Abstract
Predicting macroscopic permeability in dual-scale textile fabrics is a computationally intensive task due to the extreme heterogeneity of the Brinkman coefficient ($\beta$), which can span 14 orders of magnitude. This paper presents a deep learning approach using Fourier Neural Operators (FNO) and Tensorized Fourier Neural Operators (TFNO) to map this heterogeneous coefficient to the fluid velocity field. By introducing $H^2$ regularization to the loss function, the model captures the underlying physics more accurately than standard $L^2$ approaches. The proposed method demonstrates high accuracy, zero-shot super-resolution capabilities, and significant computational efficiency compared to traditional FFT-based solvers.
Introduction
In high-performance manufacturing, such as Liquid Composite Moulding for carbon fiber components, predicting permeability—the ease with which resin flows through fabric—is critical. This is a "dual-scale" problem involving micrometer-scale pores between individual fibers and millimeter-scale channels between woven bundles. Mathematically, this is governed by the Stokes-Brinkman equation, where the parameter $\beta$ distinguishes between these regions.
Traditional numerical solvers, while accurate, require high-performance computing resources. This research explores whether "operator learning" can provide a faster, lighter alternative capable of handling the extreme range of $\beta$, which varies by up to 14 orders of magnitude across a single fabric sample.
Methods
The researchers employed Fourier Neural Operators (FNO) and the memory-efficient Tensorized Fourier Neural Operator (TFNO). Rather than predicting permeability directly from the fabric structure, which is numerically unstable, the authors used a two-step approach:
- Neural Operator: Learns the mapping from the heterogeneous coefficient $\beta$ to the fluid velocity field $u$.
- Physics Step: Calculates macroscopic permeability $K$ from the predicted velocity field using Darcy’s law.
A key innovation is the use of $H^2$ regularization. Standard $L^2$ loss functions are insufficient because permeability calculations involve derivatives of the velocity field. The $H^2$ loss penalizes second-derivative errors, forcing the network to learn smooth, physically consistent fields.
Datasets
The study utilized a synthetic dataset of 800 pairs of input ($\beta$) and output (velocity field $u$) generated by a high-fidelity FFT-based solver (AMITEX). Input maps were created using Gaussian random fields with varying mean permeability, anisotropy, and rotation angles to simulate diverse textile architectures. Models were trained on $64 \times 64$ grids and tested on $128 \times 128$ grids to evaluate zero-shot super-resolution.
Results and Discussion
The TFNO model with $H^2$ loss emerged as the top performer, achieving a relative supremum norm error of approximately 0.0913. This accuracy remained stable even when testing on higher-resolution grids (error of 0.0971 at $128 \times 128$), proving that the FNO learned the underlying operator rather than a simple pixel mapping.
The model successfully handled the extreme heterogeneity of the input data. As shown in the distribution of $\beta$ values, the parameter spans a vast range, which typically causes numerical instability in traditional solvers. By predicting the full velocity field as an intermediate step, the model remains grounded in spatial physics, allowing it to generalize across these extreme scales.
Conclusion
This study demonstrates that physics-informed regularization, specifically $H^2$ loss, is mandatory for applying operator learning to dual-scale flow problems. While currently focused on 2D cross-sections, the FNO approach shows great promise for scaling to 3D industrial fabrics. Future work will involve testing the model on realistic 3D microstructures and actual CT scans of industrial textiles to handle real-world geometric irregularities.
Summary updated