Wood anatomy-guided machine learning linking macromechanics in plantation-grown Chinese fir

Abstract Understanding how anatomy shapes wood mechanics is essential for grading and breeding. This study develop an interpretable-SHAP-based framework providing the first conditionally independent decomposition of anatomical effects on Chinese fir ( Cunninghamia lanceolata ) performance. To mitigate data scarcity, synthetic-data generated by three generative models were evaluated for correlation, distribution and prediction. Five machine-learning models were trained on synthetic data to predict modulus of elasticity (MOE), modulus of rupture (MOR), and compressive strength parallel to grain (CSP). All generative methods produced realistic data, with Gaussian Copula performing best. The best accuracy was achieved by CopulaGAN–XGBoost (MOE, 75 %), Gaussian Copula–AdaBoost (MOR, 80 %), and Gaussian Copula–Random Forest (CSP, 91 %), outperforming models trained on real data (58 %, 64 %, 68 %). SHAP analysis identified tracheid length (15.6 %, 6.0 %, 8.3 %), wall thickness (13.8 %, 12.6 %, 7.5 %), and microfibril angle (3.4 %, 2.7 %, 4.5 %) as key traits, with microfibril angle showing the strongest interactions. Latewood versus earlywood contributions were 24.1 % versus 16.7 % (MOE), 16.6 % versus 39.9 % (MOR), and 18.1 % versus 28.8 % (CSP). Density was the most influential trait: strength (MOR, CSP) was driven by density and earlywood traits, while stiffness (MOE) depended on density and overall anatomy. These findings provide interpretable guidance for wood quality assessment, material grading and plantation improvement.