TL;DR
Get the latest gadgets delivered free — and shop member deals
- Fast, free delivery on millions of items
- Access to Prime Big Deal Days deals on October 6–7
- Prime Video, Amazon Music and more included
Terra Quantum and Empa have developed LP-FNO, an AI model that predicts three-dimensional temperature fields and melt-pool boundaries for laser welding. The teams report predictions in milliseconds and accuracy metrics measured against high-fidelity simulations; industrial use and performance on other materials have not been established in the supplied report.
Terra Quantum and Empa have developed LP-FNO, an AI model that predicts three-dimensional temperature fields and melt-pool boundaries in laser welding, according to a report published September 29. The teams say the model produces predictions in milliseconds, potentially making detailed process estimates more practical for welding optimization and process control; the reported results are based on simulations of one titanium alloy, not a demonstrated factory deployment.
LP-FNO, short for Laser Processing Fourier Neural Operator, was trained using high-fidelity thermo-fluid simulations of Ti-6Al-4V. The training data covered laser powers from 40 to 190 watts and scan speeds from 0.1 to 1 meter per second. The report says this range includes both conduction welding and stable keyhole welding, two operating regimes with different melt-pool behavior.
After training, the model generated a full three-dimensional prediction in about 8 milliseconds at standard resolution and 88 milliseconds at twice that resolution. The teams compare those times with roughly six minutes for a standard-resolution simulation and more than an hour at finer resolution. They report a speed advantage of up to 100,000 times over equivalent high-fidelity simulations. That figure is a comparison with simulation runtimes, not a measurement of faster production-line output.
The study reports about 2.5% average relative error for temperature predictions and an intersection-over-union score above 0.9 for melt-pool boundary segmentation. It also says the model could be evaluated on finer grids than those used for training without retraining. These are results reported for the study’s simulation setup; they do not establish accuracy in live welding conditions.
Faster Estimates for Welding Control
Laser welding is used in aerospace, medical device and automotive manufacturing, where the shape and temperature of the molten area beneath the laser affect weld quality and consistency. Detailed multiphysics simulations can help engineers study how process settings affect that area, but the reported runtimes make them difficult to use for immediate decisions or broad searches across possible settings.
If performance holds in industrial settings, predictions that take milliseconds could support more frequent process estimates, larger parameter searches or digital models updated alongside a physical process. Those uses are possibilities described by the company, not outcomes shown in the supplied report. The study addresses a computational bottleneck; it does not report production trials, measured improvements in weld quality, or cost savings.
From Simulation to Surrogate Model
High-fidelity models represent interacting physical effects in the material and melt pool. According to the report, a single run takes approximately six minutes at 10-micrometer resolution and more than an hour at 5-micrometer resolution. Such runtimes limit how often engineers can use these simulations during process adjustment and how many parameter combinations they can examine.
LP-FNO uses a Fourier Neural Operator to learn the relationship between laser power and scan speed and the simulated three-dimensional outcome. The researchers also reformulated the scanning problem in a frame moving with the laser and used temporal averaging. The report says this quasi-steady approach let the model represent stable keyhole behavior within the same framework as conduction welding.
The full study is reported as published in the Journal of Intelligent Manufacturing, with DOI 10.1007/s10845-026-02917-0. The report describes LP-FNO as spanning both conduction and stable keyhole regimes in a quasi-steady operator-learning framework; it attributes that first-of-its-kind characterization to the authors’ knowledge.
“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.”
— Markus Pflitsch, Terra Quantum CEO and founder
Limits of the Reported Results
The supplied report does not state whether LP-FNO has been tested against measurements from physical welding equipment, or how its predictions perform when conditions differ from the simulated training range. It also does not establish whether the reported error and boundary scores apply equally across all powers, scan speeds and welding regimes.
The reported model concerns Ti-6Al-4V. Its performance on other alloys, equipment configurations or production environments is not specified. The report gives prediction times at two resolutions but does not detail hardware, deployment requirements, or the full process needed to turn an estimate into a safe control action.
Terra Quantum describes real-time process optimization and synchronized digital twins as potential uses. The supplied material does not give a deployment schedule, identify an industrial customer, or report results from a production line. The extent to which the simulation speed advantage will translate into manufacturing benefits remains unclear.
Testing Beyond Simulated Welding
The study has been published, but the supplied report does not announce a next trial date or a commercial launch. The next evidence needed to assess industrial readiness would include validation against physical welding measurements, tests across equipment and materials, and details on how the model would connect to process monitoring or control systems.
Further evaluation could also clarify how prediction accuracy changes outside the reported training range and whether finer-grid predictions remain reliable under real operating variation. Until such results are reported, LP-FNO’s demonstrated scope remains the simulation-based study of Ti-6Al-4V across the stated laser-power and scan-speed ranges.
Key Questions
What does LP-FNO predict?
It predicts three-dimensional temperature fields and melt-pool boundaries from laser power and scan speed, based on training with simulations of Ti-6Al-4V welding.
How fast are its predictions?
The report gives approximately 8 milliseconds at standard resolution and 88 milliseconds at twice that resolution. It compares these with simulation runtimes of about six minutes and more than an hour, respectively.
Has the model been shown working on a factory line?
The supplied report does not describe a factory deployment or production trial. It presents simulation-based results and discusses industrial uses as potential applications.
Which material and welding conditions were studied?
The model was trained on simulations of Ti-6Al-4V, covering laser powers of 40 to 190 watts and scan speeds of 0.1 to 1 meter per second.
Source: rss
Evergreen bestsellers Picks
bestsellers
As an affiliate, we earn on qualifying purchases.
