Academic Awards 2025 booklet

59 Physics-Informed Learning of Neural Activity in a Brain-on-Chip Device There is a rising demand for computational efficieny, driven by expanding AI and algorithmic applications. A Brain-on-Chip (BoC), a microdevice containing living neuronal cell cultures and an interface that enables electrical stimulation and measurements, is a promising candidate for next generation energy efficient computing hardware. This study contributes to realizing this potential by developing a surrogate model of the BoC. The methodology combines the usage of measurement data and physical background knowledge of neuronal behavior. The surrogate model consists of two components: one whose structure is directly established by physics, and another that relies more heavily on data and incorporates physics into the learning process with a novel physics-informed machine learning (PI-ML) model. Including the physical background knowledge throughout the surrogate model enhances the interpretability. A numerical evaluation shows the capability of the surrogate model to accurately learn the neuronal behavior and predict unseen test trajectories. Additionally, the novel PI-ML competes with the state-of- the-art PI-ML model and shows improved model robustness and upscaling capabilities. While a gap from numerical results to experimental results remains, the developed surrogate model is an important step toward a software-based platform for experiment design, diagnostics, performance analysis, and digital replication of the BoC devices.