Predicting barrier-fluid failure for temporary abandonment of petroleum wells with a tabular foundation model

Temporary well abandonment may rely on liquid barrier fluids that must retain integrity under temperature–time exposure. It was evaluated whether supervised learning can predict barrier-fluid failure from routine laboratory measurements (pH, density, apparent viscosity) obtained after controlled time and temperature exposure conditions, including exposures beyond typical operations to elicit failures. Random Forest, LightGBM, XGBoost and CatBoost were benchmarked against a tabular foundation model (TabPFN). Models were trained with a stratified split and assessed using accuracy, precision, recall, specificity, ROC-AUC, PR-AUC, F1-score, Cohen’s Kappa, Brier Score and calibration; generalization was tested on previously unseen experiments. TabPFN achieved the highest point estimates across most of the metrics and transferred well to the hold-out set. SHAP analyses identified apparent viscosity as the most influential predictor, reflecting degradation processes associated with temperature–time exposure. These findings show the potential of a foundation model to support reliable, data-driven assessment of barrier fluid condition following temperature–time exposure in temporary well abandonment applications, complementing conventional laboratory evaluation.