Atomistic Macroscopic Simulations of Collisional Plasmas

This project will use high performance computing to simulate small amounts of very hot matter down to the level of individual atoms, using machine learning to advance current capabilities. Typical mathematical models of matter are approximate because it would almost always take too much time and computing power to solve equations that describe what every individual atom does.

This project will push through that barrier by developing methods to simulate small pieces of matter, including literally every individual atom. The goal is to provide the most accurate description of matter possible. The focus is on extremely hot matter -- plasmas -- which creates unique challenges relative to other forms of matter.

An important aspect of this new approach to simulating matter is that the simulations will observe themselves and learn from their own mistakes, so that their performance continuously improves by this process of "machine learning." Machine learning will enable advances that have never been possible in this area of research. The capabilities developed in this project will be useful in future fundamental and applied research of societal relevance, including industrial and national defense applications, and could lead to advances ranging from understanding how stars age to manufacturing more powerful computer chips to protecting communication satellites.

A deeper understanding of non-equilibrium, heterogeneous non-ideal plasmas will be developed to allow for the development of improved molecular dynamics (MD) models. Historically, the computational cost of MD has severely restricted its use to simulations of very small numbers of particles, typically only 100s to 1000s, with simulations relying on periodicity to mimic a bulk material.

This has restricted modeling to mostly equilibrium, homogeneous plasma physics. Recent advances in computational methods and machine learning offer avenues to break through this barrier. As the accessible scales increase, so do important gradients and entirely new modeling issues appear.

Force laws are the key input to MD, and scientists are now at a juncture where the problems they can tackle with MD cannot use force laws developed in the past. Further, unique to plasmas, gradients introduce new physical effects such as mesoscopic electric fields that can enhance transport, while the potential increase in computational cost associated with gradients can be severe.

In this project, both computational models and machine learning techniques for non-uniform plasmas will be obtained to address classes of problems previously beyond the reach of MD. Using these new methods, mesoscopic plasma instabilities will be atomistically explored in the presence of electric fields and transport.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.