Can measurements learn a common language?

From patterns to representations

Begin with experimental signals

XRPD traces from solid-form screening vary in peak shapes, backgrounds and preparation artefacts.

Conceptual illustration of pattern encoding and comparison. The traces and embedding positions are schematic, not experimental data or model outputs.

The scientific question

Can we learn useful representations of analytical measurements, so that comparison and interpretation become less dependent on repeated manual inspection? XRPD solid-form screening offers a concrete starting point.

The published starting point

SMolNet learns to compare pairs of XRPD patterns. The study uses 3,750 experimental measurements from 16 compounds and 24 solid forms. A leave-two-compounds-out evaluation tests transfer to unseen compounds.

What the study establishes

Training with a modified SigLIP loss improves ranking and class separability, with a reported mean AUROC of 0.98. This is not a universal accuracy claim: the de Gelder baseline achieves the best thresholded F1 score in the reported comparison.

The broader direction

The goal is to use embeddings as a general basis for analytical chemistry workflows. The workshop establishes an XRPD example, not a validated model for every measurement modality. Extending the approach calls for appropriate experimental data and evaluation for each new task.