Space weather modeling
Research models are only useful once someone other than their author can run them. Much of my work is about closing that gap.
Model staging platform
A large part of my work has been building a model staging platform on public cloud infrastructure, lowering the barrier to running research code and shortening the path from cutting-edge research to operational forecasting centers.
The canonical example is an empirical Fortran atmospheric model that historically required a compiler and local expertise to run. I wrapped it in Python as pymsis, exposed it through a serverless API (msis-api), and built the MSIS Visualizer on top of it with d3, so anyone can explore surface fields and altitude profiles of atmospheric constituents in a browser without installing anything.
Other projects follow the same architecture: VECTOR, a satellite drag calculator where users upload a spacecraft design, choose orientation and atmospheric conditions, and get a detailed drag assessment; and enlilviz, a Python solar wind visualization package that produces the animations and interactive plots space weather forecasters use operationally.
Physics-constrained machine learning
Alongside building tools for the community, I work on applying machine learning to the spatial and temporal interpolation of magnetic fields. The key idea is to avoid learning raw field values. I wrote pysecs, a Spherical Elementary Current System implementation, to reduce the problem to a set of divergence-free basis functions — which guarantees the underlying physical constraints hold no matter what the network predicts.
We then trained a deep neural network to fit those basis functions using synthetic observations from expensive HPC models that cannot run in real time. Reducing dimensionality with physical basis functions and training against large-scale HPC output has produced markedly more realistic magnetic field interpolations across North America.
This line of work continues in LiveWire, which predicts the horizontal geoelectric field an hour ahead using multi-fidelity boosted neural networks.
Storms and their consequences
Space weather stops being abstract when it takes hardware with it. I contributed to the analysis of the 2022 Starlink incident, where a modest geomagnetic storm expanded the thermosphere enough to bring down 38 newly launched satellites — a concrete demonstration of the threat storms pose to low Earth orbit operations. More recently I have worked on the I-ALiRT real-time system for NASA's IMAP mission, which turns spacecraft telemetry into forecasting inputs as it arrives.