Verifying domain knowledge and theories on Fire-induced spalling of concrete through eXplainable artificial intelligence

Abstract

• Spalling theories and domain knowledge were verified for the first time using XAI. • XAI is an effective to gain new insights into the spalling phenomenon. • Recommendations for concrete mixtures with low tendency to spall under fire are proposed. This paper adopts eXplainable Artificial Intelligence (XAI) to identify the key factors influencing fire-induced spalling of concrete and to extract new insights into the phenomenon of spalling by investigating over 640 fire tests. In this pursuit, an XAI model was developed, validated, and then augmented with two explainability measures, namely, Shapley Additive exPlanations (SHAP) and Local eXplainable model-agnostic explanations (LIME). The proposed XAI model not only can predict the fire-induced spalling with high accuracy (i.e., >92 %) but can also articulate the reasoning behind its predictions (as in, the proposed model can specify the rationale for each prediction instance); thus, providing us with valuable insights into the factors, as well as relationships between these factors, leading to spalling. Our findings indicate that there are eight key factors that heavily govern spalling: 1) presence of Polypropylene fibers, 2) degree of moisture content, 3) heating rate, 4) maximum exposure temperature, 5) silica fume/binder ratio, 6) sand/binder ratio, 7) water/binder ratio and 8) fly ash/binder ratio. While these factors were also listed by the majority of the existing spalling theories, the contribution of each factor seems to vary significantly and, most importantly, was not quantified for the most part. Thus, the validated model was then utilized to contrast and quantify the spalling-based knowledge domain and theories as collected by some of the most cited studies in this domain.