Machine learning based prediction and process optimization of SiC porous ceramic membranes sintered with NaA zeolite residue

Porous silicon carbide (SiC) ceramic membranes require extremely high sintering temperatures, limiting their large‑scale adoption in cost-effective water and wastewater treatment. Incorporating NaA zeolite residue as a sintering aid markedly lowers the sintering temperature and promotes industrial waste valorization, but the coupled effects of multiple compositional and processing parameters remain poorly understood. In this study, we developed a skip-connection multi‑path multilayer perceptron (skip@M‑MLP) framework, together with several conventional machine learning models, to predict porosity and bending strength using a curated dataset from the literature. A hybrid data augmentation strategy combining Gaussian noise and linear interpolation was applied to address small‑sample limitations, and model interpretability was achieved via Shapley Additive Explanations (SHAP) and partial dependence plots (PDPs). The skip@M‑MLP achieved the highest accuracy among all tested models (R² up to 0.8716) and revealed that activated carbon content is the dominant factor governing the porosity–strength trade‑off, followed by sintering temperature and SiC fraction. These insights link data‑driven predictions with membrane engineering, providing a theoretical basis and quantitative guidance toward the potential scalable for the low-carbon fabrication of robust SiC membranes explicitly tailored for advanced water treatment processes. • ML optimizes NaA-assisted low-temp sintering of SiC water filtration membranes. • Hybrid data augmentation overcomes small-sample limits in membrane engineering. • Skip-connection MLP accurately predicts the porosity-strength of membranes. • Interpretable AI guides cost-effective design of water treatment membranes.