Project ESPINN

Project description

Numerical simulations and data-driven machine learning methods have developed into indispensable tools in the engineering sciences and beyond in recent decades. However, the respective methodological weaknesses (high computing time and enormous amounts of simulation or measurement data required) limit their applicability, the solution efficiency, and thus the product improvement capabilities for multi-billion markets such as micro-, nano-, and power electronics.

Within the ESPINN project, a hybrid approach based on physics-informed neural networks (PINNs), is used to develop simulators that solve physical and chemical processes - supported by a few measurement data - on an atomistic to continuous scale several orders of magnitude faster than previously used numerical methods. The focus is on the implementation of PINN-based interaction potentials including their quantum mechanical effects, the resolution of 4D tensors (3D geometry and time) for parameterized diffusion processes, as well as the development of a concept-based understanding of PINNs with different NN architectures. AI and application knowledge are thus combined with explainability and transferred into four industry-relevant software solutions. These include a methodologically largely generic molecular dynamics solver, two specific PINN simulators for semiconductor technology (silicidation and photoresist processing), and a PINN analysis tool for estimating learning content and accuracy.

The ESPINN project is coordinated by Fraunhofer IISB. The institute contributes its extensive expertise in the field of silicidation and photoresist processing. Additionally, the AI-augmented simulation group at Fraunhofer IISB leverages its experience in developing PINNs at the continuum level for industry-relevant applications.

Publications

Scientific Machine Learning (SciML)-How the Fusion of AI and Physics is Giving Rise to Promising Simulation Methodologies
by A. Rosskopf et al., presentation at SISPAD 2025
Full paper

Modeling Nickel Silicidation using Physics-Informed Machine Learning
by C. Straub et al., presentation at SISPAD 2025
Full paper

Hard-Constraining Neumann Boundary Conditions in Physics-Informed Neural Networks via Fourier Feature Embeddings
by C. Straub, et al., in: Proceedings of Workshop on Machine Learning Multiscale Processes 2025