Updating Physical Field Reconstruction Model Based on Continual Learning

Physical field reconstruction plays a crucial role in domains such as environmental monitoring and engineering design. However, current research predominantly concentrates on constructing physical field models, with limited attention given to addressing continuous model updates, which are necessary in real-world applications. The concept of continual learning can be used to update physical field reconstruction models. But existing continual learning methods primarily tackle classification problems and have difficulty when applied to high-dimensional physical field regression tasks. This study extends the concept of continual learning to the high-dimensional physical field reconstruction problem by introducing a novel adaptive updating method-EHWKD-that mitigates catastrophic forgetting. EHWKD is a regularization-based approach that combines Huber reconstruction loss, EWC loss, and Wasserstein Distance Knowledge Distillation loss. This integrated approach more thoroughly mitigates the forgetting of previously learned tasks. Experimental results on three classical datasets demonstrate that EHWKD achieves a lower forgetting rate and a higher forward transfer rate compared to mainstream approaches.