A predicted property is an estimate. The physical assay is the first point where a candidate is actually measured, and disagreement between prediction and measurement is normal, not a failure of the process.
Where AI helps in the lab
- Choosing which experiments to run first, by ranking conditions by how much they would reduce uncertainty.
- Interpreting noisy results, by finding patterns across many measurements and flagging results that look inconsistent.
- Closing the loop, by feeding lab results back into the models so the next round of candidates is informed by what actually happened.
Experiment design is a search problem under constraints. Time, materials, and instrument capacity are limited, while the space of possible conditions is enormous. Ranking conditions by expected information gain means the most useful experiments run first, and each result updates the model before the next choice is made.
Assays are noisy, and a single measurement can mislead. AI helps by looking across many measurements at once, but it does not remove the need for careful experimental controls.