(615c) Hybrid-AI Based Modelling of Pressure Swing Adsorption

Modelling and optimization of pressure swing adsorption (PSA) cycles are inherently complex owing to their cyclic nature and flexible process design. The existing mathematical models based on mass, momentum and energy balances resulting in a system of nonlinear coupled partial differential equations (PDEs) are computationally expensive to solve in order to simulate PSA cycles until cyclic steady state. The burden of cyclic steady state makes simulation and optimization of processes even more complex. With the growth in number of adsorbents for a particular separation, there is an increasing need to develop faster adsorption process simulators. It is also critical that the faster models are amenable to development of various processes.

This study focuses on developing a faster modelling approach that incorporates both machine learning and underlying physical laws to accurately simulate the PSA cycle. Neural networks are utilized to fully predict the spatiotemporal profiles of different PSA constituent steps. To this end, neural network models were trained by using few spatiotemporal data points generated from the high-fidelity PSA model while enforcing the constraints of physical laws in the form of system of nonlinear PDEs [1]. As a result, the governing PDEs ensure model regularization and avoid over-fitting while increasing the generalization capabilities. Unique models were built for each type of constituent steps that are typically encountered in PSA cycles. Using the trained neural network for each constituent step, the PSA cycle is constructed and simulated until the CSS where predicted column profiles are directly used to calculate different PSA process performance indicators. The model’s performance is evaluated by comparing the CCS profiles with that of the high-fidelity PSA model at different operating conditions. At the meeting, the accuracy and reliability of this approach will be highlighted by considering the case of post-combustion CO₂ capture as an example.

References

[1] Raissi, M.; Perdikaris, P.; Karniadakis G.E. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. J. Comput. Phys. 2019, 378, 686-707.