Iterative solvers of linear systems are a key component for the numerical solutions of partial differential equations (PDEs). While there have been intensive studies through past decades on classical methods such as Jacobi, Gauss-Seidel, conjugate gradient, multigrid methods and their more advanced variants, there is still a pressing need to develop faster, more robust and reliable solvers. Based on recent advances in scientific deep learning for operator regression, we propose HINTS, a hybrid, iterative, numerical, and transferable solver for differential equations. HINTS combines standard relaxation methods and the Deep Operator Network (DeepONet). Compared to standard numerical solvers, HINTS is capable of providing faster solutions for a wide class of differential equations, while preserving the accuracy close to machine zero. Through an eigenmode analysis, we find that the individual solvers in HINTS target distinct regions in the spectrum of eigenmodes, resulting in a uniform convergence rate and hence exceptional performance of the hybrid solver overall. Moreover, HINTS applies to equations in multidimensions, and is flexible with regards to computational domain and transferable to different discretizations.
翻译:线性系统的迭代求解器是部分差异方程式(PDEs)数字解决方案的一个关键组成部分。虽然过去几十年对古典方法,如Jacobi、Gauss-Seidel、 conjugate梯度、多电离方法及其较先进的变异器等,进行了大量研究,但目前仍迫切需要开发更快捷、更有力和更可靠的解答器。根据最近在为操作员回归而进行科学深层次学习方面取得的进步,我们建议 HINTS, 一种混合、迭代、数字和可转让的异异方方方方程式。 HINTS将标准放松方法和深操作员网络(DeepONet)结合起来。与标准的数字解答器相比, HINTS能够提供更快速的解决方案,用于广泛的差异方程式,同时保持接近零的准确性。我们发现,通过机本模型分析,HINTS的单个解答器针对易位区域,导致混合求解器的总体统一趋同率,因此表现非常出色。此外, HINS适用于多门式解式方程式中的方程式,并且灵活地适用于不同的计算领域和可分解。