Temperature prediction for in-situ consolidation in the AFP process using physics-informed neural network
摘要
To address the challenges of complex heat-source input, multi-layer thermal accumulation, and high-gradient temperature prediction in automated fiber placement (AFP), this study proposes a heat-flux-driven physics-informed neural network (PINN) for multi-layer temperature-field prediction during in-situ consolidation in AFP. A non-sequential ray tracing model is first developed to obtain the spatial heat-flux distributions on the incoming tape and substrate, which are further extended into a spatiotemporal heat-flux field. A time-varying growing-domain PINN is then established based on the two-dimensional transient heat-conduction equation to predict the temperature evolution during layer-by-layer deposition. To improve the prediction accuracy in high-gradient regions, residual-based attention (RBA) and adaptive residual-based resampling (ARR) are introduced to enhance the loss weighting and collocation-point distribution, respectively. The proposed model is validated against Abaqus finite element simulations. The overall root-mean-square error, relative error, and mean absolute error are 11.78 K, 2.63%, and 8.16 K, respectively. At a representative monitoring point, the predicted temperature history achieves an R2 value of 99.3% and a mean absolute error of 5.21 K. The ablation results show that RBA significantly reduces local maximum errors, while ARR decreases the two-layer RMSE from 19.64/24.03 to 3.45/5.42 K under uniform sampling. After offline training, the trained PINN predicts the temperature field within approximately 2 s, indicating its potential to support efficient process-parameter optimization in AFP by enabling rapid repeated evaluation of thermal responses.