Web-Interfaced Molecular Dynamics Simulation Platform with Machine-Learning-Algorithms-Empowered Force-Fields

First-principles molecular dynamics yields accurate predictions but is too expensive for large materials systems, while molecular dynamics with empirical force fields could scale up to very large systems but its reliability is usually questionable. Currently it remains lacking for a web-based platform to enable the development of accurate and efficient algorithms for molecular dynamics simulations collaboratively. This problem can be solved by taking advantages of recent advances in machine learning methods to reproduce potential energy surface of first-principles calculations, and develop a GitHub-like website for sharing simulation data and improving force fields collaboratively. Current forms of neural network are several orders of magnitude faster than first-principles calculations without significant loss in accuracy. Constructing a web- interface platform with a library of neural network model-based force fields and mechanisms for data sharing and version control could be a promising way to address the problem stated above. A library of neural network model-based force fields will be developed for molecular dynamics simulation of silicon materials. The library will be integrated with a website with Git for version control and data sharing. A force field comparison mechanism will be developed and corresponding recommender machine will be provided. The successful development of the proposed system will contribute to the application of machine learning methods as powerful tools in materials science. Molecular dynamics simulation with machine learning method-based force fields will attract broad interests by boosting its application in the design and discovery of new materials with the concept of collaborative improvement. With more interests from academic and industrial communities, more revenues are expected for conducting consulting works. Moreover, several experimentally synthesizable new material designs could potentially generated, which may result in a large amount of revenue from selling or licensing the patents of these designs.