A starch slurry dries. Cracks form. The pattern of cracks depends on the solvent — water, ethanol, acetone, and mixtures produce distinct geometries. A neural network can identify which solvent was used with 96% accuracy just from a photograph of the dried residue.
The cracks are not random. Each solvent has a characteristic evaporation rate, surface tension, and interaction with the starch matrix. These physical properties translate into specific stress distributions during drying, which translate into specific crack patterns. The network isn't learning arbitrary correlations — it's reading the physics of the drying process written into the geometry of the fracture.
Note (2026-11-14; read against the paper's full text): the 96 percent figure is the paper's average accuracy over all solvents, with an uncertainty of 1 percent, but the essay's account of what was classified is looser than the paper's. The input is not the photograph itself: image analysis extracts nine morphological features (size, shape, geometry and orientational ordering) and frequency histograms of them are fed to neural network variants, and the paper says the highest accuracy came from feature sets that include the crack area distribution. The data are laboratory cornstarch slurries (20 g of liquid and 20 g of cornstarch) in one container setup: 25 samples each of water, ethanol and acetone, plus 50 water-ethanol mixtures at different concentrations, dried for 12 to 48 hours. The essay's physical account (evaporation rate and surface tension translating into stress and crack patterns) goes beyond the paper, which only says that changes in pattern features can be associated with changes of surface tension; the forensic and drug-identification uses are the essay's own.
This is forensic chemistry through fracture mechanics. The dried residue is a record of its liquid past. The information about the solvent's identity persists in the solid long after the solvent has evaporated. The liquid writes its signature in the cracks it leaves behind, and the signature is specific enough to be decoded.
The practical applications are obvious (drug identification, industrial quality control). But the conceptual point is more interesting: a destructive process (drying, cracking) is also an information-preserving process. The crack doesn't erase the liquid's identity — it encodes it. Destruction and documentation happen in the same physical act.