Development of a physics-informed machine learning framework for insight and prediction of solidification cracking formation in high-power laser beam welding

• Applicant: Dr.-Ing. Xiangmeng Meng
• Subject Area: Joining and Separation Technology
• Term: since 2026
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 582234565

In this project, a novel physics-informed machine learning (PIML) framework to predict and mitigate solidification cracking in the high-power laser beam welding of steel and aluminum will be developed. The experimental investigation, multi-physical modeling, and advanced ML approaches will be interactively combined to enhance both prediction accuracy and interpretability. Meanwhile, a quantitative, bidirectional causation chain between welding parameters, critical physical variables, and solidification crack features will be established for the first time. To achieve this, the 3D transient keyhole dynamics, heat transfer, and fluid flow will be simulated by a computational fluid dynamic model, especially focusing on reproducing the complex weld pool geometries and resolving the thermal and microstructural conditions (anisotropy and element segregation) on the solidification front. The calculated weld pool boundary will be implemented as the equivalent heat source in a thermomechanical simulation, considering strain-hardening effects, solid-state phase transformations, solidification shrinkage, and anisotropic constitutive relationships in the mushy zone, to calculate the transient stress/strain evolution. Extensive LBW experiments will be conducted on aluminum and steel alloys with varying solidification crack susceptibilities. Real-time measurements and post-weld X-ray imaging will provide not only the necessary data for model validation but also crack features for ML training. Welding parameters, simulation-derived physical variables, and experimentally measured crack metrics will be used as training, validation, and testing data for the PIML framework. An ML model with a multi-modal architecture will first be constructed to process different data types in the physical variables (scalar, sequential, and 3D data). A predictive model with clear physical interpretability will be established to correlate physical variables with observed cracking features. The hierarchy of different physical variables, i.e., their influence on the solidification crack features, will be decoupled and analyzed for the first time. Subsequently, conditional generative models with the welding parameters and material properties as the input will be developed for a rapid estimation of the relevant thermal and mechanical variables in a realistic time frame. By integrating the predictive model and the generative models, a PIML framework can be established, describing a unidirectional causation chain of welding parameters → causative physical variables→ solidification cracking formation. A multiple-objective optimization algorithm will be further incorporated, enabling the framework to select the optimized welding parameter by iteratively minimizing the difference between the calculated and desired weld geometry and crack level below the maximum allowable threshold.

• DFG Programme: Research Grants
• Co-Investigators: Dr.-Ing. Marcel Bachmann; Professor Dr.-Ing. Michael Rethmeier