Haoxiang Guan

This project is an innovation and practice program jointly organized by the University of Science and Technology of China and MSRA for junior students. In this project, I researched and implemented numerical methods for solving system dynamics equations, and created an N-body problem system that simulates the interaction of any number of particles under gravitational force. On this basis, I manually derived the force field formula of molecular dynamics, built a molecular dynamics (MD) simulation system from scratch, and verified its simulation accuracy through experiments, which was consistent with that of dedicated MD software. On the other hand, in the project, I manually derived the backpropagation formulas for various operations in neural networks, implemented the Transformer model from scratch using CuPy, and verified it on the machine translation task. Finally, I explored the application of molecular anchor models in MD simulations, successfully integrated the knowledge of the pre-trained chemical language (SMILES) model MoLFormer into the equivariant Transformer model for molecular property prediction by designing a novel attention mechanism for feature fusion, and achieved better prediction performance on the QM9 dataset.