Key points are not available for this paper at this time.
The mechanical properties of short fiber-reinforced polymer composites (SFRPCs) are governed by compositional and structural parameters, particularly the strong anisotropy induced by intricate fiber orientation distributions. These interdependencies create a high-dimensional design space, posing significant challenges for inverse design. This work proposes a collaborative paradigm that integrates interpretable machine learning with a genetic algorithm (GA) to improve the design and optimization of SFRPCs through forward prediction and inverse exploration. An ensemble surrogate model, trained on simulated data using second-order orientation distribution tensors to characterize fiber alignment, predicts homogenized elastic properties with less than 5% error validated by simulations and experiments. Leveraging the forward surrogate model's interpretability and coupled with GA, inverse exploration is performed to identify optimized parameters. Results demonstrate that the optimization strategy achieves significant error reduction, with the relative error decreased from 9.26% to 2.91% for the single-objective design task, and from 12.04% to 1.46% for the multi-objective design task. The interpretability-driven optimization framework establishes an efficient bidirectional mapping between microstructural configurations and macroscopic properties, providing superior performance in navigating high-dimensional design spaces while maintaining physical interpretability. The work provides insights into both rapid property prediction and microstructure reverse design for anisotropic composites.
Dong et al. (Tue,) studied this question.