Podcast Transcript

David Brühlmann [00:00:30]:
We’re back with Christian Witz, who’s the CEO and founder of SimVantage, continuing our conversation on computational fluid dynamics. So far, we have covered why CFD hasn’t caught on faster and where early-stage assumptions break down at scale. Ahead, Christian walks through characterizing bioreactors across scales, the FDA-backed project applying machine learning to residence time distribution and what it actually takes to bring simulation into your biotech scale-up strategy.

Nowadays we face a challenge. As you said, the unprecedented computing power. So you could be tempted to do way too much, to characterize way too much. And also a danger I’m seeing is, what do you do with all the data? You can generate tons of data, but how should you interpret that data? Or what are the questions you should be asking?

Christian Witz [00:02:39]:
So there are 2 things. The first one is that, of course, you need to validate the data. So it makes no sense that we get tons of data, but all of them is wrong. This is also one thing I would like to have as a take-home message: always ask for validations if you’re presented with fancy pictures of CFD results. Because nowadays with LLMs, you can produce fancy pictures in no time. But what you want to have is comparisons to experimental data, to literature data. And so what we are doing usually is, if we have experimental data for this specific reactor, we look at these points that were experimentally tested and see if our software gets it on point. So this is the first thing.

Then we also create a rapid response model. We call it rapid response model. It’s also called surrogate model or reduced-order model. And we are using machine learning algorithms to produce this rapid response model. And what this machine learning model basically does is to interpolate between the points that were simulated. So imagine you have like simulated at 100 RPM and at 110 RPM, and you know at each point the shear rate, for example. With the machine learning, it can interpolate and then, you know, also the shear rate distribution at 105 RPM, for example.

And if you do that, you also have first, before you do that, to split your dataset into a training dataset and a test dataset and later use when you have trained the model, the test dataset and see if you are within a certain range of error that you allow the model to have.

And this reduced-order model, and as it is in the name, is then like a small model. It’s only maybe a few megabytes in size. And this then contains like your full characterization. So the data is not as large as you would think in the end, but you’re getting rather a compact model out of that answers you in milliseconds if you give it the operating conditions. In milliseconds you get like the kLa, the shear rate distribution, and the mixing time and so on of your reactor. And this is then also how we do scale-up, but maybe we come to that later.

David Brühlmann [00:04:51]:
I just would like to address perhaps some doubts or resistance. I hear some people saying, well, we have always done it without CFD. We understand what we’re doing. Why should you still do CFD? Or differently asked, what are they missing out of?

Christian Witz [00:05:07]:
If you have done it that way and it works perfect, I never want to change it. Never touch a running system, right? But if you’re faced with challenges, like if you have a new cell type, a new kind of system, like with cell therapies where suddenly the cells are the product and you cannot add a lot of antifoam to your Pluronic, to your mixture, and suddenly your cells are much more sensitive, and then you need to get a really scale-up on the point, then I think CFD is a really good tool to support this. And I have 2 examples.

The first one is a project which we are currently doing with the Technical University of Vienna, the group of Oliver Spadiut, where we are doing together a scale-up from a shake flask to a stirred system because this is always a hard step to do. And now with our software, we can simulate the shake flasks and we can simulate the stirred systems before you do the scale-up, bring these 2 together on the computer and see which operating conditions do you need to set to your stirred tank in order to achieve the same microenvironment for the cells and therefore providing them with the best conditions to grow. And this has turned out really well and we could scale up these natural killer cells, which are hard to grow, also in stirred systems without any problems.

And the second project we had was together with the RWTH Aachen in Germany. And they used 2 methods. So they used method you learn in university, power per volume, to scale up Streptomyces. And they used our method with bringing in all the physical parameters in CFD and matching them in CFD and using then the operating conditions. And they saw a much more comparable cell morphology when scaling up with our method. And this means the cell’s microenvironment really was much closer than if you’re doing a power per volume scale-up.

And so there are certain advantages, but if your system works, go for it. If you’re facing troubles, in the back of your mind, maybe remember there is another method which you could try out really easily and in no time, and it may give you better results than your traditional method.

David Brühlmann [00:07:19]:
Because I would just like to follow up on a few things you just said. Scaling up is a challenge and a lot of people are struggling and projects have failed because of certain technical challenges along the way. And I believe CFD is a technology that helps to de-risk that. Now, there’s one big challenge. You have done that work on the shake flask, and the shake flask is a very simple system, yet it’s a very different system. I mean, I have a different, obviously, perspective than you have, so I’ll let you speak about this fluid dynamics perspective. From a bioprocess engineer side, a shake flask is rather a simple system, whereas a bioreactor is a bit more complex. You have a stirrer, you control more parameters. What were the challenges you had to solve to have a good model for the shake flask? can also make sure that then you can seamlessly scale up. We have two different perspectives.

Christian Witz [00:08:14]:
True, true, true. Because fluid dynamic-wise, I mean, you don’t have a stirrer in the shake flask, but of course you have one outside, which is your shaker. And this is usually pretty violent. And from a fluid dynamics perspective, you have these thin layers of liquid on your wall. They are very complicated to simulate, and this rotating motion can be very complex. And it has a free surface. So you can imagine when I started, the surface of my reactor always was flat, perfectly flat, because you need a different algorithm in order to have these little waves or maybe a vortex to include that in your calculation.

And then when we started, actually what we wanted was to simulate our larger air bubbles. Because you can imagine in very large reactors, and we simulated up to 160,000 liters, in very large reactors where you have non-Newtonian viscosities. There, the air bubbles can collect together and form a very large bubble that then rises to the top. And we wanted to be able to simulate these very large bubbles. And you need the same algorithm because this free surface algorithm is the same for the inside of the bubble and for this surface on the top of your reactor. So, we had to first include this free surface algorithm.

Then we validated this together with the Hochschule Niederrhein in Germany, so the University of Applied Sciences in Niederrhein. Where they tested different vortex profiles when you have different baffles in your reactor. And we did the same in simulation.

So then we progressed slowly to the shake flask where you have these really thin layers of liquid on the side of your reactor or your shake flask. And you can imagine if you have that kind of large surface, you also have to have an algorithm how the oxygen gets into your liquid. And this also has to be calculated in the right way. And also the fluid is very violently turned over again and again in the shake flask. And therefore you have a lot of shear and you need to calculate the shear correctly in order to have the correct starting point when you transfer it into a stirred system.

And so this was very complex and I can only tell a little anecdote where a colleague of mine here in the company, Michael Wagner, was giving a talk at the Himmelfahrtstagung in Germany, about his work on the shake flasks. And in the next morning at the breakfast in the hotel, he saw an elderly gentleman always looking over to his table. When he finished breakfast, this gentleman came over, introduced himself as Professor Büchs from the RWTH Aachen. And he’s like the godfather of shake flasks. And he said that he enjoyed the presentation and then that he has a lot and lots of data on the shake flask. And he really then also sent us over a lot of validation data, which was like a treasure for us. And with this data, we could then finally validate our shake flask model.

So this was now the long introduction into— yeah, shake flask is pretty complicated. It took us a long time to be able to simulate it. But if you can simulate it, get out the same parameters that you then have as an input in your stirred system, and then you just change the aeration and the stirring until you find the same oxygen concentrations, kLa values, and the same shear rate profiles. And then you can find out the operating conditions in your stirred-tank.

David Brühlmann [00:11:34]:
I’d like to zoom out and focus on other applications for CFD. You’re currently running a project on continuous bioprocessing. Tell us more about that.

Christian Witz [00:11:49]:
So this project, we’re very lucky and pretty astonished. We got this project actually from the FDA, so from the American Food and Drug Administration. And this project is concerned with continuous bioproduction and looking at the residence time distribution.

The FDA is interested in this because if you have a deviation or disturbance somewhere in the process, it’s not that obvious how many vials or amount of product you need to throw away at the end of your process. And therefore residence time distributions are important because then you can track your substance that you introduced which has higher ionic strength or something that you don’t want to have in your process. And you can track them over different unit operations till the end of your process and then you know what amount of product to throw away and which one you can keep safely and sell safely in the end.

So our task is really to find for every unit operation, and we have chosen four, so the continuous stirred tank reactor, the connectors, the filter, and the chromatography. For each of these unit operations, a first principle simulation method in order to then create the same rapid response models we created for the bioreactors. And then we can connect all these unit operations together, simulate the whole RTD chain in the end.

David Brühlmann [00:13:12]:
And since the FDA contacted you, I’m curious, Christian, is CFD data necessary in a filing or not? Do regulators think about that?

Christian Witz [00:13:23]:
So my impression is that they’re really open to the idea of filing also CFD data because you can reduce the number of unnecessary experiments when you file CFD data. And then the FDA is aware of this and they also support these initiatives where you reduce the amount of experiments, the amount of media you have to throw away and so on in order to just get the necessary data.

And what the FDA has especially is a risk-based approach, of course. So the patient risk is always the guiding principle. And if you want to know more about that, there is a guideline from the FDA, the use of and assessment of computer simulations in medical device submissions. And there is also this ASME V&V 40 standard.

Yes, it’s all for medical devices, but this credibility assessment that is described in this guideline and in this standard, you can also use this for simulations that are done in the context of manufacturing. And the FDA also will accept that because it follows a standard.

And currently the V&V 80 standard is in development, which is like the same but for manufacturing and with use cases and explanations how you do that for manufacturing. Until this is published, you can safely follow this V&V 40 standard and submit also CFD results to the FDA.

David Brühlmann [00:14:48]:
What advice would you give to a small biotech team now? They’re just about starting CMC development and they think about scale-up strategies. What kind of CFD modeling should they do? When should they do it? And what are some important questions they should ask?

Christian Witz [00:15:06]:
They can start right away when they first think about it with CFD because, as I said, we can now start with well plates and then accompany them to shake flasks and then into larger systems. So the earlier they start doing CFD, the better, because it can also then help you in two ways.

First of all, you can also select— if you’re using a CDMO, for example, CFD can help you to choose between different CDMOs because you can then tell if your process that you may have perfected in an Ambr system, if this can be safely produced in the reactors that the CDMO has available. Because with CFD you can do this on the computer, you don’t need any experiments. Therefore, you can tell in advance if your process will run safely in these reactors.

And the second thing is, and this is maybe not from a scientific perspective, with CFD you can do like a virtual scale-up. You can find the operating conditions in the pilot-scale reactors. You find in the lab reactors, which would be your best operating conditions, and of course in manufacturing. And this is something you can also then show to investors, that you have like a safe and proven path towards manufacturing. You know that these reactors will serve your process well, and you know what to set in these reactors. It reduces the risk, it reduces the time that you need, and this is all things that investors like to hear in new start-ups in the biopharma – biotech field.

David Brühlmann [00:16:55]:
As we are wrapping up, Christian, what additional questions should I have asked?

Christian Witz [00:17:00]:
I think we’ve covered pretty much everything. So maybe you should have asked if you would be able to do a CFD simulation or a scale-up.

David Brühlmann [00:16:55]:
Good question. I’m curious, will I be able to do it or not?

Christian Witz [00:17:00]:
I mean, nowadays we have like a scaler. So this is a tool which we programmed where you have your source and the target reactor. And actually, my 95-year-old grandma used it once to do a scale-up. I’m not sure if she understood what actually she is doing, but she could select a source reactor, she used the Ambr, it’s a good choice, to the STR-50, and then she transferred one set of operating conditions to the other. Of course, with some explanation from my side, but she was able to do it on an iPad.

And CFD simulations, I mean, we have now a new software which is called Bolt, and you can have a look at it on our YouTube video. It’s really easy to use. You can use it on an iPad as you would imagine a CFD software to be. Yeah, you can change a fluid level with a fingertip. And so you can really easily build your own reactor or import like a 3D geometry. You have it right in front of you, you press play, and then it’s calculated on the cloud or your own workstation, and you can play around with it. It’s really easy to use, and I’m sure you can do it, David. I’m sure.

David Brühlmann [00:18:04]:
Well, this is great marketing. If your grandmother can use CFD, so can you.

Christian Witz [00:18:12]:
That’s true. That’s true.

David Brühlmann [00:18:13]:
Yeah. Well, this has been great, Christian. What is the most important takeaway from our conversation?

Christian Witz [00:18:19]:
So the most important takeaway is that if you have a scale-up issue or if you don’t want to have one in the future, you should really consider doing CFD. I mean, there are other vendors out there, of course, as you don’t have to do it with us, but it’s really a tool that is, I think, underused and will be decisive if we are thinking about about competitiveness.

So if you’re a CDMO, in some months, years, the first CDMOs will pop up, virtual scale-up first and then a real scale-up afterwards. And you want to be one of them, right? And the same is true for pharma companies. You want to have the risk reduction and time reduction from doing a virtual process transfer first and then doing the real one. And you also want to have a digital reactor repository where all reactors are characterized and you can freely scale up between each and every So this is the future and the preparation starts now, and I think everyone should be on this currently.

David Brühlmann [00:19:21]:
Where can people get ahold of you and learn more about CFD?

Christian Witz [00:19:26]:
We have our website, which is www.simvantage.com. You can find me on LinkedIn. My name is Christian Witz. Otherwise, my email address, I think you will put it in the show notes if I can ask you to do that.

David Brühlmann [00:19:42]:
Fantastic. Thank you so much, Christian, for sharing your passion, explaining why we should use CFD. It has been a huge pleasure having you on the show today.

Christian Witz [00:19:52]:
Thank you for having me, David. It was great fun.

David Brühlmann [00:19:55]:
Simulation isn’t just an academic exercise anymore. It’s becoming part of how teams de-risk scale-up decisions and even how they talk to regulators. Christian Witz’s work, from GPU-based algorithms to an active FDA collaboration on machine learning and residence time distribution, shows how far CFD has come toward practical everyday use. Thank you for listening. If you enjoyed this conversation, please leave a review on Apple Podcasts or your favorite platform, and I’ll see you next time.