Prediction of physisorption on zeolites using Graph Integrated Adsorption Network and spline data augmentation

Abstract

This paper discusses a unified application framework termed as the Graph Integrated Adsorption Network (GIANet) for predicting gas adsorption in zeolites. The proposed method utilizes limited experimental data for predicting the adsorption performance of different zeolites. A spline‐based interpolation has been adopted to densify sparse isotherm measurements. The approach reduces training epochs by over 90 % and lowers the mean squared error by roughly 96 %. With only 1,001 parameters, GIANet attains an MAE of 0.068 mmol g−1 and an R2 of 0.997 on a 20 % random hold-out set—comparable to a material-specific Toth isotherm model and similar to much larger MatErials Graph Network (MEGNet) variants. For interpolation tests on CO<inf>2</inf> and N<inf>2</inf> adsorption, mean errors remain below 0.06 mmol g−1 (RMSE ≈ 0.04, R2 > 0.99). Extrapolation experiments—unseen temperatures, cross-framework transfers, and fully held-out isotherms—yield mean errors up to 0.11 mmol g−1 and R2 ≥ 0.89. These results suggest that GIANet's lightweight, condition-aware design offers a practical balance of simplicity and predictive accuracy for adsorption screening across diverse zeolites and operating conditions. © 2025 Elsevier B.V., All rights reserved.