Robust in-situ stress inversion in an underground powerhouse using tensor synthesis and surrogate-assisted differential evolution

Accurate characterization of the initial in-situ stress field is essential for stability assessment, support design, and surrounding rock control in deep underground engineering, yet field measurements are often highly scattered and three-dimensional inversion is computationally expensive. This study develops a robust and efficient inversion framework for a deeply buried underground powerhouse in the southeastern Tibetan Plateau. First, a three-dimensional borehole stress synthesis method is established by combining particle swarm optimization, the Huber loss, Levenberg-Marquardt iteration, and regularization to denoise multi-source measurements, suppress local outliers, and alleviate ill-conditioning in stress-tensor reconstruction. Second, a surrogate-assisted differential evolution workflow is constructed using a radial basis function network within a prediction-verification-correction active-learning loop to reduce the cost of repeated forward simulations while preserving global optimization capability. Application to the powerhouse shows that the mean relative error decreases from 14.12 to 9.34%, and the deviation variance decreases from 2.98 to 1.33 after optimization. The proposed framework improves both the reliability of inversion input data and the efficiency of field-scale stress reconstruction, providing practical support for stability evaluation and surrounding rock control in deep, geologically complex underground caverns.