机器学习势在含能材料分子模拟中的研究进展

A comprehensive review was conducted on the historical development, construction scheme, and training strategy of the machine learning potential. This novel technique approximates the potential energy surface of molecular system at the level of first-principle calculations and has been successfully applied in the molecular modelling of combustion and explosion for energetic materials, including nitramine compounds (RDX, CL-20, and ICM-102), oxidizers (AP) and high-energy particles (Al, B). In addition, the representative applications of machine learning potential on the combustion of hydrocarbon fuels were also introduced. Furthermore, the challenges and future development perspectives of machine learning potential in energetic materials were discussed. It is well demonstrated that machine learning potentials, particularly deep potential models, are highly accurate and efficient. The data-driven approach makes it feasible to empower the simulation of million atoms with a great accuracy as first-principle calculations. In the final remark, the key challenges for the further development of machine learning potentials are discussed: sampling issue for a complex potential energy surface under extreme conditions; accuracy problem in the training dataset. With 91 references.