Deep learning model of fixed bed adsorption breakthrough curve hybrid-driven by data and physical information

A deep learning model of fixed bed adsorption breakthrough curve hybrid-driven by the data and physical information (PINN_MOD) was proposed in this work. A combined method of external data constraint enhanced by penalty factors and residual-based adaptive refinement strategy was adopted to gradually optimize the neural network parameters to approximate the solution of the partial differential equation (PDE) of the dynamic binary gas adsorption process of fixed bed by minimizing the loss function. The physics-informed neural network (PINN) model was generally used to solve the forward and inverse problem of the one-dimensional single-component convection-diffusion and fixed-bed adsorption PDE models with high fidelity. However, there are convergence difficulties when it is used to solve the one-dimensional binary fixed-bed adsorption PDE model on long-time scale. In this paper, the traditional finite differential method (FDM) was first used to solve the PDE problem of one-dimensional binary fixed bed adsorption, and then the component concentration data in the spatiotemporal region obtained by FDM simulations were adopted as an external constraint of the PINN model to solve the PDE of one-dimensional binary fixed bed adsorption. Taking the CO2/N2 mixture (molar ratio 30∶ 70) adsorption models in fixed bed packed with different MOFs (CALF-20 and UTSA-16) as an example, the PINN_MOD model was used to calculate the outlet CO2 breakthrough curve of fixed bed. The FDM calculation results can be well replicated, which proves that the model can effectively obtain high-fidelity PDE solutions only relying on a small amount of external data. It is confirmed that the PINN_MOD model could obtain the high fidelity solutions by relying on only a few external data. The proposed model is expected to play an important role in the development of novel metal-organic framework adsorbents for gas separation applications. © 2023 Chemical Industry Press. All rights reserved.