Can a computer “learn” nonlinear chromatography?: Physics-based deep neural networks for simulation and optimization of chromatographic processes

A parameter estimation method for chromatographic separation process based on physics-informed neural network

Chromatographic separation processes are most often modeled in the form of partial differential equations (PDEs) to describe the complex adsorption equilibria and kinetics. However, identifying parameters in such a model requires substantial computational effort. In this work, a novel parameter estimation approach using a Physics-informed Neural Network (PINN) model is developed and tested for a binary component system. Numerical accuracy of our PINN model is confirmed by validating its simulations against those of the finite element method (FEM). Furthermore, model parameters in the kinetic model are estimated by the PINN model with sufficient accuracy from the observed data at the column outlet, where parameter fitting error can be reduced by up to 35.0 % from the conventional method. In a comparison with the conventional numerical method, our approach can reduce the computational time by up to 95 %. The robustness of the PINN model has also been demonstrated by estimating model parameters from noisy artificial experimental data.

Physics-informed neural networks to solve lumped kinetic model for chromatography process

Numerical method is widely used for solving the mechanistic models of chromatography process, but it is time-consuming and hard to response in real-time. Physics-informed neural network (PINN) as an emerging technology combines the structure of neural network with physics laws, and is getting noticed for solving physics problems with a balanced accuracy and calculation speed. In this research, a proof-of-concept study was carried out to apply PINN to chromatography process simulation. The PINN model structure was designed for the lumped kinetic model (LKM) with all LKM parameters. The PINN structure, training data and model complexity were optimized, and an optimal mode was obtained by adopting an in-series structure with a nonuniform training data set focusing on the breakthrough transition region. A PINN for LKM (LKM-PINN) consisting of four neural networks, 12 layers and 606 neurons was then used for the simulation of breakthrough curves of chromatography processes. The LKM parameters were estimated with two breakthrough curves and used to infer the breakthrough curves at different residence times, loading concentrations and column sizes. The results were comparable to that obtained with numerical methods. With the same raw data and constraints, the average fitting error for LKM-PINN model was 0.075, which was 0.081 for numerical method. With the same initial guess, the LKM-PINN model took 160 s to complete the fitting, while the numerical method took 7 to 72 min, depending on the fitting settings. The fitting speed of LKM-PINN model was further improved to 30 s with random initial guess. Thus, the LKM-PINN model developed in this study is capable to be applied to real-time simulation for digital twin.

How Can (or Why Should) Process Engineering Aid the Screening and Discovery of Solid Sorbents for CO2 Capture?

ConspectusAdsorption using solid sorbents is emerging as a serious contender to amine-based liquid absorption for postcombustion CO2 capture. In the last 20+ years, significant efforts have been invested in developing adsorption processes for CO2 capture. In particular, significant efforts have been invested in developing new adsorbents for this application. These efforts have led to the generation of hundreds of thousands of (hypothetical and real) adsorbents, e.g., zeolites and metal-organic frameworks (MOFs). Identifying the right adsorbent for CO2 capture remains a challenging task. Most studies are focused on identifying adsorbents based on certain adsorption metrics. Recent studies have demonstrated that the performance of an adsorbent is intimately linked to the process in which it is deployed. Any meaningful screening should thus consider the complexity of the process. However, simulation and optimization of adsorption processes are computationally intensive, as they constitute the simultaneous propagation of heat and mass transfer fronts; the process is cyclic, and there are no straightforward design tools, thereby making large-scale process-informed screening of sorbents prohibitive.This Account discusses four papers that develop computational methods to incorporate process-based evaluation for both bottom-up (chemistry to engineering) screening problems and top-down (engineering to chemistry) inverse problems. We discuss the development of the machine-assisted adsorption process learning and emulation (MAPLE) framework, a surrogate model based on deep artificial neural networks (ANNs) that can predict process-level performance by considering both process and material inputs. The framework, which has been experimentally validated, allows for reliable, process-informed screening of large adsorbent databases. We then discuss how process engineering tools can be used beyond adsorbent screening, i.e., to estimate the practically achievable performance and cost limits of pressure vacuum swing adsorption (PVSA) processes should the ideal bespoke adsorbent be made. These studies show what conditions stand-alone PVSA processes are attractive and when they should not be considered. Finally, recent developments in physics-informed neural networks (PINNS) enable the rapid solution of complex partial differential equations, providing tools to potentially identify optimal cycle configurations. Ultimately, we provide areas where further developments are required and emphasize the need for strong collaborations between chemists and chemical engineers to move rapidly from discovery to field trials, as we do not have much time to fulfill commitments to net-zero targets.