FDE + BayesOpt for complex pharma process development

[#1]

Hi everyone,

We’re developing a highly complex, multi-unit manufacturing process with a large number of CPPs across the various unit operations, together with continuous CQAs such as time-dependent drug release profiles.

We’ve built up quite a lot of historical data from previous DoEs, which we’re currently looking at as one combined dataset, and we also have the possibility to generate substantial additional data.

We’re considering upgrading to JMP Pro, mainly to explore whether FDE and Bayesian Optimization could work together for this type of problem.

I’d particularly like to hear from people who have used these approaches in similarly complex, multi-step process development, rather than relatively simple optimization problems.

A few things I’m curious about:

• FDE → BayesOpt: How practical is it to reduce continuous profiles to FPC scores and then use these as target Ys in BayesOpt?

• Historical DoE data: How well does BayesOpt work with data combined from several DoEs, particularly when the factor ranges and designs are different?

• Multi-unit constraints: How easy is it to deal with process boundaries and combinations of factors that are not feasible across different unit operations?

• BayesOpt vs. traditional DoE: For this type of problem, have you found the iterative approach of BayesOpt to be a real advantage compared with augmenting a conventional DoE?

I’d really appreciate hearing about real-world experiences, workflows, or lessons learned from anyone who has tackled something similar. Even if the experience was that it didn’t work as expected, that would be very useful to know.

Thanks!

[#2 · Accepted solution]

Hi @RicardoCosta,

Maybe you can get in touch with @Emmanuel_Romeu, as I think he already has worked on this exact topic of combining Bayes Opt with FDE.
The workflow you describe seems to be a reasonable one:

1. Analyze the first curves with the FDE with the target, extract FPCs.
2. Use the FPC values from the curves as responses in the BayesOpt platform, except the ones from the target that you use to set the FPC response limits values and goals (maximinze, minimize, reach target).
3. Use BayesOpt to explore and exploit your historical data, and find the next recommendations using the learning from the GP models.

Make sure to use a Validation Column on your FDE datatable to avoid data leakage during the curves modeling: Set your initial historical datapoints as training rows, and any new datapoints as test rows, to make sure the FPC values of your historical datapoints won't change at each iteration when using the FDE platform. This way, you can simply extract the new FPC values of your next experiments

About your questions:

• Historical DoE data: as long as the design spaces from the different studies are not completely disjoint (which would make the model learning more difficult), you could use these datasets to explore a broad space with BayesOpt. Note that like DoE, you need to specify the ranges for your factors. You could also use your historical data and Augment the Designs.
• Multi-unit constraints: Like for Custom DoE, you can set, modify or delete constraints. The advantage of the BayesOpt platform is that you can do it for each iterations, as well as reduce the design space (if you want to force the algorithm to only explore/optimize in a specific area). In order to set constraints, you can click on the red triangle of Augmented Acquisition Functions Profiler>Optimization and Desirability>Optimization Control Panel:
  
  

  See Solved: How do I set up a constraint when using Bayesian Optimization? - User Community
  If you have a combination of constraints or more complex constraints to enforce, the easiest way to include them in the BayesOpt platform is to create a Candidate set respecting them. You can dot it from the red triangle of Bayesian Optimization Batch Customizer, option "Generate Candidate set from Profiler settings", or create a table in JMP and deleting rows that don't respect your constraints. Once your candidate set is ready, use the red triangle of of Bayesian Optimization Batch Customizer and choose the option "Load Candidate Set from Data Table".

• This question is not so easy, and is more linked to practical considerations than statistical ones: 
  • How easy it is for you to run multiple experiments at the same time/in parallel ? What is the time to measure the responses, and how parallelized it can be ?
  • Are you measurement and experimentation conditions stable/precise enough to do multiple iterations with BayesOpt instead of doing few sequential designs ?  
  • Are there any source of noises that may change between iterations to be considered ? Any uncontrolled factors that may interfere with the Bayesian optimization process ?
  • What is the dimensionality of your use cases (how many factors ?) ? Have you alread identified the most relevant ones ? If you have a very high dimensionality (>10 factors) and/or have not completely identified the most important factors to optimize, I would recommend starting with a DoE (possibly augmented from your historical data).
  • How many responses are you measuring ? How do they compete with each others (any trade-off between responses to optimize) ? If you already know some responses are competing with each others, I think a DoE approach may be more sensible, as you will be able to really see the trade-off between the responses. With BayesOpt, the risk is to go from one optimum to the other, without fully evaluating the trade-off between the two responses.
  • What is your objective: Optimize only, or get an understanding and knowledge about your design space ? Note that with BayesOpt, you'll be focussed on the optimization part, with no or limited options to understand which factors have the biggest importance on the responses.

BayesOpt can be a real advantage if you can leverage good quality data to start, and if you already know your experimental conditions are robust and reproducible. Best situations I have seen so far is when you can combine BayesOpt and DoE approaches, as their strenghts and limitations can compensate.

Hope this answer will help you,

[#3]

Hi @Victor_G. 

Thanks for the clear explanation — it helps confirm how BayesOpt FDE can work together.

My process is highly multivariable, with several unit operations, many factors, and multiple CQAs. We have a large historical dataset from multiple DoEs, already aggregated. At this stage, the priority is extracting the most relevant insight and defining the next experiments, and we are evaluating whether an upgrade to JMP Pro is justified.

Your workflow (FDE → FPCs → BayesOpt → GP‑guided iteration) fits well. The main points I’m trying to clarify are whether BayesOpt remains reliable in a broad factor space and how well BayesOpt handles competing CQAs compared with a DoE‑centric approach, whether slow measurement cycles reduce its value, and whether heterogeneous historical DoEs should be augmented or directly modeled.

BayesOpt seems a strong complement to DoE, but I want to ensure the benefits are meaningful before committing to JMP Pro. 

If it’s not asking too much, could you or @Emmanuel_Romeu point me toward the best online references for learning more about FDE and BayeOpt — ideally material that goes beyond the basic JMP documentation?

Given the complexity of our process and the possibility of upgrading to JMP Pro, I’d like to deepen my understanding of both the theoretical foundations and the practical implementation details (especially GP‑based modeling and FPC interpretation). Any recommended tutorials, papers, videos, or training resources would be greatly appreciated.

Thanks again for your time and for the clarity of your explanations — much appreciated.

[#4]

Hi Ricardo,

- For the FDE/BayesOpt component, it is relatively simple to convert the FPC scores in a loop with BayesOpt and FDE - the workflow is essentially 1) Run the experiment with BayesOpt 2) Gather functional data 3) Run FDE and bring out your 'difference from target' value 4) Put into BayesOpt and get your next experiment. This runs in a loop. To add a shameless self-plug - the FDE Model Screen tool helps to quickly pick the best FDE model for your data. I can comment that I've spoken recently with customers who've used this in developing formulations to meet set dissolution profiles, using the target difference (instead of F2) as part of their BayesOpt process with success.

-Historical DoE Data - essentially BayesOpt would look as wide as the data is for the low/high limits of the factors, you can choose to restrain this in the platform, but you give the BayesOpt platform a wide search area for it to explore - it might then suggest to add addition 'exploration' runs, or go straight for 'exploitation' (optimisation).

- BayesOpt v. DoE - Victor has given you a really nice, expansive answer, I just want to add that it is important to think of BayesOpt as an additive experimental tool, not a replacement for DoE. It can be more flexible to restrictive experimentation and constraints, and can provide a more efficient approach to reaching an optima. But if your goal is to have a fully expansive understanding of your system, it may be better to either do a DoE, or hybrid BayesOpt + DoE approach (ie start with BayesOpt to find optimal value, then augment with a DoE in a smaller range to get robustness values).

Thanks!
Ben

[#5]

Yes, indeed, I have submitted an abstract on combining FDE and BayesOpt for the next Discovery Summit in March 2027. My example is relatively simple, but it illustrates the trade-off between the responses to be optimized that @Victor_G mentioned. To this end, I recommend not being too ambitious, and setting minimum and ideal targets to be achieved with similar weighting.

Combining several sets of preliminary data is entirely possible, but you may need to check what the Space Filling Design recommends, compare this with what your data covers, and possibly add a few points to get closer to the recommended Space Filling Design.

You need to determine which FPC is the most relevant, as it is not necessarily the largest. For example, you may focus on a small peak corresponding to an impurity. You may also need to optimize several FPCs and assign weights to them, which could be based on the eigenvalues; however, this is not necessarily the case if, once again, an impurity corresponding  less significant FPC proves to be critical. We can set up a call if you want to with @Ben_BarrIngh your SE covering your account.

[#6]

Hi all, Thank you @Ben_BarrIngh. @Emmanuel_Romeu and @Victor_G  for the time you took to share such detailed and thoughtful guidance. The perspectives you each brought — from the practical FDE/BayesOpt loop were extremely valuable.

Before moving to the next steps on our side, I want to take the opportunity to learn substantially more about both FDE and BayesOpt

Your comments made it clear that a deeper understanding of model selection, target definition, FPC weighting, and GP‑driven iteration will be important for applying these tools correctly — especially as we evaluate whether upgrading to JMP Pro is the right decision.

Thanks again for the clarity, the examples, and the openness to continue the discussion. I truly appreciate the support.