A priori error estimate of deep mixed residual method for elliptic PDEs (Q6182315)

The paper presents a groundbreaking theoretical analysis of the deep mixed residual method (DMRM) in the context of solving elliptic partial differential equations (PDEs). DMRM, a novel deep learning-based approach, has demonstrated remarkable efficiency and accuracy in previous empirical studies, yet its theoretical foundations remained unexplored until now. The authors embark on a rigorous examination of DMRM's convergence properties, offering the first comprehensive a priori error estimate for this method. Their analysis unveils a significant finding: the convergence of neural network solutions is ensured with the augmentation of training samples and network size, without imposing any constraints on their ratio. This pivotal insight not only solidifies the reliability of DMRM but also highlights its scalability, a crucial attribute in practical applications.