Molecular simulation is a method that can reproduce the movement of molecules, and has an extremely wide range of uses such as biological substances, macromolecules, and materials, and is used for the development of new materials and the elucidation of pathology.On the other hand, when performing a large molecule or a long-time simulation, there is a drawback that the calculation is difficult because a large-scale calculation resource is required.

　In this research, we modeled molecular simulation as stochastic time evolution and proposed a new model of deep learning (one of artificial intelligence) for stochastic time evolution.There was a problem that errors were accumulated in the repeated time evolution, but since the proposed model has an error reduction mechanism, it became possible to repeat the time evolution.

　This time, we conducted an application experiment to the phenomenon that the entanglement of polyethylene, which is a polymer, is disentangled.The phenomenon could not be reproduced with the short-time molecular simulation data, but when looking at the long-term data predicted by the proposed model from the short-time data, the phenomenon of untangling is firmly reproduced, and it becomes normal diffusivity. It turned out that there was.

　With this achievement, companies and research institutes conducting research using molecular simulations will be able to reduce the amount of simulation that requires calculation, which will greatly improve the efficiency of research and development.Furthermore, it is expected to be applied to various time series data such as natural language processing, economic data, and motion data.The results of this research were published on the site of the 2018nd Association for the Advancement of Artificial Intelligence (AAAI-4) on April 26, 32 (local time).

Paper information:[AAAI Publications, Thirty-Second AAAI Conference on Artificial Intelligence] Multi-step time series generator for molecular dynamics