Abstract: To address the challenge of water contamination in space telescopes, four common zeolite molecular sieve materials: ZSM-5, ZSM-22, MCM-41, and SAPO-11 were investigated in this study. Using a combined simulation facility that provides atomic oxygen exposure and ultraviolet radiation, the adsorption capacities of these zeolites for water molecules under various environmental conditions were evaluated. Furthermore, a predictive model for the adsorption capacity of contaminants was developed using radial basis function (RBF) neural networks, combining experimental data with machine learning techniques. The results demonstrate the model’s effectiveness in predicting the adsorption performance of molecular sieves, achieving R² values greater than 0.99 and mean absolute error and root mean square error values below 10^-5. This performance surpasses that of other models, including Long Short-Term Memory (LSTM) neural networks, Convolutional Neural Networks (CNN), and neural networks trained with backpropagation (BP) algorithm.The establishment of this model overcomes the time-consuming and labor-intensive challenges of studying molecular sieve adsorption performance through experimental methods alone and lays a foundation for the development of more complex data prediction models.