Where panel codes stop: or how to use higher-fidelity codes earlier

Early in most aircraft programmes, broad aerodynamic shape questions get answered with a vortex-lattice method (VLM) or a panel code. I love potential codes - I've written them for funfor fun?Yes. I'm that sort of nerd. They run in seconds, they fit in a design loop, and for a conventional wing at the design point they are usually close enoughhow close?At cruise on this airframe, all three with tails: VLM CL 0.395 (25% above Flow360), panel 0.218 (31% below), Flow360 0.316.. The trouble is that "close enough" is means they leave some physics out - and almost every aircraft ends up having deleterious behaviour that wasn't in the ADB ADB?What we call an ADB/Aerodynanmic Database is usually a collection of .csvs and scripts, not an actual Database - we'll be changing that soon, too. or predicted in earlier design, because it wasn't findable with the tools being used.

In my experience higher-fidelity tools like Flow360 aren't used early enough not for want of the information, but for a few reasons:

• RANS (not to mention LES, DDES) is typically too slow to fit in the design looptoo slow?Flow360 is much faster than any other CFD I've used, but it's often the prep/post that takes the bulk of the time.
• The geometry definitions used by preliminary designers, VLM, and panel codes are often incompatible with the watertight OML definitions needed for CFD
• CFD is traditionally a siloed discipline and handoffs become files and geometries between people - so comparison of results between the parametric definitions desired at preliminary design is a nightmare

There is a better way - and I joke that Flow360 is so easy to use that even I can use it. Is that a yardstick?I've been using CFD since about 2007, and I've *written* my own solvers - I'm not a CFD novice, but I really found doing CFD unenjoyable at earlier parts of my career. If something takes more than about ten minutes to set up, then I am likely to get distracted. Why I moved away from using CFD was the complexity and nuance of setup, and the demarcation of the simulation from the sorts of questions I wanted to ask.

I've run tens of thousands of Flow360 sims since I started at Flexcompute, and my interests lie less in one particular solution, and its individual precision and accuracyaccuracy?My colleagues are doing a great job showing Benchmarks of Flow360. This post is showing how we can use it for predictive trends - and I'll be sharing Flight Dynamics (and the accuracy of *those*) results next week., and more in the variation of results with things I either do to the geometry, or on results post-processed at a later date. Loops that traditionally require multiple people and difficult handoff.

To make the point concrete, I took one parametric airframe geometry?using an in-house fully-differentiable parametric aircraft geometry engine, Loom, contact me if you want to give it a go!and ran it three ways: a vortex lattice, a panel method, and Flow360 RANS. The geometry is identical in all three. It comes from the same parametric model, so I can be sure any difference is the method, not the CAD or the way the geometry was turned into the flow model.

Lift: unsurprising

VLM and the panel method are straight lines, at 0.050 and 0.047 per degree. Flow360 is not. Its lift slope is 0.056 per degree between 2.9° and 5.4°, so VLM is 10% low and the panel method 17% low, rising to 0.064 per degree between 5.4° and 7.9°, where they are 22% and 27% low. That curvature is vortex liftsee itThe vortices on the real aircraft: Flow360 rotation planes at 2.9°, 5.4° and 7.9°, and the extra suction they put on the wing that a panel code misses.: the swept leading edge and the chined forebody shed vortices that sit over the upper surface and add suction. Neither potential method can produce it - at least not in the time constraint I set myself to get this done. But that feature was easy to find in Flow360.

The two simpler methods also disagree with each other on the offset. VLM sees the camber and twist of a thin sheet; the panel method sees a thick body and a different load distribution. At the cruise angle, 5.4°, Flow360 gives CL = 0.316. VLM over-predicts by about 0.08 (25%), and the panel method under-predicts by about 0.10 (31%).

All three include the tails: the potential methods dont tell a consistent story here. In the VLM the two tails add CL 0.016 and lift the right way, a little under Flow360 (0.029 against 0.034 per radian, per tail). In the panel method the tails lift the wrong way: their own lift falls as the angle rises, and that is most of why its line sits so low. The faint lines on the plot are both methods with the tails taken off.

With a system that doesn't have commonality of geometry, this would be three different models - and ensuring you were comparing like with like would be a nightmare. Believe me. I've done it. Confession.Aside from very small model, I've _never_ made an AVL model myself. I always used other peoples.

Drag: always more interesting

A panel method gives induced drag only which is…which is most of the lift-dependent drag, so that's why they're good for stability purposes: about 87 counts at 5.4°, 67% below Flow360's totalthe splitAt cruise: 87 counts induced (panel, with tails) against Flow360's 263 (67 friction and 196 pressure), so the panel figure is 33% of the total.. Flow360 gives 263 counts at 5.4°: 67 counts of skin friction and 196 counts of pressure drag. By 7.9° the total is 568 counts, and almost all of the increase is pressure drag, as the vortex system grows and the flow over the aft body and tips changes.

In a preliminary-design loop that means the drag-polar shape, the best L/D angle and the margin before the drag rise aren't really available. You can do what most folks do and bolt on a handbook build-up for friction and form drag, but the pressure drag that depends on this particular shape is exactly what a handbook can't know.

What only CFD shows you

Forces are only half the story. RANS gives the flowfield itself: where the vortices form, where they go, and what they do to the tails and the aft body. That's what tells a designer *why* a change helps or hurts, not just whether it does.

Making CFD fast enough for preliminary design

These days CFD isn't as slow as old we all collectively still think it is - but even with Flow360, it still can't match the VLM or panel speed. A RANS run is minutes to hours, a VLM run is a second.

The trick to using in the design loop is to Tailor your condition matrix to yield the most information at any condition, and use it in conjunction with the lower-fidelity methods in a workflow that usually has to be written from scratch for every aircraft Workflow.CFD-corrections to lower-order methods are a tried-and-tested workflow on many aircraft programmes - but in my experience, the workflow is normally bespoke to a single programme. Flow360, Tailor, and Thread make it a generalised process, repeatable, and fast..

So I ran Flow360 on 24 cases, 16 "candidates" from the same parametric model at a range of angles of attack, and trained a learned model (Tailor) on them. It predicts forces, surface pressure, and the full flowfield for shapes we never ran, in a browser, as you move the design sliders. In reality, for real preliminary design I woudl run more - but I set myself a challenge to see what I could get done in a few hoursWhy the time constraint?I enjoy writing and I enjoy defining new workflows, but the point of this exercise is to show Flow360 being used as part of prelinary design - not to actually do a prelminary design study.

Tailor is tested by leaving one whole candidate out, training on the rest, and predicting the shape it has never seenhow?Each dot is a real Flow360 run, predicted by a model that never saw that shape.. Across all 16 shapes, the held-out lift error is 0.029 (9% of the average CL) and the drag error 28 counts (9% of the average drag), both RMS. Run by run, the median error is 5% in lift and 7% in drag. Near the aircraft, its predicted pressure field is closer to the true Flow360 result than the flow of the nearest real run: a Cp error of 0.031 (3% of the freestream dynamic pressure) against 0.050 (5%). Every prediction carries its own uncertainty, and you can always snap back to a real Flow360 run to check.

The learned model does not replace CFD. It is built from it, and it is only as good as the runs behind it What ML isOne of my oft-repeated lines is that Machine Learning is *fancy curve fitting*, so super at interpolation, and caution should be heeded with extrapolation. But it moves CFD-quality information, including the vortex lift and the pressure drag that VLM and panel methods miss, into the stage of design where decisions are cheapest to change.

None of this says a learned fit can substitute for CFD. Quite the opposite: as these figures show, the fit is often quite far from the CFD. But that is often good enough for preliminary design, and that same fit can help inform how your large-density ADB gets built.

The takeaway

• Use VLM and panel methods for what they are good at: fast, linear trends on attached-flow wings.
• Use CFD where the physics is nonlinear: vortex lift, pressure drag and the shape of the drag polar decide the design on swept, blended and chined configurations.
• Use lower fidelities AND Flow360 together: you can do preliminary design and avoid surprises that crop up at flight test.
• Train on CFD, then explore: a learned model from a modest number of Flow360 runs gives you the design space at preliminary-design speed, with errors you can measure. This then gives you a geometry to do more design, and can immediately run more CFD on the OML.

If your preliminary design still runs on potential-flow tools alone, we'd like to show you what the same loop looks like with Flow360 behind it. Get in touch.