AI surrogates return a flow field in seconds, but they are trained on CFD and fail quietly outside their data. What they do well and what engineers should learn.

DEEB-FLOW illustration of a velocity volume rendering in a ventilation case
交互式概念示意图:隔离病房的天花板送风射流与床头低位排风口。横向移动指针可降低送风温度:超过某个临界温差后,较重的射流会在房间内脱离天花板并落到床上,其速度远超吹风不适限值;而一个仅用贴附射流工况训练的假设替代模型(虚线)仍在预测沿天花板的贴附流动。脱离过程遵循一条标度律;数值为示意性质。

本文暂无您所用语言的版本,以下为英文内容。

The question behind the headlines

Students and young engineers now ask whether CFD is still worth learning, since AI will soon be doing it. It is a fair question, and the headlines encourage it. Trained models return a flow field in seconds where a solver takes hours.

Here is where I think AI in CFD stands, written by someone who runs simulations and has no AI product to sell. In short, surrogate models are already useful for one kind of problem, they are trained on CFD and are only as good as it is, and the engineers most at risk are the ones who never learned to check a result.

What an AI surrogate is

Three things get lumped together, and they are different.

  • Data-driven surrogates. You run many CFD simulations across a range of designs or operating conditions, then train a neural network to map inputs to outputs. Afterwards the network predicts a new case almost instantly. It has learned to interpolate your simulations.
  • Physics-informed neural networks. The governing equations are built into the training, so the network is penalised when its output violates them. Raissi and co-authors introduced them in their modern form in 2019.
  • Machine learning inside the solver. Examples are learned turbulence closures, learned corrections for coarse meshes and learned initial guesses that speed up convergence. Here the conventional solver still does the work and the model assists it.

The first is where the headline speed-ups come from, and it is the one people mean when they talk about replacing CFD.

Where it works today

In 2026 Aerospace America reported on a panel with speakers from NVIDIA, Neural Concept and Leonardo Helicopters. The examples included a 540-fold speed-up in ground-noise prediction and an AI model that replaced 24-hour CFD runs for aluminium casting. These are figures reported by the speakers, and I have not seen the underlying data.

The pattern behind such results makes sense. A design team exploring a fixed family of shapes, at conditions inside a known envelope, can afford to compute a set of cases once and then query the surrogate as often as it likes. Early-stage design, optimisation loops and real-time monitoring all fit that description.

Where it fails

It fails outside what it was trained on. A surrogate knows its training set and nothing else about physics. Ask it about a geometry, a flow regime or a boundary condition that the training data did not cover, and it will still return a smooth, plausible field. It has no way to tell you that the answer is wrong.

A conventional solver fails loudly. It diverges, or the residuals stall, or the mass balance does not close. A neural network fails quietly, and quiet failure is the dangerous kind in engineering.

  • It needs CFD to exist. Every training case is a simulation someone had to run and verify. The same report notes that proprietary training sets are sometimes as small as 20 simulations. A surrogate trained on unverified CFD reproduces its errors at high speed.
  • Physical consistency is not guaranteed. A data-driven model can return a field that does not conserve mass. Physics-informed approaches reduce this risk without removing it.
  • Certification is still being worked out. For anything safety-related, regulators want to know why a result can be trusted. According to the same report, European guidelines for certifying such models are expected in 2027.

So will AI replace CFD?

My view is that AI moves CFD to a different place in the workflow. The solver becomes the source of truth that trains and checks the fast model. That makes the quality of the underlying simulations matter more, because their errors are now copied a thousand times. Verification and validation, the unglamorous part of the work, becomes the valuable part.

What shrinks is routine repetition, such as the fortieth variation on a geometry an organisation has simulated for ten years. If that is all an engineer does, a surrogate will do it faster.

What stays is the judgement around it. That means deciding what to model, knowing which physics matter, noticing that an answer is implausible and knowing where the training data ends. Those skills take the longest to learn, and no surrogate has them.

What I would do as an engineer

  • Learn the fundamentals properly. They are what lets you tell a right answer from a confident one, whichever tool produced it.
  • Learn enough machine learning to ask good questions. Ask what the training range was, how the model was tested outside it and what it conserves.
  • Treat every surrogate prediction as a hypothesis until it has been checked against a full simulation at a few points, including the edges of the design space.
  • Keep your simulation data organised. Clean, documented, verified cases are an asset now in a way they were not five years ago.

What I would do as a manager

Ask one question of anyone proposing to replace simulation with a trained model. What happens when we ask it something it has not seen?

If the answer is a clear description of the training envelope and a check that flags queries outside it, you are talking to someone serious. If the answer is a speed-up figure, keep the solver licence.

To the student who asks whether to learn CFD, my answer is yes. The people who will use AI well in this field are the ones who would have known the answer was wrong without it.

来源与延伸阅读

相关服务

模型审核与定制研究

明确模型能够支撑哪些结论、不确定性存在于何处,以及还需核实哪些内容。

了解服务
全部文章