Abstract
We present Neural Spectral Methods, a technique to solve parametric PartialDifferential Equations (PDEs), grounded in classical spectral methods. Ourmethod uses orthogonal bases to learn PDE solutions as mappings betweenspectral coefficients. In contrast to current machine learning approaches whichenforce PDE constraints by minimizing the numerical quadrature of the residualsin the spatiotemporal domain, we leverage Parseval's identity and introduce anew training strategy through a \textit{spectral loss}. Our spectral lossenables more efficient differentiation through the neural network, andsubstantially reduces training complexity. At inference time, the computationalcost of our method remains constant, regardless of the spatiotemporalresolution of the domain. Our experimental results demonstrate that our methodsignificantly outperforms previous machine learning approaches in terms ofspeed and accuracy by one to two orders of magnitude on multiple differentproblems. When compared to numerical solvers of the same accuracy, our methoddemonstrates a $10\times$ increase in performance speed.