Physics-Informed Data-Driven Machine Learning Cross-Scale Cold Atmospheric Plasma Chemistry

Cold atmospheric plasma (CAP) relies on plasma chemistry, especially the reactive species, and physical emissions such as the radio frequencies and UV-VIS photons to become broadband treatment tools that can be applied to biomedical, environmental, and material enqlneering1. However, at the current stage, the acquisition of the complete chemical compositions, including all the species concentrations, in the CAP plume is still an impossible task2, Therefore, we propose a general method that uses a machine-learning technique to solve such a cross-scale problem. The idea is to let an artificial neural network guess all species concentrations in the CAP plume at a steady state while using Fourier-transform infrared spectroscopy (FTIR) to measure the concentrations of some species. Next, let the neural network's prediction pass through a chemical simulation for a few time steps to find the scale of change of all the species concentrations. Since the CAP system is working at a steady state, any variation of species concentrations in the simulation are error. Also, any disagreement between the neural network's prediction and the FTIR results are error. Once the total error converged to an acceptable low value, we consider such a prediction is an accurate solution to the problem, with its uniqueness proved by the 2
^nd
law of thermodynamics.