A Shallow U-Net for Enhancement of Multi-Scale Localization in Fiber Reinforced Polymer Matrix Composite Representative Unit Cells

This study addresses key limitations of existing micromechanics closed-form models, with emphasis on their implementation within the NASMAT ecosystem. A shallow U-Net is introduced to enhance the precision of local stress and strain fields computed using the Generalized Method of Cells (GMC) for fiber-reinforced polymer matrix composites (PMCs). The network is trained on High-Fidelity Generalized Method of Cells (HFGMC) data so that the surrogate explicitly corrects GMC's lack of shear-normal coupling while improving the localization of the fields. Simultaneously, the innovative contracting/expansive architecture preserves the full spatial resolution of the underlying representative unit cell (RUC), unlike a standard convolutional neural network (CNN) surrogate. The optimization process encompasses both structural and hyperparameter fine-tuning, in which a broad set of design factors are varied and evaluated based on their influence on the learning and simulation steps. The resulting model achieves a low validation error (MSE ≈ 5.11 × 10⁵) and significantly refines localization. In particular, the measured linear regression coefficients (R² ≥ 0.965) underline the improvements in shear components, with slopes rising from 0.3604 to 0.9514 (stress) and from 0.6312 to 0.9640 (strain) relative to the ideal target (slope = 1.0, R² = 1.0). Despite the higher fidelity, the average simulation wall-time remains low (GMC+ML ≈ 0.1526 s) compared with HFGMC (≈ 0.19099 s). Overall, the proposed surrogate shows strong potential for multi-scale analysis and damage modeling applications.