POLARIS: Exoplanet Imaging ML Benchmark

• POLARIS is an astronomical archive and machine-learning benchmark that uses uniformly reduced SPHERE/IRDIS polarized-light data to enhance direct imaging of exoplanetary systems.
• It frames reference star selection as a semi-supervised representation learning and classification problem with less than 10% manually labeled samples.
• The Diff-SimCLR framework integrates diffusion priors into contrastive learning, achieving high accuracy (up to 93%) in distinguishing circumstellar disks from reference stars.

Searching arXiv for the benchmark paper and closely related exoplanet imaging context. POLARIS, expanded as POlarized Light dAta for total intensity Representation learning of direct Imaging of exoplanetary Systems, is an astronomical archive and machine-learning benchmark positioned at the intersection of high-contrast imaging of exoplanetary systems and modern representation learning. It is built from the full public SPHERE/IRDIS polarized-light archive from 2014–2024 and is designed to automate the selection of disk-free “reference” stars for reference differential imaging, thereby improving the recovery of faint circumstellar disks and, ultimately, exoplanets at scale. The benchmark frames reference selection as a representation-learning and classification problem over uniformly reduced polarimetric observations, with less than 10 percent manual labeling (Cao et al., 4 Jun 2025).

1. Scientific setting and observational motivation

Direct imaging of exoplanets and circumstellar disks is a high-contrast imaging problem in which an extremely faint, compact, or extended signal lies adjacent to a bright star. For Sun-like stars, the planet–star contrast is typically $\sim 10^{-6}$ for massive, Jupiter-like planets imaged in the near-IR today and $\sim 10^{-10}$ for an Earth twin in visible light, beyond current ground-based capabilities. Even with coronagraphs and extreme adaptive optics, the stellar point spread function and speckles dominate the residual image, so the astrophysical signal remains comparable to residual noise (Cao et al., 4 Jun 2025).

Within this setting, three differential-imaging strategies are central. Angular Differential Imaging exploits field rotation, but it can self-subtract extended disk structures and distort their morphology. Reference Differential Imaging instead subtracts the point spread function of a different, disk-free reference star, preserving extended circumstellar structure much better and therefore remaining essential for disk characterization. Polarimetric Differential Imaging, particularly with SPHERE/IRDIS, measures Stokes $Q$ and $U$ and reconstructs polarized intensity; scattered light from dust is polarized, whereas stellar light is mostly unpolarized, so disks become visible with fewer artifacts. POLARIS addresses the operational bottleneck created by this workflow: expert humans still inspect PDI products and total-intensity images to decide whether a system is free of circumstellar structure and can serve as a reference, a procedure that is labor-intensive, subjective, and poorly matched to archives containing millions of images.

2. Archive construction and data products

POLARIS is derived from a decade of public dual-polarization imaging observations with SPHERE/IRDIS on the ESO Very Large Telescope. During a typical 1–2 hour sequence on one star, IRDIS cycles through the four polarization states $\{Q^+, Q^-, U^+, U^-\}$, producing $4n$ individual exposures for $n$ cycles. The archive is uniformly reduced with an adapted IRDAP pipeline: bad-pixel and flat-field corrections are applied, stars are centered using dedicated centering exposures, sky and thermal background are removed, ordinary and extraordinary beams are combined to reconstruct Stokes parameters, PDI post-processing yields the final polarized-intensity products, and bad exposures are manually removed before re-running IRDAP to enforce consistent quality (Cao et al., 4 Jun 2025).

The archive has two main layers: postprocessed PDI images used for labeling and representation learning, and the underlying preprocessed total-intensity exposure sequences used for background modeling and subtraction.

Data product	Scale and format	Function
PDI-postprocessed “Q files”	921 images; 96 manually labeled; FITS, JPEG, NumPy	Disk/reference classification
Preprocessed exposure sequences	75,910 exposures linked to the 921 Q products; FITS, NumPy	Total-intensity background modeling
ML-ready crops	$256 \times 256$ central crops	Representation learning input

The original Q images are $1024 \times 1024$ pixels, but machine-learning experiments use the central $256 \times 256$ crop. Total-intensity exposures are stored as $\sim 10^{-10}$0 per original file and, for machine learning, cropped to $\sim 10^{-10}$1 by stacking the polarization cycles. Pixel values in total-intensity exposures span roughly $\sim 10^{-10}$2 counts s$\sim 10^{-10}$3 pixel$\sim 10^{-10}$4; negative values are set to zero, a log transform stabilizes dynamic range, and images are rescaled to $\sim 10^{-10}$5. For some tasks, including single-frame PDI classification, images are normalized to $\sim 10^{-10}$6 and saved as JPEGs. The full archive exceeds 200 GB, whereas the Q-image subset occupies only a few hundred MB.

3. Labeling strategy, task design, and evaluation protocol

A defining feature of POLARIS is its sparse supervision regime. Only 96 of the 921 Q images are manually labeled, using a binary scheme in which target denotes a system with clearly visible circumstellar disk structure in polarized light and reference denotes a non-detection of such structure. The remaining 813 unlabeled PDI images are assigned labels through unsupervised feature learning followed by spectral clustering, with the PDI image acting as a high-signal summary from which the associated total-intensity exposure sequence inherits its label. This suggests a deliberate benchmark design for semi-supervised and unsupervised learning, rather than a conventional fully supervised archive (Cao et al., 4 Jun 2025).

The benchmark supports four connected tasks. The primary task is reference-versus-disk classification on single PDI Q images. A second task is unsupervised representation learning over all 921 PDI images, with a 32-dimensional latent representation used for downstream classification or clustering. A third task is background reconstruction for total-intensity imaging: automatically identified reference sequences are used to train generative models that learn the distribution of stellar point spread functions, then estimate the background of a target sequence for subtraction. A fourth task evaluates zero-shot classification by large vision-LLMs.

The principal metric is classification accuracy,

$\sim 10^{-10}$7

reported as the mean over folds in 10-fold stratified cross-validation. For clustering, predicted clusters are matched to ground truth via the Hungarian algorithm before computing accuracy. This evaluation protocol places classical downstream classifiers and unsupervised clustering methods on a common footing.

4. Baseline models and the Diff-SimCLR framework

POLARIS evaluates several families of models that produce 32-dimensional image representations. The baseline self-supervised and unsupervised methods are a masked autoencoder, DeepCluster, and SimCLR. The central methodological contribution is Diff-SimCLR, described as a latent-enhanced contrastive learning framework that injects information from a conditional denoising diffusion probabilistic model into SimCLR’s representations. A DDPM is first trained on POLARIS images for 300 epochs with learning rate $\sim 10^{-10}$8 and batch size 16. For each image, the first $\sim 10^{-10}$9 reverse-process states,

$Q$0

are extracted as a latent trajectory, and the paper reports that $Q$1 gives the best trade-off between informativeness and noise. These diffusion priors are encoded with the same ResNet backbone used for the augmented image views, concatenated with per-view features, and optimized with the standard InfoNCE loss (Cao et al., 4 Jun 2025).

Quantitatively, Diff-SimCLR gives the strongest supervised downstream performance on the 96 labeled Q images under 10-fold stratified cross-validation. With a linear SVC on 32-dimensional features, Diff-SimCLR reaches 93.00% accuracy, compared with 86.46% for SimCLR, 85.00% for the masked autoencoder, and 74.00% for DeepCluster. In unsupervised spectral clustering, the accuracies are lower but still comparable across the best methods: 77.33% for Diff-SimCLR, 77.78% for SimCLR, 77.00% for the masked autoencoder, and 74.89% for DeepCluster.

Zero-shot large vision-LLMs are also evaluated with a domain-specific prompt that asks the model to act as an exoplanet astronomer and classify a PDI image as target or reference. Among the tested systems, Gemini-2.0-Flash reaches 75.21% accuracy, GPT-4.1 reaches 75.00%, and GPT-4o reaches 67.71%, while open-source LVLMs perform around 49–52%. The benchmark therefore separates task-specific representation learning from generic multimodal reasoning rather clearly.

5. Polarimetric representation learning and astrophysical use

POLARIS uses the IRDIS PDI products called “Q files” as the principal supervisory signal because they encode morphology in a form where stellar light is largely canceled and polarized dust scattering remains visible. The machine-learning representation is therefore not learned from raw Stokes cubes but from single-channel grayscale PDI products whose contrast structure already reflects the polarimetric measurement process. In many disk studies, Stokes quantities are transformed into radial coordinates such as $Q$2 and $Q$3; POLARIS instead centers its benchmark on the IRDAP-derived Q products as released in the archive (Cao et al., 4 Jun 2025).

Several domain-specific preprocessing choices are operationally important. The central $Q$4 crop is used around the star; the central 8 pixels are blocked by the coronagraph; and for generative background reconstruction an 80-pixel radius is often masked so that the model emphasizes the point spread function halo instead of saturated central structure. After background imputation, residuals are rotated to a standard sky orientation using the parallactic-angle metadata in parangs.fits, with angle $Q$5 degrees.

Astrophysical validation is obtained by training a variational autoencoder on automatically identified reference exposures and using it for point spread function reconstruction in total-intensity data. The reconstructed backgrounds show realistic stellar point spread function structures, including Airy rings, and residuals after subtraction recover known disks such as V351 Ori, HD 37400, and PDS 70-like disks with morphological detail consistent with published PDI images. In the representation space itself, t-SNE and PCA visualizations show that Diff-SimCLR features separate target systems from reference stars even without labels, and the target cluster exhibits hints of substructure associated with different morphologies and inclinations, including spirals, rings, and edge-on disks. This suggests that the learned representation captures more than a binary disk/non-disk distinction.

6. Position within high-contrast imaging, limitations, and reuse

POLARIS differs from prior high-contrast imaging workflows by combining scale, uniform reduction, label structure, and explicit representation-learning benchmarks. Existing post-processing methods such as KLIP, PCA-based methods, nonnegative matrix factorization, REXPACO, and 4S are usually evaluated on small numbers of systems, whereas POLARIS packages the entire public SPHERE/IRDIS PDI archive from 2014–2024, processed through a single validated pipeline and organized for supervised, semi-supervised, and unsupervised learning. The paper characterizes it as the first uniformly reduced, high-quality exoplanet imaging dataset, a form of resource described as rare in astrophysics and machine learning (Cao et al., 4 Jun 2025).

The release is public on Zenodo in two versions, with v2.0 described as the latest version consistent with the Croissant metadata in the appendix. It includes 96 labeled PDI-postprocessed images, 813 unlabeled PDI images from 2014–2023 with more to be added as public data accumulate, and all corresponding preprocessed exposure sequences from 2014–2024. The reduction workflow is based on an adjusted IRDAP pipeline, and the machine-learning baselines, including MAE, DeepCluster, SimCLR, Diff-SimCLR, and the VAE, are implemented in PyTorch on NVIDIA A5000 GPUs.

The benchmark has several explicit limitations. The labeled subset emphasizes bright protoplanetary disks; fainter debris disks are present in the archive but may be mis-labeled as references during automatic propagation. The assumption that non-detection in polarized light implies non-existence in total intensity is safe for disks but not for planets, which may be bright in total intensity while remaining essentially unpolarized. The current release is SPHERE/IRDIS-only, so generalization to Gemini Planet Imager, CHARIS, VAMPIRES, Roman’s CGI, or the Habitable Worlds Observatory remains plausible rather than demonstrated. The authors identify better handling of faint debris disks, explicit planet detection and characterization, incorporation of physical priors such as temporal coherence across epochs and orbital motion, and cross-instrument generalization as natural next steps. A broader implication stated in the benchmark is that automated reference selection could reduce dedicated reference-star observing time by up to ~50%, allowing large public archives of previously observed stars to be reused more systematically for future high-contrast campaigns.