Scientific machine learning
On this page
The problem
A mass spectrum contains evidence about a molecule, but interpreting that evidence as a molecular graph is an inverse problem. Different structures can produce ambiguous observations.
What I’m exploring
Spec2Graph investigates molecular graph prediction from mass spectra using spectral diffusion models. Related projects in chemical-graph-series, mol-modalist and ChemGPT-R examine other ways to represent molecules and use machine learning for chemistry.
Design decisions
My chemistry background gives the problem domain a concrete meaning: a generated graph needs to be chemically meaningful, and a prediction needs evidence beyond a plausible-looking output.
The inverse problem’s ambiguity matters. A model’s prediction should be evaluated under the dataset and task conditions rather than treated as a unique recovered structure.
Evidence and status
The public repositories contain the experimental implementations and project descriptions. The work is experimental; dataset conditions and structural ambiguity are part of the evaluation problem.