← Journal2 min read

Joining Matter42

Founding engineer on a team teaching models to read defects in 2D materials from cheap measurements.

I've joined Matter42 as a founding engineer.

The problem

2D materials like WSe₂ are only as good as their defects allow. A monolayer's electronic and optical behavior depends heavily on how many vacancies it has, so labs need to know the defect density of what they've made.

The standard way to find out is transmission electron microscopy. TEM is excellent, and it's also slow, expensive, and not something you want to run on every sample. Raman spectroscopy, on the other hand, is fast and cheap, and defects do change a Raman spectrum. The catch is that the relationship is subtle, and there's very little labeled data connecting the two.

The approach

If you can't collect enough labeled data, you simulate it. We're building a simulation-to-ML pipeline, starting with WSe₂ monolayers:

  • Generate defected structures across a range of vacancy densities.
  • Pick potentials we can trust. We benchmark machine-learned interatomic potentials like ORB v3 and MACE against DFT phonon references, using κSRME to choose models whose force constants are physically reliable, not just low-error on energies.
  • Simulate the measurements. Phonopy phonon spectra act as a Raman proxy, and abTEM simulates HRTEM and STEM images for training defect-segmentation models.
  • Learn the mapping. Train models that infer defect density directly from Raman features.

The goal is simple to state: let a lab take a quick Raman measurement and get a defect density it can trust, saving TEM for the cases that actually need it.

Why this, why now

A lot of what I've built on Ouro over the past couple of years, including DFT services, MLIP tooling for relaxations and phonons, and generative crystal models, turns out to be the exact toolkit this problem needs. It's rare to find a role where the side project you've been obsessing over is the job description.

It's also a chance to work on something where the output is a measurement a lab uses on Monday, not a candidate in a spreadsheet. I'm still building Ouro alongside this, and I'll write about how those two fit together as it takes shape.