Machine learning molecular simulations with physical constraints

点击翻译按钮获取中文摘要

Molecular simulations have become a powerful approach, highly complementary to experiments, to interrogate the structure and dynamics of macromolecular and colloidal systems. The resulting datasets are however typically highly dimensional, complex and noisy, making the extraction of useful information difficult. Artificial intelligence techniques such as deep learning are remarkable in their capacity of extracting and presenting useful information even in situations where humans would struggle. This project aims at the development of artificial intelligence-based methods for the analysis of molecular simulations and generation of models that are faithful to available experimental data, and consistent with physical laws. We will investigate the usage of generative models to enhance the sampling of molecular conformational space. These models will be exploited to discover new states and identify the specific transition path between two known states. Potential applications of the developed computational tools range from protein folding and colloidal self-assembly to formulation science.