Machine Learning-Assisted Multi-Objective Optimization of Geothermal Well Placement Through Three-Dimensional Fracture-Network Modeling

Selecting injection and production wells in a fractured geothermal reservoir is a multi-objective problem because designs that increase heat recovery can also accelerate thermal decline or impose excessive hydraulic demand. This study develops an original simulation and optimization workflow that is independent of thermo-hydro-mechanical modeling. The forward model couples steady single-phase Darcy flow with transient local thermal nonequilibrium heat transport in a three-dimensional voxelized fracture network. Separate fluid and rock energy balances represent advective heat removal, phase-specific conduction, and interphase heat exchange. A Latin hypercube design generated 720 simulations across three stochastic fracture networks. Random forest, extremely randomized trees, gradient boosting, and multilayer perceptron surrogates were evaluated on a fixed independent test set. Ninety active-learning simulations were added near the preliminary Pareto region. The selected gradient-boosting surrogate achieved test-set coefficients of determination of 0.815 for final production temperature, 0.881 for logarithmic thermal drawdown, and 0.994 for cumulative heat. NSGA-II then minimized pressure drop and thermal drawdown while maximizing cumulative heat. To prevent surrogate error from determining the reported optimum, all 120 surrogate-Pareto candidates were re-simulated, producing a 99-design exact nondominated front. On the reference grid, the compromise design recovered 2035.0 GWhth over 20 years, compared with 940.0 GWhth for the baseline, while pressure drop decreased from 3.194 to 0.296 MPa. Final production temperature was 151.0 °C and no 10% thermal breakthrough occurred within 20 years. Mass and energy residuals were and , respectively. Grid tests showed stable cumulative heat but substantial hydraulic sensitivity caused by voxelized fracture connectivity. The results therefore demonstrate a reproducible screening method, not a site forecast. Full re-simulation of the final candidates provides a defensible separation between machine-learning acceleration and physics-based evidence.