A customer wants a quote on a custom heat exchanger.
With Fluent alone
An engineer sets up and solves a case before sales can answer.
With a surrogate model
The sales engineer types the sizes and reads the pressure drop and outlet temperature.
AI surrogate models for CFD
A full 3D Fluent run takes hours. A quote, a design review or an operating decision needs hundreds of answers by tomorrow. We validate your case, run the training simulations with PyFluent on our cluster, and hand you a trained model that predicts pressure drop, temperatures and forces for designs nobody has simulated yet.
Prediction console
Shell-and-tube heat exchanger, shell side
Pressure drop
12.0 kPa
Outlet temperature
32.2 °C
Illustration only: a simple formula running in your browser, not a trained model. Yours is trained on your own validated Fluent runs and comes with its measured error.
Where the hours go
Every new question waits in the same queue: geometry, mesh, setup, solve, and an engineer to read the result. These are the questions that wait longest.
With Fluent alone
An engineer sets up and solves a case before sales can answer.
With a surrogate model
The sales engineer types the sizes and reads the pressure drop and outlet temperature.
With Fluent alone
40 cases to set up, solve and post-process.
With a surrogate model
40 predictions in under a minute, then Fluent confirms the best three.
With Fluent alone
A request to the simulation team, and a wait.
With a surrogate model
An answer on the operator's screen while the decision is still open.
With Fluent alone
Running them all would take years of solver time.
With a surrogate model
An optimizer searches all 10,000 on the model and sends the finalists to Fluent.
Fluent stays in the loop: the model answers the routine questions, and full runs check the cases that decide something.
What a surrogate model is
A surrogate model, also called a metamodel or response surface, is a machine learning model that learns how your simulation's outputs respond to its inputs. We run a planned set of Fluent cases across the ranges you care about, collect the results, and train the model on them. Once trained, it predicts a new case in a fraction of a second.
It's only as good as the runs behind it. Inside the trained ranges, we measure its error against Fluent runs it never saw and give you the number. Outside them, its answers can't be trusted, and the validation report says where that line is.
Inputs you choose
flow rates, sizes, angles, temperatures
Planned Fluent runs
one validated setup, every run checked
Trained model
error measured on runs it never saw
Predicted outputs
pressure drop, heat transfer, forces, uniformity
How a model is built
Automation multiplies whatever it's given. A wrong setup run 500 times gives 500 wrong answers, so we validate first and automate second.
We agree the inputs, their ranges, the outputs the model must predict and the accuracy you need. You get a fixed quote.
You approve: inputs, ranges, outputs
One Fluent case is set up, put through a mesh independence study and checked against your test data, a published result or your own simulation.
You approve: the validation plot
We plan which cases to run so they cover the design space evenly. A common starting point is ten runs per input, with more added where the model is still unsure.
You see: the run plan
PyFluent sets up and solves every case the same way on our 512-core cluster. Each run's convergence is checked, and runs that fail are fixed or dropped before they reach the training data.
You see: the run log
We fit and compare model types, such as Gaussian processes, gradient boosting and neural networks, and hold back runs the model never sees. Its error on those runs is the accuracy we report.
You get: the measured error, per output
The model arrives as a Python package, a desktop tool or a web app, with documented code and a walkthrough session. Questions and extensions after delivery are part of the job.
You get: the model, the code, the data
What we can build
Most projects combine two or three. A surrogate model starts with an automated Fluent workflow, and optimization usually runs on a surrogate.
You get
A trained model that predicts your outputs from your inputs in under a second.
Useful when
You need hundreds of answers across a known design range.
Fluent needed to use it?
NoYou get
A surrogate behind a simple form or web page, for people who don't run CFD.
Useful when
Sales, quoting or operations need answers without waiting for engineering.
Fluent needed to use it?
NoYou get
An algorithm searches your design space for the best design against your targets and manufacturing limits, for one objective or several.
Useful when
You know what "better" means, such as lower pressure drop or more uniform flow, but not which design gets there.
Fluent needed to use it?
Only to confirm the winnersYou get
Your recurring simulation as one command: geometry import, meshing, setup, solve and a standard report.
Useful when
The same kind of case comes back every week and is set up by hand each time.
Fluent needed to use it?
YesYou get
Sensitivity studies, parameter sweeps and what-if scenarios, run from a table and compared in one report.
Useful when
You need to know which parameters matter before you design around them.
Fluent needed to use it?
No, you receive the resultsYou get
An in-house tool where engineers enter dimensions and operating conditions, and Fluent meshes, solves and reports in the background.
Useful when
Several engineers run the same validated model, and not all of them are Fluent experts.
Fluent needed to use it?
YesWhere surrogates pay off
Each card shows what goes into the model and what comes out. We also work on energy and biomedical devices.
HVAC and data centers
Heat transfer equipment
Oil, gas and process
Turbomachinery
Process equipment
Automotive and marine
A manufacturer quotes custom heat exchangers, and every quote needs a thermal check that waits for an engineer. We validate one Fluent model of the product family, run a few hundred cases across its sizes and flow rates, and train a surrogate on them. Sales engineers get a form that returns pressure drop and outlet temperature in seconds. Full simulations stay for the orders that get signed.
This shows how a typical project runs. It isn't a named client's result.
Platform or partner
Several software platforms train AI models on simulation results. They work well for teams that already have hundreds of clean, consistent runs and the engineers to produce more. Most teams have a handful of cases, set up by different people over several years.
Already using a platform? We can produce validated Fluent training runs for it.
What you receive
your-model/
Cost and timing
We don't list a price, because two projects rarely cost the same. After the feasibility check you get a fixed quote and a timeline, and if the scope changes, we requote before we charge.
We read your brief, estimate how many runs the model needs, and tell you honestly if a surrogate is the right tool.
Get a free feasibility checkQuestions
A machine learning model trained on CFD results. It predicts the outputs of a new case, such as pressure drop, temperatures, forces or flow uniformity, from its inputs in a fraction of a second, without running the solver. It's valid inside the ranges it was trained on.
It depends on how many inputs the model has and how strongly the outputs change across their ranges. A common rule of thumb starts at about ten runs per input, so 50 to 60 for five or six inputs, then adds runs where the model is still unsure. Wide ranges and strongly nonlinear flow push the total into the hundreds. We estimate the number in the free feasibility check.
We hold back Fluent runs the model never sees during training and report its error on them, output by output, before delivery. You set the accuracy target in the scope. The validation report also states the ranges the model is valid in, since outside them its error isn't measured.
Not to use the trained model: it runs in plain Python. The training runs happen on our cluster, or on your own systems if your data has to stay inside. An automated workflow or a custom engine that runs Fluent itself needs Fluent wherever it runs.
Often, yes. We check them first: runs made with different meshes, physics models or convergence criteria can't be mixed blindly, and a single under-converged run can bend the whole model. Results that pass join the training set and cut the number of new runs.
You do. You receive the trained model, its documented source code, the training data and the scripts that produced it. We sign an NDA before any technical detail moves, and your existing template works for us.
Whatever suits the people who'll use it: a Python package your engineers call from their own scripts, a desktop tool, or a web page where sales staff or operators enter a few values. Each comes with documentation and a walkthrough session.
The model knows the ranges it was trained on. When a new size, material or operating condition comes in, we run the extra cases with the same automated workflow and retrain. The scripts are yours too, so your team can do it.
No. PyFluent is the open-source Python interface Ansys publishes for Fluent. Fluent still solves the physics, while PyFluent lets a script drive meshing, setup, solving and results, so hundreds of cases run the same way without anyone clicking.
It depends mostly on how long one Fluent run takes and how many runs the model needs. The feasibility check comes within 24 to 48 hours, and the quote names the timeline for every stage. Behind it are 15+ years of Ansys Fluent projects.
Free feasibility check
A few lines are enough. A senior CFD engineer reads them and replies within 24 to 48 hours with what's possible, roughly how many Fluent runs it takes, and a fixed price.
Worth including
Rather talk first?
Working on a thesis or a paper? Our CFD Consulting page covers research projects
AI surrogate models for CFD
Send us the inputs and outputs in a few lines. We'll reply with an estimate and a fixed price, free.
Ask for the estimate© 2011 - MR CFD LLC. All Rights Reserved.
MR CFD Support
Choose the way that suits you best.
Live chat didn't load in this browser. Please message us on WhatsApp or by email instead.
Or message us
Scan this code with WeChat to add us.
wxid_s63x2eminh0a12
In WeChat, tap + and then Add Contacts, and paste this ID.
AI Assistant
Ask me anything about MR CFD
How can I help?
Ask away, I'll point you the right way.
What I can do on this page