Optimization of ReaxFF Reactive Force Field Parameters for Cu/Si/O Systems via Neural Network Inversion with Application to Copper Oxide Interaction with Silicon

• Published in: Journal of physical chemistry. C 2023-10, Vol.127 (41), p.20445-20458
• Main Authors: Roshan, Kamyar Akbari, Talkhoncheh, Mahdi Khajeh, Sengul, Mert Yigit, Miller, David Jonathan, van Duin, Adri C. T.
• Format: Article
• Language: English
• Subjects: C: Physical Properties of Materials and Interfaces
• ISSN: 1932-7447; 1932-7455
• DOI: 10.1021/acs.jpcc.3c03079

The presence of transition metal oxide impurities introduced during crystal formation or during the fabrication process may lead to a significant yield loss in microelectronics and device manufacturing. To enable a large-scale molecular dynamics study of the effects of copper oxide impurities inside silicon on the structural evolution and mechanical properties of Cu/Si/O systems, one needs to understand the diffusional characteristics of copper and oxygen compounds next to the silicon lattice. In this work, we introduce an accelerated deep learning-based reactive force field parametrization platform. In this platform, we train a deep neural network to learn the production of ReaxFF outputs, given a set of force field parameters. Subsequently, the trained neural network is used, as an alternative to ReaxFF, by means of the neural network inversion algorithm to seek the inputs to the neural network (force field parameters) that produce the experimental and quantum mechanics reference property values of the system. We compared the performance of the neural network inversion optimization algorithm with that of the previously used brute force search method by looking at the total optimization time and the total reduction of the discrepancies between the results of molecular dynamic simulation and the reference property values within the force field training set. The neural network inversion algorithm significantly reduces the average optimization time, which directly translates into less computational resources required for the optimization process. Moreover, we compared the quality of the force fields optimized by both algorithms in describing the chemical properties of the Cu/O systems, including the heat of formation and the relative phase stability. We demonstrated that the results of the force field, optimized using the proposed neural network inversion algorithm, align more closely with the reference chemical properties of Cu/O systems within the force field training set than those optimized by the brute force algorithm. We used this platform to develop a Cu/Si/O ReaxFF reactive force field by training on density functional theory (DFT) data, including heat of formation values for various Cu/Si/O materials. The developed force field was further used to perform molecular dynamics simulations on models with up to 3542 atoms to study atomistic interactions between copper oxide compounds and silicon by looking at the diffusional behavior of copper and oxygen