Generalized, Adsorbent-Agnostic, Artificial Neural Network Framework for Rapid Simulation, Optimization, and Adsorbent Screening of Adsorption Processes

Adsorption-based processes, such as pressure swing adsorption, because of their cyclic nature and the presence of multiple design variables, are challenging to simulate and optimize. Here, we introduce the machine-assisted adsorption process learning and emulation (MAPLE) framework, a generalized data-driven surrogate model that fully emulates operation of an adsorption process at the cyclic steady state. This framework is based on a dense feedforward neural network trained with a Bayesian regularization technique. This framework is generalized for a wide range of inputs which include the adsorbent properties, the Langmuir adsorption isotherm parameters, and operating conditions. The MAPLE framework can then be used to predict key performance indicators such as product purity, recovery, energy, and productivity for any arbitrary adsorbent and operating condition almost instantaneously. The key differentiator of the current work lies in the incorporation of the adsorption isotherm as an input, which is specifically beneficial for rapid screening of large adsorbent databases. A CO2 capture case study is used to illustrate this methodology. A relatively small computational training effort (2000 core hours) is required to train MAPLE with an average prediction accuracy (test R-Adj(2) >= 0.995) for all the four key performance indicators across a wide range of adsorbents/adsorption characteristics. The simulation and optimization time is brought down from 1500 core hours per adsorbent to <= 1 core min per adsorbent. The ability of this framework for rapid screening of adsorbents for postcombustion CO2 capture is also illustrated.