress, computational materials design still faces challenges, especially when simulating large systems needed to describe defects, interfaces, or amorphous states with the accuracy of density functional theory (DFT) or beyond.[1] To address these limitations, machine learning (ML) methods have become increasingly popular in recent years. In this talk, I will present how we’ve developed robust data generation strategies that support the creation and benchmarking of new ML models.[2] I’ll focus on methods for large-scale quantum-chemical bonding analysis and workflows for ML interatomic potentials. We’ve shown that quantum-chemical bonding properties can be used in ML models to predict phononic properties.[3] This enables us to validate several expected correlations— such as the link between bonding strength and force constants—on a large scale. Additionally, we’ve built an automated training framework for machine-learned interatomic potentials (autoplex).[4]Despite significant progress, computational materials design still faces challenges, especially when simulating large systems needed to describe defects, interfaces, or amorphous states with the accuracy of density functional theory (DFT) or beyond.[1] To address these limitations, machine learning (ML) methods have become increasingly popular in recent years. In this talk, I will present how we’ve developed robust data generation strategies that support the creation and benchmarking of new ML models.[2] I’ll focus on methods for large-scale quantum-chemical bonding analysis and workflows for ML interatomic potentials. We’ve shown that quantum-chemical bonding properties can be used in ML models to predict phononic properties.[3] This enables us to validate several expected correlations— such as the link between bonding strength and force constants—on a large scale. Additionally, we’ve built an automated training framework for machine-learned interatomic potentials (autoplex).[4] Initial workflows include random structure searches, which are well-suited for general-purpose potentials, as well as workflows tailored to ML potentials with accurate phonon properties. While atomistic simulations are highly effective for certain material properties, others— like magnetism or synthesizability—remain difficult. In these cases, it’s promising to benchmark established ab initio methods against chemical heuristics or to develop new ML models based primarily on experimental data.[5,6]…

