Physics-Informed Surrogate Modelling for Dissimilar FSW

Friction stir welding (FSW) of dissimilar materials — particularly aluminium alloys and high-strength steels — involves a large process parameter space (rotation speed, traverse speed, plunge depth, tool geometry) with non-linear, coupled effects on the resulting microstructure and mechanical properties. Running full experimental campaigns to explore this space is costly and time-consuming. Surrogate modelling offers a computationally efficient alternative: build a model that approximates the process–property relationships and use it to guide experiments intelligently.

This project develops a physics-informed surrogate modelling framework for dissimilar FSW, grounded in the process-structure-property chain established from experimental thesis work on AA6082-T6 / DP780 steel joints.

Why Gaussian Process Regression

GPR is well-suited to this problem for two reasons. First, it provides uncertainty estimates alongside predictions — critical when making decisions about which experiments to run next. Second, it naturally incorporates prior physical knowledge through the choice of kernel function, allowing the model to reflect known smoothness and correlation structure in the FSW process space.

Framework components

• GPR surrogate model trained on process parameter inputs (rotation speed, traverse speed, plunge depth) with ultimate tensile strength (UTS) and hardness as outputs
• Bayesian optimisation using the surrogate’s uncertainty estimates to propose the next-best experiment (exploration vs. exploitation trade-off)
• Process-structure-property chain encoding the physical relationships between IMC layer thickness, grain refinement, and joint efficiency as constraints

Stack

Python · scikit-learn · GPy · NumPy · Matplotlib · SciPy

Links

• GitHub: Deep7285/Physics-Informed-Surrogate-Modelling-for-Dissimilar-FSW
• Related paper: (Kumar et al., 2026)