Prediction of Hydraulic Fracturing Initiation Pressure in Natural Gas Hydrate-Bearing Sediments Using a Deep Neural Network

Hydraulic fracturing offers a potential solution to the challenges of low productivity and sand production in natural gas hydrate (NGH) extraction. However, accurately calculating the fracturing initiation pressure Pb remains challenging due to the structural complexity of NGH-bearing sediments. This study compiles existing experimental data and develops a predictive model for Pb using a deep neural network. The effects of architecture, configuration, and regularization on model performance were systematically analyzed. Results indicate weak correlations between influencing factors and Pb with absolute correlation coefficients below 0.3. Key factors such as the number of neurons and hidden layers, batch size, batch normalization, loss functions, optimizers, and activation functions markedly influence model performance. Neither L1/L2 regularization, dropout, nor their combinations substantially improve the model. Only in specific cases do certain regularization tricks enhance particular metrics. The optimized model achieves test performance metrics of MSE = 19.34 MPa2, MAE = 3.16 MPa, and R2=0.81. At the same time, by comparing the performance of models using Dataset A and Dataset B, it was found that injection parameters (Qi, μi) and geometric scale effects (Vs, d50) have a certain influence on the fracture initiation pressure in natural gas hydrate-bearing sediments. Expanding the dataset is expected to further improve model performance.