The development of new computer chips and the manufacturing of tiny electronic components involve extremely complex processes. To understand and optimize them, research and development currently rely on time-consuming computer simulations. While these simulations are highly accurate, they often require many hours or even days of computing time. Data-driven methods, such as conventional neural networks, can perform these tasks faster, but they require large amounts of measurement or simulation data and often produce results that cannot be physically explained. Both factors limit their practical use, especially in areas where reliability and safety are critical.

The ESPINN project employs a new technology called Physics-Informed Neural Networks (PINNs). These special AI models combine the best of both worlds: they learn from data while simultaneously incorporating known physical laws. This results in models that are physically accurate and significantly faster than traditional simulations. PINNs can, for example, simulate chemical reactions or diffusion processes on scales from atomic to macroscopic, and up to a thousand times faster once the learning process is complete. In addition, the project develops a novel analysis tool to evaluate the reliability and explainability of these AI models, providing insight into why a model produces a particular result.

ESPINN aims to advance the development and practical application of PINNs in industrial simulation processes. The project develops four interconnected software solutions that cover different aspects of semiconductor manufacturing. One is a flexible molecular dynamics simulation tool that utilizes PINNs to accurately calculate atomic interactions significantly faster than before. Building on this, a PINN simulator for silicide formation helps optimize metal contacts with semiconductors, which is crucial for microchip performance. Another key component is a PINN model for simulating photoresist processes, which improves the precision of lithography and structuring. Finally, an analysis and quality assurance tool evaluates the performance and reliability of all models, providing a kind of “seal of approval” that facilitates the use of PINNs in sensitive applications such as medicine, safety, and aerospace.

Fraunhofer SCAI contributes extensive experience in physics-based and data-driven simulation methods for virtual materials design at the quantum level. In addition to the Tremolo-X software, which enables quantum dynamic simulations for semiconductor development, SCAI has successfully applied PINN models to predict material properties of oxide glasses. The institute also brings expertise in data-driven molecule design and in predicting molecular properties using machine learning models.

The Fraunhofer-internal program PREPARE funds the project for three years. Fraunhofer SCAI cooperates closely with Fraunhofer HHI and Fraunhofer IISB.

Project duration: 05/2024 until 04/2027