Woven composites are integral to industries such as aerospace, automotive, and civil engineering, where high structural integrity and material efficiency are critical. These materials, known for their outstanding strength-to-weight ratios and adaptability, are vital in crafting lightweight and durable designs. The complex meso-scale structures of woven composites, coupled with geometrical configurations, microscale properties, and manufacturing imperfections like fiber waviness and misalignment, make it difficult to accurately predict mechanical behaviors under varied loading conditions. Consequently, theoretical predictions and actual material behavior differ considerably.

Conventional modeling techniques like finite element analysis, while providing deep insights, suffer from high computational costs and limited scalability due to the inherent variability of composite materials. Surrogate neural network models offer a promising alternative with reduced computational demands and improved adaptability. However, their effectiveness is often limited by the quality and representativeness of the training data, with traditional single fidelity datasets failing to generalize across real-world conditions [1].

This research introduces a pioneering surrogate model that employs domain adaptation for predicting inelastic behavior in woven composites, a first in this field to our knowledge. By utilizing contrastive learning, the model minimizes discrepancies between full-field and mean-field samples, aligning them based on stress characteristics [2]. This approach not only enhances prediction accuracy but also addresses the critical issue of negative transfer in surrogate models. A negative transfer of knowledge from a source domain to a target domain is mitigated by advanced domain adaptation techniques, which ensure that the transfer of knowledge augments rather than diminishes the model's performance. By demonstrating robust performance even with limited data, our model provides a comprehensive solution to the challenges of variability and model reliability. In our evaluations, we found significant improvements in predictability and usability across a variety of applications, which is an important advancement in material science.