Can physics-based learning see planets hidden in JWST glare? FLIP improved sensitivity by up to 15-fold—but only on four datasets
A differentiable physical model recalibrated JWST's stellar point-spread function and recovered the known planet 51 Eridani b at 4.3 micrometres, where a standard pipeline had not detected it. The gains are promising, but they come from four reprocessed datasets and tuned injection-recovery tests, not a blind planet survey.
The short version
- What happened
- FLIP jointly optimizes a physical model of star position and flux, optical-path errors before and after the coronagraph and detector response, using automatic differentiation.
- Who it affects
- Publicly funded telescope observations may yield more science when improved modelling recovers signals hidden in archived data.
- What remains unproven
- Four reprocessed JWST NIRCam datasets are not a prospective or blind planet survey.
Keep in mind
Four reprocessed JWST NIRCam datasets are not a prospective or blind planet survey.
FLIP used one telescope roll while the KLIP comparison used both rolls, so the input configurations were not identical.
Learning rates, stopping points, masking and other processing choices involved researcher judgement.
No dataset reached the fundamental photon-noise limit, and performance was not uniformly superior at larger separations.
Source analysis · 1 linked source. Primary analysed source date: 4 November 2026.
Read the full report · 9 min The evidence, methods, implications and limitations
The direct finding: a physical model recovered a signal that empirical subtraction missed
A Nature paper published on 4 November reports the first on-sky use of a differentiable physical model to subtract stellar glare from real JWST coronagraphic images. The method, Physics-based Learning for Imaging Planets, or FLIP, reprocessed four NIRCam datasets: HIP 65426 in the F444W filter, 51 Eridani in F410M and F430M, and TYC 5899 in F444W. It recovered the already known giant planet 51 Eridani b in F430M at a reported signal-to-noise ratio of 6.3, where the original standard analysis had produced a non-detection. This was a new image of a known planet at that wavelength, not discovery of a new world.
The largest sensitivity claim was also specific, not universal. For the TYC 5899 dataset, which has no known companion, injection-recovery tests indicated roughly 15 times better sensitivity than the standard KLIP method at 0.5 arcseconds and about five times at 1 arcsecond. For 51 Eridani F430M the gains were about five times at 0.5 arcseconds and 2.5 times at 1 arcsecond. FLIP also improved the recovered signal-to-noise ratio for HIP 65426 b and 51 Eridani b in F410M. The paper supports a strong technical advance on these observations, while the phrase 'up to 15-fold' should never be read as an average across telescopes, targets or distances.[1]
What FLIP learns: corrections to an optical model, not a catalogue of planet appearances
Coronagraphs block much of a star's light, but tiny pointing, mirror and instrument changes reshape the residual point-spread function. Standard reference differential imaging combines empirical images of another star and subtracts that estimate from the science target. When the telescope state changes, the reference no longer matches perfectly and residual starlight can bury a faint companion. FLIP begins with STPSF, the physical simulator used for JWST point-spread functions, and adjusts parameters that represent the star's precise position and brightness, optical-path differences before and after the coronagraphic mask, and response variations across detector pixels.
Automatic differentiation supplies gradients for those parameters, and the authors use the Adam optimizer with an L1 loss and cosine learning-rate schedule. They first calibrate the model on a reference-star observation, putting the simulation close to the instrument state, then tune it against the science image to absorb later drift. That distinction matters when calling the method AI: it is a constrained optimization system grounded in telescope physics, not a generative model trained to imagine planets. The planet signal should remain in the residual after a better stellar model is removed. The scientific value therefore comes from combining a causal instrument model with efficient gradient-based fitting.[1]
The distinction that matters
How FLIP turns a telescope model into a cleaner planet search
The method calibrates a physical simulator before subtracting the star, then checks signal preservation with synthetic planets.
Reference image
Pre-calibrate the JWST optical model
Start near the instrument state
Science image
Tune position, wavefront and detector parameters
Match changed stellar glare
Subtraction
Remove the optimized stellar model
Leave candidate companion signals
Injection tests
Recover fake planets across separations
Measure sensitivity and signal loss
Confirmation
Repeat observations and independent reductions
Separate planets from artefacts
The comparison covers four datasets—and does not give both methods identical inputs
Researchers compared FLIP with Karhunen–Loève image processing, or KLIP, using the published hyperparameters considered optimal for each earlier analysis. KLIP combined reference-star differential imaging with angular differential imaging across both telescope roll angles, then derotated and mean-combined the images. For simplicity, FLIP processed only one roll per dataset. This is not a conventional advantage for FLIP—it used less of the science data—but it also means the comparison is not a strictly matched one-algorithm swap. Future work could apply FLIP to both rolls and test whether combining them yields further improvement or changes signal loss.
The target set intentionally spans three stellar types, two coronagraph masks, three filters and different intervals between reference and science observations. That breadth is useful, but four datasets remain a small denominator. Two systems contain known planets, while TYC 5899 provides a difficult no-known-companion case where synthetic planets can be injected. A method that works across these conditions deserves follow-up; it does not yet establish performance in a prospective survey, on crowded scenes, on ground-based instruments or on the many brightness and stability regimes astronomers encounter. The claim is about reprocessed JWST NIRCam observations, not all exoplanet imaging.[1]
Injection-recovery guards against overfitting, but researcher choices still matter
A flexible model can accidentally fit away a real planet together with the star. The team addressed that risk by injecting simulated off-axis planets before subtraction and tracking their recovered signal every roughly 300 optimization iterations. Sensitivity tests used five to seven separations and one to three injections per separation, with each planet run separately at four to 20 times the five-sigma fundamental-noise contrast limit. Signal-to-noise typically rose, plateaued and then declined as overfitting increased, so optimization stopped around the average plateau or peak. The resulting contrast curves accounted for small-sample statistics, coronagraph throughput and algorithmic signal loss.
That is a serious validation design, but it is not fully automatic or free of tuning. Learning-rate scales were selected by coarse order-of-magnitude searches and adjusted for convergence; a final stage reduced rates to limit overfitting. The paper says results were not strongly sensitive to modest rate changes or the exact fake-planet choice, yet the stopping objective can depend on whether astronomers want broad-field sensitivity or maximum recovery at one location. Manual bad-pixel handling and a Gaussian smoothing step also enter the pipeline. Independent teams should reproduce the reductions from the released code and pre-register tuning choices on held-out targets before treating the reported gains as routine.[1]
The result beats a stability limit, not the fundamental noise limit
The authors simulated 50 pairs of reference and science observations with realistic optical-path drift and star–coronagraph misalignment to estimate the contrast KLIP would reach if telescope instability were the limiting factor. Real KLIP performance was generally within a factor of two of that estimate, while FLIP often went below it. This supports the proposed mechanism: a model that adjusts the optical state can overcome mismatch that empirical subtraction cannot. However, FLIP did not reach the fundamental photon-noise limit in any observation. Residual detector effects, degeneracies among fitted parameters, computation and signal loss still constrain performance.
Some edge cases also resist a celebratory summary. FLIP's signal loss was higher than KLIP at some larger separations, where background noise rather than starlight subtraction dominated. HIP 65426 sensitivity was not sampled inside 0.6 arcseconds because bright planet lobes contaminated the region. An inconclusive residual signal appeared beside 51 Eridani b in F430M, and a nearby star in TYC 5899 was distorted. These observations do not overturn the close-in gains, but they show why discovery claims need morphology checks, multiple epochs and independent reductions rather than a threshold crossing alone.[1]
What changes for astronomy—and what evidence would change this assessment
For astronomers, better subtraction can extract more value from observations already stored in public archives and improve the scientific return from scarce telescope time. Close-in sensitivity is especially important because smaller angular separations can correspond to more Solar-System-like orbital distances. The public JWST programmes, code and example data make replication possible. Still, this paper does not show an Earth analogue, establish a planet's atmosphere or justify claims that model-based subtraction alone solves future habitable-world imaging. Any candidate found through FLIP would still require repeat observation and astrophysical confirmation.
Confidence would rise with blinded prospective tests on withheld JWST targets, multiple independent reduction teams, matched one-roll and two-roll comparisons, and pre-specified injection grids and stopping rules. A survey should report both recoveries and false positives, computational cost, uncertainty in calibrated instrument parameters and performance across stellar brightness, filters and reference strategies. The assessment would weaken if apparent gains disappear on held-out data or if residual artefacts repeatedly mimic companions. This is consequential peer-reviewed research, but it is not Breaking: the paper improves an analysis method and recovers a known planet; it does not report an urgent discovery or active hazard.[1]