Neural network descriptors for predicting adsorption and deactivation in energy storage tanks

To enable energy storage technologies based on adsorption in vehicle tanks, this study presents Artificial Intelligence (AI) models to predict energy deactivation curves, using data from 17 activated carbon samples. A multiscale approach is employed to address known bottlenecks in the application, particularly uncertainties in material characterization and the scarcity of experimental data. These limitations were overcome through the careful selection of experimental N 2 adsorption isotherms at 77 K and the use of molecular simulation to obtain pore size distributions (PSDfull and PSDrep) and single-component isotherms for the gases present in natural gas. The PSDs and C1–C4 isotherms were used to generate energy deactivation curves via a thermodynamic model of a storage tank. These deactivation curves served as output data for training the AI models based on artificial neural networks (ANN), optimized using Particle Swarm Optimization (PSO) method. Three models (A, B and C) were developed using PSDfull, PSDrep, and C1–C4 isotherms as input descriptors, respectively. All models showed high predictive performance, with R 2 values ranging from 0.97 to 0.93. To the best of our knowledge, this is the first study to use PSDs as descriptors in AI models. The fact that the use of PSDrep (model B), as the only descriptor, obtained a high R 2 value (0.97), is particularly significant because it considerably simplifies and reduces uncertainties in the development of databases involving activated carbons. Furthermore, thermodynamic tanks have complexity comparable to PSA units and their variants, indicating that the methodology can be readily transferred to the development of AI models for PSA design in industrial applications.