Abstract: Geothermal resources hosted in sandstone reservoirs offer a clean, reliable, and renewable energy source that can support regional energy transitions under China’s carbon peaking and carbon neutrality targets. However, complex geology, pronounced reservoir heterogeneity, and exploration uncertainty hinder the accurate regional assessment of geothermal-resource potential. In the Northwestern Shandong Plain, 10 geological variables relevant to geothermal potential were compiled, and feature selection was performed using Pearson correlation analysis and variance inflation factors. Subsequently, a vision transformer Kolmogorov–Arnold network (ViT-KAN) model was developed to assess the geothermal-resource potential of sandstone reservoirs by replacing the multilayer perceptron module in the transformer encoder of the ViT with a KAN. Model performance was evaluated from the complementary perspectives of classification discrimination and probabilistic prediction accuracy using six metrics: area under the receiver operating characteristic curve (AUC), accuracy (ACC), recall, mean squared error (MSE), root mean squared error (RMSE), and mean absolute error (MAE). Shapley additive explanations (SHAP) and counterfactual analysis were integrated to identify key factors governing the model predictions. The depth to the base of the geothermal reservoir was strongly and positively correlated with reservoir thickness, with a Pearson correlation coefficient of 0.88; therefore, it was excluded during feature selection to reduce redundancy among the input variables. On the test set, the ViT-KAN model achieved an AUC of 0.936, an accuracy of 0.840, and a recall of 0.907, thus exceeding corresponding ViT values of 0.914, 0.802, and 0.887, respectively. The test set MSE, RMSE, and MAE of the ViT-KAN model were 0.161, 0.401, and 0.311, respectively, compared with 0.201, 0.448, and 0.344 for the ViT. These results indicate that the ViT-KAN model provides better classification discrimination and higher probabilistic prediction accuracy compared with the ViT alone. The predicted probabilities were classified into five geothermal potential levels using the natural-breaks method. Although the two models yielded broadly consistent spatial patterns, the ViT-KAN model provided a more detailed delineation of areas with high geothermal potential. Under the ViT-KAN model, areas with very low and low geothermal potential constituted 60.9% of the study area, whereas areas with medium, high, and very high geothermal potential constituted 18.1%, 12.2%, and 8.8%, respectively. Areas with high geothermal potential were concentrated primarily in Decheng District, Lingcheng District, and Pingyuan County in the northern region of the study area. The predicted potential classes were generally consistent with the spatial distribution of known wells, and 87.61% of the geothermal wells were located in areas classified by the ViT-KAN model as exhibiting very high geothermal potential. The reduced proportion of non-geothermal wells in areas with medium and high geothermal potential suggests that the ViT-KAN model can identify areas with low geothermal potential and mitigate exploration risk more effectively. SHAP and counterfactual analyses consistently identified reservoir temperature, porosity, and reservoir thickness as the most influential variables. Higher values of all three variables generally increased the predicted geothermal potential, while reducing them individually or jointly shifted a sample from high to low potential class. The agreement between the two interpretation methods enhances the credibility of the model predictions. These findings provide a scientific basis for regional geothermal-resource assessment and the prioritization of exploration targets.