A novel approach to modeling breakdown pressure dynamics using machine learning

This study presents a novel machine learning (ML)-based approach for predicting breakdown pressure (BP) in hydraulic fracturing using experimental data. Unlike traditional analytical models that rely on simplified assumptions, ML models can capture complex nonlinear relationships between BP and its influencing factors. However, a key limitation in BP prediction stems from dataset constraints, particularly the scale differences between experimental setups and real-world formations. To mitigate these limitations, this research utilizes a unique dataset of 144 BP data points, incorporating various rock mechanical properties, injection parameters, and fluid properties. Additionally, a separate analysis of pressurization rate, based on 32 additional experimental data points, was conducted to better understand its effect on fracture initiation—an aspect often overlooked in ML-based studies. The dataset includes critical parameters such as injection rate, confining pressure, tensile strength, Young's modulus, permeability, unconfined compressive strength, Poisson's ratio, porosity, wellbore radius, and fracture geometry ratio. Five ML models—LightGBM, CatBoost, XGBoost, Kolmogorov-Arnold Network (KAN), and TabNet—were trained and evaluated. TabNet achieved the highest predictive performance (R² = 0.94) due to its attention-based feature selection and deep-learning-based representation learning. Model performance was assessed using mean absolute error (MAE) and mean squared error (MSE) to ensure robustness. To further enhance model interpretability, SHapley Additive exPlanations (SHAP) and TabNet's attention mechanism were used to explicitly assess feature importance, providing insights into the relative influence of different parameters on BP predictions. Additionally, advanced feature-handling techniques were employed to address categorical variables automatically, ensuring minimal preprocessing bias. The findings demonstrate the scalability of ML models for BP prediction using experimental data, reducing reliance on costly and time-consuming laboratory testing. By incorporating advanced interpretability techniques, systematic pressurization rate analysis, and robust ML architectures, this research provides a more accurate, data-driven approach for optimizing hydraulic fracturing designs.