Our team has been spending a lot of time building evaluations for machine learning systems.

We have learned some lessons and wrote them down arxiv.org/abs/2503.10837

I'm learning materials according to mine.

Thanks, @morinryan.bsky.social for creating this service!

With this, he not only obtains SOTA performance, but he could also reproduce corrections of incorrectly assigned structures in the literature, i.e., spot mistakes.

However, chemists like creating new compounds that might not be in PubChem.
Thus, he extended his system with a GA that evolves molecules to best match chemical structures.

In the first step, he aligns encoders for molecules and spectra using contrastive learning.
Based on this, he can retrieve compounds for any combination of spectra.
In practice, he searches PubChem using spectral embeddings.

Chemists often combine many different techniques to elucidate structures.

Adrian has been building a system that mimics this using models and genetic algorithms.