Benchmarking Factorisation Techniques in Neural Operators for Smart Energy Systems

This project, led by Bitnet, aims to benchmark, design, and scale novel Neural Operator-based frameworks for predictive maintenance and digital twin applications in energy-critical systems—extending from hydrogen fuel cells to electric vehicle (EV) batteries. Building upon prior developments in the Horizon Europe OPEVA project and BitNet’s in-house research, we will focus on predicting voltage, current, and state-of-health (SOH) degradation using high-fidelity, physics-informed AI models trained on large-scale time-series sensor data.

The core innovation of this proposal lies in the design of factorisation-aware Neural Operator architectures that not only model degradation dynamics, but also capture and explain sensor behavior across both electrochemical and thermal domains. We plan to evaluate a range of emerging approaches—including Fourier Neural Operators (FNOs), DeepONets, and Transformer-based operators—with specific emphasis on their scalability, interpretability, and suitability for real-time deployment.

We will investigate how different factorisation techniques—such as Fourier-mode pruning, low-rank tensor compression, and attention sparsification—affect model generalisation, memory usage, and training efficiency across GPU nodes. These models will be benchmarked using multi-million-step time-series datasets from both hydrogen fuel cell and EV battery platforms, processed with NVIDIA Modulus and custom PyTorch pipelines on EuroHPC JU infrastructure.

By expanding the project scope to cover both fuel cell and EV battery use cases—and targeting sensor-level interpretability—the project supports EuroHPC JU’s strategic goals in AI acceleration, digital twin readiness, and sustainable transport. Our results will inform best practices for deploying factorisation-aware neural solvers in industrial energy systems, advancing the state-of-the-art in scientific machine learning.