AI4S NotesNotes for the space
between model and experiment.
Seven original essays on how to read scientific machine learning. Follow the evidence from a model output to a claim that can be tested.
What turns a prediction into a discovery claim?
A useful prediction changes the next experiment. The strength of the discovery claim depends on what that experiment can rule out.
A dataset split is a scientific experiment
The test set defines the kind of unfamiliarity a molecular model must handle. Choose that unfamiliarity before comparing scores.
Reading a protein structure as a functional hypothesis
A predicted structure can organize an experiment without settling what a protein does in its biological setting.
The distance between a promising material and a usable sample
Candidate ranking is only one decision in a synthesis workflow. A useful evaluation follows the material into the laboratory.
What a physical constraint buys a learned surrogate
A small equation residual is evidence about one test. Reliable use also depends on boundaries, regimes, and accumulated error.
Why weather skill needs a separate climate argument
Short-range predictive skill and long-term climate claims require different evidence, even when they describe the same atmosphere.
Keeping controls inside an AI experiment loop
An adaptive system needs a record of its decisions and controls that survive adaptation. Otherwise, iteration can outrun inference.