AI-accelerated CFD (Computational Fluid Dynamics): how does byteLAKE’s CFD Suite work?

Published in

Becoming Human: Artificial Intelligence Magazine

7 min readApr 19, 2021

byteLAKE’s CFD Suite (AI-accelerated CFD)

CFD Suite is a collection of AI models (Artificial Intelligence) to significantly accelerate the execution of CFD simulations (Computational Fluid Dynamics).

Computational Fluid Dynamics acceleration with Artificial Intelligence

CFD Suite is a data-driven solution designed to enable straightforward and simple integration with leading CFD CAE (Computer-Aided Design) software tools and workflows. Thanks to a close collaboration between byteLAKE and CFD software providers, CFD Suite can act as an add-on that allows engineers to ultimately cut the time to results of complex and expensive simulations. CFD Suite is also compatible with leading open-source CFD software. Hardware-wise, it is a cross-platform solution and supports both CPUs and GPUs with FPGAs being on the roadmap. Furthermore, CFD Suite addresses industries’ requirements of supporting CPU-only architectures and therefore has been optimized in terms of performance and scalability for all Intel technologies.

It has been experimentally proven that the CFD Suite can reduce the time to results at least by a factor of 10x and keep the accuracy of predictions above 93%.

Key features include:

– CFD solvers accelerated with AI / Artificial Intelligence
(time to results significantly shorter & better speedup than with hardware accelerators)

– Augmentation of numerical solvers with complementary AI models
(faster analysis, reduced cost of trial & error experiments, 100% compatible with existing tools)

– Solver specific AI models for Digital Twin industrial scenarios
(generalized AI models for different use cases, geometries, and parameters’ values)

– Designed to scale up fast
(solver specific AI model within 2–4wks vs. 1–2yrs efforts for hardware adaptations)

CFD Suite = at least 10x faster CFD simulations at 93% (and higher) accuract

From simulations to predictions

Typically, it takes 3 simple steps to get started with CFD Suite:

CFD Solver selection: pick the one you’d like to accelerate

a) Typical candidate: simulations that take too long

Setting the targets for AI Model(s)

a) What level of acceleration do you expect? Reduce time/iterations? Anything else?

b) What is the accepted accuracy? Within 85–95%? Higher? Lower?

c) Should AI generate only the result of the simulation or intermediate steps as well?

d) Consultancy to assess the solver’s input/output (types, ranges, etc.)

AI Model training

a) Based on the above, byteLAKE will train the CFD Suite to accelerate your CFD Solver(s)

b) CFD Suite will be 100% compatible with your workflow
= no need for any changes in infrastructure or data formats/types

New AI models are constantly added to the collection which gradually increases the number of CFD simulations that can be handled by the CFD Suite off-the-shelf.

The roadmap below shows the direction of the CFD Suite development.

Training

CFD Suite deployment starts with the AI model(s) training. Therefore, past CFD simulations are required (usually between 8–50 such simulations) to produce accurate predictions and ensure AI models proper generalization level. Generalization enables predictions across a wide range of various input parameter values incl. geometries.

Inferencing (predictions)

Once the AI models are trained, they can be run as part of the CFD CAE software in a form of an add-on. As the current version of the CFD Suite (March 2021) is data-driven, the prediction is organized alongside these steps:

1. First, the initial part (i.e., 3–10%) of the simulation is done with a CFD solver

2. Then the CFD Suite takes over the results and its AI model(s) predict the results, accelerating time to results

3. Note that CFD Suite can work with 3D data, producing 100% compatible results with the client’s existing CFD software tools. Therefore, results predicted by CFD Suite can be taken for further processing and analysis right away and no further formatting is necessary.

CFD Suite AI models are produced per solver and can accurately predict results for a family of simulations.

Once the AI models have been trained, the users can freely modify input parameters and still get accurate and fast results. Breaking down the inferencing (prediction) part into steps leads to the following structure of the prediction process:

AI models can work in 3 modes

· ACCURATE — focused on the accuracy of the predictions,

· MEDIUM — offering a trade-off between accuracy and acceleration

· FAST — maximizing reduction of time to results.

Currently, steady-state CFD simulations are supported with transient being on the roadmap.

In the example presented above (FAST mode), the process requires 3.2% of a conventional CFD simulation done by a CFD solver. Afterward, CFD Suite takes over the results, does the data formatting for AI prediction purposes, predicts the outcome of the simulation, and re-calculates the data back to the configured formats. Overall, the acceleration is about getting the results at least 28 times faster for the FAST mode.

On the inside, CFD Suite is based on the reduced ResNet network organized as residual blocks. This is combined with byteLAKE’s proprietary AI models.

CFD Suite delivers:

• AI Model(s) compatible with your existing workflow

– No need to change any data types, replace toolchains, etc.

– No hardware/infrastructure upgrades or changes needed

– Cross-platform compatibility (CPU-only, CPU+GPU, PC/ data center/ cloud, etc.)

•AI Model(s) easy to use

– Highly compatible solution ready to take data as CFD Solvers do

– Easy adaptation. Working the same way, the standard CFD Solvers do

– Minimum 10x faster results compering to CFD Solvers

– Highly adaptive solution ready to follow changes in CFD Solvers

•AI Model(s) enabling new possibilities

– Much faster simulations

– Significantly reduced cost of trial & error experiments

Example case study: chemical mixing

The goal was to compute the converge state of the liquid mixture in a tank equipped with a single impeller and a set of baffles. Based on different settings of the input parameters, the simulation predicted the following set of quantities

– the velocity vector field (U)

– pressure scalar field (p)

– turbulent kinetic energy (k) of the substance

– turbulent dynamic viscosity (mut)

– turbulent kinetic energy dissipation rate (epsilon)

Here are the example results for the chemical mixing case. We used a mesh size of <2M cells and the simulation required 5,000 (five thousand) steps to converge into a final state (steady-state simulation). We used a tool called MixIT, provided by Tridiagonal Solutions to generate the results headlined as “CFD”. The results visible on the right-hand side (headlined as “CFD Suite (AI)”) have been generated by byteLAKE’s CFD Suite using the ACCURATE mode.

In that case, further analysis of the results led us to the following conclusions:

CFD Suite = at least 10x faster CFD simulations at 93% (and higher) accuract

Learn more

Learn more about CFD Suite by visiting our website www.byteLAKE.com/en/CFDSuite. We constantly update our product with new features so follow our blog post series where we share the latest updates. Also, consider joining dedicated LinkedIn and Facebook groups to stay in touch with the CFD Suite community. You might as well be interested in listening to our panel discussion about how CFD Suite has been successfully used in the CFD/chemical mixing space.