---
title: "A bioprocess digital twin needs live data"
slug: bioprocess-digital-twin-live-data
date: 2026-11-09
author: "Invert Team"
category: industry
summary: "Will Olughu explains why a bioprocess digital twin needs live data, how to plan the data architecture, and what sponsors should agree with CDMOs."
url: https://invertbio.com/blog/bioprocess-digital-twin-live-data
---

# A bioprocess digital twin needs live data

**Invert Team** · November 9, 2026 · Industry

_Will Olughu explains why a bioprocess digital twin needs live data, how to plan the data architecture, and what sponsors should agree with CDMOs._

<style>
.blog-post-body figure.twin-fig{margin:30px 0 34px;}
.blog-post-body figure.twin-fig-lead{margin-top:0;}
.blog-post-body figure.twin-fig iframe{display:block;width:100%;border:0;background:transparent;}
.blog-post-body figure.twin-fig .h-stack{height:385px;}
.blog-post-body figure.twin-fig .h-rail{height:380px;}
.blog-post-body figure.twin-fig .h-alias{height:340px;}
.blog-post-body .twin-takeaways{background:var(--sage);border:0.5px solid var(--border);border-radius:10px;padding:22px 26px;margin:30px 0;}
.blog-post-body .twin-takeaways ul{margin-bottom:0;}
@media(max-width:640px){
.blog-post-body figure.twin-fig .h-stack{height:590px;}
.blog-post-body figure.twin-fig .h-rail{height:340px;}
.blog-post-body figure.twin-fig .h-alias{height:305px;}
}
</style>

<figure class="twin-fig twin-fig-lead">
<iframe class="h-rail" src="/demos/bioprocess-digital-twin-live-data-figures.html?fig=rail" title="A model on run day 12, fed by manual exports and then by a live data path" loading="eager" scrolling="no"></iframe>
</figure>

A process model can help you plan a run or understand a completed batch. Using it to recommend a change while the process is running requires a connection to current process data.

That connection is central to how [Will Olughu](https://www.linkedin.com/in/will-olughu-882b3752/), then global CMC lead at Ipsen, defines a digital twin. On Bioprocess Online’s *Better Biopharma* podcast, he put it this way:

> "You can't say something is a twin if it's not ingesting the same data."

For teams building a twin, that means planning how data will reach the model before building it. For sponsors that outsource manufacturing, it also means agreeing with the CDMO on access to that data and ownership of the model.

## Digital twin vs. process model or simulation: what is the difference?

Olughu describes a digital twin as a virtual copy of a physical process. In CMC work, that usually means a set of equations describing the process, with one requirement: the twin "has to be online." It receives data from the process while the run continues.

He illustrates this with a bioreactor feeding pH and dissolved oxygen data into a hybrid model. In his example, the model detects a pump fluctuation and predicts an effect on the glycosylation profile. It then tells the team to compensate at day 12, before the product goes out of specification.

Without a connection to the run, the model cannot respond to that fluctuation as it happens. In Olughu’s words, a model that is not online "is static." He describes it as "almost backward looking."

Offline models still have value. His teams use them for scale-up, tech transfer to CDMOs, and in silico process characterization. He calls each one a "potential digital twin." But he says very few teams, "if there's any," use a true digital twin today.

Between an offline model and a twin, there is also the digital shadow. A common way to tell them apart is how data moves between the process and the model:

- **Digital model.** No automatic data flow. Someone moves the data in and out by hand.
- **Digital shadow.** Process data flows into the model automatically, but the model's output does not flow back to the process.
- **Digital twin.** Data flows both ways. The model receives live process data, and its recommendations go back to the process.

Olughu's glycosylation example has both directions: live data reaches the model, and the model tells the team when to compensate. A shadow is also a useful step toward a twin. In the shadow mode described below, the twin runs beside the process, and no one acts on its recommendations yet.

## Why bioprocess models can remain disconnected

A model may need data from several unit operations, each with its own software. Olughu describes bioreactor data in one system and chromatography data in another, with no connection between them.

When systems cannot share data, someone has to export the files and prepare them for the model. This [fragmentation of bioprocess data](/blog/why-bioprocess-data-fragmentation-is-slowing-down-the-industry) leaves the model dependent on when that work gets done.

Until the next update, the model is working with an older picture of the process. If the data arrives after the team needs to act, it is too late to inform that decision.

Separate systems also limit how much of the process a twin can cover. A bioreactor twin needs data connections to downstream operations to extend beyond the bioreactor. Olughu identifies these unit-operation silos as the biggest barrier to twins in GMP.

## Map the data architecture before you build the model

Asked where to start, Olughu put data architecture ahead of the hybrid model and virtual process characterization. His advice is to understand how data will reach the model before you build it.

He describes a path from equipment controllers through supervisory control and data acquisition (SCADA), a historian, and a data lake to the twin. Mapping that path shows which systems need to connect and where delays could occur.

<figure class="twin-fig">
<iframe class="h-stack" src="/demos/bioprocess-digital-twin-live-data-figures.html?fig=stack" title="The data path to a digital twin as five layers, from equipment and control to the twin" loading="lazy" scrolling="no"></iframe>
</figure>

Each piece of equipment added to the process needs a connection to that path. Standards such as [OPC UA](/blog/what-is-opc-ua-bioreactors) let bioreactors publish process data as the run happens, but the data still has to reach the model in time to be useful.

The decision you want the twin to support determines how much delay you can tolerate. That is why planning the data architecture belongs at the start: it establishes whether the model will have current data when the team needs it.

This is the part of the work Invert takes on. [Invert](/product/integrations) brings data from historians, bioreactors, and file exports into one workspace. It maps each source's names to shared definitions and records where every value came from. The model team can then start from consistent data, not from a new round of exports.

## Why sponsors and CDMOs should care now

A CMC leader does not need a working GMP twin to start making these decisions. Data access and model ownership are easier to address when planning the work with a CDMO, before the architecture and contract are settled.

**For sponsors, the CDMO usually controls the systems supplying the data.** That makes [integrating bioprocess data across sites, systems, and CDMOs](/blog/how-to-integrate-bioprocess-data-across-sites-systems-and-cdmos-invert-bioprocess-ai) a question of access as well as technical connections.

Olughu estimates that "nine out of ten times" a sponsor uses the CDMO’s modeling platform. The sponsor’s data helps improve the CDMO’s algorithm, yet he says MSAs usually do not settle ownership. If the sponsor changes CDMOs, it may be unable to take the model and what it learned along with the program. His advice is to agree on data and model ownership before signing the MSA.

**For [CDMOs](/solutions/cdmo), data access affects how sponsors can use the model.** Olughu says he would rather run one engineering batch than three. Modeling is part of that ambition, and he expects sponsors to ask more questions about who owns the digital process as AI use grows.

Agreeing how sponsors will receive process data gives both sides a basis for that work. If the goal is to support decisions during a run, they need to agree on the format and timing of access.

**A shared vocabulary makes data usable across sites and scales.** Olughu calls for an open bioprocess data standard: "a pH code should be the same" in a 50-liter and a 2,000-liter bioreactor.

When sites and vendors name the same parameter differently, teams must resolve those differences during [tech transfer](/blog/bioprocess-tech-transfer-navigating-the-data-dilemma). [Harmonizing bioprocess metrics across instruments and runs](/blog/harmonizing-bioprocess-metrics-across-instruments-and-runs) gives them a shared definition to work from.

<figure class="twin-fig">
<iframe class="h-alias" src="/demos/bioprocess-digital-twin-live-data-figures.html?fig=alias" title="Column names from different scales and sites mapped to one definition" loading="lazy" scrolling="no"></iframe>
</figure>

Olughu estimates that standard, connected data and a digital process could shorten tech transfer from nine months to nine weeks. He presents this as a possibility, not a measured result.

## What comes after the data path

Once teams understand how data will reach the model, Olughu describes the next steps:

1. **Hybrid model.** Combine mechanistic models of physical behavior, such as mass transfer and fluid dynamics, with machine learning to reduce error in less fully understood cell biology.
2. **Virtual process characterization.** Use in silico experiments to identify critical process parameters and edges of failure, then confirm the model through lab work.
3. **Shadow mode.** Run the twin alongside the GMP process without using it to make decisions. After each batch, review whether its predictions were correct and explainable.
4. **Human in the loop.** In commercial manufacturing, have a qualified person review the twin’s recommendation and decide whether to act.

Shadow mode shows why the data architecture matters even before a twin informs a decision. To assess its predictions, the team needs to compare them with what happened in the batch and align the two in time. Sponsors and CDMOs should agree on access to the data needed for that review.

## Start where it is practical

Olughu describes today’s "pockets of excellence" as single unit operations early in process characterization, outside GMP. Teams can learn from a twin connected to a bioreactor there with fewer validation and change-control complications.

He also cautions that "you cannot simulate your way out of a bad process". Teams need to address problems in the process itself and decide what level of model performance is good enough for their current stage.

For a sponsor considering a twin, a practical starting point is to map how data from the CDMO’s systems would reach the model. That gives the team concrete access requirements to discuss before signing the MSA or investing in the model.

<div class="twin-takeaways">

**Key takeaways**

- Under Olughu’s definition, a process model becomes a digital twin when it receives live data from the running process.
- Separate systems can leave a model waiting for someone to export and prepare the data it needs.
- Map how data will move from equipment to the model, including whether it will arrive in time for the intended decision.
- Agree with the CDMO on data access and model ownership before signing the MSA.
- Shared definitions for process parameters reduce the translation work needed for modeling and tech transfer.

</div>

## FAQ

**What is a digital twin in bioprocessing?**
Olughu defines a digital twin as a virtual copy of a bioprocess that receives live data from the physical process. With current data, it can update its predictions and recommend adjustments while the run continues.

**What is the difference between a digital twin and a process model?**
A process model can simulate and extrapolate offline. Under Olughu’s definition, it becomes a digital twin when connected to the live process. He calls an offline model a "potential digital twin."

**What is the difference between a digital shadow and a digital twin?**
A digital shadow receives live data from the process, but its output does not go back to the process. A digital twin also sends its predictions or recommendations back, so the team can act on them during the run.

**What do you need before you build a bioprocess digital twin?**
Map how process and analytical data will reach the model, and how quickly it needs to arrive to support the intended decisions. If a CDMO runs the process, agree on data access and ownership before signing the MSA.

**Who owns a digital twin built at a CDMO?**
Olughu says MSAs often leave ownership unresolved. He recommends that sponsors agree on ownership of the data and any model trained on it before signing.

---

*Listen to the full conversation: [Implementing AI and Digital Twins in Bioprocessing with Will Olughu, Ph.D.](https://www.bioprocessonline.com/doc/implementing-ai-and-digital-twins-in-bioprocessing-with-will-olughu-ph-d-0001) on Bioprocess Online’s Better Biopharma podcast.*

*Invert brings data from historians, bioreactors, and file exports into a shared workspace. For sponsors working with CDMO exports, it maps values to shared Library definitions and keeps track of each value’s source. [See how Invert brings data together.](/product/integrations)*