Parallelization & Practical Settings
Beyond the parameter space and objectives, a few practical settings can significantly affect how quickly your optimization converges. This article covers parallelization (number of simultaneous measurements), batch-constrained parameters, and expert context.
Parallelization (Number of Simultaneous Measurements)
The parallelization setting controls how many experiments the optimizer suggests in each iteration. It should match how many experiments you can run simultaneously in your lab.
The Key Principle
Set the parallelization to match your lab setup — specifically, the number of experiments you can run at the same time. If you can run 8 experiments in parallel, set it to 8. If you can only run one experiment at a time, set it to 1. This is the most important factor: the parallelization should reflect how you actually do your experiments so that you get the most frequent feedback from the optimizer.
More frequent iterations mean the model learns faster and gives you better recommendations sooner. The optimizer benefits from receiving as much data as possible per iteration, so use the highest parallelization your lab allows — but never more than you can actually measure simultaneously.
Examples
You have a 24-well plate reactor → set parallelization to 24.
You run experiments one at a time on a single bench setup → set parallelization to 1.
You have two instruments that can each run 4 samples → set parallelization to 8.
Let your experimental setup and limitations dictate this number. The optimizer will adapt to whatever parallelization you choose.
Parallelization-Constrained Parameters
A parallelization-constrained parameter takes the same value for all suggestions within a single iteration. The optimizer can change it between iterations, but within one iteration, every experiment shares the same value.
When to Use
Use parallelization-constrained parameters when the experimental setup truly fixes some factors per iteration:
Equipment — all experiments in an iteration run on the same reactor or instrument
Material lots — one catalyst lot is prepared for an entire iteration
Operators or shifts — a human factor that does not change within an iteration
Only mark a parameter as parallelization-constrained if it physically cannot change between experiments in the same iteration. Do not use it as a way to reduce the search space — the optimizer handles that naturally.
Expert Context
Expert Context lets you share domain knowledge with the AI. It influences the initial experimental design (first batch of suggestions) and is also referenced by recommendation insights throughout the entire experiment. Expert context can only be set during experiment setup — once the experiment starts, it is fixed. Note: For experiments with constraints, expert context does not impact the initial sampling (which is handled by the constraint solver), but it still affects the insights shown throughout the experiment. See the full article for detailed guidance. Here are the key tips for writing effective context:
What to Include
Optimality regions — where in the parameter space good or bad outcomes are expected, even if only from intuition or literature. Example: "Yield tends to peak at high temperature and low residence time."
Range of objectives — typical or acceptable ranges for each objective. Example: "Yield is usually 60-90% for this class of reactions; we need selectivity above 0.8."
Similarity to known problems — reference classic formulations, benchmark problems, or prior studies. Example: "Similar to a standard catalyst screening setup."
What to Avoid
Project names, internal codes, or long narrative that does not add predictive value.
Information that contradicts your formal parameter space or objectives.
Vague statements — "Temperature around 150 °C" is better than "moderate temperature."
Keep it concise and signal-rich. Each field has a 600-character limit. Focus on information that helps the AI reason about where good solutions lie.
Summary
Setting | Guideline |
Parallelization | Match to number of simultaneous measurements your lab setup allows |
Initial design phase | Depends on the problem: 1 per numerical variable, 1 per categorical option without properties, and 1 per property (for categorical variables with properties) |
Parallelization-constrained parameters | Only for factors that physically cannot change within an iteration |
Expert context | Set during experiment setup only. Influences initial design and insights throughout. Focus on optimality regions, objective ranges, and similarity to known problems |