Key points are not available for this paper at this time.
Leveraging available experimental results on the development of temperature dependent material properties, enabled the creation of 96 finite element analysis models. Those examined the heat transfer process through concrete walls exposed to the standard fire curve ISO834, while monitoring the developed temperature trends on the non-exposed face. The analysis results were used for training and testing an artificial neural network on 23 independent geometrical and material property variables and one dependent variable (total of 24 variables), predicting the ultimate temperature developed on the unexposed face of the wall. Game theory SHAP algorithms were employed to “explain” the ANN results and allow a quantitative interpretation of the contribution of each factor to the final predictions of the network. The intuitive assumption that the thickness of the wall and imposed temperature development profile are the dominant factors is confirmed and numerically quantified, with contributions of 30.78 % and 22.43 % calculated respectively. Factors describing the development of material properties cumulatively account for 32.56 % of the predicted temperatures while the corresponding contribution of initial and final material property figures is 6.98 %. This not only gives an insight to the most influential factors affecting the thermal behaviour of walls exposed to elevated temperatures but also highlights the need for careful consideration on how materials are classified and used in real life applications. • Game theory SHAP algorithms for increasing explainability of ANNs. • Quantifying contribution of temperature dependent material properties. • Combination of experimental and FE simulation results for training ANN algorithm. • Walls of different concrete mixes exposed to standard fire curve ISO834. • ANN development protocol and accuracy assessment methodology.
Bakas et al. (Tue,) studied this question.