Adsorption equilibrium of ammonia and water on porous adsorbents at low pressure: Machine learning-based models

Ammonia is a harmful gas in semiconductor cleanrooms. Activated carbons are widely used to remove ammonia. However, the existence of ambient moisture significantly impacts ammonia adsorption on adsorbents. Therefore, it is necessary to understand their adsorption performance on activated carbons to improve the product yield in semiconductor cleanrooms. Although many ammonia and water adsorption performance studies have been conducted, there has been little experimental research on low-pressure ammonia and water adsorption performance. In this study, several traditional isotherm models were applied to fit experimental isotherm data. The double-site (DS) Langmuir isotherm model provided good performance for ammonia adsorption on M-1, M-3, and M-4. The Toth isotherm model provided the best fit for ammonia adsorption on M-2. The Dubinin-Serpinsky (DS) isotherm model provided the best fit for water adsorption on the four investigated materials even at low pressure. In regard to predicting adsorption capacity, the previous prediction approach may limit the accuracy of the models by using a single set of characteristics to predict the adsorption capacity at different pressures. To solve this, an attention-based recurrent neural network (RNN) is proposed in this study. The attention mechanism is able to assign different weights to each material characteristic for different partial pressures. Due to this strength, the attention-based RNN realizes the prediction of adsorption isotherms by only taking material characteristics as inputs. The mean absolute percentage error (MAPE) of the attention-based RNN model in the prediction of adsorption capacity was 4.09%-18.68% and 3.68%-20.58% for ammonia and water, respectively, indicating that the well-trained model provides a reasonable prediction. The results predicted by the attention-based RNN model are consistent with the experimental data in terms of adsorption isotherm type, further confirming the reliability of the RNN model.