As the optimization progresses, SDLabs provides two ways to understand what the algorithm has learned: Recommendation Insights (LLM-generated commentary) and Predictions (direct model queries). Both are available on the Overview tab.
Recommendation Insights
When the optimizer finishes a recommendation cycle, an insights sidebar becomes available next to the recommendations table. Toggle it using the sidebar button. The sidebar has two tabs:
Recommendation Insights — AI-generated text explaining why the optimizer chose these particular parameter combinations. The text appears with a typewriter animation when new insights are generated.
Expert context — the domain knowledge you provided during experiment setup. This tab is disabled if no expert context was entered.
Insights are generated by an LLM that interprets the optimizer's decisions in the context of your experiment. If you provided expert context during setup, insights will reference and comment on it throughout the entire experiment — not just during the initial design phase. Insights are available once the initial optimization has completed (i.e., after the initial exploration phase) and are refreshed after every recommendation cycle.
Predictions
The Predictions section lets you query the predictive model for any parameter combination — not just the ones the optimizer suggested. This is useful for testing hypotheses, validating intuition, or exploring regions of the design space.
How to use it:
Expand the Predictions accordion on the Overview tab.
Enter values for each parameter in the input row. Numerical parameters show their valid range as a placeholder. Categorical parameters show a dropdown.
Click Query to get the model's predicted measurement values, including uncertainty ranges.
You can also click Load suggestions to auto-fill the prediction table with the current recommendations, then adjust individual values.
Add multiple rows to compare predictions side by side. Predictions are not stored — they exist only during your current session.
When Predictions Are Available
Predictions require a predictive model. During the initial exploration phase, the Predictions section is disabled and shows how many iterations remain before the model is ready.
Once a model exists, the section header shows which iteration the model was built on. If the model is out of date (new data has been submitted since the last recommendation cycle), an alert offers the option to Retrain Model.
Prediction Explanations
After querying a prediction, you can request a Prediction explanation chart by clicking the chart icon on the prediction row. This generates a waterfall chart showing how each parameter contributes to shifting the predicted value from the dataset average to the current prediction.
Explanation charts are available when the experiment has at most 9 parameters and 6 objectives.
Good to Know
Insights are based on the latest completed optimization, not real-time. If you submit new measurements, insights update after the next recommendation cycle.
The prediction model is probabilistic — it reflects the algorithm's current uncertainty. Predicted values may sometimes seem unrealistic for exploratory points. This is expected behavior and reflects the model's uncertainty, not a literal expectation.
During model updates, predictions are temporarily unavailable. They return once the update completes.
If the model update fails, you can retry via the Retry Retraining button, or submit new measurements to start a regular recommendation cycle.