Jan Trieschmann¹, Rouven Lamprecht¹, Tobias Gergs¹, Christian Stüwe¹, Luca Vialetto^1,2, Sahitya Yarragolla^1,3, Finn Zahari¹, Richard Marquardt¹, Thomas Mussenbrock³, Hermann Kohlstedt^1
1Kiel University, Kiel, Germany
²Stanford University, Stanford, CA
³Ruhr University Bochum, Bochum, Germany
Emerging technologies for micro-/nanoelectronics devices, in particular with application in novel neuromorphic computing, heavily rely on plasma processing. Engineering such devices continuously requires a more precise process control. This may be attained with a physical knowledge-based design of related plasma processes, taking into account the properties of the corresponding devices. Modeling and simulation of such surface-facing process plasmas, paired with measurement data of the fabricated devices, may enable physical interpretation and guide the process design. Despite the ever-increasing compute power, however, a consistent simulation at all levels is difficult due to the extremely complex dynamics of multi-component plasmas interacting with bounding surfaces. Hierarchical coupling of the manifold physico-chemical processes may be realized through unbiased data-driven surrogate models. These are derived from high fidelity data obtained from physical models at the lower scales (e.g., atomistic simulations of the surface kinetics). Moreover, data-driven approaches correlating experimental device characteristics with local conditions during processing may open a path in understanding and optimizing the plasma process. Thus, a data-driven link between global process quantities (e.g., pressure, voltage, current) and microscopic quantities (e.g., thin film composition, electrical properties) may be devised. The outlined approach is discussed at the example of thin film memristive devices, processed through reactive sputter deposition. A model framework including plasma simulations, a surface kinetics model, and a data-driven correlation of memristive device properties is described. Therein, individual model components are replaced or augmented by machine learning surrogate models, fostering a correlation and interpretation of the device characteristics with respect to intrinsic plasma parameters. We suggest that such a hybrid physical and data-driven model is a versatile and widely applicable tool in plasma processing.