Nonlinear learning-enhanced MLP-driven design of PIMs gas separation membranes beyond Robeson upper bounds

This study pioneers the integration of multilayer perceptron (MLP) with interactive design for predicting gas separation performance in Polymers of Intrinsic Microporosity (PIMs), leveraging their rigid, contorted architectures to transcend Robeson upper bounds. A curated dataset of 389 PIMs/PIM-polyimides spanning six gases (He, H₂, O₂, N₂, CO₂, CH₄) enabled development of an MLP model achieving unprecedented accuracy (mean R² = 0.969, RMSE = 0.156 Barrer), outperforming K-Nearest Neighbors regression (KNN), Gradient Boosting Decision Trees (GBDT), Random Forest (RF) and Gaussian Process Regression (GPR) models by ∼50 % error reduction. SHAP interpretability analysis decoded three fundamental design principles governing PIMs performance: backbone contortion, ladder connectivity, and hydrophobicity optimization, guiding creation of an interactive web platform that accelerates PIMs design through real-time permeability prediction. The platform demonstrates <2 % prediction variance across experimental benchmarks, reducing development cycles from years to weeks while maintaining solution processability. By combining MLP's nonlinear mapping of microporosity-permeability relationships with actionable design feedback. The open-access tool's forthcoming multi-model ensembles and quantum-ML integration establish a new paradigm for data-driven membrane innovation, bridging computational discovery with industrial-scale gas separation challenges.