Non-destructive terahertz assessment of mechanical fragility in palm-leaf manuscripts using interpretable machine learning

Conservation of palm-leaf manuscripts (PLMs) requires tools that can estimate mechanical fragility without additional sampling once calibrated against reference mechanical data. This study develops a conservation-oriented, calibration-based workflow that integrates terahertz time-domain spectroscopy (THz-TDS), supervised regression, cross-validated model evaluation, and SHapley Additive exPlanations (SHAP) for quantitative prediction of tensile strength. A total of 223 independent palm-leaf specimens representing different deterioration states were measured by THz-TDS, and reference tensile-strength labels were obtained by destructive testing along the longitudinal fiber direction for model calibration. Absorption-coefficient spectra were modeled using partial least squares regression (PLSR), random forest (RF), and extreme gradient boosting (XGBoost), with performance evaluated through cross-validation and a held-out test set. The optimized RF model achieved the best test-set performance (R 2 = 0.931, RMSE = 3.284 MPa, MAE = 2.807 MPa, MAPE = 14.078%). SHAP analysis identified absorption coefficients at 0.568 THz, 0.569 THz, and 0.205 THz as the most influential predictors. The adjacent variables at 0.568 and 0.569 THz are interpreted as a correlated predictive spectral region whose physical origin remains hypothetical. These results support a calibration-based proof of concept for tensile-strength estimation under the investigated specimen, instrument, acquisition, and environmental conditions. Once appropriately calibrated, THz-TDS acquisition and model prediction can be applied non-destructively.