Insider Brief
- Terra Quantum and Empa developed LP-FNO, a Fourier Neural Operator model that predicts three-dimensional temperature fields and melt-pool boundaries from laser welding parameters.
- The model was trained on high-fidelity simulations of Ti-6Al-4V across laser powers of 40–190 watts and scan speeds of 0.1–1 meter per second.
- Once trained, LP-FNO generates full 3D predictions in about 8 milliseconds at standard resolution and 88 milliseconds at twice the resolution.
PRESS RELEASE — Terra Quantum, a global leader in quantum technologies, and Empa, the Swiss Federal Laboratories for Materials Science and Technology, announced LP-FNO (Laser Processing Fourier Neural Operator), an artificial intelligence surrogate model that predicts full three-dimensional melt-pool dynamics in laser welding up to 100,000 faster than traditional multiphysics simulation. Removing the computational bottleneck that has long blocked real-time process control and digital twin deployment in industrial laser processing.
The Simulation Bottleneck in Laser Welding
Laser welding is a precision manufacturing process used across aerospace, medical device, and automotive production, where control of the melt pool, the zone of molten metal found beneath the laser beam, is central to weld quality and consistency. High-fidelity multiphysics simulations have been the primary tool for understanding and optimizing the process, but their computational cost has made real-time application impractical. A single simulation run using current gold-standard tools takes approximately six minutes at standard 10 µm resolution and over one hour at finer 5 µm resolution, placing process control, large-scale parameter optimization, and uncertainty quantification out of reach for industrial operators.
What LP-FNO Does
LP-FNO learns the relationship between laser process parameters, power and scan speed, and the resulting three-dimensional temperature fields and melt-pool boundaries, using a Fourier Neural Operator architecture that mixes information globally across the physical domain in a single forward pass. The model was trained on high-fidelity thermo-fluid simulations of Ti-6Al-4V titanium alloy, spanning a process window of 40 to 190 watts of laser power and scan speeds from 0.1 to 1 meter per second, covering both conduction and stable keyhole welding regimes. Once trained, LP-FNO produces full 3D predictions in approximately 8 milliseconds at standard resolution and 88 milliseconds at twice the resolution.
A key element of the approach is a quasi-steady reformulation of the laser-scanning problem. By transforming the transient simulation into a reference frame that moves with the laser beam and applying temporal averaging, the research team converted an inherently time-dependent problem into a form suitable for operator learning. This enabled LP-FNO to accurately represent stable keyhole welding dynamics, the most physically complex regime, involving deep vapor depressions, recoil-pressure-driven surface deformation, and strong laser-absorption variation, within the same framework used for conduction-mode welding.
“The ability to compress hours of physics simulation into milliseconds is not an incremental improvement, it is the kind of change that makes entirely new applications possible,” said Markus Pflitsch, CEO and Founder of Terra Quantum. “With LP-FNO, industrial operators can run process optimization in real time, build digital twins that stay synchronized with the physical process, and explore parameter spaces that were previously too expensive to probe. This is what deploying AI on the toughest problems in manufacturing looks like.”
Key Findings
- Speed: up to 100,000 times faster than equivalent high-fidelity multiphysics simulation
- Temperature accuracy: approximately 2.5% average relative error
- Melt-pool prediction: intersection-over-union (IoU) score above 0.9 for melt-pool boundary segmentation
- Regime coverage: to the authors’ knowledge, the first surrogate model to span both conduction and stable keyhole welding regimes within a quasi-steady operator-learning framework
- Resolution invariance: a model trained on coarse-resolution data can be evaluated on finer grids without retraining, delivering accurate super-resolved predictions
- Architecture benchmark: LP-FNO demonstrates the most favourable accuracy–efficiency trade-off among the neural architectures tested, including fully connected coordinate networks, U-Net, and DeepONet variants
The full study is published in the Journal of Intelligent Manufacturing (DOI: 10.1007/s10845-026-02917-0).
“This partnership brought together Empa’s expertise in laser processing, materials science, and process-modelling know-how with Terra Quantum’s capability in advanced machine learning and neural operator methods,” said Dr. Elia Iseli, Head of Light-Matter Dynamics Group, Empa. “The accuracy with which LP-FNO reproduces the manufacturing-relevant outputs of the high-fidelity models, including temperature fields, melt-pool geometry, and phase interfaces, highlights the predictive power of the surrogate approach. This level of agreement reflects the close integration of machine-learning development and expertise in laser process modelling throughout the project.. It is a result that could not have been achieved without this partnership.”
“Fourier Neural Operators are architecturally well-suited to this class of problem because they learn in spectral space, mixing information across the full physical domain at each layer rather than propagating it locally through convolutional neighborhoods,” said Florian Neukart, Chief Technology Officer at Terra Quantum. “That global perspective is what gives the model its resolution-invariant property: the spectral weights learned on coarse training data generalizes naturally to finer evaluation grids. Combined with the quasi-steady reformulation, this gave us a single trained model that handles both conduction and keyhole regimes with consistent accuracy across the process window.”
To learn more about Terra Quantum, please visit: https://terraquantum.swiss/
