(375aa) Simultaneous Optimization Via Physics-Informed Neural Network Solvers with Compartment Models for Reactor Design

Multiphysics simulation lies in its ability to capture the complex interactions between different physical domains such as fluid dynamics, species transport, and chemical reactions. Computational fluid dynamics (CFD) was utilized to simulate these phenomena, providing remarkably precise solutions. However, its computational burden for solving partial differential equations simultaneously makes it impractical for chemical reactor design and optimization. Thus, physics-informed machine learning (PIML) has recently emerged as a surrogate model.¹ PIML modifies its loss function to embed physical laws within its architecture, thereby enabling accurate solutions.^2,3 Moreover, PIML provides enhanced generalization and inference speed due to its mesh-free approach. Nonetheless, PIML still has a main challenge in resolving its non-robustness on complex multiphysics problems. Consequently, we aim to propose a hybrid approach for multiphysics simulation based on PIML and compartment model to ensure both accuracy and efficiency.

In this study, we developed a real-time optimization framework to explore optimal geometry configurations and operating conditions of polymer and crystallization reactor systems. Firstly, PIML is trained in an unsupervised way to predict flow fields, implementing the Navier-Stokes equations, initial conditions, and boundary conditions into the network. Additionally, it incorporates geometric and operational parameters to provide immediate prediction under various settings via single network training.⁴ Then, flow fields are divided into axial and radial compartments to couple fluid dynamics with chemical reaction kinetics, assuming perfect mixing throughout every compartment.⁵ Due to its data-free approach, this multiphysics simulation model offers unprecedented versatility and a significant reduction in computational cost. Moreover, connection to optimization algorithms allows it to approach the optimal point of geometry configurations and operating conditions far more swiftly. Our new methodology represents a remarkable advancement in multiphysics simulations and demonstrates the efficacy of PIML in modeling and optimization for chemical reactor systems. In conclusion, this research opens up a promising avenue for future applications in reactor engineering, highlighting the potential of our approach to digital twin technologies.

1. 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.

2. Karniadakis, G. E., Kevrekidis, I. G., Lu, L., Perdikaris, P., Wang, S., & Yang, L. (2021). Physics-informed machine learning, Nature Reviews Physics, 3, 422-440.

3. Raissi, M., Yazdani, A., & Karniadakis, G. E. (2020). Hidden fluid mechanics: Learning velocity and pressure fields from flow visualizations, Science, 367(6481), 1026-1030.

4. Hennigh, O., Narasimhan, S., Nabian, M. A., Subramaniam, A., Tangsali, K., Fang, Z., & Choudhry, S. (2021). NVIDIA SimNetâ„¢: An AI-accelerated multi-physics simulation framework, International Conference on Computational Science, 447-461.

5. Tajsoleiman, T., Spann, R., Bach, C., Gernaey, K. V., Huusom, J. K., & Krühne, U. (2019). A CFD based automatic method for compartment model development. Computers & Chemical Engineering, 123, 236-245.