Development of a physics-informed machine learning framework to predict process porosity formation and to optimize welding parameters in high-power laser beam welding

• Applicant: Dr.-Ing. Xiangmeng Meng
• Subject Area: Joining and Separation Technology ; Joining and Separation Technology
• Term: since 2022
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 506270597

In this project, a newly developed machine learning (ML) strategy, named physics-informed machine learning (PIML), will be utilized to predict the process porosity in the high-power laser beam welding of steel and Al. In the PIML framework, the experimental investigation, multi-physical modeling, and the ML algorithm will be interactively combined. Meanwhile, a bi-directional and quantitative chain of causation, i.e., welding parameters ↔ physical variables ↔ porosity formation, will be created.Thus, a multi-physically coupled numerical model will be developed to solve a 3D transient problem of laser-material interaction, and the induced heat transfer, fluid flow, and free surface deformation. This allows for a quantitative description of the temperature distribution, velocity field, and keyhole dynamics. Laser beam welding experiments and the accompanying measurements and characterizations will be performed to obtain not only the validation data of the numerical model but also the porosity data for the model training.Two modules of the PIML framework will be established by using the welding parameter, calculation results, and the measured porosity distribution as the training, validating, and testing data. Firstly, the causative physical variables will be extracted from the multi-physical model. A predictive model of porosity formation with the physical variables as input will be developed, namely giving the correlation of physical variables → porosity formation. The hierarchy of different physical variables on the porosity distribution pattern and ratio will be decoupled and analyzed for the first time. Secondly, another machine learning (ML) model, using the welding parameters and material thermo-properties as the input, will be developed to realize a fast prediction of the physical variables in a realistic time frame. A multiple-objective optimization algorithm will be incorporated into the ML model, enabling the ML model to select the optimized welding parameter. Eventually, the two modules will be integrated to build the complete PIML framework, by which not only the occurrence of porosity under certain process parameters could be predicted, but also the optimized welding parameter could be automatically selected to produce the desired weld geometry with the lowest porosity level.

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