To address the limited in situ stress data and poor nonlinear fitting of existing methods, a Particle Swarm Optimization (PSO)–XGBoost inversion approach is proposed. XGBoost effectively models complex relationships between finite element results and measured stresses, leveraging its strong nonlinear mapping and suitability for small samples. PSO globally optimizes XGBoost hyperparameters, utilizing its fast convergence and global search capability. Combined with 5-fold cross-validation, this avoids empirical tuning errors and enhances generalization. The model uses finite-element-based stress-response values as inputs and calculates in situ stress data derived from hydraulic fracturing interpretations as targets. Engineering applications show that the PSO-XGBoost model outperforms common methods, achieving superior prediction accuracy and generalization with fast convergence. This offers a high-precision inversion approach for small-sample conditions, supporting engineering design and safety assessment.
Tian et al. (2026) studied this question.