A framework for fire performance assessment of mass timber structures : numerical and machine-learning approaches

This thesis advances the evaluation of fire performance in mass-timber structures, from individual components to a complete building system, through the development of novel numerical modelling approaches and the integration of machine-learning techniques. Firstly, timber compartment fire scenarios are generated based on the one-zone fire model and long short-term memory (LSTM) network. A user-oriented program, Compartment fire predictor, is developed to enable customized input parameters. The predictive capability of the predictor is comprehensively validated through benchmarking against large-scale compartment fire tests. Subsequently, for glulam members exposed to fire, an equivalent section temperature (EST) is proposed to simplify the non-uniform temperature field of the section into a single representative temperature that captures the strength/stiffness degradation. The modelling of the glulam column and beam with EST is validated based on experimental test results. The reliability assessment is carried out for the glulam column under standard fire exposure. A component-based glulam beam-column connection model is proposed using nonlinear springs to represent the responses of shear planes, simplifying the modelling procedure and improves the computational efficiency. A multi-fidelity neural network (MFNN) is then introduced to integrate the high-fidelity model and component-based model. The MFNN manages to balance the trade-off between the modelling accuracy and calculation speed, enabling sufficiently accurate prediction of connection response using only a limited number of high-fidelity simulations. Next, a numerical modelling approach for the heat transfer analysis considering the layer fall-off of the CLT panel under fire is developed. The developed numerical model is calibrated against the experimental standard fire test. Based on this framework, the probabilistic thermal response of CLT panels is evaluated, and the reliability of the post-protection factor is assessed. Finally, a prototype ten-storey timber frame building model is developed by integrating component models developed above. Compartment fire scenarios are formulated to account for the layer fall-off of the CLT ceiling under fire. The efficiencies of candidates for the intensity measure are assessed. The fragility assessments are carried out for thermal and structural responses, respectively.