(542a) Surrogate Modeling of 3D Flow Dynamics and Chemical Reactions Using Physics-Constrained Deep Learning

As product purity requirements and environmental regulations become more stringent, the need of high-fidelity chemical process models grows. A simple model assuming perfect mixing is widely employed for a reactor such as CSTR, however, inhomogeneities due to imperfect mixing may occur and greatly affect the rate and dynamics of systems [1-3]. The inhomogeneous state can greatly affect product quality, such as causing cell death in bioreactors [4] or creating hot spots in polymerization reactors [5]. Therefore, for the correct interpretation of chemical process, a high-fidelity model that can accurately simulate the flow inside the reactor should be used.

Computational Fluid Dynamics (CFD) techniques have been extensively used to create high-fidelity models of reactors, owing to its capability to simulate three-dimensional flow and reflect kinetics changes accordingly [4, 5]. However, employing CFD techniques necessitates a significant amount of computational power in order to solve systems of partial differential equations (PDEs) such as mass and momentum balance equations. Additionally, the majority of classic PDE solvers rely on mesh-based techniques such as the Finite Element Method (FEM) and the Finite Volume Method (FVM), which require the entire mesh be rebuilt whenever the geometry of the target system changes. These issues significantly restrict the use of the CFD model for real-time or structure optimization.

To address the aforementioned issues, cost-effective surrogate models have been investigated. A recently proposed model of this type is the physics-informed neural network (PINN) [6]. PINN is a neural network that is trained in such a way that the residuals of governing equations are minimized. The PINN has advantages, including the ability to learn PDE solutions without using a mesh and to predict flow information in real time following training. Lu et al. (2019) demonstrate with the efficiency of PINN in learning benchmark PDEs and solving inverse problems [7]. Choi et al. (2020) show that PINN is capable of learning CFD models with moving reference frames as well as chemical reactions [8].

However, Lu et al. (2019) and Choi et al. (2020) incorporate the boundary condition (BC) and the initial condition (IC) into the loss function via a soft constraint. Because NN with soft constraints cannot be forced to produce an output that satisfies BC and IC, the output may not be physically realistic. As a result, to ensure the physical feasibility of the output, BCs and ICs should be applied to the neural network in a ‘hard’ manner. Sun et al. (2020) demonstrated the application of hard constraints to simple flow and learnt flow according to parametric geometry [9], while Berg and Nyström (2018) demonstrated the use of hard constraints to more complex geometry [10]. However, both studies addressed mainly difficulties involving two-dimensional geometry.

The purpose of this study is to develop a framework for surrogate modeling of a three-dimensional, non-isothermal reactor with non-linear chemical reactions. BCs such as non-slip conditions on reactor wall and ICs were applied in the form of hard constraints, with another neural network learning the distance to the boundary. A CSTR with a van de Vusse reaction was chosen as the target system. The trained hard-constrained NN prediction results were compared to the CFD model for validation. The surrogate model successfully learnt the flow information of the high-fidelity CFD model and decreased the prediction speed by several orders of magnitude when compared to the CFD model. This method can be employed for the real-time optimization of chemical reactors.

References

[1] Villa, C.M., Dihora, J.O. and Harmon Ray, W. (1998), Effects of imperfect mixing on low-density polyethylene reactor dynamics. AIChE J., 44: 1646-1656. https://doi.org/10.1002/aic.690440716

[2] Qi, N., Zhang, H., Zhang, K., Xu, G., & Yang, Y. (2013). CFD simulation of particle suspension in a stirred tank. Particuology, 11(3), 317–326. https://doi.org/10.1016/j.partic.2012.03.003

[3] Strizhak, P., Ali, F., & Menzinger, M. (1998). Inhomogeneities of CSTR on a Macroscale Due to Spatial Dependence of Micromixing Time: The BZ Reaction. https://pubs.acs.org/sharingguidelines

[4] Mishra, S., Kumar, V., Sarkar, J., & Rathore, A. S. (2021). CFD based mass transfer modeling of a single use bioreactor for production of monoclonal antibody biotherapeutics. Chemical Engineering Journal, 412. https://doi.org/10.1016/j.cej.2021.128592

[5] Fathi Roudsari, S., Ein-Mozaffari, F., & Dhib, R. (2013). Use of CFD in modeling MMA solution polymerization in a CSTR. Chemical Engineering Journal, 219, 429–442. https://doi.org/10.1016/j.cej.2012.12.049

[6] Raissi, M., Perdikaris, P., & Karniadakis, G. E. (2019). Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 378, 686–707. https://doi.org/10.1016/j.jcp.2018.10.045

[7] Lu, L., Meng, X., Mao, Z., & Karniadakis, G. E. (2019). DeepXDE: A deep learning library for solving differential equations. https://doi.org/10.1137/19M1274067

[8] Choi, S., Jung, I., Kim, H., Na, J., & Lee, J. M. (2022). Physics-informed deep learning for data-driven solutions of computational fluid dynamics. Korean Journal of Chemical Engineering, 39(3), 515–528. https://doi.org/10.1007/s11814-021-0979-x

[9] Sun, L., Gao, H., Pan, S., & Wang, J. X. (2020). Surrogate modeling for fluid flows based on physics-constrained deep learning without simulation data. Computer Methods in Applied Mechanics and Engineering, 361. https://doi.org/10.1016/j.cma.2019.112732

[10] Berg, J., & Nyström, K. (2018). A unified deep artificial neural network approach to partial differential equations in complex geometries. Neurocomputing, 317, 28–41. https://doi.org/10.1016/j.neucom.2018.06.056