Temperature field prediction based on a 3D-U-net model

The solidification process of castings involves intricate physical transformations and is critical in the final quality of castings. The temperature field during solidification serve as fundamental indicators for defect prediction and process optimization. Conventional approaches rely on numerical simulations. However, they are constrained by their low computational efficiency, complex modeling requirements, and high computational cost. Therefore, this study adopted deep learning as an alternative method for predicting temperature fields. A temperature field simulation program developed in Python was used to calculate the temperature field and generate training and prediction datasets. A three-dimensional U-net network structure was established. It took the initial temperature field, geometric shape, time, and thermophysical parameters as inputs to predict the dynamic temperature field of casting models with varying geometries and aluminum alloys under different pouring temperatures. The proposed model achieves an average absolute error of 1.99 °C. Under an absolute error tolerance of 7.5 °C, the average accuracy reaches 94.73%, the average correlation coefficient is 0.9998, and the average prediction time for the temperature field is 0.47 s. Based on the prediction results, shrinkage defects in castings were predicted, and the effectiveness of the model was preliminarily verified through experimental testing, indicating the feasibility of deep learning in predicting the temperature field of castings.