Computational and machine learning modelling approaches for weld quality predictions in friction stir welding of high-density polyethylene

Friction stir welding (FSW) has gained recognition as an effective technique for joining thermoplastic materials. This study presents a comprehensive investigation to predict FSW characteristics such as weld temperatures, vertical axial forces, material flow and defect formation in FSW of high-density polyethylene (HDPE) using both computational and machine learning approaches, validated by experimental results. A three-dimensional Coupled Eulerian-Lagrangian (CEL) numerical model is developed to simulate the HDPE welding process under varying process parameters of rotational and welding speeds, and bottom preheating conditions. The CEL numerical model is validated against experimental data, showing markedly good agreement in predicting void formation, thermal and mechanical responses. Results indicate that preheating the bottom surface proved effective in reducing thermal gradients and stabilizing vertical force fluctuations. Quantitatively, bottom-surface preheating at 60 °C reduced peak plunge force from 1600 N close to 1100 N and raised peak temperatures from 70 °C (without preheat) to about 100 °C, hence smoothing through-thickness gradients and stabilizing the axial force distribution. In parallel, two multilayer perceptron artificial neural networks (ANNs) are trained on experimental data to predict temporal temperature and axial force responses with R2 between 0.91 and 0.99, demonstrating high accuracy and strong generalization across the tested FSW parameter space. The computational CEL simulations accurately predicted weld zone morphology, including stir zone, thermo-mechanically affected zone and heat-affected zone, however the combined CEL–ANN framework enables rapid generation of process maps that highlights defective and defect-free welding windows, with optimal conditions obtained at 1500 rpm and 30 mm/min with 60 °C preheating, and poor weldability observed at 1000 rpm without preheating due to wormhole formation. These findings provide process-driven design guidelines for developing robust HDPE joints and demonstrate how integrating computational modeling with machine learning enables efficient optimization, offering valuable insights into the key characteristics of the FSW process in thermoplastics and supporting its broader adoption in industrial applications.