Sartorius Innovation Forum 2026

Answered by Franziska Jenner

The Ambr-integrated capacitance probe applies single-frequency measurement to monitor viable cell volume. We have not tried multi-frequency scanning in Ambr bioreactors yet but are evaluating this in larger scale.

Answered by Barry Heavey

Often I observe the KPIs are not defined very well in biopharma industry. I regularly see companies using simple KPIs like “right first time” in areas such as handover of a process from PD to commercial. Minimizing risk (regulatory and quality) always wins out over setting specific and aggressive operation targets for operational excellence. If PPQs are successful then they have achieved RFT, but they may have handed over a process with low titre, suboptimal process capability and the receiving site has many deviations and investigations to deal with in future and long cycle times. This was sometimes accepted because new product launches were not very frequent and a lot changes in the technology of manufacturing in between. So clients focus on reducing risk and not being on the critical path for launch, rather than setting aggressive targets and monitoring lots of KPIs for the next process. They then accept that they will have to invest in developing a 2nd generation process and in this case they will start to focus more on broader KPIs and aggressive targets. But this has to change as the “pace is accelerating” – more new drugs and biobetters coming so ideally clients develop processes that are “Great first time” not just “right first time” and “greatness” is clearly defined with operational excellence KPIs. Then these clients will be able to spend less time investing in 2nd,3rd and 4th generation processes

Answered by Barry Heavey

This is happening all the time in other industries. Think about the automotive industry and the French company Valeo – they develop and contract manufacture complex electronic componentry for car manufacturers (e.g. sensors). They also have to be involved with their clients in the design of exciting new cars, so they know what components will be needed. Highly IP sensitive work. Companies like Valeo manage their R&D work on their most important products in product lifecycle management (PLM) systems. Their clients do the same with their pipeline products (CAD files, specs and tolerances, bill of materials, assembly strategy). The two companies need to provide each other with limited, cloud based access to their PLM systems to ensure they are aligned in every way through the lifecycle from product ideation to prototyping to mass production. But they need to firewall so that Valeo don’t have their IP leaked to other vendors that the client works with and the client doesn’t have their designs leaked to the other clients Valeo work with.

I am not a cybersecurity expert but I take this as an inspiration that this is possible. I am sure it is not perfect, there have probably been “leaks” of data and IP in the automotive industry with this IT architecture. But there are always leaks, even without IT system integration. The pharma industry is just more slow moving and conservative than the automotive industry

Answered by Barry Heavey

I am a big fan of the concept of the “molecule steward” who is responsible for all aspects of the lifecycle of a molecule process optimization – DS, DP and analytical methods. They report on molecule KPIs and targets to senior management and work with all the multidisciplinary team who have contributed, are contributing or could contribute to performance (PD, Reg, MS&T, Ops, eng, QA, QC, supply chain, commercial).

If the molecule steward see a big problem or big opportunity to improve a process (we could intensify upstream, adopt continuous downstream and improve our release assats), they should be able to get the buy-in and resources from senior management to pull together and oversee the multidisciplinary team to make those improvements. They can provide regular updates to senior management to ensure the multidisciplinary team stays focused. But if another molecule steward makes a stronger case for their molecule, then the multidisciplinary resources might be pulled away towards that molecule.

Answered by Adrien Lugari

Cell therapy manufacturers require GMP-grade media with strong regulatory documentation, including CoAs and traceability. Serum-free, Xeno-free and Animal component-free media are highly preferred as they improve safety and consistency. Suppliers must also demonstrate the ability to scale globally with secure supply chains, redundant manufacturing capacity, and dual-sourcing strategies to minimize production risk with low lot-to-lot variations. As regulations around cell therapy is constantly evolving and may vary per country, recurring review of regulatory requirements is advised.

Answered by Sherwin Ting

1. Control pH via CO₂ stripping/gas-phase control (adjust sparge gas composition) rather than liquid base, since CO₂ diffuses more gently than a bolus of base.

2. If base is unavoidable, use diluted base with small dosing increments.

3. Consider Na₂CO₃ or NaHCO₃ instead of NaOH — gentler local pH shock.

4. Reduce the need for correction by limiting lactate production upstream (glucose-limited feeding), so less base is needed overall. Also remember that pH setpoint should be achieved before adjusting pCO₂ levels.

Answered by Sherwin Ting

Common acceptance criteria are: viable cell density and viability profiles trending within the historical/clinical batch range (often ±20%, or overlapping 95% confidence bands), matched growth rate and metabolite (glucose/lactate) trajectories, and comparable titer/quality attributes at harvest.

Statistical equivalence testing (e.g., multivariate/PCA comparison of time-course data) is increasingly used rather than single-point comparisons.

Answered by Sherwin Ting

Agree — this is essentially how large-scale perfusion is already done in practice, for both ATF and TFF. Rather than building one very large module (which worsens the axial pressure drop/Starling effect we discussed), the industry scales by numbering up smaller parallel modules.

It's feasible, but adds engineering complexity: flow balancing across parallel units, synchronization, and equal fouling/aging across modules over the run.

Answered by Sherwin Ting

I think this is a process development related question on recirculation timings. The extra length is only needed because conventional rTFF cycles slower than ATF, giving Starling recirculation time to foul before switching — it's an engineering/pump-speed limitation, not a fundamental physical requirement.

If you can match the recirculation times, a longer hollowfiber is not required. As for footprint, if we need a larger hollowfiber, we can also consider make them longer (taller). This way footprint is the same.

Answered by Aissa, Tarek

The main difference is that Native DeltaV STR and Biobrain STR integrated with DeltaV use two fundamentally different automation concepts.

Native DeltaV STR bioreactor automation is built directly on DeltaV technology. The bioreactor appears as a native DeltaV unit operation and can be fully customized by the customer within the DeltaV environment.

Biobrain STR integrated with DeltaV is controlled by the Sartorius Biobrain automation platform. DeltaV acts as the supervisory layer and communicates with Biobrain via OPC UA. DeltaV recipes can orchestrate the process, while Biobrain executes the equipment phases and control strategies.

Sartorius provides integration libraries, graphics, faceplates, and phase templates to simplify the integration effort and significantly reduce engineering time.

Native DeltaV STR is always an ETO project. Standard Offering will be Biobrain STR.

Answered by Vaibhav Patil

While the underlying mathematical modeling and validation rigor are indeed similar, the dramatic acceleration in soft sensor deployment comes directly from two operational bottlenecks: data acquisition strategy and physical infrastructure.

1. Leveraging Existing Time-Series Data:
Building a Raman or NIR model requires generating entirely new, highly specific spectral datasets. This involves designing and executing deliberate, time-consuming wet-lab experiments strictly to map varying analyte concentrations to spectral signatures.

In contrast, soft sensors rely solely on existing time-series process data. You can leverage the rich data from your routine process development (PD) and pilot-scale batches. You bypass the experimental data-generation phase entirely because the model utilizes the historical process variations you have already captured.

2. Non-Invasive, Hardware-Free Deployment:
Soft sensors function entirely non-invasively without requiring manual offline samples or physical process interventions. Deploying Raman involves procuring physical probes, engineering port installations, establishing sterilization protocols, and conducting specialized operator training.

A soft sensor is deployed purely at the software level, eliminating physical lead times, hardware maintenance, and capital expenditure.

Answered by Barry Heavey

Often I observe the KPIs are not defined very well in biopharma industry. I regularly see companies using simple KPIs like “right first time” in areas such as handover of a process from PD to commercial.

Minimizing risk (regulatory and quality) always wins out over setting specific and aggressive operation targets for operational excellence.

If PPQs are successful then they have achieved RFT, but they may have handed over a process with low titre, suboptimal process capability and the receiving site has many deviations and investigations to deal with in future and long cycle times.

This was sometimes accepted because new product launches were not very frequent and a lot changes in the technology of manufacturing in between. So clients focus on reducing risk and not being on the critical path for launch, rather than setting aggressive targets and monitoring lots of KPIs for the next process.

They then accept that they will have to invest in developing a 2nd generation process and in this case they will start to focus more on broader KPIs and aggressive targets.

But this has to change as the “pace is accelerating” – more new drugs and biobetters coming so ideally clients develop processes that are “Great first time” not just “right first time” and “greatness” is clearly defined with operational excellence KPIs.

Then these clients will be able to spend less time investing in 2nd, 3rd and 4th generation processes.

Answered by Barry Heavey

This is happening all the time in other industries. Think about the automotive industry and the French company Valeo – they develop and contract manufacture complex electronic componentry for car manufacturers (e.g. sensors).

They also have to be involved with their clients in the design of exciting new cars, so they know what components will be needed. Highly IP sensitive work.

Companies like Valeo manage their R&D work on their most important products in product lifecycle management (PLM) systems. Their clients do the same with their pipeline products (CAD files, specs and tolerances, bill of materials, assembly strategy).

The two companies need to provide each other with limited, cloud based access to their PLM systems to ensure they are aligned in every way through the lifecycle from product ideation to prototyping to mass production.

But they need to firewall so that Valeo don’t have their IP leaked to other vendors that the client works with and the client doesn’t have their designs leaked to the other clients Valeo work with.

I am not a cybersecurity expert but I take this as an inspiration that this is possible. I am sure it is not perfect, there have probably been “leaks” of data and IP in the automotive industry with this IT architecture. But there are always leaks, even without IT system integration.

The pharma industry is just more slow moving and conservative than the automotive industry.

Answered by Barry Heavey

I am a big fan of the concept of the “molecule steward” who is responsible for all aspects of the lifecycle of a molecule process optimization – DS, DP and analytical methods.

They report on molecule KPIs and targets to senior management and work with all the multidisciplinary team who have contributed, are contributing or could contribute to performance (PD, Reg, MS&T, Ops, eng, QA, QC, supply chain, commercial).

If the molecule steward see a big problem or big opportunity to improve a process (we could intensify upstream, adopt continuous downstream and improve our release assats), they should be able to get the buy-in and resources from senior management to pull together and oversee the multidisciplinary team to make those improvements.

They can provide regular updates to senior management to ensure the multidisciplinary team stays focused. But if another molecule steward makes a stronger case for their molecule, then the multidisciplinary resources might be pulled away towards that molecule.

Answered by Aissa, Tarek

This is definitely possible. Third Party Equipments can be connected to Biobrain Operate and also Mixed Mode is possible.

Answered by Adrien Lugari

Cell therapy manufacturers require GMP-grade media with strong regulatory documentation, including CoAs and traceability.

Serum-free, Xeno-free and Animal component-free media are highly preferred as they improve safety and consistency.

Suppliers must also demonstrate the ability to scale globally with secure supply chains, redundant manufacturing capacity, and dual-sourcing strategies to minimize production risk with low lot-to-lot variations.

As regulations around cell therapy is constantly evolving and may vary per country, recurring review of regulatory requirements is advised.

Answered by Aissa, Tarek

Yes. This is possible. Values can be ingested via OPC UA.

Currently we are developing a Model Execution Engine for Biobrain Control, that is able to execute models directly on the Instruments as well. Launch is planned beginning of next year.

Answered by Vaibhav Patil

Yes. We have tried several approaches beyond PLS/OPLS.

The important point is that the hybrid modelling approach is not tied to one particular machine-learning algorithm. What matters is whether we can develop a reliable relationship between the process inputs and the biological rate we want to represent.

For example, we have explored:
- Simple regression and multilinear regression
- DoE-based equations
- Machine-learning models
- Models developed using R or Python
- Neural-network based approaches

These can be used to estimate things such as:
- Specific product formation rate, qP
- Specific substrate consumption rate
- Specific metabolite production or uptake rate
- Other process-relevant biological rates

Once we have a suitable equation, it can potentially become part of the hybrid model, regardless of where that equation was developed.

This also gives us an interesting option. We can try several candidate equations and compare them rather than assuming that one modelling technique will always be the best.

In practice, we can build different hybrid model configurations and see which one gives the best combination of accuracy, robustness and practical usefulness.

Answered by Christopher Montoya

This is probably one of the most important parts of the exercise. It can look simple initially, but choosing the right level of model complexity is sometimes the tricky part.

I normally start with the key variables that we already understand to have a strong biological or process impact.

For example:
Glucose, Glutamine, Glutamate, Lactate, Ammonia, pH, Temperature, Dissolved oxygen and Osmolality.

We then look at how the model performs and gradually add other variables where there is a good scientific reason to do so.

Depending on the process, these could include:
• Additional metabolites or ions
• Gas-flow parameters
• Oxygen or CO₂ fractions
• Pressure-related parameters
• Other process variables

We can also create new features if there is a useful biological quantity that is not directly measured but can be estimated from the available data.

So the approach is usually iterative rather than trying to build the perfect model in the first attempt.

In our experience, it can take several iterations to arrive at a model that is sufficiently robust and useful for the intended application.

We also have hands-on training where users can go through this model-building process themselves, test different structures, understand where the model works or fails, and improve it step by step.

The ultimate goal is not to build the most complicated model. It is to build a fit-for-purpose digital representation of the process that we can actually use for simulation, scenario testing and optimisation.

Answered by Vaibhav Patil

For example, if a conventional study requires 20 runs, we could potentially investigate whether around 5 to 8 strategically selected experiments are enough when combined with existing process knowledge and historical data.

In some cases, we may be able to go even lower if we already have a strong mechanistic understanding and relevant historical data.

The key point for me is that we are not simply trying to reduce experiments. We are trying to make each experiment more informative.

Instead of asking only, “How many experiments do we need?”, we can ask:
“Which experiments will give us the maximum additional information and improve the model most effectively?”

That is where I see the real value. In suitable cases, a significant reduction in experimental workload may be possible, but I would always determine this based on the specific process and available data rather than treating 50% reduction as a guaranteed number.

Answered by Vaibhav Patil

Yes. We have also explored artificial neural networks.

In one application, an end user used a feed-forward neural network to estimate a target parameter such as qP. At a fundamental level, a neural network is still a mathematical function of the input variables.

For example, a simplified representation could look like:

qP = 0.7 × max(0, 0.5x₁ − 0.2x₂ + 0.1) − 0.4 × max(0, 0.1x₁ + 0.8x₂ − 0.5) + 0.3

This is only an illustrative example. The actual model had more variables and a more complex structure.

One interesting learning from our projects is that more advanced AI does not automatically mean a better model, especially when the number of batches is limited.

With a small dataset, simpler approaches such as multilinear regression or PLS/OPLS can sometimes perform better than neural networks. As the amount of historical data increases, however, neural networks and other advanced AI approaches become more interesting.

In one application with roughly 25 to 30 batches, the neural-network model performed better than some of the other machine-learning approaches we evaluated.

So I would look at it this way. We should use the most appropriate model for the data we have, rather than simply the most sophisticated algorithm available.

And as the data volume grows, there is definitely scope to bring more advanced AI approaches into the hybrid modelling framework.

Answered by Vaibhav Patil

This is probably one of the more challenging questions, and I would not expect one model to automatically explain every media formulation.

In one of our projects, we looked at six different media formulations. We could not develop one common model that adequately represented all six media.

Instead, we developed separate digital twins for each medium, and these worked well within their respective domains.

That still gives us a lot of value.

Once we have these models, we can virtually test different:
- Process conditions
- Feeding strategies
- Intensification strategies
- Operating conditions

across the different media before deciding which experiments are actually worth running.

So instead of experimentally testing every possible combination, we can first use the digital twins to screen a much larger virtual space and then take the most promising conditions into the lab.

Looking ahead, it would certainly be interesting to investigate a more generalized model that incorporates media composition directly into the model. That would require enough data across different media and a good understanding of how media changes affect cellular behaviour.

So today, media-specific digital twins are already a practical approach. With more data, there is an interesting opportunity to move toward more generalized models.

Answered by Vaibhav Patil

No. qP was mainly used in the presentation because it is an easy way to explain the concept within a limited amount of time.

The same approach can be extended to several other biological rates.

For example:
• Specific product formation rate
• Specific substrate consumption rates
• Specific metabolite uptake rates
• Specific byproduct formation rates
• Oxygen uptake related rates

These data-driven relationships can then be integrated with the mechanistic part of the model.

For example, substrate concentration and process conditions can influence consumption rates. Those rates then influence metabolite balances and cellular behaviour. Similarly, product formation and byproduct generation can be linked to the evolving process state.

So we can build a multi-rate hybrid model, rather than making only qP data-driven.

There is, of course, a practical boundary to how deeply we want to model the biology.

The current approach is mainly focused on the process and cellular-behaviour level, rather than trying to reproduce every intracellular mechanism such as individual enzymatic reactions, detailed TCA-cycle kinetics or complete intracellular metabolomics.

That level of detail may be useful for some scientific questions, but it is not always necessary for getting practical value from a digital twin.

For me, that is actually one of the strengths of the approach. We can combine what we already know mechanistically with data-driven relationships where the data can add value, and build the model to the level of detail that is appropriate for the business or process question we are trying to answer.

This version should feel more natural coming from you in a follow-up with a scientist or process-development person. It also leaves several natural openings for the next conversation, particularly around their data, their specific modelling question, and what could be tested virtually before committing to experiments.

Answered by 홍종광 교수

네, 평균값을 입력하는 경우와 CFD에서 얻은 공간적 분포를 이용하는 경우에는 분명 차이가 있을 수 있습니다. 특히 세포의 대사 반응이 DO, glucose, pH 등의 평균적인 환경에서 계산한 세포 반응과 서로 다른 local environment에서 계산한 반응의 평균은 같지 않을 수 있습니다.

다만 여기서 한 단계 더 고려해야 할 부분이 있습니다. 실제 바이오리액터에서는 세포가 특정 mesh나 위치에 고정되어 있는 것이 아니라 유동을 따라 계속 이동하고 있고, 혼합이 충분히 잘 이루어지는 바이오리액터라면 한 세포가 지속적으로 나쁜 환경에 머무르기보다는 서로 다른 영역을 반복적으로 이동하게 됩니다. 따라서 단순히 “바이오리액터 내부에 얼마나 나쁜 local environment가 존재하는가”만 보는 것보다, 세포가 그 영역에 얼마나 자주 들어가는지, 얼마나 오래 머무르는지, 그리고 다시 정상적인 환경으로 돌아오기까지 얼마나 걸리는지가 더 중요할 수 있습니다.

예를 들어 평균 DO가 동일한 두 bioreactor가 있다고 하더라도, 하나는 대부분의 영역에서 DO가 균일한 반면 다른 하나는 매우 낮은 DO 영역이 존재할 수 있습니다. 두 번째 reactor에서도 세포가 저산소 영역을 매우 짧게 통과한다면 biological effect가 크지 않을 수 있지만, circulation이나 mixing이 느려 저산소 영역에 오래 머물거나 반복적으로 노출된다면 metabolism이나 growth에 상당한 영향을 줄 수 있습니다.

따라서 CFD와 cell mechanistic model을 통합할 때는 단순히 average value vs. local distribution 만 비교하기보다는, local distribution + cell trajectory + exposure time/history를 고려하는 것이 더 mechanistic한 접근이라고 생각합니다. CFD의 Eulerian–Lagrangian tracking을 활용할 수 있습니다.

Answered by 홍종광 교수

디지털 트윈이나 스마트팩토리 구축의 핵심은 공정에서 발생하는 데이터를 잘 모으고, 그 데이터를 기반으로 예측모델을 구축한 다음, 예측 결과를 실제 공정 제어에 활용하는 것이라고 생각하면 될 것 같습니다.

예를 들어 바이오리액터에서는 DO, pH, 온도, agitation, gas flow 같은 데이터가 계속 발생하고, cell density, glucose, lactate 같은 배양 데이터가 있고, 추가적으로 많은 영양물질, 또는 세포가 배출하는 대사물질의 데이터를 가능하면 실시간으로 확보할 수 있습니다. 먼저 이런 데이터를 하나의 batch와 시간축으로 연결해서 지속적으로 축적하는 것이 필요하고, 그 데이터들 사이의 모델을 만드는 것이 필요합니다. 그것이 예측모델이고, 이를 만들기 위하여 확보한 데이터베이스를 학습하는 AI 모델 또는 machine-learning model도 사용할 수 있고, 세포 성장이나 대사, mass balance처럼 우리가 알고 있는 현상을 수식으로 표현한 mechanistic model도 사용할 수 있습니다. 최근에는 두 가지를 결합한 hybrid modeling도 많이 연구되고 있습니다. AI만으로 모델을 만들려면 일반적으로 많은 데이터가 필요한데, bioprocess에서는 실제 생산 데이터를 충분히 확보하기 어려운 경우가 많습니다. 이때 기존의 물리적·생물학적 지식을 mechanistic model로 활용하고 데이터 기반 모델을 결합하면, 제한된 데이터에서도 예측 성능과 모델의 활용성을 높일 가능성이 있습니다.

CFD도 이런 mechanistic information을 제공하는 하나의 방법이 될 수 있습니다. CFD로 mixing, oxygen transfer, shear stress와 같은 reactor 내부의 물리적 환경을 계산하고, 이를 cell growth나 metabolism model, 그리고 실제 공정 데이터와 결합하면 보다 통합적인 예측모델로 발전시킬 수 있습니다.

그리고 궁극적으로 중요한 것은 예측 자체가 아니라 예측을 이용한 사전 제어입니다. 예를 들어 현재 데이터를 이용해서 몇 시간 뒤 DO demand가 증가하거나 glucose가 부족해질 것으로 예측한다면, 문제가 실제로 발생한 뒤 대응하는 것이 아니라 agitation, gas flow 또는 feeding을 미리 조정할 수 있습니다.

그래서 스마트팩토리의 발전 단계를 단순화시켜보면, Data collection and integration → Prediction → Proactive control로 볼 수 있습니다. 소규모 single-use 공장이라면 처음부터 완전 자동화를 목표로 하기보다는, 우선 장비와 분석기기에서 나오는 데이터를 자동으로 통합하고, 그 위에 AI와 mechanistic model을 결합한 예측모델을 구축한 뒤, 충분히 검증되면 사전 제어까지 단계적으로 연결하는 것이 현실적인 접근이라고 생각합니다.

Answered by 홍종광 교수

현재 구상하고 있는 SDL은 우선 바이오신약 개발 과정에서 후보물질을 대규모로 빠르게 스크리닝하기 위한 스크리닝 SDL (Self-Driving Lab)입니다. 단순히 장비를 자동화하는 것이 아니라, 실험부터 분석, 데이터 해석, 다음 실험 결정까지 사람의 개입을 최소화하여 연결하는 것이 목표입니다.

구성을 보면, 24-well plate 기반의 세포배양과 transfection, 세포배양 및 배양성능 평가, 단백질 정제, 품질분석까지 자동으로 연결하는 workflow입니다. 여기서 얻은 데이터를 바탕으로 우수한 후보를 선정하고, 해당 후보가 확대배양이나 다음 평가 단계로 이어지는 closed-loop system을 구축하고 있습니다.

두 번째는 bioreactor 기반의 자율 공정개발 workflow입니다. 배양 중 사람 개입없이 autosampling을 수행하고, 세포 및 대사체 데이터와 pH, DO 등의 공정 데이터를 자동으로 수집한 뒤 단백질 정제와 품질분석까지 연결합니다. 그리고 한 라운드의 배양 결과를 바탕으로 다음 라운드의 배양조건을 결정하고 다시 실험하는 closed loop를 구축하는 방향입니다.

결국 SDL의 핵심은 개별 장비 자체보다는 실험–분석–데이터 수집–의사결정–다음 실험을 하나의 자동화된 loop로 연결하는 것이라고 생각합니다. 몇 개의 SDL제조사에서 제안하는 기기들이 있으나, 사용자 필요에 따라 장비 추가가 가능합니다.