Explainable machine learning-based prediction of blast loads on structural surfaces in two-dimensional spatial coordinates

• Artificial neural networks (ANN) were used to predict blast loads on structures. • Shapley Additive Explanations were used to explain the ANN’s prediction process. • Logic followed in the prediction processes adheres to fundamental blast theories. • Scientific consistency of the prediction processes strengthens ANNs’ reliability. • ANN-based framework visualised blast load predictions on a structure’s surface. High-fidelity analysis tools are widely used for shock wave pre-propagation and modelling in blast engineering. However, their time-consuming nature highlights the need for efficient alternatives. In recent years, Artificial Neural Networks (ANN) have gained significant traction in blast engineering research for their remarkable accuracy in predictive modelling. However, the work done so far has primarily focused on framework development, rather than validating the procedure. Hence, this manuscript focuses on establishing the transparency and trustworthiness of ANNs through Explainable Machine Learning (XML). Additionally, an effective visualisation method for shock wave propagation was developed in this research, addressing a gap, as only 1–2 studies have previously focused on blast load visualisation. This manuscript reports the initial work done in the ongoing development of an ANN-based predictive tool forecasting crucial blast loading parameters: peak incident pressure, impulse, reflected pressure, and time of arrival. The ANNs were developed using datasets of 500–1500 blast events derived from experimentally validated numerical simulations and they demonstrated predictive accuracies with an R 2 value exceeding 0.99. XML techniques were employed to interpret the prediction process, confirming the ANNs' adherence to blast principles. Using a Python framework, the ANNs were extended to visualise the distribution of blast loads on 2D spatial coordinates across a structure’s surface, achieving a maximum error range of −8 % to 6.5 % compared to validated complex high-fidelity model prediction. By validating the reliability of these ANNs through rigorous visual outputs, this research pioneers the first step in developing an alternative predictive tool to conventional shock wave propagation techniques.