[論文レビュー] Solving and learning advective multiscale Darcian dynamics with the Neural Basis Method

Yuhe Wang, Min Wang|arXiv (Cornell University)|Feb 19, 2026

Model Reduction and Neural Networks被引用数 0

ひとこと要約

tldr: The Neural Basis Method (NBM) provides a projection-based neural solver for coupled Darcy flow–transport, with an operator-learning extension (NBM-OL) that delivers accurate, robust results and massive speedups in parametric settings.

ABSTRACT

Physics-governed models are increasingly paired with machine learning for accelerated predictions, yet most "physics--informed" formulations treat the governing equations as a penalty loss whose scale and meaning are set by heuristic balancing. This blurs operator structure, thereby confounding solution approximation error with governing-equation enforcement error and making the solving and learning progress hard to interpret and control. Here we introduce the Neural Basis Method, a projection-based formulation that couples a predefined, physics-conforming neural basis space with an operator-induced residual metric to obtain a well-conditioned deterministic minimization. Stability and reliability then hinge on this metric: the residual is not merely an optimization objective but a computable certificate tied to approximation and enforcement, remaining stable under basis enrichment and yielding reduced coordinates that are learnable across parametric instances. We use advective multiscale Darcian dynamics as a concrete demonstration of this broader point. Our method produce accurate and robust solutions in single solves and enable fast and effective parametric inference with operator learning.

研究の動機と目的

• Motivate a physics-conforming, projection-based alternative to loss-weighted physics-informed learning for PDEs.
• Develop Neural Basis Method (NBM) that uses a predefined neural basis and a PDE-constrained projection to solve PDEs.
• Extend NBM to parametric operator learning (NBM-OL) for efficient many-query scenarios.
• Demonstrate NBM and NBM-OL on advective multiscale Darcian dynamics (Darcy flow–transport).
• Show robustness, accuracy, and significant speedups in single-query and parametric settings.

提案手法

• Construct physics-conforming neural vector and scalar bases with a fixed neural basis to form a finite-dimensional approximation space.
• Solve for solution coefficients via a PDE-constrained least-squares projection (linear in coefficients) to enforce the operator and boundary conditions.
• Use energy-consistent weighting in a mixed least-squares formulation to preserve Darcian scaling and local conservation.
• Apply upwind control-volume stabilization for the transport/advection step.
• Advance Darcy flow with implicit time integration and Picard linearization, and handle transport with multiple substeps to resolve fronts.
• Extend to NBMA-OL by learning parameter-dependent coefficient maps in the fixed neural basis space, enabling fast online evaluation; optionally apply POD compression.

実験結果

リサーチクエスチョン

• RQ1 Can a projection-based neural basis framework provide stable, interpretable enforcement of PDE operators compared to loss-weighted PINNs?
• RQ2 Does the Neural Basis Method yield accurate, robust solutions for advective multiscale Darcian dynamics and retain operator structure under basis enrichment?
• RQ3 Can NBM-OL learn parametric dependence of Darcy–transport operators with strong out-of-distribution generalization and substantial speedups?
• RQ4 How does energy-consistent weighting and mixed formulation affect mass conservation and multiscale accuracy in Darcy-flow–transport problems?
• RQ5 What is the performance of NBM and NBM-OL in single-query vs. parametric many-query regimes?

主な発見

• NBM achieves accurate and stable single-instance solutions for advective multiscale Darcy–transport problems, with spectral-like convergence and competitive or superior fieldwise errors compared to FVM and vanilla PINN.
• Energy-consistent weighting and mixed least-squares formulation preserve physical scaling and local conservation, improving multiscale accuracy.
• In multiscale permeability settings, NBM maintains robustness and achieves spectral-like residual decay as basis size grows, with errors staying controlled across permeability contrasts up to 10^4.
• NBM–OL learns parametric Darcy flow and transport operators, showing good in-distribution and out-of-distribution generalization with sub-percent relative L2 errors and linear residual-error tracking.
• NBM–OL delivers massive speedups over FVM in parametric evaluation, quantified as orders of magnitude faster online evaluation.
• Across multiple scenarios, NBM–OL maintains accurate tracer fronts and concentration fields under coupling and hyperbolic dynamics