Importance-driven bottleneck model: a multi-stage deep learning approach for analyzing swelling-shrinkage behavior of wood species and their mechanical properties

Artikel i vetenskaplig tidskrift, 2026

Experimental data were collected from defect-free specimens of 57 wood species in Central Europe, encompassing their structural, physical, chemical characteristics, and mechanical properties. The dataset highlights the complex relationships between physical properties (shrinkage–swelling behavior in three orthotropic directions), chemical composition (holocellulose, lignin, extractive content), structural characteristics (density, fiber length), and mechanical properties (compression, tensile and bending strength, hardness, and impact resistance), emphasizing the need for efficient and interpretable methods to identify their correlations. To address this issue, we developed an importance-driven bottleneck model, by revisiting the concept bottleneck model, which consists of a system of three connected neural networks; auxiliary, input, and output, that collaboratively work to investigate the relationship between the physical, structural, chemical, and mechanical features. The input and output networks are linked by a concept c bottleneck layer, with its features defined by analyzing wood’s structural characteristics and composition via feature importance analysis conducted in the auxiliary network. Density and fiber length were identified as the most important concept c features, yielding testing  scores above 0.70 for predicting mechanical properties. Predictive accuracy increased to over 0.85 when a third component, either holocellulose or lignin content, was included in the bottleneck. By adding all five chemical and structural features used in concept c layer, the prediction accuracy was slightly increased. While the importance-driven bottleneck model with three or more concept c features was narrowly outperformed by an end-to-end network using shrinkage–swelling as direct inputs, it offered a superior balance between performance and interpretability. This mathematical strategy, forcing information through a concept bottleneck, creates an interpretable physics-inspired AI framework.