Hey there,
It’s Robin from CFD Engine and I wanted to step outside the OpenFOAM-ParaView-AWS box this week & share three next-level CFD-related innovations. Three things for which my first reaction was “this is nuts” 🤯
I can’t pretend I know how they work, their implications, limitations (or whether they’re actually for real) but taken at face value, they represent some noteworthy developments.
We have an overgrown GPU with crazy vital statistics, a super-speedy CFD code & some bonkers AI – let’s go…
Wafer-Scale Engine
Introduced in 2019 the Cerebras Wafer-Scale Engine is the biggest chip you can make from a single silicon wafer & it’s a beast
A chip the size of an iPad, with 850,000 cores, 40GB RAM and 20 PB/sec of memory bandwidth (hint: that’s a lot).
But, before you get your purchase order ready, this is a specialist bit of kit. The cores are similar to those found in a GPU, perfect for deep learning, but not for running OpenFOAM.
A quick glance at the Cerebras website shows they’re firmly focussed on machine learning. But, putting lots of cores close to memory & letting everything talk to each other very quickly, sounds like exactly the sort of architecture we’d be interested in.
They published a paper where it ran CFD (or did it? I think they simulated a simulation 🤔 ) But in theory, GPU-only CFD codes could run on it & they’ve started making it available in the cloud, so maybe someone will take it for a spin?
Lot’s of cores & fast interconnects should make things pretty quick, but would it be able to keep up with this next Speedy Gonzales?
Flow360
Flow360 from Flexcompute is a next-gen RANS solver that solves things fast – like 50million cells in 3 minutes fast.
You have questions, so do I, but I don’t have many answers I’m afraid. Here’s what I’ve got scraped together from their website.
It’s a proprietary code put together by some big brains, from some big universities. It solves RANS, DDES & LES simulations on conventional grids, running on their cloud platform & it does it MUCH quicker than anything else & just as accurately.
An astute friend pointed out that it would probably take longer to upload your next case than to solve it 🤔
I have some reservations about cloud-only solutions, especially with proprietary codes & I’m not sure I need a solution that quickly. But they’re undeniably impressive numbers & it would make mincemeat of your next optimisation DoE.
I’ll be interested to see where this goes – I’ll keep watching.
Sketch2Mesh
This is the combination of a few technologies – an attention grabber for what deep learning can do when applied to simulation & geometric data.
Sketch2Mesh from Neural Concept is a demo of an application that allows a user to sketch a car with a handful of lines (the sort of thing you’d draw on a napkin) which it converts into a 3D model of the car & predicts its aerodynamics – in real time.
This, plus Gravity Sketch would be amazing 🤯
I don’t think anyone is suggesting that this is going to replace the traditional design cycle anytime soon, but the underlying techniques are definitely interesting.
Using an archive of CFD simulations to train a model to relate shapes to flow structures/pressures/loads etc is an interesting approach. Using that model to predict the performance of new shapes, without having to parameterise them first, is compelling.
Obviously this relies on having piles of old CFD data available to train the models, but there are plenty of companies in that position. I wonder what they could uncover by letting these techniques loose in the archives?
This article explains the approach in a little more detail – worth adding to your reading list.
Your turn
There’s so much innovation going on in & around our space, these were just three things that recently floated across my feed & that I thought were pretty crazy.
What was the last innovation you saw that provoked a similar reaction?
Or maybe you have some more info on the ones I mentioned?
Drop me a note & fill me in – I like getting your replies 🙏
Sim-u-later,
