Beyond our client work, we carry out applied research at the crossroads of artificial intelligence, numerical simulation and computing systems. Research and client work enrich each other: our research feeds into our solutions, and the problems we encounter with our clients open up new avenues of research.
Conventional numerical solvers are accurate but computationally expensive. We develop surrogate models, such as PINNs (Physics-Informed Neural Networks) or Neural Operators, that approximate their results in a fraction of the time, enabling large design-space exploration, optimisation and real-time simulation.
Paper Neural Operators for Spark Plasma Sintering Simulation Journal of the Japan Society of Powder and Powder Metallurgy, 2025A purely data-driven model can produce physically inconsistent results, especially outside its training domain. We work on hybrid approaches that embed physical laws at the heart of learning, to obtain models that are more reliable, need less data and can solve inverse problems, such as identifying physical parameters from measurements.
We study the design of hybrid digital twins that combine conventional numerical models (finite elements…), real data from sensors installed on industrial systems, and deep learning algorithms. Our work focuses on developing advanced deep learning architectures to build high-fidelity digital replicas, particularly for industrial processes.
Paper Toward A Spark Plasma Sintering Digital Twin Using A Deep Learning-Based Approach Euro PM2024, Malmö. Keynote Paper Award.Choosing the infrastructure for an AI system means predicting its performance, energy consumption and cost before buying it. We develop analytical simulators that model these quantities, from workstation to server, and calibrate them against measurements on real hardware. This work contributes to more frugal AI, sized to actual needs.
European project
Hexo Signal takes part in this European research project on physics-informed machine learning: neural networks coupled with particle-based numerical simulation methods.
PIML-Particles · HORIZON-MSCA-2026-SE-01-01
EPMA — Euro PM2024
Contact
Feel free to get in touch.
contact@hexosignal.com Contact us