Machine Learning-Enabled Initial Productivity Prediction of Hydraulically Fractured Horizontal Wells in Volcanic Reservoirs

Initial production capacity is a critical metric for evaluating the deliverability and economic viability of new wells. Given the highly heterogeneous nature of volcanic reservoirs, a strongly non-linear mapping relationship exists between physical parameters and initial productivity. To investigate this relationship under controlled parameter combinations, this study constructed numerical reservoir models based on an orthogonal experimental design to generate a simulation-derived productivity dataset, and established three initial productivity prediction models: Multiple linear regression (MLR), PSO-LightGBM, and PSO-TabNet. Particle Swarm Optimization (PSO) was incorporated to search the internal hyperparameters of both LightGBM and TabNet. Subsequently, 15 new parameter combinations outside the orthogonal experimental scheme were utilized as an independent test set to validate model performance within the numerical simulation scenarios considered in this study. The results show that the non-linear machine learning models captured the simulated multi-factor productivity response more effectively than MLR, and PSO-TabNet achieved the lowest test-set error among the three evaluated models, with a Mean Absolute Percentage Error (MAPE) of 3.28%. The proposed framework therefore demonstrates promising predictive capability for the numerical simulation scenarios considered in this study. However, because the current dataset is generated from numerical simulation rather than direct field production, further validation using field data and additional baseline models is required before broader operational deployment.