Abstract: Solvation, charge transport and reactions in complex environments play a vital role in many atmospheric, material, and chemical processes. In addition to experiments, computational tools can significantly contribute to our fundamental understanding of these phenomena. For example, first principles molecular simulations based on density functional theory (DFT) have significantly contributed to uncover molecular level picture of many complex chemical phenomena and new material design. However, the use of DFT (let alone more accurate quantum calculations) is limited to small system size (thousands of atoms) and time scales (few pico seconds) due to high computational cost. Recent advancement in machine learning has opened a new window to overcome this problem. Modern machine learning techniques can be used to learn the interatomic forces generated by quantum level of theory and thereby extend the time scale and length scale of standard AIMD simulations significantly.
In this talk, I will discuss our recently developed protocol to learn quantum level interatomic forces efficiently via an artificial neural network. I will show two instances where machine learning accelerated molecular simulations to discover unprecedented molecular details. First, I will discuss in detail a case study of modeling chemical reactions in complex heterogeneous environment where solvation, polarization and reactions play a key role. Secondly, I will discuss our ongoing work on modeling charge and defect transport in strongly correlated perovskites rare earth nickelates. Finally, I will discuss current limitations and prospective avenues to improve those so that machine learning models would be used on a routine basis in molecular simulations.