Interpretable Multimodal Machine Learning for Sustainable Electret Materials: Deciphering Process-Structure-Property Relationships in PLA Melt-Blown Nonwovens via Integrated CNN and CatBoost

The sustainable development of high-performance polylactic acid (PLA) electret melt-blown nonwovens, a promising biodegradable alternative to conventional polypropylene filters, has been hindered by the inefficient trial-and-error optimization of multivariate process parameters. To address this challenge, we propose an interpretable multimodal machine learning (ML) framework that integrates a multi-input convolutional neural network (CNN) with the CatBoost algorithm. This innovative pipeline synergistically combines process parameters and microstructural features extracted from scanning electron microscopy (SEM) images to predict key electret performance metrics, including surface potential and charge density, with exceptional accuracy (R2 > 0.90). SHapley Additive exPlanations (SHAP) analysis further quantifies complex nonlinear and synergistic interactions among 13 input features, surpassing traditional black-box models by providing physically grounded insights. Our analysis identified an optimized processing window, a medium–high die temperature (210–225 °C), and hydraulic pressure (3–4 MPa), which balances charge implantation with thermal degradation, thereby enhancing charge stability. This interpretable ML framework reduces overall development expenditures by more than 80% (calculated by comparing the 480 experimental runs required for conventional single-factor scanning with the 83 runs used in our reduced factorial design). Beyond lowering the material, labor, equipment, and computational costs, the workflow supports a transition from trial-and-error optimization to a more knowledge-driven and sustainable strategy for material development. Additional long-term efficiencies arise from the SHAP-guided elimination of unproductive regions in the design space.