機械学習を用いたモンテカルロ分子シミュレーションの高速化

抄録

Monte Carlo molecular simulation is a powerful and commonly used method for investigating molecular behavior. State updating algorithm greatly affects the efficiency of the simulations. In this work, we propose a novel method to construct highly efficient state updating algorithm using machine learning. Our proposed method contains the following three components: (1) Continuous Normalizing Molecular Flow(CNMF) method, which create various probability distributions which suitable for proposal distribution of Metropolis-Hastings algorithm, (2) a self-learning molecular flow MetropolisHastings (SL-MFMH) algorithm which is unbiased sampler with combining CNMF method, (3) a reverse process selflearning optimization (RPSO) method, which optimizes the efficiency of the SL-MFMH method. We performed experiments to evaluate the efficiency improvement of our method. The results show that the proposed method has about 10 times better performance than the original MH algorithm. We also found that the average displacement of the atoms does not decrease with decreasing acceptance rate using our method, which differs from the conventional method.