Paper 14093-64
Paper 14093-64
Extrapolating laser weld penetration with machine learning, trading accuracy for generalisation
16 April 2026 • 11:30 - 11:45 CEST | Curie A (Niveau/Level 1)
Abstract
Optimising laser keyhole welding, a process critical for high-volume industries, is challenging due to its complex, interdependent parameters and physical phenomena. Traditional modeling and experimental optimisation are often ineffective.
In this study, we evaluated machine learning models (deep learning TabPFN and a simpler Multi-Layer Perceptron) for predicting weld penetration and generating processing maps. While deep learning (R² = 0.94) and optimised MLP (R² = 0.92) excelled within training ranges, they struggled with extrapolation. Surprisingly, a basic MLP (R² = 0.87) demonstrated superior performance beyond the training data.
Presenter
Icam, site de Strasbourg-Europe (France), ICube (France)
I am a Teaching and Research Faculty member and Director of Research at ICAM Strasbourg-Europe, where I contribute to the development of research strategy, academic excellence, and student training.
My scientific work is conducted within the ICube laboratory and focuses on light–matter interaction, with a strong emphasis on laser machining and laser-based manufacturing processes. My research interests include materials behavior under laser irradiation, process optimization, and advanced manufacturing technologies.
In parallel, I teach materials physico-chemistry and eco-design, and I supervise multidisciplinary student engineering projects, particularly in the aerospace sector. I am strongly committed to linking research, education, and industrial applications.