Cross-scale multimodal modeling of wood mechanical properties: Prediction and mechanistic insights

Accurate evaluation of wood mechanical properties is fundamental to structural design and material grading in construction applications. However, it remains challenging because of wood's intrinsic cross-scale heterogeneity and the limitations of conventional evaluation methods. This study proposes a unified cross-scale multimodal framework that integrates macroscopic physical parameters with microscopic anatomical information to predict key mechanical properties of wood. To evaluate model robustness and generalizability across diverse wood species, a systematic benchmarking scheme comprising a 10 × 10 evaluation matrix of vision backbones and regression algorithms is established. The evaluation results demonstrate that the proposed framework consistently outperforms single-modality baselines, achieving stable predictive performance for compressive strength parallel to grain (CSG), modulus of rupture (MOR), and modulus of elasticity (MOE), with R 2 values exceeding 0.73 and reaching 0.8050. Interpretability analysis further reveals distinct cross-scale governing mechanisms among different mechanical properties. Macroscopic physical attributes provide the primary explanatory basis for CSG and MOE. In contrast, microscopic anatomical features play a substantially greater role in MOR prediction, contributing 38.52% of the total explanatory weight because bending failure is highly sensitive to localized stress concentrations. Furthermore, integrating anatomical features from three orthogonal sections provides a comprehensive representation of the three-dimensional wood structure. This mitigates single-plane biases and ensures more robust predictive performance across diverse species. Overall, the proposed framework provides a practical and interpretable approach to cross-scale prediction and mechanism discovery, with potential applications in high-throughput wood grading and data-driven quality assessment for construction.

Hali tarjima qilinmagan