From your experimental question to the right design

Process development often starts with minimal variation in the parameters. A setting is tried, the result is looked at, and the next setting is chosen from it. One factor is moved while the others are held, then another, which the literature calls one-factor-at-a-time experimentation, OFAT. Other runs are exploratory, performed on a hunch, an idea or something read in a paper. Both are the natural way to work on a process that is very new, when the question is to find a baseline process to start working with.

The downside shows up later. Because the conditions were tested without a structure, the runs are difficult to include in a systematic modelling approach. The levels do not line up, interactions between factors were never varied together, and the runs rarely fit the design a later campaign is built on. Statistical DoE software fits its statistics to the design it generated, so experiments that were not part of one usually stay outside it, siloed in a folder while the designed campaign starts with minimal prior knowledge.

In Novasign Studio none of those runs are wasted. You enter the runs you have already performed as your own run table, with the levels you actually chose. Studio combines them with any design you generate next, so the early runs train the model alongside the designed experiments and are considered when the digital twin predicts what to run after that.

Design of experiments (DoE) is the starting point of many new campaigns and new drug candidates, where no or only a limited amount of information is available. Some teams come to it as the continuation of their first exploratory runs, others start with it instead of them. Which factors to investigate, and at which levels, has traditionally been decided by scientists and engineers from previous work and a review of the literature.

Which design you need depends on what you want to find out, how many factors are in play and how many runs you have. Studio helps you make that call. Describe your process and your goal to the AI Assistant, and it proposes the factors, their levels and a design to start from, and explains why it fits. The proposal appears in your workflow as a suggestion you accept, change or dismiss.

If you already know the design you want, you build it yourself. You define the factors with their levels and units, pick the design, and the run table appears in your workflow. Studio keeps track of the suggested, the performed and the predicted experiments in one place, so a campaign runs without intermediate exports and imports. The runs of this campaign are the basis data the hybrid model is built on.

Which design fits depends on the question. A screening design such as the fractional factorial investigates many factors in few runs. A central composite design studies fewer factors at more levels, when the curvature of the response matters. Studio builds both, as well as the full factorial.

See the design before you book a vessel

Model-based DoE (mDoE) uses the model you already trained to decide which experiments are worth performing next. The digital twin predicts how conditions nobody has run would behave, and only the targeted runs that carry the most information for your question are performed, whether that question is an optimum, the characterization of a design space or the edge of failure. Each of them then feeds the model.

This is the right approach once experiments have been performed, for example after a designed campaign. A hybrid model trained on them is run as a digital twin over the conditions of interest, including setpoints and process durations that were never tested, and returns a predicted response for each of them. The model is the hybrid model from your earlier runs, extended with every experiment the loop adds.

From those predictions Studio produces the model-guided suggestion. The conditions inside your boundaries are ranked by the objective you are after, and the few that stand out are the experiments worth performing next. Each of them is worth exactly one experiment, because it is the cheapest way to find out whether the model understands the region you want to run in. At every station you can inspect the result, accept Studio’s suggestion or adapt it before anything is booked in the lab.

Two outcomes are possible, and both are useful. Either the run confirms the predicted promising condition, a high product titer at acceptable quality for instance, and you finish the campaign or concentrate the remaining experiments on that part of the space. Or the prediction is not confirmed, and a suggested good condition turns out to be a bad one. The run joins the dataset, the model is retrained on all of it, and it is sharper exactly where it was wrong.

Testing several conditions within one run yields more information for the model to learn from. Instead of three experiments at static conditions, you run one experiment with three conditions as consecutive stages. Nothing is dropped from the design. Every condition is still run, several of them inside one vessel, and you set the limits your process can live with, such as how many stages a run may hold, how many runs the campaign may book and which factors are safe to move part-way through a run. The hybrid model built on your earlier campaigns is extended with every stage. Intensified DoE fits process characterization best, once the edge of failure is known, because every stage can then be placed inside conditions the process tolerates.

A hybrid model then attributes each measurement to the stage that produced it. It walks the run in the time increments your data was recorded in and treats temperature, pH, stirrer speed and dissolved oxygen as settings that drive the rates, which is why a level change part-way through a run is something the model reads rather than something that confuses it.

This works when one condition is met. The process has to settle again after a shift. If the temperature is moved from 37 to 35 °C and back to 37, then after a waiting period the process should behave as if the step to 35 had never happened. The literature calls the failure of this the memory effect, and in mammalian cell culture it reports about two days per stage for the cells to sense the change and adapt. Where a factor leaves a lasting mark, it stays a factor between runs rather than within one.

In the Arkeon program, two intensified campaigns ran in continuous 2 L reactors, the first to find a stable temperature and pH, the second to set the process parameters for scale-up. The hybrid model learned from the transition after each shift as well as from the steady states, and it predicted the steady state better when those transient phases were part of its training data.

A continuous process, shifted stage by stage

The measured titer peaked at 35 °C and pH 6.5, 4 g/L against about 2.5 g/L at the centre point, 60% higher. The product concentration in the plot is the model’s prediction for the run, and the model learned the direction of every shift from the one run.

A filtration step, screened in one run