Thermal field estimation in CFRTP composites using an attention-enhanced U-Net

This study presents a surrogate model based on the convolutional U-Net architecture to predict the thermal field in a carbon fibre-reinforced thermoplastic tape at the microscale during brief and localized heating. Leveraging microstructure data within a machine learning framework, the proposed model aims to enhance the accuracy of temperature field predictions at a low computational cost. The incorporation of a co-attention mechanism to handle image channels of different nature significantly improves precision, resulting in a strong correlation between the model’s predictions and the ground truth obtained from the numerical solution of the heat equation. This capability enables rapid assessment of diverse microstructures, facilitating optimization and real-time applications in manufacturing settings.