Artificial neural networks as a decision-support system for predicting the quality attributes of thermally modified wood

Abstract

Thermal modification represents a key, environmentally safe method to enhance the dimensional stability and biological resistance of timber. However, the complex, non-linear interactions among coupled hydrothermal process variables often render traditional linear empirical models inadequate. This study addresses this limitation by developing a robust multi-objective decision-support model driven by a Multi-Layer Perceptron (MLP) Artificial Neural Network (ANN). A curated meta-dataset comprising 200 validated experimental data points from peer-reviewed literature (2000–2025) was compiled across softwood and hardwood species. The proposed MLP framework integrates five input parameters (Wood Taxa/Species Type, Density, Temperature, Duration, and Medium) to simultaneously forecast three primary quality attributes: Mass Loss (\(\:ML\)), Equilibrium Moisture Content (EMC), and Modulus of Rupture loss (\(\:{\Delta\:}\text{MOR}\)). Rigorous 5-fold cross-validation on an unseen independent test set (\(\:n=40\)) yielded high predictive fidelity (\(\:{R}^{2}>0.925\) across all outputs), outperforming baseline Multiple Linear Regression (MLR, \(\:{R}^{2}=0.695\text{--}0.742\)), Support Vector Regression (SVR), and Random Forest (RF) algorithms. A SHapley Additive exPlanations (SHAP) model interpretability analysis pinpointed treatment temperature (\(\:49.5\%\)) and processing duration (\(\:27.9\%\)) as the primary kinetic drivers of cell-wall thermochemical degradation. This computational framework provides an accurate virtual prototyping tool for smart timber manufacturing, eliminating destructive sampling while respecting established operational boundaries.