Podcast Transcript

David Brühlmann [00:00:32]:
Should you use computational fluid dynamics for your scale-up? Or is it an overkill for a team like yours? It’s a question a lot of CMC leaders sit with. Simulation sounds powerful, but also complicated, expensive, and perhaps out of reach. Christian Witz built his career answering exactly that question. From academic GPU algorithms to founding SimVantage, he has made CFD practical for everyday bioprocess teams. Let’s find out if it’s right for yours.

Welcome, Christian. It’s good to have you on today.

Christian Witz [00:02:26]:
Thank you, David, for having me. It’s great to be here.

David Brühlmann [00:02:30]:
A huge pleasure. Share something that you believe about bioprocess engineering that most people disagree with.

Christian Witz [00:02:39]:
I don’t know about the disagree part, but what always my saying is that a CFD is needed by everyone. It’s a tool which you should have available like Excel, which is also complicated by the way. But between you and the perfect CFD is only the user interface, which is often not tailored to bioprocesses. So I don’t know about the disagree part. I’m sure that most people see CFD as something special. But I think that CFD should be a tool in your toolkit like Excel, and you should be able to use it every day for your process development, for your transfers, and so on.

It’s only the user interface which makes it complicated, and this is what we have done in the last years to just tailor the user interface for CFD to bioprocesses, and therefore everyone can use it, I’m sure.

David Brühlmann [00:03:30]:
I’m excited about our conversation, Christian, because I believe CFD is important. We’re gonna spend more time on it, why that is important, and also unpack the whole thing. Let’s first start about yourself. I’d be curious, tell us what drew you into CFD and then eventually into biotech, and what were some interesting pit stops along the way?

Christian Witz [00:03:54]:
So my journey didn’t start in biotech. I studied in Upper Austria in a small Austrian city, and it was a university which was mostly concerned about natural resources, iron and steel, wastewater treatment. So my initial contact with simulation was in the context of heat pump. So the thing that is now in everyone’s mind about using electricity to heat your house. There was a new kind of heat pump. My task in my bachelor thesis actually was to simulate the process in this newly developed heat pump and then to make a control software for it. And this went well until the program was finished and I showed it to my professor and the professor noticed that the inventor of the heat pump just forgot a contribution of kind of energy that is needed to operate the heat pump. And then suddenly the efficiency dropped to levels that were nothing new anymore. And so in the end, the concept was scrapped.

And then I did my master’s thesis on falling particles that are affected by airflow and how they interact during this fall. And therefore also separating larger from smaller particles because of the different drag. Then during my master thesis, I was sure that I wanted to continue in the simulation field.

And then an opportunity here in Graz opened up and the task was to transfer an existing program that was used for bioreactor simulation to the graphic card, which was very new at the time that you can calculate something with the graphic card. So this was my PhD thesis or the task of my PhD thesis. And I transferred the program, which was written in Fortran, so like an ancient programming language, to the modern C++ and CUDA, which you then can interact with the graphic card.

And then when we did all this, we were a team with a postdoc and me. When we did all this, we noticed that the program doesn’t really work. So there were some shortcuts, there were some bugs in it, and so we had basically not to start from scratch, but remove large parts of the program and then start with simple test cases like an empty bucket of water. In the beginning, it wasn’t even able to simulate the resting fluid in an empty bucket. And then I slowly stabilized it there and then added one stirrer blade and built it up from that and then added bubbles, added substances like dissolved oxygen. And then Philipp joined, which is the co-founder of SimVantage. And I always joke that now he has to remove all the bugs that I introduced in this program, and he still does it today.

Finally, I ended up in the biotech field, and I have to be honest, I was quite astonished because in the iron and steel industry, in the oil and gas industry, it’s standard that you fully characterize your equipment. So this is like the cash cows in your company, the steam cracker and all of that stuff. And it’s fully characterized in terms of simulation so that you have the last bit of efficiency, you get it out of this device.

And then when I started in the biotech biopharma industry, I was quickly astonished that this is really not existing. I mean, partly because it’s a very complicated simulation case. So you have this multiphase flow with a phase being the air bubbles and one phase being the liquids. And this is a challenging simulation task. Then you have all these substances included. And then in the end, of course, biology. This is also a hard thing to simulate and we cannot fully do it until now with first principles. Maybe we’ll come to that later.

And so my task basically was to also introduce these methods into the biopharma field and the biotech field. And I have to be honest, I had a hard time. I always said that it seems, so my impression, maybe wrong, was that people here in this field are not used to work interdisciplinary. So they have kind of their own language, lots and lots of abbreviations.

In the beginning, I didn’t know what USP and DSP was. And I can remember when I asked the question to a colleague, can you tell me what this CHO is? She said, Chinese hamster ovaries. And I said, no. Really, tell me, I need to know what this is. And he said, yes, Chinese. Okay. It was a steep learning curve, I have to be honest.

David Brühlmann [00:08:32]:
Oh, I imagine you came from a completely different field of the industry and then moving into biotech. That’s definitely a challenge. I’m curious, Christian, you did a lot of work in academia before you decided to spin out your company. And what was this trigger and why did you think you should focus on the bioprocessing industry?

Christian Witz [00:08:53]:
So the focus was just because the method which we are using to do these simulations is the lattice Boltzmann method. And you don’t need to get into the details, but let me say that the method is good because it fits very well to the graphic cards. And when an algorithm and the hardware fits well together, then it gets very fast. And for this simulation method, it’s very easy to simulate devices which have like a rotating part included. And it’s also easy then to introduce other algorithms like for the bubbles and the substances, just as I said before. And therefore this was a perfect niche. So there were not a lot of activities going on at the time. It was a hard simulation problem and our algorithm fit very well to these problems. So these were the reasons why we got into this field.

And the reason why we created a spinoff was that at least in Austrian university, if you are not a professor, after some time you have to leave university. You don’t get like an extended contract or an unlimited contract. And then the whole working group for bioprocess simulation would have been, I don’t know, dissolved or the people there would have a hard time if I would have left at this point in time. They were doing their master’s thesis, bachelor’s thesis, partly PhD thesis. And it was like, I don’t want to leave all these people alone.

In the beginning, I wasn’t really sure if we can do that. So we had industry partners that were Novartis Sandoz at the time when they were still together, one company. And I always said, okay, I don’t know if they will support this. Actually, they were very supportive. So they became actually our first customer. They also supported us in giving us the freedom to use this program for the foundation of a company. And also the Technical University of Graz was very supportive because they also had the task to create more spin-offs.

So in Austria, there is a huge discussion that we have a lot of technology, a lot is developed, but oftentimes these technologies don’t make it into a real company or a real use case where you can earn money with. And so it was a good time. So we got the very good conditions to buy this technology from university. We had the need. And also there was this program, the Spin-off Fellowship of the Austrian Research Promotion Agency. I think they copied it from ETH Zurich, if I’m not mistaken. And this was a great program. There you’re supported with the foundation of your company. You get these boot camps where people from tax advisors and business developers, other founders, they speak with you, they give you tips and tricks to found a company.

And then in the end, so you cannot found the company during this program, which is 1.5 years. In the end, you’re really well prepared to do your first steps and to set it up and succeed.

David Brühlmann [00:11:53]:
Let’s address the elephant in the room. Do people in biotechnology or bioprocessing rarely use CFD?

Christian Witz [00:11:58]:
As I said, it’s because it has this aura of it’s complicated. You need a dedicated person to do it. In the best case, this person has like a physics PhD or a mathematics PhD that they know all these differential equations. I mean, if you think about it, you wouldn’t like hire an IT technician in order to use a computer at your company. company, right? Everyone can use it. And then, and the same should be true for CFD as well.

Of course, what I don’t leave out is that there are and there were some companies which like took advantage that this CFD was not really much used in this field, and they took advantage of that by providing low-quality work in terms of CFD. And this is something I heard firsthand from people from a company that they really had a bad experience and then they stopped it for some years.

And all these together, I think, are the reasons why it is not well used, together with the fact that it’s really a hard problem. So this is more like scientific.

You can tackle this hard problem with simplifications. So imagine you have your bioreactor, and please stop me, I can talk for hours about this. Imagine you have your bioreactor, which is like a huge cylinder. And one of the things that people would do at the time, because they couldn’t simulate the whole cylinder because they didn’t have the calculation power, they would cut out cake slices out of this cylinder and then say, okay, the boundaries or the cut surfaces of this cake are periodic. So if you add a lot of cake pieces around, you get the full cake, right?

But this is really not a way to tackle the flow inside a bioreactor because this flow inside a bioreactor has huge 3D flows that are wrapping around the whole reactor. And you do not really get these huge 3D flows when you take this cake piece approach. You could use simplifications that then led to wrong results, which in turn produced bad outcomes for the use of CFD initially. And I think this is one of the reasons why it had really a hard time in the beginning getting established.

David Brühlmann [00:14:12]:
Yeah. So, no doubt about that. It’s a difficult problem to solve. You need computing power. My question is then, how are you solving this problem? How are you making sure that people can use that every single day? Because what I’m hearing is and you’re making a case for this, this is such an important tool, you should be using it frequently. So how are you gonna make this happen?

Christian Witz [00:14:35]:
Two things. First of all, we have calculation power, which is of course unmatched in history, available at our fingertips. So when we started our software, we rolled it out in the cloud and we had like Google Cloud in the background and we can have hundreds of GPUs from Google within a second. And therefore the individual GPUs are so fast that one GPU is enough, but you can imagine that you can do a whole series of different cases like different stirring speed, different gas flow rates, different viscosities, all within the time which one simulation takes because you can do it in parallel. This is one factor.

So everyone has this available. There are GPU clusters within larger companies and also smaller companies can just buy a gaming PC with a GPU and do calculation on this, or they just buy the computing resources from the cloud.

And the second thing is also the algorithms have developed. So you can imagine when I started, you didn’t really have an algorithm which fit to our base algorithm, this lattice Boltzmann algorithm, to transport substances around. So if you think about dissolved oxygen, which dissolves from your bubbles, in the beginning we didn’t have a technique to transport this oxygen around in the reactor. And this got developed, this got published and therefore now we have this full set which you need in order to describe all the processes inside a reactor.

And all this together also bringing in new ways of designing user interfaces. So nowadays with LLMs, you can beautifully design user interfaces without the need to be an interface designer. And you can quickly iterate them. You hand it over to friendly customer or to academic partner, get feedback on the same day because you handed it over in the cloud. And then you change user interface again until you get a good fit. And all this together really makes rapid progress possible.

David Brühlmann [00:16:39]:
When people develop a process and then scale up, usually they start in a spin tube or in a shake flask, then they go into an Ambr 250, perhaps later a pilot run, and eventually to the large-scale bioreactor. So how do you go about this? Do you characterize each reactor individually? Can you transpose or transfer some learnings? What is your strategy?

Christian Witz [00:17:04]:
So this depends if you characterize it fully. It depends really on the fact if you’re doing it once because you’re like a startup which has one new drug candidate and you want to scale it up, then it makes sense to do like a consulting project with us or get a short license to our software and do it once and find the operating points that you want for the target reactors where you transfer it to.

The second thing would be if you’re like a company which has their in-house reactors, or if you are a CDMO especially where you bring in a lot of processes where you know this scale-up process will be followed a lot of time with different densities, with different viscosities, with different operating conditions, then it makes sense to characterize every reactor or every device like the shake flask or the spin tube individually for the whole operating spectrum. Usually do 100 to 200 simulations for each of the devices, for each of the Ambr you’re using, or for each of your manufacturing reactors.

And then you can transfer every point to every reactor. So then you get total freedom. You can really say, okay, I have this process in the shake flask, how do I transfer it into the Ambr? And within seconds, because you have done all the work beforehand, within seconds it tells you, okay, set this stirring rate, set this aeration conditions in the Ambr, and then you get the same conditions, your cells will feel the same, and then from the physical point of view, you get this perfect scale-up. And you can then do the same from Ambr to pilot, from pilot to manufacturing. This is how I would approach that.

David Brühlmann [00:18:41]:
And do you need to redo that with every new molecule, or could you basically rely on this initial characterization and then just say, hey, this is plug and play?

Christian Witz [00:18:54]:
Yeah, basically what changes, like you have a different media composition, you have a different cell type, you’re using now E. coli or— so you can change a lot. And for us, it only changes density, viscosity, surface tension, salt concentration inside your fermentation broth or cell culture media. And these parameters are then handed to the scale-up software and it does then the scale-up with these parameters.

So as we’re on the physics side and not on the biological side, for us it makes no difference. And if we have characterized the reactor fully, so also including, for example, the viscosity and the density changes, then you can simply have every type of scale-up that you can imagine.

David Brühlmann [00:19:39]:
What are some important choices people need to make as they’re going about down this path to characterize bioreactors? Would you advise them?

Christian Witz [00:19:49]:
So, one limitation that we have is that it only makes sense to characterize the reactor with a certain amount of CFD simulations. You can imagine if you have the stirring speed in a certain range, the aeration rate in a certain range, and every other parameter you add, like the volume and the viscosity and so on, it exponentially increases the number of simulations you need. Because for every parameter you add a whole new volume, a whole new field of new points in this space where you can operate the reactor.

And therefore it’s often a discussion if you really need like all the viscosities that you can imagine from water to honey, or is it like a few discrete steps that you want to have included? Or if you think about reactor volume, for example, is it really necessary to have all from empty to maximum full? to include that, or can you like limit it to a sensible range?

And this is basically it. But if you say, okay, I want to have the full characterization, we can also do that.

David Brühlmann [00:20:54]:
That’s where we’ll pause for now. Christian Witz has already touched on the reasoning behind CFD’s slow adoption, and there’s more ahead when we continue. If this conversation is sparking new questions about your scale-up strategy, consider leaving a review on Apple Podcasts or wherever you’re listening from. Thank you for tuning in today, and I’ll see you in part 2.