lightcurve-fitting

Lightcurve fitting library for astronomical transients, written in Rust.

Provides three complementary fitters that operate on multi-band photometry (ZTF g/r/i):

Fitter	Input	Method	Key outputs
Nonparametric	magnitudes	Gaussian-process interpolation	peak mag, t0, rise/decay timescales, FWHM, GP derivatives, von Neumann ratio, power-law decay index
Parametric	fluxes	PSO model selection → SVI or Laplace uncertainty	best-fit model (Bazin, Villar, TDE, Arnett, Magnetar, …), posterior means & uncertainties
Thermal	magnitudes	Blackbody color fitting via PSO	log temperature at peak, cooling rate

Quick start

use lightcurve_fitting::{
    build_mag_bands, build_flux_bands,
    fit_nonparametric, fit_parametric, fit_thermal,
    UncertaintyMethod,
};

// times, mags, mag_errs, bands: parallel Vecs from your photometry table
let mag_bands  = build_mag_bands(&times, &mags, &mag_errs, &bands);
let flux_bands = build_flux_bands(&times, &mags, &mag_errs, &bands);

let (np_results, trained_gps) = fit_nonparametric(&mag_bands);
let p_results  = fit_parametric(&flux_bands, false, UncertaintyMethod::Laplace);
let t_result   = fit_thermal(&mag_bands, Some(&trained_gps));

GPU acceleration (optional)

When built with --features cuda, the parametric fitter uses CUDA batch PSO to fit all sources simultaneously on a GPU. This provides significant speedup at scale (see Benchmarks).

cargo build --release --features cuda

Requires CUDA toolkit (tested with 12.x). The build system auto-detects nvcc via the CUDA_PATH environment variable or standard locations.

Building

cargo build --release

Testing

Run the unit and integration tests:

cargo test

Run GPU-specific tests (requires CUDA):

cargo test --features cuda --test test_gpu -- --nocapture

Benchmarks

Throughput benchmarks measure scaling across two axes:

• Source-count scaling: fixed 30 pts/band, 10–1,000 sources
• Point-count scaling: fixed 100 sources, 10–500 pts/band

Run the full benchmark suite:

# With GPU
cargo test --release --features cuda --test bench_throughput -- --ignored --nocapture

# CPU only
cargo test --release --test bench_throughput -- --ignored --nocapture

# Generate plots
python3 benchmarks/plot_throughput.py

# Submit as SLURM job
sbatch benchmarks/run_bench.sh

Results are written to benchmarks/throughput_results.csv. See docs/benchmarks.md for full methodology and results.

CI

Two GitHub Actions workflows are included:

• test.yml — runs cargo test on every push to main and on PRs.
• throughput.yml — builds in release mode, runs the throughput benchmark, uploads wall_time.txt as an artifact, and on PRs compares against the main baseline. Fails if wall time regresses by more than 10%.

License

GPL-3.0