Rational selection of stope structure parameters is crucial for maintaining stope stability and enhancing mining economic efficiency. This study? addresses the limitations of conventional empirical analogy methods (prone to subjective evaluation biases) and standalone numerical simulation approaches (complex modeling procedures and high computational workloads) in stope parameter optimization. We propose an artificial intelligence-based stope parameter optimization method integrating Dung Beetle Optimizer (DBO), Multilayer Perceptron (MLP), and Support Vector Machine (SVM). The DBO algorithm enhances MLP and SVM performance, while the SVM-MLP hybrid framework mitigates the risk of MLP converging to local optima. Utilizing stope parameter datasets and stability evaluation grades from representative mines, the DBO-MLP-SVM model was trained. A case study of a limestone mine in southwestern China demonstrated that stope stability classification results from the DBO-MLP-SVM model indicated instability at 7m and stability above 8m. FLAC3D numerical simulations further validated these findings by analyzing stope displacement, pillar deformation, and surface subsidence, confirming the critical pillar dimension threshold of 8m. The consistency between AI model predictions and numerical simulation results underscores the method's reliability. Finally, through quantitative comparison with traditional methods, the stope production capacity achieved by intelligent collaborative optimization reached 167.44 t/d, representing a 51.80% increase in production capacity. With comparable loss and dilution rates, this approach effectively balanced the dual objectives of production efficiency and safety assurance, further validating the superiority of the model. This research provides a novel hybrid intelligent framework for data-driven stope parameter optimization in mining engineering.