Skip to content

Research

I did my PhD in computational plasma physics at the University of Maryland, then left research for industry. A decade later, AI has made it possible to come back. Not as a full-time researcher, but as someone who can run real physics investigations on the side.

Over the past year, what I hand over to AI has grown in three steps: first a coding milestone, then a parameter scan, and now a whole physics problem.

Plasma turbulence, with AI

In late 2025, working largely solo with Claude, I rebuilt GANDALF, my PhD-era code for kinetic plasma turbulence, in JAX. I didn't read the code. I checked the physics it produced instead. The code is on GitHub, along with a paper describing it and how it was verified.

This spring I used the code to look for the dissipative anomaly in kinetic plasma turbulence. Across a 50× scan in collisionality, the dissipation rate stayed flat, and a paper draft followed. A closer look showed the runs were only weakly turbulent. The flat rate was the known linear phase-mixing result, already in the literature, including a 2015 paper of my own. It wasn't evidence of an anomaly. In November I closed the study without submitting the paper.

Now

My current study builds on a recent paper by Chandran, Mallet and Meyrand on the gyrokinetic helicity. It asks whether conserving that quantity changes how free energy cascades to fine scales in velocity space.

This time I describe the physics problem and leave the method to the model. The study runs as a loop of Claude sessions, each starting fresh, with the repository as the only memory. It stops and asks me only for a short list of reasons, such as reaching its GPU budget.

Write-ups