Satoshi Hamaguchi, Osaka University, Osaka, Japan
The process developments for thin-film coating, surface modification, and semiconductor manufacturing have become so complex that there is much room for improvement in their efficiency with data-driven approaches. A trial-and-error approach by experienced engineers based on the knowledge of existing processes and materials can be replaced with a more systematic approach based on synthetic knowledge inferred by machine learning (ML). Currently, a large amount of data on specific processing tools is typically collected and used for process optimization and control for those tools. However, if such data are associated with the underlying physics mechanisms, they may be used to optimize different process tools and, possibly, to develop new processes. Computer models of processing tools with physics-based process models, i.e., digital twins of processing tools, allow such physics-informed data-driven approaches to solving complex problems of process development and control. One of the main obstacles to developing such digital twins is the lack or shortage of fundamental physics parameters such as reaction rate constants. A data-driven approach assisted by artificial intelligence (AI)/ML techniques may also be able to infer such physics parameters. In this presentation, starting with a brief overview of the current status of data-driven plasma science, the author will discuss several subcomponents of a digital twin of a plasma processing tool. As in typical plasma systems, a plasma process tool model involves multi-scale physics and may be divided into a macroscopic bulk plasma model and microscopic or even nano-scale surface reaction models. Each model also consists of several subcomponent models, and at the most fundamental level, gas-phase and surface chemical reactions are governed by quantum mechanics. Using some examples, the author will discuss how “surrogate models” of the digital twin’s subcomponents, constructed with experimental data augmented with first-principle numerical simulation data, may allow reliable real-time simulation of a process tool or inference of new process conditions.