(213d) Combining Machine Learning and CFD Simulation to Evaluate Design and Operating Conditions Faster

For the process industry, the digital tools for design and operation of equipment is ripe for supplementation with machine learning. Design engineers can benefit from the constant stream of innovations in the ML community when productizing new equipment, while operators of the equipment can be enhanced in productivity from on-demand information when these ML models are deployed for their easy access. There are several themes by which machine learning can cause a material impact: augmented user experiences, optimization, better utilization of physical measurement, and time savings via faster insights.

Utilization of a limited number of experiments and complimenting them with simulated data can result in significant time savings and valuable insights into engineering processes. Specifically, how simulation experts, a widely used engineering method, can realize large time savings when employing ML models in their sensitivity and design exploration, which can help reduce time to market. Further, how physical measurement of in-operation equipment can be combined with simulations with as input to (ML powered) executive digital twins (xDTs) to provide the operator with simulation-level detail of performance from limited discrete physical measurement points.

In this study, we perform multiple batches of computational fluid dynamics (CFD) simulations to characterize the performance of a 20 cubicmeter aerated stirred bioreactor with respect to changes in the geometric design as well as the operating conditions in terms of stirrer speed and aeration rate. The final goal is to optimize the reactor in terms of energy consumption and productivity. More significantly, we then explore a series of three independent scenarios whereby unique ML-based ROMs are created and analyzed. First, we conduct an optimization study in a hybrid modality: machine learning models are used when possible, during the optimization study in conjunction with traditional simulation for the design points. Second, we create and leverage ML-based ROMs for adjacent design characterization akin to typical robustness, reliability, and sensitivity analysis. Finally, we showcase the time and hardware savings when using AI-based (dynamic) sampling for creating simulation datasets from user-specified objectives when only a small number of simulations are budgeted to run.