I recently visited two customers (bioprocess) consecutively for AI/ML-related demos, and interestingly, both of them raised the same question. Both customers are currently exploring digital twin approaches — specifically, generating simulation data to supplement limited experimental datasets.
1. Have you seen any cases — or do you have experience — generating simulation data generated from a mechanistic model ?
For some context, a mechanistic model is a mathematical representation of a process based on fundamental physical, chemical, and biological principles — rather than empirical data. It allows us to generate simulated process data purely from first principles, without running actual experiments.
In bioprocess purification, typical examples include:
- Chromatography → van Deemter equation, mass transfer equations
- Ultrafiltration/Diafiltration (UF/DF) → Darcy's law, concentration polarization models
Any insight would be helpful for me.
On my end, when customers asked about building mechanistic models, my initial suggestion was to implement them in Python and integrate Python with JMP. but i am wondering if there are more deeper ways to help them.
2.
Since real experimental data is limited, the plan is to generate a large dataset using a mechanistic model. However, this simulation data will inevitably carry some error compared to real experiments.
What are the recommended approaches to correct or compensate for this error — so that the AI model trained on this data is still reliable?
Any known methods or case studies in JMP would be appreciated!
Hi @HyeWonJung
Assistant is excellent at simulating data with JSL. I've been pleasantly surprised across a wide range of disciplines.
Suggest setting the Perspective to something like Chemical Engineer then giving it prompts requesting simulation of the first principle models like you have described.
@russ_wolfinger
Thanks for the quick response. I wasnt' aware that the Assistant’s simulation capability could extend to something like building virtual labs—this is a very compelling direction for us. I’ll share this with the team and follow up, as it could become a strong topic for our discovery discussions this year. We’re also seeing leading biopharma companies starting to explore concepts like virtual bioreactors.
I had a follow-up question from a practical standpoint.
In many real-world scenarios, customers don’t have sufficient experimental data, so they tend to rely on mechanistic models to generate simulated data. However, there is usually a noticeable gap between simulated outputs and real experimental observations.
What is the most effective way to bridge this gap when building a realistic virtual lab?
I am wondering if there are actual cases or how companies actually bridge the gap between mechanistic simulation and real-world variability when building something like a virtual bioreactor.
Any insights are helpful !
@HyeWonJung
Regarding gap-closing: This is almost surely necessary and is a natural way to keep experts in the loop. I've found AI coding LLMs to also be very good with follow-up prompts to make and fine-tune adjustments to original code.
It is sounding like the final envisioned code for a virtual bioreactor could end up be complex and large -- well beyond basic JSL platform calls. In this case would make sense to not use Assistant and move instead to a true coding IDE like Visual Studio Code + AI chat like Github Copilot. My personal favorite is Claude Code and I've had some major successes with it. A Python + JSL hybrid code architecture would probably be good, leveraging the strengths of each language.
Overall the amazing capabilities of modern AI coding makes constructing a virtual bioreactor much more attainable than it was just a few months ago.
