Fig. 2.

Illustration of the U-Net-based neural network architecture designed for reconstructing 2D temperature fields from partially observed data. The process initiates with a ’Spatial Propagator’ that pre-processes the input masked temperature fields, followed by a U-Net for the prediction of the complete temperature distribution. The figure also presents three distinct masking approaches used in our experiments: ’Random’, ’Ring’, and ’Grid-edge’, alongside an exemplar system showcasing a randomly distributed conductivity and heat-source field, and its resulting steady-state temperature field as simulated for dataset creation.