Quantum AI Solves Manufacturing Simulation Bottleneck

An innovative artificial intelligence surrogate model, LP-FNO, has been developed by Terra Quantum and Empa, significantly accelerating the simulation of laser welding processes. This AI predicts full three-dimensional melt-pool dynamics up to 100,000 times faster than conventional multiphysics simulations, overcoming a long-standing obstacle to real-time process control and digital twin implementation in industrial laser processing.

Laser welding, crucial for aerospace, medical devices, and automotive manufacturing, relies on understanding and controlling the melt pool. Previously, high-fidelity simulations, essential for optimization, were computationally intensive, taking up to an hour for detailed analysis. LP-FNO, trained on titanium alloy simulations, now delivers predictions in mere milliseconds, even at higher resolutions.

The AI model employs a quasi-steady reformulation and a Fourier Neural Operator architecture to efficiently learn the complex relationship between laser parameters and melt-pool behavior. This approach accurately models both conduction and stable keyhole welding regimes, a first for surrogate models. Experts anticipate LP-FNO will unlock new applications by enabling real-time process optimization and synchronized digital twins.

The study, published in the Journal of Intelligent Manufacturing, highlights LP-FNO's accuracy, with an average relative error of approximately 2.5% for temperature and an IoU score above 0.9 for melt-pool boundary segmentation. The model's resolution invariance is another key feature, allowing accurate predictions on finer grids without retraining, demonstrating a superior accuracy-efficiency trade-off compared to other tested neural architectures.