MUSA-PINN: Multi-scale Weak-form Physics-Informed Neural Networks for Fluid Flow in Complex Geometries

Weizheng Zhang, Xunjie Xie, Hao Pan, Xiaowei Duan, Bingteng Sun, Qiang Du, Lin Lu · 2026

Read MUSA-PINN: Multi-scale Weak-form Physics-Informed Neural Networks for Fluid Flow in Complex Geometries by Weizheng Zhang, Xunjie Xie, Hao Pan, Xiaowei Duan, Bingteng Sun, Qiang Du, Lin Lu (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-08465.txt.

Introduction

Zhang and colleagues solve incompressible flow in triply periodic minimal surface geometries with MUSA-PINN, a multi-scale weak-form physics-informed network. Physics-informed networks minimize point residuals, but that local focus fails to spread global information through winding channels, causing unstable gradients and broken mass conservation. MUSA-PINN instead writes Navier-Stokes constraints as integral conservation laws over layered spherical volumes. It imposes continuity and momentum through flux-balance residuals at three scales. Tests claim it beats rivals, cutting errors by up to 93% while keeping mass conservation.

Catalog record

CREATORWeizheng Zhang, Xunjie Xie, Hao Pan, Xiaowei Duan, Bingteng Sun, Qiang Du, Lin Lu
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE62949
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.08465

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

MUSA-PINN: Multi-scale Weak-form Physics-Informed Neural Networks for Fluid Flow in Complex Geometries Weizheng Zhang, Xunjie Xie, Hao Pan, Xiaowei Duan, Bingteng Sun, Qiang Du, Lin Lu Abstract While Physics-Informed Neural Networks (PINNs) offer a mesh-free approach to solving PDEs, standard point-wise residual minimization suffers from convergence pathologies in topologically complex domains like Triply Periodic Minimal Surfaces (TPMS). The locality bias of point-wise constraints fails to propagate global information through tortuous channels, causing unstable gradients and conservation violations. To address this, we propose the Multi-scale Weak-form PINN (MUSA-PINN), which reformulates PDE constraints as integral conservation laws over hierarchical spherical control volumes. We enforce continuity and momentum conservation via flux-balance residuals on control surfaces. Our method utilizes a three-scale subdomain strategy-comprising large volumes for long-range coupling, skeleton-aware meso-scale volumes aligned with transport pathways, and small volumes for local refinement-alongside a two-stage training schedule prioritizing continuity. Experiments on steady incompressible flow in TPMS geometries show MUSA-PINN outperforms state-of-the-art baselines, reducing relative errors by up to 93% and preserving mass conservation. 1 Introduction Scientific machine learning has advanced rapidly in recent years, offering new paradigms for modeling and solving complex physical processes. Yet in engineering fluid dynamics, conventional computational fluid dynamics (CFD) remains the dominant tool ( 26 ; 30 ) , with costs that become prohibitive in parameterized design and optimization loops. High-efficiency heat exchangers exemplify this challenge: their performance relies on geometrically intricate three-dimensional flow passages, and practical design exploration often requires repeated simulations across many geometric variations. Such simulations via CFD typically require high-quality volumetric meshes, but mesh generation for slender, tortuous, and topologically complex channels is time-consuming and frequently demands substantial expert effort and manual tuning ( 13 ) . Physics-informed neural networks (PINNs) provide a mesh-free alternative by representing the solution field with neural networks and training with physics constraints from governing equations and boundary conditions ( 24 ) . PINNs are attractive for fast inference, differentiable modeling, and inverse problems with sparse observations ( 17 ; 25 ) . For high-dimensional fluid simulations, PINNs have the potential to significantly improve computational efficiency, in contrast to traditional CFD methods whose computational cost increases dramatically with problem dimensionality ( 14 ) . However, applying PINNs to industrial-grade internal flows with tortuous, high-aspect-ratio, and topologically complex channels remains challenging. Empirically, the accuracy and stability of standard PINNs often degrade substantially as geometric complexity increases, even for steady incompressible flows. Figure 1 provides a qualitative comparison of flow streamlines on an industrial liquid-cooling plate, contrasting CFD with a standard PINN baseline and MUSA-PINN. In this paper, we make the observation that a key reason for this failure is structural: the dominant training signal in standard PINNs is pointwise residual minimization, which cannot match with the global nature of conservation laws and produces a local–global inconsistency in complex transport geometries. In long and winding passages as found in heat exchangers, where information propagation along streamlines and across junctions is weak under purely local supervision, such locally small residuals frequently translate into physically inconsistent behavior. (a) (b) (c) Figure 1 : Flow streamlines on an industrial liquid-cooling plate. (a) CFD reference, (b) PINN-PE baseline (degrades in this complex geometry), and (c) MUSA-PINN. MUSA-PINN achieves a 14.91\% relative \ell_{2} error in velocity w.r.t. CFD. A natural remedy to the above problem is to impose conservation in an integral sense. Prior attempts include enforcing global flux constraints between inlets and outlets ( 14 ) or adopting variational formulations with element-wise quadrature ( 19 ) . Yet, global boundary constraints are sparse and may leave the interior under-constrained, while element-based variational formulations typically require meshing and may introduce additional computational burdens or stability issues on complex geometries. This motivates a mesh-free formulation that enforces conservation over interior regions, so that physical information can propagate reliably through tortuous channels instead of being constrained only at scattered points or the outer boundary. In this work, we propose MUlti-Scale weAk-form PINN (MUSA-PINN), a hybrid strong–weak form physics-informed framework for steady incompressible flow in complex geometries. Inspired by the classical Finite Element Analysis using weak form integral equations to improve convergence, our main idea is to reformulate governing equations into flux-balance constraints over control volumes using the Gauss divergence theorem, turning volumetric residuals into surface integrals of physically meaningful fluxes. These surface integrals are efficiently approximated by Monte Carlo sampling on control-volume boundaries, avoiding volumetric meshing while providing a more structured and globally informative training signal than pointwise residuals. To make the weak-form constraints effective across different industrial channels, we further introduce a multi-level subdomain placement strategy: large-scale control volumes encourage long-range conservation across transport pathways; medium-scale control volumes trace a geometry skeleton, so conservation signals follow flow channels; and small-scale control volumes refine local details and stabilize near-boundary behaviors. Together, these components provide dense conservation supervision throughout the domain and mitigate the local–global mismatch that limits conventional PINNs in complex internal flows. We evaluate MUSA-PINN on flow channels extracted from Triply Periodic Minimal Surface (TPMS) structures, which serve as representative heat-exchanger passages due to their complex geometry and favorable transfer characteristics ( 42 ) . Experiments on multiple TPMS variants demonstrate that MUSA-PINN significantly improves solution accuracy and physical consistency over state-of-the-art baselines, and ablations verify the contributions of the weak-form and multi-scale designs. We further discuss extensions toward more complex industrial components (e.g., liquid-cooling plates and DualMS channels ( 44 ) ) to highlight the scalability of the proposed approach. Our main contributions are summarized as follows: • Weak-form conservation via control volumes. We introduce a divergence-theorem-based formulation that enforces mass and momentum conservation through surface flux integrals over spherical control volumes, providing a volumetric mesh free and physically structured supervision signal beyond pointwise residuals. • Multi-scale subdomain placement for complex channels. We design a multi-level strategy that improves coverage and information propagation in long, tortuous, and branching internal passages, bridging global consistency and local fidelity without manual domain decomposition. • We benchmark on geometrically intricate TPMS heat-exchanger channels and conduct systematic comparisons and ablations, demonstrating consistent gains in accuracy and recovered flow structures, and outlining scalability to more complex industrial geometries. 2 Related Work PINNs solve PDEs by parameterizing solution fields with neural networks and enforcing governing equations and boundary/initial conditions via automatic differentiation ( 24 ; 5 ) . Recent studies have improved and analyzed PINNs training, including advances in architectures, optimization strategies, sampling schemes, and theoretical perspectives on when and why PINNs succeed or fail ( 17 ; 7 ; 29 ; 20 ) . These developments position PINNs as a flexible learning-based solver family that can be deployed across diverse physical systems ( 2 ; 11 ; 31 ; 10 ) . PINNs for Navier-Stokes equations. PINNs have been applied to incompressible Navier–Stokes flows under various settings, including data-scarce or label-free formulations and enhanced representations ( 16 ; 4 ; 1 ; 9 ) . While substantial progress has been demonstrated on many 2D benchmarks and simple 3D domains, steady flow prediction in 3D complex channels remains challenging due to intricate geometry, long-range coupling, and optimization instability. We target this regime and propose a weak-form, multi-scale formulation tailored to complex 3D geometries. Strong-form PINNs and Weak-form PINNs. Strong-form PINNs minimize pointwise PDE residuals, offering a

Read the complete text