Speaker
Description
For the optimization of semiconductor components, a detailed understanding of low-resistance ohmic contacts between metal and semiconductor is essential. This process is known as silicidation. On a macroscopic level, it can be described mathematically by coupled reaction-diffusion equations. A promising approach to solving such systems of non-linear differential equations is the use of physics-informed neural networks (PINNs). We will showcase the advantages of using PINNs in comparison to traditional solvers. The focus of this talk is on the choice of the representation model in the PINN method. Firstly, we will review various neural network architectures, including the recently introduced Kolmogorov-Arnold networks (KANs) and their adaptations. We will provide a thorough comparison of the performances of the different architectures for the silicidation problem. Secondly, we discuss how to incorporate different physical priors into the neural network's architecture. Such physical priors may include boundary conditions, initial conditions, or constraints on the range of the solution's output. Hard-constraining physical priors allows one to reduce the number of optimization tasks and can thus improve the training performance. In particular, we present a new method of hard-constraining discontinuous initial conditions adapted to diffusion problems. This work was supported by the Fraunhofer Internal Programs under Grant No. PREPARE 40-08394.
