A machine learning-enabled framework for multiscale modeling of surface processes in low-temperature plasma processing

Manufacturing of next-generation semiconductor devices requires precise control over plasma-surface interactions (PSI), which are inherently multiscale in nature, spanning a large range of length scales (Å to cm) and time scales (ps to sec). Better understanding and control of the overall process and development of new technologies requires a systematic framework to model the physics of these multiscale interactions and connect the disparate scales. In this work, a general framework is proposed to model PSI at the atomic and surface scale, where each species is characterized by the states it occupies and the transitions across these states. The dynamics are characterized by a set of master equations, with unknown transition probabilities, where additional physics, such as subsurface diffusion, are introduced [1]. Representing the transition probabilities as neural networks in relation to system parameters like ion energy and fluence, the resulting system of ODEs is solved in a neural ODE framework and trained with MD simulation data [2]. The framework has been tested on Si RIE and ALE, with further validation on literature results [2]. The master equation description can be applied to any dynamical process, with any transition probability represented as neural network, and the entire system solved in a neural ODE framework, for longer length and time scales, inaccessible via MD alone.

[1] Oehrlein et al, J.Vac.Sci.Technol.B42,000000(2024)

[2] Vella et al[AM1] , J. Vac. Sci. Technol. B. 40(2):023205, 2022.