From prediction to design: Optimized design of waste wood composites via GAN-augmented interpretable machine learning

Machine learning (ML) models rapidly and accurately predict material properties at low computational cost. An ML approach combined with Generative Adversarial Networks (GAN) for data augmentation predicts the mechanical properties of waste wood-based composite. Using phytic acid (PA) and tannic acid (TA) as green additives in combination with isocyanate (MDI), waste wood-based composite was prepared, and their modulus of elasticity (MOE) and modulus of rupture (MOR) were systematically tested. To address the challenge of limited experimental data, GAN were employed to augment the mechanics data. The quality of the generated data was validated using metrics such as Fréchet Inception Distance (FID), Wasserstein Distance, and Kolmogorov-Smirnov (KS) statistics. Comparing ML models before and after data augmentation, XGBoost performed optimally for both MOE and MOR prediction after GAN enhancement, with testing set R² values reaching 0.7846 and 0.9742, respectively, while overfitting was significantly alleviated. SHAP interpretability analysis further revealed that TA concentration had the most pronounced effect on MOE, while adhesive content was the key factor determining MOR. Through response surface methodology (RSM) analysis, the optimal process formulations for the MOE and MOR of waste wood-based composites were determined to be TA = 7.20 wt%, PA = 8.87 wt%, MDI = 7.85 wt%, and TA = 4.28 wt%, PA = 8.97 wt%, MDI = 7.85 wt%. This study demonstrates GAN-based data augmentation effectively enhances the predictive capability of ML models for small-sample material data, offering a data-driven pathway for the high-value utilization of waste wood resources and the design of high-performance composites.