JWST coronagraph images get a differentiable physics model for starlight removal: up to an order of magnitude better sensitivity and the first 4.3-micron image of 51 Eridani b
Synopsis
Directly imaging exoplanets requires subtracting the star's glare from the image, and standard practice builds that model from reference-star images, so telescope pointing errors and thermal drifts between reference and science observations limit sensitivity. FLIP implements a physics-based instrument model in PyTorch and uses automatic differentiation to optimize millions of parameters in stages, pre-calibrating on a reference star and then tuning on the science image to compensate those instabilities. On JWST/NIRCam coronagraph data, FLIP achieves 2 to 15 times better sensitivity than the standard KLIP method and makes the first detection of 51 Eridani b at 4.3 micrometers in F430M at signal-to-noise 6.3.
Interpretation
Starlight subtraction for direct exoplanet imaging no longer has to rely only on reference-star images; a differentiable physical instrument model can fit the stellar point spread function directly on the science image. The most successful prior approaches built the point spread function model empirically from reference observations taken at other times, wavelengths or stars, and those methods are limited by differences between reference and science observations caused by telescope instabilities. On JWST/NIRCam coronagraph data, FLIP beats standard KLIP in every observation, with improvements up to an order of magnitude relative to published analyses, and detects the previously undetected 51 Eridani b in F430M at signal-to-noise 6.3.
The method writes the star, the optical system and the detector as one differentiable physical model, then uses gradient descent to make the simulated star pattern match the real science image. The model starts from the nominal STPSF optical models maintained by the Space Telescope Science Institute and is implemented in PyTorch so automatic differentiation can optimize its free parameters. Optimization runs in two stages: full calibration on a reference star, then tuning only star position and flux plus the wavefront error after the focal-plane mask on the science image; on a single NVIDIA A30 GPU, stage one takes 10 to 15 minutes and stage two takes 10 to 30 minutes.
Perspective
The demonstration is limited to real on-orbit JWST/NIRCam coronagraph data, four datasets covering HIP 65426, 51 Eridani and TYC 5899 across the F444W, F410M and F430M filters and two coronagraph masks; the authors propose extension to MIRI, the Roman Space Telescope, the Lazuli Space Telescope and ground-based telescopes, but those extensions are not shown here.
FLIP does not yet reach the photon noise limit in any observation, and its models have not been systematically calibrated to return reliable estimates of the instability parameters with uncertainties; in the TYC 5899 data FLIP sometimes shows higher signal loss than KLIP at larger separations, and the authors list a more sophisticated detector model, disentangling parameter degeneracies and hyperparameter exploration as improvement paths.