FLIP improves sensitivity in tested Webb exoplanet images

A study published November 4 reports that FLIP improved sensitivity to planets in tested James Webb Space Telescope (JWST) NIRCam coronagraph images. Its most striking result is a signal from the known giant planet 51 Eridani b in an image taken at 4.3 μm, with a signal-to-noise ratio (SNR) of 6.3. The wavelength falls in a carbon-dioxide absorption band; atmospheric characterization is future work, not a result of this image. The study compares FLIP with KLIP, a standard method for subtracting starlight from telescope images.

Read each sensitivity gain with its target, filter and separation: the results varied across the observations.

What FLIP found at 4.3 μm

The 51 Eridani b result is an image in NIRCam’s F430M filter. At the planet’s location, the study reports about four times better sensitivity than KLIP; an earlier published analysis had reported no detection in that filter. Here, 4.3 μm is the wavelength of the observation, not a measurement of the planet’s size or atmospheric composition. The paper says future work will use the photometry, or measured light, for atmospheric characterization.

The result concerns a known giant planet. The study also tested sensitivity with simulated planets inserted into images, which is a different kind of result from observing a companion.

How FLIP separates starlight from a planet’s signal

A telescope’s point-spread function (PSF) is the pattern it produces when imaging a point of light, such as a star. Changes in the telescope between reference-star and science observations can make ordinary image subtraction leave some of the star’s glare behind. FLIP uses a physical model of the telescope and detector to fit that light in the images.

The researchers built the differentiable model in PyTorch from STPSF optical models and wavefront-sensing information. Automatic differentiation calculates how changes in the model’s parameters affect its output, and an Adam optimizer adjusts those parameters. FLIP first calibrates the model against a reference-star observation, then fits a smaller set of parameters to the science image to account for differences such as pointing and wavefront changes.

To guard against fitting away a planet’s signal, the researchers insert simulated planets into the science image and track their SNR as the model is optimized. They stop when that SNR peaks or levels off. The end-to-end fit took about 35 minutes on one NVIDIA A30 GPU in the study.

FLIP follows earlier data processing. The observations were processed with JWST pipeline version 1.18.1 and spaceKLIP version 2.2.1 before the FLIP analysis.

Results vary by target and measurement

Across the tested datasets, the study reports sensitivity improvements of 2 to 15 times over standard techniques. The table separates SNR results for known companions from sensitivity tests using simulated planets; those metrics describe different things.

For TYC 5899, the 15-fold figure applies at a separation of 0.5 arcseconds from the star, while the gain at 1 arcsecond is about fivefold. Both figures come from simulated-planet sensitivity tests. They do not describe an observed planet around TYC 5899.

The demonstration covers NIRCam coronagraph images

The researchers tested four datasets from three systems: HIP 65426 in F444W; 51 Eridani in F410M and F430M; and TYC 5899 in F444W. All were JWST NIRCam coronagraph observations. The paper discusses extending the approach to MIRI or other observatories as a future possibility, not a demonstration in this study.

The tested reductions did not reach the fundamental photon-noise limit. Systematic calibration of the recovered telescope-instability estimates, including their uncertainties, remains future work.