Abstract: With the development of deep resources and the extension of underground engineering to the deep Earth, the accurate prediction of rock mechanical properties under high temperature has become the key to ensuring engineering safety. The strength, modulus and failure mode of rocks under high temperature exhibit highly nonlinear evolution characteristics, which are difficult to be fully described by traditional theoretical models. Existing data-driven methods are mostly modeled based on statistical correlation, with the common limitation of “emphasizing correlation over causality”, resulting in weak interpretability and insufficient generalization ability, which can hardly meet the practical requirements of prediction accuracy for deep engineering. On this basis, an intelligent analysis framework integrating causal discovery, multi-task sparse learning and counterfactual explanation is proposed. Firstly, through physics-informed causal graph learning combined with prior knowledge of rock thermodynamics, key features causally related to strength are screened from data to eliminate interference from irrelevant variables. On this basis, a multi-task sparse learning network is constructed to realize the collaborative prediction of rock strength, elastic modulus and failure mode. The parameter sharing mechanism is used to mine the inherent coupling relationship among mechanical parameters, and sparse constraints are adopted to focus on core causal features. Finally, a counterfactual reasoning module is introduced to quantify the causal effects of variables and their action paths, enabling the model to have transparent decision logic. The results show that the coefficient of determination (R²) of strength prediction on the test set reaches 0.935 with the mean absolute error (MAE) of only 6.2 MPa, which is significantly better than traditional machine learning and causal inference models (R² = 0.721-0.893, MAE = 8.7-18.6 MPa). In out-of-distribution intervention experiments, the model shows strong robustness, and the causal paths revealed by counterfactual reasoning are highly consistent with rock thermodynamic mechanisms. This study results provide an analytical tool with high accuracy, strong generalization and mechanism transparency for deep high-temperature rock engineering.