Expert context is where you tell the optimizer what you already know about your problem. It is a single free-form text field that travels with the experiment from setup through every iteration of the run. The AI reads it whenever it proposes new conditions and uses it as domain grounding for its reasoning.
Good expert context shortens the path to good results: the optimizer starts closer to promising regions, makes more sensible hypotheses, and produces insights that connect back to your domain. Empty expert context still works — the optimizer falls back on the variable bounds alone — but you give up a lot of free signal.
What It Is
A single text area, up to 4,000 characters, attached to each experiment.
Free-form: write in paragraphs, lists, sections — whatever reads well for the problem.
Editable at any point — during setup, during a run, after a pause, and at any iteration.
Optional, but strongly recommended.
For guidance on what to write, see Writing Effective Expert Context.
When the Algorithm Uses It
Expert context is now read by the optimizer at every stage of an experiment, not only at the start:
Initial design phase. Before enough data has been collected for the surrogate model to take over, the optimizer uses a hybrid sampling strategy: a portion of the warm-start points are placed in regions your expert context highlights as promising, while the rest are sampled broadly across the variable space. This balances using your knowledge with making sure no part of the search space is ignored.
Model-guided phase. Once the optimizer has enough data to train a surrogate model, expert context is still read on every iteration. It is provided to the AI candidate generator as domain grounding, helping it formulate hypotheses about which region of the search space is worth probing next. The surrogate-model-driven suggestions and the context-grounded suggestions are combined into the batch you see.
Insights. The natural-language explanations attached to each iteration are written by an LLM that receives your expert context as part of its prompt on every iteration. Without expert context the LLM can only describe the geometry of the search — which points were proposed and where they sit in the variable space. With expert context, it can reason about why those points are sensible in your domain: it references your variables by their physical meaning, connects proposals to the interactions you described, and frames trade-offs in language a domain expert would recognize. Writing strong expert context is the single biggest lever you have on the quality of the insights you see.
Crucially, the AI does not blindly follow expert context. As data accumulates, the optimizer weighs your context against what the experiments actually show. If the context contradicts the data, the data wins. This protects the optimization from a confidently-written but wrong piece of context.
Editing Expert Context
You can edit it during a run. Changes apply to the next iteration the AI generates. Already-generated recommendations are not retroactively changed — the platform shows a banner reminding you of this when you edit during a running experiment.
You don't need to start over. Updating expert context does not reset the data, the model, or the iteration history. The experiment continues from where it was.
It can be used to steer the search. Beyond providing background, expert context can be edited mid-run to focus the optimizer on a region, request more exploitation, or revise an intuition that the data has invalidated. See Updating Expert Context During a Run for guidance.
Experiments with Constraints
For experiments that define explicit constraints on the variables, the initial design is handled by the constraint solver, so expert context does not shift the warm-start point distribution in the same way. However, expert context still:
Feeds into the model-guided phase candidate generation.
Shapes the AI-generated insights throughout the experiment.
There is no reason to skip expert context on constrained experiments — it still does meaningful work.
When Should I Use It?
Always, if you have any prior knowledge about the problem. The cost of writing a few paragraphs is small; the benefit is that early iterations are less likely to be wasted on regions you already know don't work, and later iterations have a richer grounding for hypothesis generation.
Even when you have no firm prior expectations — for instance, when exploring a brand-new chemistry — it is worth writing at least a short description of the problem: what is being optimized, what the variables represent, and what the system actually is. A few sentences of plain framing is much better than an empty field. It anchors the AI in your domain, lets the insights talk about your system rather than abstract variables, and costs you almost nothing to write. Reserve "empty" for the case where you genuinely cannot describe the problem at all — which is rare.
Where to Go Next
The rest of this section is organised so each article covers a different angle. Read them in order if you're new to the platform; jump directly to a case study if you want to see a real expert context evolve through a run.
Writing Effective Expert Context — what to put in, what to leave out, short illustrative snippets for a physical-system and a chemistry-reaction problem, and seven short use-case patterns you can mix and match.
Updating Expert Context During a Run — using expert context to steer an in-flight optimization, when to revise intuitions, and how the AI balances your context against the data.
Case Study: Optimizing a Suzuki Coupling Reaction — an end-to-end single-objective reaction-yield optimization. Shows the expert context at setup, mid-run after data refines the picture, and late in the run during exploitation.
Case Study: Optimizing a Vegan Cheese Formulation — an end-to-end multi-objective formulation problem with mass-balance and fat-cap constraints. Shows how to encode trade-offs between competing objectives and how the context evolves as the Pareto front emerges.