Transferable machine learning model for multi-target nanoscale simulations in hydrogen-carbon system from crystal to amorphous

Abstract

Carbon materials especially with hydrogenation have attracted wide attention for their novel physical and chemical properties and broad application prospects. A systematic theoretical simulation method accurately describing atomic interactions for hydrogen-carbon systems is crucial for the design of carbon-based materials and their industrial applications. Multiphases of hydrogenated carbon materials, from crystal to amorphous, with covalent network and diverse chemical reactions bring huge difficulties to construct a general interatomic potential under various conditions. Here, we demonstrate a transferable active machine learning scheme with separated training of sub-feature spaces and target-oriented finetuning, and construct a general-purpose pre-trained machine learning potential (MLP) for hydrogen-carbon systems. The pre-trained MLP is further efficiently transferred to three target spaces of deposition, friction and fracture with scale reliability. This work provides a robust tool for the theoretical research of hydrogen-carbon systems and a general scheme for developing transferable MLPs in multiphase systems across compositional and conditional complexity.