Fast and accurate GPU-based CAE software is possible, according to M-Star and the pharmaceutical industry.
The CAE industry has had a lot of overarching goals over the last decade. Though not an exhaustive list, two that often come up are:
- Simulation democratization—the idea of bringing simulation technology to everyone.
- Simulation computational improvements—the aim to produce accurate results at faster speeds.
One candidate that could meet these two goals is GPU-based simulation, which offers results at the blink of the eye. This speed enables users to quickly play around and pick up the software, as well as iterate designs. However, traditionally software like this has traded accuracy for those improved speeds. As a result, they are good divining rods that point engineers towards more optimal designs, but they are not a great tool when accurate evaluations are needed later in the product development lifecycle.
A fully coupled transient CFD-DEM simulation, produced using M-Star CFD, which shows particle dissolution within an agitated tank. “A lot of codes do CFD. Sometimes these codes talk with other software—with varying degrees of success. Within M-Star, all physics are fully coupled and integrated directly within a single package,” said John Thomas, president of M-Star Simulations. “One-way coupling, two-way coupling, four-way coupling, non-uniform size distributions, non-spherical particles, particle reactions, particle break-up, particle swelling, particle dissolution … it’s all there and seamlessly integrated. And, since the code is bred to run on GPUs, the results are fast and beautiful. This multiphysics simulation probably took a couple of hours to run. Dissolution is a huge application space of M-Star in biopharma—an industry underserved by current vendors.”
Enter M-Star Simulations’ GPU-based CFD software. During a demo by John Thomas, president of M-Star Simulations, on the surface the software looked like other simulation applications. Import CAD, produce a mesh, run the simulation and start post-processing. However, what was surprising was that within five minutes Thomas had done all of that while modeling a transient mixer (with and without 500,000 particles). Most impressive was that the results only took seconds to compute.
After witnessing GPU-based simulations before, this alone might only earn a raised eyebrow—but Thomas claimed that the results were accurate and could be produced using an off-the-shelf gaming computer.
As a cherry on top of it all, he supplied a list of academic papers, co-written with pharmaceutical giants, to back up these claims. Now this was interesting.
What is M-Star?
M-Star is an internal CFD simulation software that has been programmed from the ground up to produce accurate simulations, explained Thomas.
“Because it was built from the ground up to run on GPUs, we can fully exploit GPU architectures to realize orders-of-magnitude improvements in speed and fidelity,” said Thomas. “On paper, people don’t usually comprehend what orders-of-magnitude improvements in speed really means. This difference isn’t like Ford versus Chevy. It’s more like the difference between jetliners and covered wagons. Although both can be used for cross-country transport, they present two radically different paradigms of travel. That’s what we’re doing with simulation.”
A simulation depicting the response of an HVAC system with a circulation flow rate managed by a PID controller. Thomas explained, “In this system, off-gas vapors escape the hood and pool near the ceiling. A sensor positioned above the hood records the time-evolution of the concentration, triggering the response of the HVAC purge. Once off-gas concentrations are reduced to acceptable levels, the purge stops. This is a full transient integration of the fluid flow simulation with an external PID controller to see how the entire system, with control theory and flow physics, behaves in practice. This runs in a couple of hours on two 3090 GPUs. I don’t think people would attempt to solve this problem using conventional CFD applications. You might try to get a picture of the steady state flow field and maybe a mixing time, but a complete integration of fluid field, PID controller, dynamics response is not practical using the finite volume/element approach. People from that background think that what we did would take two months building a model and six-month running it on a thousand CPUs. What’s the big deal? The big deal is that we can discuss the system over breakfast and have the solution by dinner—all on a desktop workstation with two GPUs.”
“I’m not trying to simulate steady-state flow or minimize residuals. I’m marching through space and time,” he added. “And it’s all happening on a rinky-dink RTX 2060 GPU—which isn’t very powerful. But if I hit ‘go,’ I’m not waiting for this simulation to solve. I’m watching it advance in real-time on my machine. I’m solving this, while running it, while post-processing it—and all of it in 3D.”
Okay, it’s fast. But how accurate is M-Star compared to traditional CFD technology? Ansys Discovery, the only comparable GPU-based CAE application of this type (that I know of), can produce simulations with similar speed and it has a wider scope of physics options. However, it is an early in development, first-pass simulation tool. Currently, you would not use it as the sole, and final, simulation application to assess a design. Think of it as a compass needle that points towards an optimal solution as true North.
Thomas explained that M-Star offers capabilities and accuracy on a similar scale to Ansys Fluent but at those GPU speeds. “Every output we predict, when possible, we make an immediate comparison against measured data. As the simulation is going along, I’m looking at things like force, torque and the transient flow field. If I have a probe, I’m looking at the time evolution, the velocity in the simulation and comparing that against measured data. This isn’t a calculation, it’s a digital twin I compare directly to physical data.”
So, how is this all possible? Thomas explained that “because we’re using a more powerful computer architecture, we can use better foundational transport algorithms. Out of the gate, we’re using LES and particle/bubble tracking, so this isn’t about how quickly I can minimize residuals. We’re using an approach that captures a spectrum of physics that doesn’t exist in RANS-based tools.”
The preprocessing process of the M-Star workflow also caught my eye. Typically, I’m used to seeing a mesh of varying cell sizes. But the ones used by Thomas were quite uniform. Frankly, it would have made many CFD engineers faint.
“Much of the user’s time is spent on meshing,” said Thomas. “People don’t use simple meshes where everything is uniform and Cartesian because it probably will break. The problem scales so unfavourably with system size that they can’t hope to run dozens or hundreds of millions of grid points. So, people invest time to make the problem tractable and making the mesh not one cell bigger than it needs to be in order to make their simulation run as efficiently as possible.”
“At M-Star,” he argued, “we are up to our ankles in speed. Which means that rather than me trying to spend days and weeks massaging a mesh to make it acceptable to the problem statement and make it tractable with respect to the software, I choose the finest mesh I need, and I go get lunch.”
In other words, Thomas is looking to spend 15 minutes on setting up the simulation and mesh, move onto another task and come back to high resolution results in a few hours.
What the Industry Thinks About M-Star
The M-Star team sifts through the literature to find fluidic data, then recreates those experiments, as a simulation, within the software to verify its precision and accuracy. And M-Star isn’t afraid to share this data itself. For instance, they have worked with heavy-hitting pharmaceutical organizations to produce academic papers that utilized M-Star technologies. Some of these pharmaceutical organizations include:
“The name of the game is trust,” said Thomas. “These pharmaceutical manufacturers, they need to know they can trust the product because they use it to make very high-value drug products. So, we have papers that systematically go through the code, making sure it agrees with measured data.”
These pharmaceutical companies no doubt have the budgets to use any CFD technology. They choose M-Star, said Thomas. That speaks volumes. But don’t take his word for it, take AstraZeneca’s:
“We used M-Star CFD’s software to simulate the dynamic time evolution process of the UF/DF protein concentration step,” said AstraZeneca to Evan Flach, associate director of sales at AIChE. “Concurrent volume reduction, protein concentration and viscosity increase were simulated with retentate tank outlet and return flow with permeate draw off over a 6-fold volume reduction step. Real-time video rendering of the entire concentration step provides insight into the fluid dynamics such as shear at the air/liquid interface, mixing time and tendency to short circuit the retentate return line to the tank outlet.”
AstraZeneca added, “The M-Star CFD simulation rendering allows engineers and scientists to see via video and graphics quantitative time-evolving estimates of shear forces at the air/liquid interface, effect of volume/viscosity on mix times, and visual representation of flow patterns in the vessel during concentrations. This fluid dynamic insight can be used to optimize operating conditions in the UF/DF step to minimize protein damage and improve homogeneity in the process.”
As for competition, Thomas says that M-STAR doesn’t like to think much about them. But, when he does, it’s with quite a lot of humor. “Customers tell us, ‘We already have this tool,’” he said. “We walk in for a demo; we usually leave with a pretty excited new M-Star user.” Or in other words, a purchase order.
“Head-to-head, we have never lost a sale to larger vendors in this field,” said Thomas. “We’re often solving the same physics as these other tools, but at setup and run-times that are orders-of-magnitude faster.”
M-Star Focuses on the Needs of Its Customers
Thomas’s demo continued to impress, especially when he said, “I have particles interacting with the fluid flow, I have two-way coupling between the two. The particles are being suspended. They’re moving around in real-time, and tandem, with the fluid solution. So, I’m solving all these particle dynamics interactions to construct simulations to predict fully transient 3D models of a particle laden flow.”
And I could talk about more examples, but the tool and company had other aspects that need mentioning that were more standard but noteworthy. For instance, its Cloud capabilities.
“M-Star runs on multiple GPUs and scales well across multiply GPUs,” said Thomas. “Typically, a multiple GPU resource is a cloud-based resource. Clients will install M-Star on AWS and then they will boot up an eight GPU resource to run a very large simulation. When they’re done with it, they just shut that resource down.”
He added, “You can also submit to the cloud directly in the M-Star GUI. It’s straightforward for me to add scripts to my pre-processor to take my jobs to the cloud.” The results in the cloud can then be streamed back to the workstation.
In essence, users can install it on the cloud and run M-Star off browser, or they can install it on their computer and send data back and forth to the cloud. Thomas even notes that some users will take the M-Star technology, strip it of its GUI to use whatever they find comfortable. It is quite customizable to the needs of the customer.
Additionally, M-Star has a very tight relationship with its customers. Many of them have been with the company since the beginning. As a result, those early adopters were able to help guide the development of the tool from its initial offering to what exists today.
As for why those initial customers were patient with M-Star even though they had the budgets to purchase more established software? Thomas says it’s because they recognized the end game was to their advantage. They were helping to develop a powerful simulation code that was lightening fast and optimized to their processes.