The design of the stress-applying part (SAP) in polarization-maintaining fibers (PMF) is a typical multi-variable, multi-objective optimization problem. The coupled effects of SAP diameter, B₂O₃ molar fraction and SAP edge-to-core distance determine attenuation, polarization-maintaining capability, mechanical reliability, and radiation resistance. Conventional trial-and-error design is therefore inefficient, especially when multiple performance constraints must be satisfied simultaneously.

Based on 500 experimental datasets, predictive models were established for eight performance indicators, including attenuation, post-irradiation loss, tensile strength, nd value, extinction ratio, temperature-induced extinction-ratio variation, beat length, and grinding crack probability. Three ensemble learning algorithms, i.e., random forest (RF), XGBoost, and gradient boosting regressor (GBR), were systematically compared via 5-fold cross-validation. SHAP analysis was employed to reveal the dominant effects of design variables. Furthermore, a large language model (LLM) was introduced as a natural-language interface and explanation layer to parse user requirements, diagnose infeasible constraints and provide minimum-relaxation suggestions.

The optimal models achieved R² values ranging from 0.76 to 0.99 with an average of 0.906, indicating that the nonlinear mapping between SAP parameters and performance metrics can be reliably captured. SHAP analysis revealed that the edge distance dominated attenuation-related indices, the SAP diameter governed mechanical and polarization-related properties, and the B₂O₃ molar fraction exerted the strongest effect on grinding crack probability. For 30 natural-language test cases, the LLM achieved 100% accuracy in indicator recognition and inequality-direction determination, and 85.0% exact-value accuracy. Across seven representative design scenarios, the proposed framework reduced the expected number of trial experiments by 91% on average and by up to 99.8% in the four-constraint scenario. When no feasible solution existed, the system could identify the bottleneck constraint and quantify the minimum relaxation required to recover feasibility.

A collaborative ML-LLM framework was developed for the inverse design of PMF SAPs. The method integrates quantitative prediction, mechanism-oriented interpretation and natural-language-driven interaction, thereby converting conventional trial-and-error design into an efficient data-driven screening process. This approach provides a practical pathway for intelligent design of polarization-maintaining fibers and may also be extended to other inorganic non-metallic materials with strongly coupled process-property relationships.