28 June 2026
New paper: Bayesian model selection for off-axis strength
My new paper, On the selection of off-axis strength models for anisotropic materials: a Bayesian approach, has been published open access in Structures.
The paper tackles a common problem in empirical modelling: a more complex model can look better when judged only by goodness-of-fit, but the extra parameters may not improve prediction for new data. Using a Bayesian framework with leave-one-out cross-validation, the study compares competing off-axis strength models across seven datasets covering glass fibre-reinforced polymers, carbon fibre-reinforced polymers, bamboo composites, and wood.
The main result is that the simpler three-parameter Generalized Hankinson model consistently outperforms the more complex four-parameter Li-Wei-Wang model for most datasets. The paper also shows why reporting individual specimen data matters: specimen-level analyses are much better at distinguishing between competing models than analyses based only on summary statistics.
Elsevier’s AI reading assistant described the contribution as a reproducible framework that penalises unnecessary model complexity, quantifies uncertainty, and challenges the assumption that a more general empirical model is automatically a better one. That is the heart of the paper: better model selection is not about adding parameters, but about improving prediction.