Laser thermotherapy possesses precise and high-efficiency advantages for malignant tumor ablation. Accurate prediction of tissue thermal damage ranges and optimal selection of therapeutic parameters are core scientific challenges restricting the clinical popularization of thermal therapy. Taking rat subcutaneous glioma xenografts as the research subject, this study numerically solves the Pennes bioheat transfer equation via the finite difference method and quantitatively assesses tissue thermal damage based on the Arrhenius thermal damage integral model. The influences of thermophysical and laser parameters on the formation mechanism and spatial distribution characteristics of tumor thermal damage are systematically investigated. On this basis, seven representative machine learning algorithms are comprehensively compared in terms of prediction accuracy and applicability, and an ensemble learning (EL) model with controllable prediction errors is further established. The results demonstrate that the proposed EL model outperforms all single baseline algorithms in generalization ability and robustness, with a mean absolute error (MAE) of 0.000504 and a root mean square error (RMSE) of 0.000583, exhibiting substantially superior prediction accuracy. Compared with two mainstream ensemble frameworks, extreme gradient boosting (XGBoost) and light gradient boosting machine (LightGBM), the developed EL model achieves an outstanding fitting performance with a coefficient of determination (R2) up to 0.9998. Furthermore, an EL-based parameter inversion framework integrated with the sled dog optimization (SDO) algorithm is constructed for thermal damage optimization. With the preset tumor thermal damage depth defined as the constraint condition, comparative experiments are conducted against genetic algorithm (GA), Fata Morgana algorithm (FATA), gray wolf optimizer (GWO), differential evolution (DE), and particle swarm optimization (PSO). The validation results reveal that the hybrid EL-SDO collaborative optimization model can stably obtain the optimal key parameters for laser thermotherapy (including laser power and irradiation time), with the maximum parameter inversion error limited within 0.1 mm for arbitrary target thermal damage depths. This work provides a robust theoretical foundation and technical reference for individualized and intelligent treatment planning of clinical laser thermotherapy.