Geothermal Resource Characterization in the Lower Friulian Plain: Integrating Artificial Intelligence and Physics-Based Modelling

Geothermal resource characterization in sedimentary–carbonate basins is often hindered by sparse, heterogeneous, and incomplete datasets, limiting the reliability of conventional geochemical, geophysical, and numerical approaches. This challenge is particularly evident in the Lower Friulian Plain (NE Italy), a geologically favorable yet underexplored geothermal province characterized by thick Meso–Cenozoic carbonate platforms, complex fault-controlled hydrothermal circulation, and severe data imbalance in temperature and hydrochemical observations. This PhD thesis develops an integrated framework that combines artificial intelligence (AI) techniques with physics-based modelling to transform fragmented regional datasets into a physically consistent interpretation of geothermal behavior from shallow aquifers down to depths of 1200 m. The research is structured into four interconnected stages. First, machine-learning models were benchmarked on a strictly complete subset of the hydrochemical–temperature database to assess the predictive value of geochemical indicators. Among six tested algorithms, extreme gradient boosting achieved the highest accuracy (R² ≈ 0.99), confirming that bicarbonate, magnesium, electrical conductivity, and depth encode meaningful thermal information. Second, the problem of widespread missing data was addressed using a convolutional autoencoder trained to reconstruct temperature and hydrochemical variables across the entire regional dataset. The reconstructed fields were further analyzed using Isolation Forest anomaly detection, revealing spatially coherent hydrothermal anomalies aligned with major fault systems and structurally elevated carbonate platforms. Third, a physics-guided Bayesian neural network framework was developed to reconstruct the deep-temperature field of the Lower Friulian Plain under strong vertical data imbalance. By integrating imputed regional data, deep borehole logs, and auxiliary geospatial features, heteroscedastic Bayesian models with Monte Carlo dropout were trained under progressively stronger physical constraints, including monotonicity of vertical temperature gradients and gradient-band limits consistent with conductive heat transfer in carbonate systems. The resulting models produce a continuous three-dimensional temperature field with quantified epistemic and aleatoric uncertainty, maintaining physical coherence even in data-poor regions and successfully validating against independent well data. Finally, a fully coupled three-dimensional hydrothermal model of the Grado geothermal doublet was developed and calibrated using observed hydraulic heads and borehole temperature profiles. Long-term simulations demonstrate that sustainable operation is achieved for discharge rates up to approximately 100 L/s combined with well spacings of at least 600–700 m, while higher extraction rates or reduced spacing lead to accelerated thermal decline and efficiency losses. Overall, this thesis demonstrates that AI and physics-based modelling act as complementary, scale-dependent tools for geothermal characterization. AI methods enable regional-scale reconstruction, prediction, and uncertainty quantification under data scarcity, while numerical modelling provides physically grounded validation and reservoir-scale insight. The proposed framework is transferable to other heterogeneous and data-limited geothermal systems and supports more reliable, sustainable geothermal development strategies.