Enhancing concrete frost resistance prediction with an explainable neural network

Abstract

Concrete in frigid northern regions is prone to the detrimental impacts of freeze-thaw cycles, leading to the development of cracks and spalling, which affects structural safety and durability. This research introduces a novel method for predicting concrete frost resistance utilizing an explainable machine learning model named Generalized Additive Model with Interactions Networks (GAMI-Net). This model provides both precise predictions and comprehensive explanations. The dataset, consisting of 781 sets of freeze-thaw test data encompassing influencing factors and evaluation indices of concrete frost resistance (including compressive strength, dynamic elasticity modulus, and mass loss rate), is employed for model development. The prediction outcomes showcase the superior accuracy of the GAMI-Net model when compared to other machine learning models like EBM and XGBoost. The model achieves a root mean square error (RMSE) of 0.0584, mean absolute error (MAE) of 0.0386, and an R-squared (R2) of 0.9281 for compressive strength prediction. Field experiments confirm the model's efficiency in predicting concrete performance with minimal deviation from predictions. GAMI-Net's interpretability offers insights into feature importance and individual factor contributions through global and local interpretations. The model excels in visualization, rendering the results easily understandable and conferring a distinct advantage in terms of clarity. This approach holds the potential to guide concrete mix ratio optimization, ultimately fortifying the frost resistance of concrete structures.

Keywords

• Frost Resistance of Concrete
• Machine Learning
• GAMI-Net Model
• Interpretability