(615f) Experimental Validation of an Adsorbent Agnostic Machine-Assisted Adsorption Process Learning and Emulation (MAPLE) Framework

The machine-assisted adsorption process learning and emulation (MAPLE) framework, a generalized data-driven surrogate model based on a deep neural network, can fully emulate an adsorption process cyclic steady state. The training data was generated using a detailed 1D finite volume model with inputs sampled using a Latin Hypercube (LHC) approach. The model has multiple inputs such as cycle configurations, step time, pressure, the Langmuir parameters, particle density, and can predict outputs such as extract or light product purity and recovery. As such, the model is trained to the adsorbent agnostic, so that it can predict the performance of the cycle for a set of arbitrary adsorbent and process parameters [1]. The main advantage of this approach is that once trained the model is an excellent tool that can be used for adsorbent screening. In previous studies, we have validated the framework using numerical simulations [1,2] . In this work we provide the experimental validation.

In this work, the MAPLE model is trained for a Skarstorm cycle that is suitable for raffinate purification. In order to test the efficacy of the model, a case study of air separation to produce high purity O₂ is considered. Two commercial adsorbents, Li-LSX and Zeolite 13X were chosen validation. A series of characterization experiments allowed the description of the isotherms of N₂ and O₂ on these adsorbents. Their Langmuir isotherm parameters were taken as the input to the MAPLE model and multi-objective optimizations to maximize purity and recovery of the process for each of these sorbents were performed by coupling MAPLE with an optimizer. Several points on the Pareto curves were chosen and the corresponding operating parameters were translated into experiments on a two-column lab-scale PSA rig. The results show that the MAPLE optimization framework can correctly predict performance. This study paves the path for the reliable use of the MAPLE framework for process optimization and adsorbent screening.

[1] Pai, K.N., Prasad, V. and Rajendran, A., 2020. Generalized, Adsorbent-Agnostic, Artificial Neural Network Framework for Rapid Simulation, Optimization, and Adsorbent Screening of Adsorption Processes. Ind Engg. Chem. Res., 59(38), pp.16730-16740.

[2] Pai, K.N., Prasad, V. and Rajendran, A., 2021. Practically Achievable Process Performance Limits for Pressure Vacuum Swing Adsorption-Based Post- combustion CO₂ Capture. ACS Sus. Chem. Engg., 9, pp.3838-3849.