PREDICTING BENDING STRENGTH OF PHILIPPINE BAMBOO SPECIES USING ARTIFICIAL NEURAL NETWORKS

Abstract

Bamboo has gained attention as a sustainable structural material because of its advantageous strength-to-weight ratio and rapid renewability. However, the variation in its mechanical properties across species and geometric characteristics presents challenges for reliable structural design, while multispecies data-driven prediction of bamboo bending strength remains limited. This study reported standardized bending test data and developed an artificial neural network (ANN) model using pooled experimental data from five bamboo species. Four-point bending tests were conducted on four species following ISO 22157, while data for one additional species was obtained from a previously published study. The pooled dataset comprised 313 specimens. Mean bending strengths ranged from 64 to 84 MPa, with coefficients of variation of 20-26%, demonstrating material variability. The ANN used outer diameter, wall thickness, moisture content, density, thickness-to-diameter ratio, and the density-thickness product as predictors. The optimized model achieved a test R² of 0.89, RMSE of 6.76 MPa and MAPE of 3.08%. Repeated validation across multiple random seeds produced a mean test R² of 0.807 ± 0.094, providing an additional assessment of predictive performance across alternative data partitions. Permutation feature importance indicated that the trained ANN relied most strongly on wall thickness. The resulting framework may support preliminary bending-strength estimation within the geometric and physical domain represented by the pooled dataset.