Active Learning for Efficient and Scalable Dataset Creation for Large-Scale Atomistic simulations

AI Technology: Machine Learning | Deep Learning 

Machine Learning Interatomic Potentials (MLIP) are transforming atomistic simulations by combining near quantum mechanical (QM) accuracy with the computational efficiency of empirical force fields. However, the creation of MLIP requires extensive datasets, generated through resource-intensive QM calculations, and poses challenges in ensuring dataset reliability and compactness. Current databases primarily focus on stable equilibrium configurations, lacking the out-of-equilibrium structures necessary for large-scale, non-equilibrium simulations. 

To address these challenges, this project proposes a refined framework for automated, efficient, and transferable dataset generation called Smart Configuration Sampling. SCS employs active learning strategies, utilizing ensemble model deviations to guide the iterative selection of high-value atomic configurations. To enhance generalizability, the project incorporates advanced clustering and stratified sampling based on atomic environment similarity, reducing reliance on specific model architectures. This ensures compact datasets with minimal redundancy and maximum transferability. 

After an initial phase for developing the new sampling feature, the project focuses on systematically generating datasets relevant to tribo-chemical systems, including carbon-, silicon-, metal-based and 2d-materials. By filling the current gap in tribology-specific datasets, this framework will significantly advance simulation accuracy and efficiency, offering a transformative tool for material design and tribological research.