(100c) High Fidelity Real-Time 3D Mixing Simulation

Mixing is an important process in most chemicals, process, consumer products, bio-pharma, and food companies. Upfront simulations of mixing provide information on mitigating failure modes, time and power constraints, improving process reliability, and reducing time-to-market. Current state-of-the-art approach is to perform 3D Computational Fluid Dynamics (CFD) study of mixing by experts. CFD Models generate large sets of 3D high-fidelity data, usually at the expense of long computation times. CFD also requires a certain degree of expertise regarding the process, the physics, and the CFD tool used. Only an experienced analyst may possess the knowledge required of all three (process, physics, and CFD tool). A plant engineer may have the knowledge in the first two (process and physics), while an operator may only have the first (process). This creates challenges for companies to use the benefits of CFD analysis throughout their processes rather than just ideation and troubleshooting stages.

One approach to democratizing simulation is creation of custom workflows or templates. Mixing simulations are highly repeatable making them ideal for simulation automation and the deployment of custom workflows or mixing-simulation templates. These templates are still running 3D CFD simulations in the background. These custom templates, however, do remove the barrier to CFD deployment since they streamline the typical CFD steps; pre-processing, set up, and post-processing. Use of these templates makes it easier for anyone to perform the analysis and generate reports of pertinent CFD results. Though the process is streamlined and automated, it is still not be conducive for providing real-time/close-to-real-time guidance to operators for decision making.

This talk will cover an approach to compress the power of 3D high-fidelity physics analysis into lightweight, easy to use models: Reduced Order Models (ROMs). This approach is currently gaining traction in the context of digitization journey that most companies are adopting. ROMs provides 1. faster response times (as fast as real time), and 2. democratization of modeling and simulation by significantly reducing the computational time and expertise requirements to perform CFD analysis in order to extract meaningful insights. ROM is a simplification of a high-fidelity computational model that preserves essential behavior and dominant effects for the purpose of reducing solution time and storage capacity required for the more complex model.

Here, we present the creation and the use (consumption) of a ROM of CFD simulations of a mixing tank. There are multiple methods for creating ROMs depending on the complexity of the underlying physics and the process. The method presented here is different than approaches like response surfaces, where you get output signal corresponding to input signal based on created correlations. We will be creating full 3D nonlinear ROMs which will be predicting results on every location of the domain. This will not only allow for getting those output signals but also provide capabilities to visualize the results on the 3D domain.

In this paper a series of simulations were performed for 3 different parameters: Impeller speed (100 RPM -500 RPM), Liquid Density (750 Kg/m³-1600 Kg/m³) and liquid viscosity (1cP -10cP) to train the solver. A DOE was created using optimal space-filling algorithms and a total of 24 different cases were simulated. Once the solver is trained, reduced order models for flow characteristics within the mixing tank and shear stress on tank, baffle, and impellers were created. Each CFD simulation was run with Multiple Reference Frame (MRF) approach using Ansys Fluent. The mixing tank considered for the analysis has 2 types of 4-blade impellers, a sparger and 4 baffles. For performing CFD analysis the domain was divided into 1.2 million grid points. Each run took 30 mins on 4 CPUs. After this initial computational investment, the physics-trained solver then created ROMs which can be executed within seconds on a single CPU. The results from the ROM were verified against simulation data and the results match well (Figure 1). Results in Figure 1A shows wall shear stress on impeller/shaft. Whereas 1B shows same results obtained from the ROM. It is seen that results are matching well with similar hot spots locations. Figure 2 compares velocity at a center plane and supports the observations made from Figure 1. If needed, the error can be further reduced by increasing training data and/or decreasing the required tolerance.

This three-dimensional non-linear ROM can be used for real-time monitoring and to perform ‘what-if’ analysis within seconds without compromising on accuracy. This modelling of mixing, from 3D high-fidelity models to ROMs, paves the way to digitize the asset and perform real-time monitoring as well as predictive maintenance. This ROM can be represented at the system level as a Functional Mockup Unit (FMU), thereby allowing upstream and downstream components (could be other FMUs built similarly) interact. At the system level these mixing tank FMUs can be connected to controllers as well.

Most mixing processes in the industry involve the addition of various ingredients into the mixer at different times and different temperatures to create product. Precise temperature control is important since the final product quality depends on set temperatures at intermediate steps. Cooling fluid flow within the mixing vessel walls, jacketed piping within the mixer itself, and external parallel circuits with heat exchangers are common methods to control temperature. The temperature control mechanisms are usually achieved using some PLCs. These controllers currently use first principle models or 0-D models or empirical formulas to drive the system. These models are based on some sensor readings in the system. A physics based digital twin can enhance the performance of the controller logic by providing multiple virtual sensors as well as information about mixing continuously. This will help provide suggestions to operators or invoke automation to achieve a consistent yield. Current work will show complete workflow of taking a 3D detailed physics simulation of a mixing tank and bring it all the way to system level model to drive a controller giving plant managers/operators a way to virtually design/optimize the operations. The complete system can not only be modelled to provide information like remaining life and time-to-maintenance. But, can then be connected with sensor data on any IOT platform to make an operational digital twin or smart manufacturing plant.