AGFNN: A smart platform for uncertainty-aware prediction of freezing time in artificial ground freezing

Accurate and efficient prediction of freezing time ( t f ) is essential for the design and safety evaluation of artificial ground freezing (AGF) systems. Traditional numerical simulations, while accurate, are computationally intensive and often impractical for iterative design or real-time decision-making. This study proposes a surrogate modeling framework based on a neural network ( NN ) architecture to predict t f across a wide range of soil conditions and design parameters. The surrogate model is trained on a large dataset generated from transient finite element simulations and incorporates Monte Carlo dropout (MCD) to enable uncertainty quantification. The NN -based surrogate model demonstrates high predictive accuracy and rapid inference, enabling reliable estimation of t f with associated confidence intervals. To enhance model transparency, the Shapley Additive Explanations (SHAP) method is also employed to evaluate the relative importance of each input parameter in influencing the predicted output. The trained model is deployed as a standalone desktop application named AGFNN with an intuitive graphical user interface (GUI), supporting both single-case manual input and batch-mode Excel-based predictions. The platform streamlines the prediction workflow for engineering users, enabling uncertainty-aware decision-making without the need for numerical modeling expertise. Results demonstrate that the proposed method significantly reduces computational cost while maintaining interpretability, offering a robust and accessible tool for design optimization and risk-informed decision support in AGF projects.