Overcoming Geothermal Data Gaps with AI: Predictive Modeling and Imputation in the Lower Friulian Plain

Accurate prediction of aquifer temperatures is essential for optimizing low-enthalpy geothermal systems. However, hydrogeochemical datasets often suffer from significant missing values, particularly at greater depths, hindering reliable geothermal resource assessment. This study focuses on the lower Friulian Plain (north-eastern Italy), where over30% of temperature data is missing beyond 600 meters depth. Initially, a clean subset of the data was used to predict reservoir temperatures using six machine learning algorithms, including eXtreme Gradient Boosting (XGBoost), decision tree, generalized regression neural network, extreme randomized trees, radial basis function, and elastic net. XGBoost outperformed all others, achieving an R² of 0.9930, RMSE of 0.788, and MAE of 0.587, as confirmed by Wilcoxon signed-rank test. Sensitivity analysis highlighted bicarbonate as themost influential feature (0.51), followed by magnesium, electrical conductivity (0.11 each), and water depth (0.08).Following this, a deep learning autoencoder (ConvAE) was employed to impute missing values across the full dataset. ConvAE demonstrated high reconstruction accuracy (R² > 0.95) even under synthetic noise, enabling more comprehensive modelling. This two-step AI framework, first using ML for temperature prediction on clean data, then DL for imputation, offers a robust methodology for enhancing geothermal exploration and planning in data-scarce regions.