An integrated physics–machine learning framework for time-dependent injectivity decline and fracture growth in horizontal water-injection wells

• Hybrid physics–ML model simulates injectivity decline and fracture growth • Fracture propagation governed by mechanical admissibility within ML space • Dynamic evolution traced through static surrogate response space • Throughput classification enables interpretable uncertainty analysis • Filtration damage, pressure support, and operating time control fracture growth Water injection is widely used for reservoir pressure maintenance, waterflooding, and produced-water disposal; however, sustained injectivity is often compromised by particle-induced formation damage and unintended fracture propagation. Suspended solids accumulate near the wellbore and fractures, increasing bottom-hole pressure, reducing injectivity, and potentially initiating fractures that alter flow distribution. Predicting the coupled evolution of particulate-induced damage and fracture growth in horizontal injectors remains challenging because analytical models cannot represent transient flow behavior, while high-fidelity simulators are computationally expensive for dynamic workflows. This study presents an integrated physics–machine learning framework to simulate injectivity decline and fracture half-length evolution in horizontal water-injection wells subjected to particulate-laden injection. Machine-learning surrogate models trained using tree-based regressors (CatBoost, LightGBM, and XGBoost) on more than 30,000 high-fidelity reservoir simulations act as hydraulic response operators, providing bottom-hole pressure, fracture-tip pressure, and flow partitioning for static states of a horizontal well intersected by a single transverse fracture across a wide range of reservoir, well, and operating conditions. The time-dependent solution is obtained by sequentially tracing mechanically admissible fracture-length states within the surrogate response space as time and filtration damage evolve. Simulation scenarios organized into throughput-based injection systems enable interpretable uncertainty analysis. Results show that fracture initiation and growth are governed by near-well filtration damage, reservoir pressure dissipation capacity, cumulative particulate injection, and operating duration. The framework provides a computationally efficient and physically consistent tool for fracture-risk assessment and injection design.