Accelerating Quantum Chemistry Simulations Using GPU-Optimized Deep Reinforcement Learning Algorithms

Abstract

A distributed deep reinforcement learning framework based on GPU cluster was constructed to solve the computational problems of large-scale quantum chemistry simulation. The framework includes technologies such as parallel strategy optimization, asynchronous gradient update, and dynamic scheduling tasks. Efficient communication algorithms are adopted to maximize the utilization of hardware resources and achieve scaling. Experimental evaluation of representative quantum chemistry problems was performed on a multi-node GPU platform. The results show that the average GPU utilization rate exceeds 92% due to the near-linear scale expansion to 128 GPUs. This improves the simulation throughput and efficiency. The system has good performance and convergence behavior for a variety of workloads. The results show that these GPUs, matched with advanced improved distributed DRL models, can break thru the traditional bottleneck of using supercomputers to solve complex physical and chemical problems, so as to complete all these dazzling and complex molecules faster and better.