What question did this study set out to answer?

The aim is to develop a multi-fidelity design framework that predicts effective properties of composite structures while minimizing costs.

January 25, 2026

Multi-fidelity design and optimization framework of damage-tolerant fiber-reinforced composite structures via CNN-assisted homogenization

Puntos clave

The aim is to develop a multi-fidelity design framework that predicts effective properties of composite structures while minimizing costs.
Homogenized 2D unit cells were assembled into larger structures via layout matrices.
Neighborhoods of each unit were encoded and inputs fed into convolutional neural networks to predict property deviations.
A lightweight model was created for rapid training and efficient inference in resource-constrained environments.
The framework eliminated over 92.42% of expensive full-fidelity simulations, achieving a 22-fold speed-up.
The model trained in under 30 minutes on a single GPU while maintaining over 80% accuracy compared to high-fidelity models.
Convergence studies confirmed robustness and scalability across new configurations.

Resumen

Purpose Neighbor interactions among representative volume elements induce systematic perturbations in their effective stiffness and Poisson's ratio when these units are embedded in larger composite structures. However, classic homogenization treats each unit's effective properties as invariant, regardless of the surrounding microstructure. To bridge this gap at minimal cost, we introduce a multi-fidelity homogenization framework augmented by a convolutional neural network correction. Design/methodology/approach First, 2D unit cells are homogenized and then assembled into large structures via layout matrices. For each unit, its neighborhood is encoded into a two-channel input and passed to high-capacity neural network models, which predict the local deviations from baseline effective properties. Then, the homogenization results of each unit cell within an assembled structure are corrected correspondingly. Findings This approach eliminates more than 92.42% of expensive full-fidelity simulations, yielding a 22-fold speed-up. To enable deployment in resource-constrained or high-throughput settings, we then systematically prune and reorganize the neural network architectures into a lightweight model that trains in under 30 min on a single graphics processing unit and achieves five times faster inference with comparable accuracy (over 80% of the high-fidelity accuracy can be recovered). Convergence and generalization studies confirm that the model remains robust and scalable across unseen configurations, paving the way for rapid structural design and optimization campaigns. Originality/value This study introduces a convolutional neural network-enhanced multi-fidelity framework for designing and optimizing fiber-reinforced composite structures intended for segmental failure. Unlike previous works focused on material behavior and fabrication, this research preliminarily addresses the unexplored gap of integrating design, performance prediction, and geometry-driven optimization. The framework enables fast, accurate and scalable exploration of local geometry effects, delivering a damage-tolerant, energy-absorbing composite structure that enhances resilience.

Me gusta

Guardar