Improving Fire Resistance Prediction of Glulam Timber Columns Using Gaussian Process Surrogate Modelling

Abstract

The growing application of mass timber in structural engineering has increased the need for reliable assessment of structural fire performance beyond conventional fire resistance ratings. Current Eurocode provisions are largely based on standard fire curves, such as ISO 834, which consider only the heating phase. However, experimental research has demonstrated that glued laminated timber (glulam) columns can fail during the cooling phase of natural fires, a phenomenon referred to as delayed failure. This behaviour is linked to the combustible nature of timber, its low thermal conductivity causing slow heat penetration, and the rapid degradation of mechanical properties at elevated temperatures. To account for the complete fire exposure, including heating and cooling, the concept of burnout resistance is introduced. This metric evaluates whether a structural member maintains its load-bearing capacity throughout the entire fire event.

This thesis aims to improve the understanding and prediction of delayed failure and burnout resistance of glulam timber columns subjected to natural fire scenarios. Two main objectives are addressed: (i) assessing the capability of non-linear thermo-mechanical finite element modelling in SAFIR to reproduce delayed failure under realistic fire exposures, and (ii) investigating Gaussian Process Classification (GPC) as a surrogate modelling technique for efficient parametric evaluation of burnout resistance.

A combined numerical–surrogate framework is developed. A parametric finite element model is constructed in SAFIR for isolated spruce glulam (GL24h) columns with square cross-sections, four-sided fire exposure and eccentric axial loading. Model predictions are compared with available full-scale experimental data from natural fire tests. The numerical simulations provide binary failure outcomes that serve as input for GPC models, which are enhanced using an active learning strategy to efficiently refine the estimated failure boundary. Surrogate models are developed progressively, from low-dimensional representations to a six-dimensional parameter space incorporating structural, fire and material variables.

The simulations reproduce the structural response during the heating phase with reasonable accuracy. However, predictions of delayed failure during the cooling phase are less reliable, primarily due to limitations in thermal modelling of sustained internal temperatures. Improved agreement is achieved when internal temperature histories better match experimental observations. The results demonstrate that GPC can approximate numerical generated failure boundaries and support efficient exploration of interacting variables, provided that the underlying numerical model is sufficiently accurate. Overall, the findings highlight the importance of accounting for cooling phases and improving the fire resistance assessment glulam timber columns.