PCG-Informed Neural Solvers for High-Resolution Homogenization of Periodic Microstructures

TL;DR

CGiNS introduces a PCG-informed neural solver for high-resolution 3D homogenization of periodic microstructures. It tightly couples geometry-aware sparse 3D convolutions, explicit periodic boundary encoding via Peri-mapping, and a global displacement constraint with a lightweight PCG refinement embedded in the decoder, trained using a minimum potential energy-based loss. On TPMS, PSL, and Truss datasets, it achieves relative elasticity-tensor errors below 1% and delivers 2x–10x speedups over GPU multigrid solvers while scaling to $512^3$ voxels. The approach generalizes across different base materials and multiphase configurations without retraining, and its solver-in-the-loop design preserves physical consistency and stability even on complex topologies, enabling efficient high-throughput screening and topology optimization. Limitations remain to linear elasticity with periodic BC and memory demands at the largest resolutions, with future work aiming at nonlinear materials, mixed boundary conditions, and memory-efficient extensions.

Abstract

The mechanical properties of periodic microstructures are pivotal in various engineering applications. Homogenization theory is a powerful tool for predicting these properties by averaging the behavior of complex microstructures over a representative volume element. However, traditional numerical solvers for homogenization problems can be computationally expensive, especially for high-resolution and complicated topology and geometry. Existing learning-based methods, while promising, often struggle with accuracy and generalization in such scenarios. To address these challenges, we present CGINS, a preconditioned-conjugate-gradient-solver-informed neural network for solving homogenization problems. CGINS leverages sparse and periodic 3D convolution to enable high-resolution learning while ensuring structural periodicity. It features a multi-level network architecture that facilitates effective learning across different scales and employs minimum potential energy as label-free loss functions for self-supervised learning. The integrated preconditioned conjugate gradient iterations ensure that the network provides PCG-friendly initial solutions for fast convergence and high accuracy. Additionally, CGINS imposes a global displacement constraint to ensure physical consistency, addressing a key limitation in prior methods that rely on Dirichlet anchors. Evaluated on large-scale datasets with diverse topologies and material configurations, CGINS achieves state-of-the-art accuracy (relative error below 1%) and outperforms both learning-based baselines and GPU-accelerated numerical solvers. Notably, it delivers 2 times to 10 times speedups over traditional methods while maintaining physically reliable predictions at resolutions up to $512^3$.

Figures (11)

• Figure 1: Overview of the CGiNS framework.This framework is illustrated using periodic lattice structures with a target resolution of $n$. Given a periodic microstructure, CGiNS voxelizes the input domain and applies soft-voxel padding to enhance boundary fidelity. The geometry and base material tensor are processed through a sparse convolutional encoder to extract multi-resolution features $\mathcal{Q}^l$. At each level $l$, a U-Net block refines these features, which are then decoded into displacement predictions $u^l$ and corrected via global translation to satisfy displacement constraints. A $t$-step preconditioned conjugate gradient (PCG) smoother eliminates high-frequency errors. Coarse-to-fine trilinear interpolation propagates feature information across levels, while energy-based losses $\mathcal{L}^l$ are computed at each resolution for self-supervised training. All computations operate on sparse tensors to ensure scalability to high-resolution structures.
• Figure 2: Accuracy–runtime and residual–runtime comparisons across resolutions. We compare CGiNS (red star) with two GPU-based solvers—AmgX (blue) and Homo3D (green)—on voxel grids of size $64^3$, $128^3$, $256^3$, and $512^3$. Top row: Relative error of the effective elasticity tensor $\delta(\%)$ versus wall-clock time, illustrating the accuracy–runtime tradeoff. Bottom row: Residual norm $\Vert r \Vert$ versus time, showing convergence behavior. All plots use logarithmic axes.
• Figure 3: Impact of PCG-coupled training on projection energy and convergence behavior. (a) Initial projection energy $\Vert e\Vert_{\tilde{K}}^2$ computed on 100 microstructures using four initialization methods: Zero (classical solver baseline), PH-Net, CGiNS-0 (trained without PCG), CGiNS-8, and CGiNS-10. Samples are sorted in ascending order based on CGiNS-10 energy. (b) PCG residual norm $\Vert r\Vert$ over 10 iterations for a representative sample. CGiNS-10 consistently achieves the lowest initial projection energy and fastest convergence, confirming the effectiveness of joint network–solver training.
• Figure 4: 2D illustration for Peri-mapping data structure. (a) Periodic tiling of a microstructure in space, where the domain repeats periodically. (b) To emulate periodic boundary conditions within a single simulation cell, CGiNS remaps edge nodes (e.g., red to red) across domain boundaries, allowing convolution kernels to access wrapped-around neighbors. This ensures explicit preservation of boundary continuity during feature extraction. The schematic shows a 2D example but applies equally to 3D.
• Figure 5: Soft voxel padding for boundary-aware representation. Left: Original voxelized microstructure. Right: Soft voxels (shown in translucent beige) are added around the boundary to preserve geometric context and enhance convolutional feature continuity near edges, while incurring minimal memory overhead.