Code
I build differentiable, GPU-native tools for stellar population inference. All of it is open source — contributions, issues and feature requests are genuinely welcome.
CERIDWEN · github.com/Espe13/ceridwen
A GPU-native, gradient-based SED-fitting framework written in JAX. CERIDWEN (Comprehensive SED Estimation Routine Involving Data-driven Weight calculatioNs) reformulates stellar population synthesis as a fully differentiable computation, so that posteriors over stellar mass, star-formation history, chemical enrichment, dust and redshift can be sampled with Hamiltonian Monte Carlo rather than slower gradient-free methods.
The forward model is modular: pre-computed SSP grids, flexible non-parametric star-formation and chemical-enrichment histories, nebular continuum and line emission, per-age-bin dust attenuation with energy-balance IR re-emission, and a differentiable observer-frame conversion including a JAX-native luminosity-distance integrator and IGM transmission. Python-side toggles resolve at trace time, so each spectrum compiles to a single XLA kernel.
sedpy-jax · github.com/Espe13/sedpy_jax
A JAX-friendly port of sedpy providing differentiable filter projection and a library of dust-attenuation laws. It supplies the photometry and attenuation machinery used by CERIDWEN, so any law registered in sedpy-jax is immediately available to the fitter with no extra plumbing.