Roßkopf, AndreasAndreasRoßkopfCheng, XuepengXuepengChengStraub, ChristopherChristopherStraubTenbrinck, DanielDanielTenbrinck2025-11-132025-11-132025-09-24https://publica.fraunhofer.de/handle/publica/49934410.1109/SISPAD66650.2025.11186115The combination of physical modeling and machine learning, known as Scientific Machine Learning (SciML), is enabling a new generation of simulation methodologies. In this work, we demonstrate the potential of SciML for simulating silicide formation in Ni-SiC systems, a process highly relevant to TCAD and device engineering. Four neural network architectures - standard MLPs, enhanced mMLPs with residual connections, interpretable Kolmogorov-Arnold Networks (KANs), and Chebyshev KANs (cKANs) - are benchmarked, each trained solely on the governing physical laws. All models tested achieve accurate results and enable rapid evaluation, providing large speedups over traditional solvers. Among these, the mMLP yields the best accuracy. These findings underscore the strong potential of SciML for efficient and accurate TCAD simulations, paving the way for scalable, data-integrated modeling of complex material interactions in semiconductor technology.enconference paper