PINN-MG: a multigrid-inspired hybrid framework combining the iterative method and physics-informed neural networks

doi:10.1007/s10483-026-3419-9

摘要/Abstract

Abstract:

Iterative methods are widely used for solving partial differential equations (PDEs). However, the difficulty in eliminating global low-frequency errors significantly limits their convergence speed. In recent years, neural networks (NNs) have emerged as a novel approach for solving PDEs, with studies showing that they exhibit faster convergence for low-frequency components. Building on these complementary frequency-convergence characteristics of iterative methods and NNs, and drawing inspiration from multigrid methods, we propose a hybrid solving framework consisting of a combination of iterative methods and NN-based solvers, termed physics-informed neural network multigrid (PINN-MG, abbreviated as PMG). In this framework, the iterative methods eliminate local high-frequency oscillation errors, while PINNs correct global low-frequency errors. Throughout the solving process, high- and low-frequency components alternately dominate the error, with each being addressed by the iterative methods and PINNs, respectively, thereby accelerating the convergence. We validate the proposed PMG framework on the linear Poisson equations and nonlinear Helmholtz equations. The results show significant acceleration of the PMG built on the Gauss-Seidel (GS), pseudo-time, and generalized minimal residual (GMRES) methods. A detailed analysis of the convergence process validates the rationality of the framework. To further evaluate the advantages of PMG, we apply it to indefinite Helmholtz equations, a class of problems in which traditional solvers often diverge due to the divergence of the low-frequency components. The PMG framework effectively overcomes this divergence, improving both the stability and the convergence speed. We propose the PMG framework as a data-free hybrid solver that does not rely on any pretraining and, more importantly, provides a unified mechanism to tightly couple the NN methods with classical iterative solvers, achieving an organic and iterative integration of the two paradigms.

Key words: physics-informed neural network (PINN), iterative method, hybrid computing architecture, spectral bias

中图分类号:

O241.8

. [J]. Applied Mathematics and Mechanics (English Edition), 2026, 47(8): 1855-1878.

Daiwei DONG, Wei SUO, Jiaqing KOU, Weiwei ZHANG. PINN-MG: a multigrid-inspired hybrid framework combining the iterative method and physics-informed neural networks[J]. Applied Mathematics and Mechanics (English Edition), 2026, 47(8): 1855-1878.

图/表 22

Method	GS	2-level multigrid	3-level multigrid	4-level multigrid	5-level multigrid	PMG
Time/s	8 777	2 229	526	138	37	84
Speed-up	–	$× 3.93$	$× 16.69$	$× 63.60$	$× 237.22$	$× 104.48$

Method	GMRES	2-level multigrid	3-level multigrid	4-level multigrid	5-level multigrid	PMG
Time/s	274	62	39	37	36	83
Speed-up	–	$× 4.42$	$× 7.02$	$× 7.40$	$× 7.61$	$× 3.30$

Method	Pseudo-time	2-level multigrid	3-level multigrid	4-level multigrid	PMG
Time/s	556	537	170	53	213
Speed-up	–	$× 1.03$	$× 3.27$	$× 10.49$	$× 2.61$

[1]	SMITH, G. D. Numerical Solution of Partial Differential Equations, 3rd ed., Clarendon Press, Oxford (1985)
[2]	EYMARD, R., GALLOUËT, T., and HERBIN, R. Finite volume methods. Solution of Equation in ℝⁿ (Part 3), Techniques of Scientific Computing (Part 3), 713–1018, Elsevier, Amsterdam (2000)
[3]	REDDY, J. N. An Introduction to the Finite Element Method, McGraw-Hill, New York (1985)
[4]	TREFETHEN, L. N. and BAU, D. Numerical Linear Algebra, Society for Industrial and Applied Mathematics, Philadelphia (1997)
[5]	SAAD, Y. and SCHULTZ, M. H. GMRES: a generalized minimal residual algorithm for solving nonsymmetric linear systems. SIAM Journal on Scientific and Statistical Computing, 7(3), 856–869 (1986)
[6]	BRANDT, A. and LIVNE, O. E. Multigrid Techniques: 1984 Guide with Applications to Fluid Dynamics, Revised Edition, Society for Industrial and Applied Mathematics, Philadelphia (2011)
[7]	TROTTENBERG, U., OOSTERLEE, C. W., and SCHULLER, A. Multigrid, Elsevier, FL, United States (2000)
[8]	HACKBUSCH, W. Multi-Grid Methods and Applications, Springer Science & Business Media, Berlin (2013)
[9]	STÜBEN, K. Numerical Analysis: Historical Developments in the 20th Century, Elsevier, FL, United States (2001)
[10]	LIU, Y. L., ZHANG, W. W., and KOU, J. Q. Mode multigrid — a novel convergence acceleration method. Aerospace Science and Technology, 92, 605–619 (2019)
[11]	RAISSI, M., PERDIKARIS, P., and KARNIADAKIS, G. E. Physics-informed neural networks: a deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 378, 686–707 (2019)
[12]	SIRIGNANO, J. and SPILIOPOULOS, K. DGM: a deep learning algorithm for solving partial differential equations. Journal of Computational Physics, 375, 1339–1364 (2018)
[13]	E, W. N. and YU, B. The deep Ritz method: a deep learning-based numerical algorithm for solving variational problems. Communications in Mathematics and Statistics, 6(1), 1–12 (2018)
[14]	LIAO, Y. L. and MING, P. B. Deep Nitsche method: deep Ritz method with essential boundary conditions. Communications in Computational Physics, 29(5), 1365–1384 (2025)
[15]	SAMANIEGO, E., ANITESCU, C., GOSWAMI, S., NGUYEN-THANH, V. M., GUO, H., HAMDIA, K., ZHUANG, X., and RABCZUK, T. An energy approach to the solution of partial differential equations in computational mechanics via machine learning: concepts, implementation and applications. Computer Methods in Applied Mechanics and Engineering, 362, 112790 (2020)
[16]	XU, Z. Q. J., ZHANG, Y. Y., LUO, T., XIAO, Y. Y., and MA, Z. Frequency principle: Fourier analysis sheds light on deep neural networks. Communications in Computational Physics, 28(5), 1746–1767 (2025)
[17]	XU, Z. Q. J., ZHANG, Y. Y., and LUO, T. Overview frequency principle/spectral bias in deep learning. Communications on Applied Mathematics and Computation, 7(3), 827–864 (2025)
[18]	XU, Z. Q. J., ZHANG. Y., and XIAO, Y. Training behavior of deep neural network in frequency domain. Proceedings of the Neural Information, Springer, Berlin (2019)
[19]	RAHAMAN, N., BARATIN, A., ARPIT, D., DRAXLER, F., LIN, M., HAMPRECHT, F., BENGIO, Y., and COURVILLE, A. On the spectral bias of neural networks. Proceedings of the 36th International Conference on Machine Learning, PMLR, Victoria, BC, Canada, 5301–5310 (2019)
[20]	BASRI, R., JACOBS, D., KASTEN, Y., and KRITCHMAN, S. The convergence rate of neural networks for learned functions of different frequencies. Advances in Neural Information Processing Systems, 32, 4761–4771 (2019)
[21]	BASRI, R., GALUN, M., GEIFMAN, A., JACOBS, D., KASTEN, Y., and KRITCHMAN, S. Frequency bias in neural networks for input of non-uniform density. Proceedings of the 37th International Conference on Machine Learning, PMLR, Victoria, BC, Canada, 685–694 (2020)
[22]	LEE, J., XIAO, L. C., SCHOENHOLZ, S. S., BAHRI, Y., NOVAK, R., SOHL-DICKSTEIN, J., and PENNINGTON, J. Wide neural networks of any depth evolve as linear models under gradient descent. Journal of Statistical Mechanics: Theory and Experiment, 2020(12), 124002 (2020)
[23]	JACOT, A., GABRIEL, F., and HONGLER, C. Neural tangent kernel: convergence and generalization in neural networks (invited paper). Proceedings of the 53rd Annual ACM SIGACT Symposium on Theory of Computing, ACM, New York (2021)
[24]	DU, S. S., ZHAI, X., PÓCZOS, B., and SINGH, A. Gradient descent provably optimizes over-parameterized neural networks. International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA (2019)
[25]	BIETTI, A. and MAIRAL, J. On the inductive bias of neural tangent kernels. Advances in Neural Information Processing Systems 32, Curran Associates, Inc., Vancouver, Canada (2019)
[26]	ARORA, S., DU, S. S., HU, W., LI, Z., and WANG, R. Fine-grained analysis of optimization and generalization for overparameterized two-layer neural networks. Proceedings of the 36th International Conference on Machine Learning, PMLR, Victoria, BC, Canada, 322–332 (2019)
[27]	TANCIK, M., SRINIVASAN, P. P., MILDENHALL, B., and FRIDOVICH-KEIL, S. Fourier features let networks learn high frequency functions in low dimensional domains. Proceedings of the 34th International Conference on Neural Information Processing Systems, ACM, New York, 7537–7547 (2020)
[28]	WANG, S. F., YU, X. L., and PERDIKARIS, P. When and why PINNs fail to train: a neural tangent kernel perspective. Journal of Computational Physics, 449, 110768 (2022)
[29]	CUI, C., JIANG, K., LIU, Y., and SHU, S. Fourier neural solver for large sparse linear algebraic systems. Mathematics, 10(21), 4014 (2022)
[30]	ZHANG, E. R., KAHANA, A., KOPANIČÁKOVÁ, A., TURKEL, E., RANADE, R., PATHAK, J., and KARNIADAKIS, G. E. Blending neural operators and relaxation methods in PDE numerical solvers. Nature Machine Intelligence, 6(11), 1303–1313 (2024)
[31]	HU, J. and JIN, P. A hybrid iterative method based on MIONet for PDEs: theory and numerical examples. Mathematics of Computation, 95, 2103–2127 (2026)
[32]	SUO, W. and ZHANG, W. W. A novel paradigm for solving PDEs: multi-scale neural computing. Acta Mechanica Sinica, 41(6), 324172 (2024)
[33]	CHIU, P. H., WONG, J. C., OOI, C., DAO, M. H., and ONG, Y. S. CAN-PINN: a fast physics-informed neural network based on coupled-automatic-numerical differentiation method. Computer Methods in Applied Mechanics and Engineering, 395, 114909 (2022)
[34]	BARAKAT, A. and BIANCHI, P. Convergence and dynamical behavior of the ADAM algorithm for nonconvex stochastic optimization. SIAM Journal on Optimization, 31(1), 244–274 (2021)
[35]	LIU, D. C. and NOCEDAL, J. On the limited memory BFGS method for large scale optimization. Mathematical Programming, 45(1), 503–528 (1989)