Adapting Exoplanet Detectors to Each New Observation Without Erasing the Planet

Photographing a planet next to its star means finding a point of light that can be a million to ten billion times fainter than the star, inside a field of "speckles": blobs of leftover starlight that look just like planets. Neural detectors trained on archives of past telescope data now beat classical methods. The catch is that every new observation comes with its own atmospheric seeing, stellar brightness and optical drift, so its speckles follow a different distribution. The usual machine-learning fix for this mismatch is to adapt the model on the unlabeled new data, either by aligning features with an adversarial "domain discriminator" or by minimizing the model's uncertainty. Here that fix is dangerous. The one thing that distinguishes the new observation is exactly what we are looking for, and an adaptation objective that cannot tell planet from speckle will quietly teach the model to ignore faint planets.

This note proposes a way out that uses how these observations are taken. In angular differential imaging, the telescope keeps its optics fixed while the sky slowly rotates across the detector. Speckles stay put and a real planet traces a known arc. If we feed the observation to the detector with the wrong rotation angles (for example, with the angles reversed), the speckles are unchanged but any real planet no longer follows a trajectory the detector expects. The result is a "null cube": target-domain noise with no coherent planet signal in it. We inject synthetic planets into this null cube and then adapt a small set of normalization parameters of an archive-pretrained detector. The adaptation is ordinary supervised learning, with no discriminator and no minimax game. The claim is that this per-observation, planet-blind adaptation matches or beats retraining a detector from scratch for each target, at a small fraction of the compute. We also claim it improves on a frozen archive model, and that it keeps false-alarm rates calibrated where statistics-only and adversarial adaptation do not.

Why this problem, and what already exists

The field has split into two camps that each pay a price. deep PACO first whitens each observation with a local patch-covariance model. It then trains a U-Net detector from scratch on synthetic planets injected into temporally shuffled frames of that same observation. It reports 2 to 5 times better sensitivity than PACO alone. Because it is fit to one observation, it learns nothing across the archive and has to be retrained per target. Archive-trained models take the other side. MODEL&CO trains on 322 SPHERE-IRDIS cubes and reports an eight- to tenfold contrast gain over PACO at short separations. ExoMILD trains a learnable patch-statistics model with a U-Net prior on 220 SHINE observations and reports AUC 0.761 in multi-spectral mode, against 0.693 for PACO and 0.515 for PCA/KLIP. Mono-spectral ADI mode reaches an AUC of 0.645 (Deep learning for exoplanet detection and characterization by direct imaging at high contrast). Both models are frozen at test time. MODEL&CO handles cross-observation variation with heuristic temporal masking and still shows vulnerability to outlier observations. These papers motivate the proposal: archive learning helps, per-observation fitting helps, and nobody combines them.

Generic test-time adaptation supplies candidate methods and also constrains which ones are safe. Tent adapts only the scale and shift parameters of normalization layers by minimizing prediction entropy. That parameter restriction is what we reuse. The entropy objective has two problems here: it is undefined for flux-reconstruction and likelihood outputs, and it rewards confident mistakes, which in this setting means confidently calling a faint planet "speckle." DUA and MemBN recalibrate normalization statistics without any loss, and MemBN adds a memory queue for tiny batches. Statistics carry no notion of which pixels are signal, but they are the natural cheap competitor. Adversarial alignment is the textbook alternative for unlabeled shift, and it is known to be unstable. Task and alignment gradients conflict and representations collapse (Domain-Robust VQA With Diverse Datasets and Methods but No Target Labels; Reversing Gradients in Adversarial Domain Adaptation for Question Deduplication and Textual Entailment Tasks). Stable single-objective replacements such as MMD have beaten it in practice (Domain Adaptation for Enterprise Email Search). Our position is sharper than "use a stabler alignment loss." Once a planet-free target-domain sample exists, the adaptation problem has labels, and no alignment game is needed.

Core Claim

Write the target observation as a cube of frames with parallactic angles $\theta$. The null cube $N$ is the same frames paired with angles $-\theta$ (reversed about the mean). This breaks any real companion's trajectory while leaving speckle statistics, temporal correlations and detector noise intact. Unlike deep PACO's temporal shuffling, frame order is preserved. We keep a detector $f$ pretrained on the archive and adapt only its normalization affine parameters $\phi$:

$$\min_{\phi}\; \mathbb{E}_{s}\big[\ell\big(f_{\phi}(N \oplus s),\, s\big)\big] \;+\; \lambda\, \mathrm{Cal}\big(f_{\phi}(N)\big)$$

Here $s$ is a synthetic companion with random position and contrast, injected along a trajectory consistent with the reversed angles. $\oplus$ denotes injection, and $\ell$ is the detector's native training loss (detection plus photometry). $\mathrm{Cal}$ penalizes the gap between the empirical exceedance rate of the detection map on the uninjected null cube and the nominal false-alarm rate at a few thresholds. $\lambda$ weights that term. After adaptation, we run $f_\phi$ on the real cube with the true angles.

This matters for three reasons. The objective is ordinary minimization, so it avoids minimax instability by construction. Its only "labels" are injections into planet-free data, so there is no pressure to absorb a real companion. It also turns the synthetic-to-real gap into something measurable: injected planets are simulated, speckles are real, and we check the residual gap on known real planets.

Research Program

The program has three parts. First, build the pretrained detectors: a MODEL&CO-style reconstruction network and an ExoMILD-style statistical network, both trained on the SHINE F150 archive. We define a held-out set that is shifted on purpose, split by observing epoch, seeing and stellar magnitude, so the gap between archive and target is real rather than random.

Second, characterize the null cube. The key physical limit is small field rotation near the star. With few pixels of arc, reversed-angle trajectories overlap the true one, so the null contains residual planet flux. We will measure planet leakage into $N$ as a function of separation and total rotation, using injected planets as ground truth. We will mask separations where leakage exceeds a set fraction, and report that mask honestly as the method's inner limit.

Third, study the adaptation itself. We will compare which parameters to adapt (affine only, affine plus statistics, full network) and how many injections are needed. Compute per target will be reported in GPU-minutes against deep PACO's from-scratch retraining.

A secondary question is whether starting $\phi$ from environmental telemetry helps, for example with a small network mapping wind speed and temperature to initial affine parameters. This follows Half-sibling regression meets exoplanet imaging, which reports that telemetry improved detection metrics by factors of 1.1 to 4.0 in a linear model, and the FiLM conditioning of Domain Conditional Predictors for Domain Adaptation. This part is only an ablation, not the core claim.

Baselines

The naive reference is the archive-pretrained detector run frozen on the shifted held-out cubes. Beside it we run the classical references PCA/KLIP and PACO on identical inputs. These anchor the scale: ExoMILD's reported AUC is 0.761 against 0.515 for KLIP in multi-spectral mode, and 0.645 in mono-spectral ADI.

The strongest cross-observation baselines are MODEL&CO and ExoMILD themselves, trained on the same archive split. MODEL&CO's reported result is an eight- to tenfold contrast increase over PACO at short separations. Beating these baselines means a higher true-positive rate at a fixed false-alarm rate on the shifted held-out cubes, and deeper 5-sigma contrast curves, with the gain concentrated on the most-shifted observations.

The strongest per-observation baseline is deep PACO, retrained from scratch on each held-out target with its original shuffled-injection recipe. Its reported gains are 2 to 5 times sensitivity and up to 0.5 magnitudes in contrast over PACO. We run it with its published protocol, and also with a compute budget equal to ours, to separate "better" from "cheaper." Beating it means matching its detection performance at equal compute and exceeding it at small separations, where a single cube lacks the diversity to train from scratch.

The statistics-only baselines are DUA and MemBN applied to the same pretrained detector's normalization layers, using frames of the target cube as the stream. Their reported gains are on natural-image corruptions only; for example, DUA cut CIFAR-10C level-5 error from 49.2% to 26.8%. There is no reported result on Lacuna for either method on astronomical data cubes.

The adversarial baseline is gradient-reversal alignment between archive and target features. We give it the same tuning budget as our method and report its variance across seeds, not just its best run. An MMD alignment variant is run alongside as the stable non-adversarial comparison.

Ablations isolate the key ingredient. We adapt on the raw target cube with true angles (identical otherwise, so real planets are present in the adaptation data), on a shuffled cube (deep PACO's null), and on the reversed-angle null. We also drop the calibration term and compare full fine-tuning with affine-only updates.

All methods share the backbone, archive split and injection generator. Hyperparameters are tuned on a validation set of observations disjoint from test. Tuning budgets are equal in number of trials.

Why not just…?

Why not just retrain per observation, as deep PACO does? This is the real threat, since it already uses the target's own speckles and injections. Our method differs by keeping what the archive taught the detector and changing under 1% of parameters. The archive advantage is largest near the star, as MODEL&CO's small-separation contrast gains suggest. The equal-compute and published-protocol comparisons settle this directly. If deep PACO matches us at equal compute everywhere, the archive prior buys nothing and we will say so.

Why not just recompute normalization statistics on the target (DUA/MemBN)? It is cheaper and needs no injections. But statistics cannot distinguish a bright planet's pixels from speckles, and nothing in them calibrates false alarms. The input whitening that PACO-style front ends already apply may also leave little for statistics to fix. The statistics-only baseline, run on top of whitened inputs, tests whether our injection-supervised objective adds anything beyond re-centering. If it does not, the simpler method wins.

Why not just align archive and target features adversarially or with MMD? Alignment objectives reward making the target look like the archive. A target containing a real companion differs from the archive exactly at the companion, so alignment pushes toward erasing it. This is the invariance-destroys-faint-features failure in the adversarial literature, on top of its known optimization instability. The raw-cube ablation and the known-planet check below measure this erasure directly for both GRL and MMD.

Why not just fine-tune the whole pretrained network on target injections? This is the obvious middle ground. The risk is overfitting to one cube's few hundred frames, which is what forces deep PACO's shuffling tricks. The affine-only versus full fine-tuning ablation decides it. We expect full fine-tuning to be competitive at large separations and worse near the star.

Proposed Experiments

Evaluation uses three settings. The first is shifted held-out SHINE F150 cubes with blind injections following the Exoplanet Imaging Data Challenge protocol. The second is the public EIDC datasets that deep PACO used. The third is the HR 8799 archival sequences, whose four real planets test the synthetic-to-real gap in the injected companions. Metrics are ROC AUC, true-positive rate at fixed false-alarm rates, 5-sigma contrast versus separation, and empirical versus nominal false-alarm rate on planet-free regions. We also report photometric relative error and astrometric RMSE (ExoMILD reports 0.11 pixels and 0.51 relative flux error), compute per target, and seed variance.

Two checks decide the proposal. The first is planet preservation. On HR 8799 cubes, adapting on the reversed-angle null must leave the signal-to-noise of the known planets unchanged or improved. Adapting on the raw cube, and adversarial or MMD alignment, are predicted to lower it measurably. If the null-cube version also lowers it, the premise fails. The second is shift dependence. The improvement over the frozen archive model should grow with the measured shift between a target and the archive, for example seeing difference or feature-statistic distance, and should vanish on in-distribution targets. A flat gain would mean we are just adding capacity, not adapting.

A third, diagnostic check maps the inner working angle. We report where null-cube leakage becomes material and whether the method's advantage over deep PACO survives inside it. That region is where both the scientific payoff and the physical limit of the idea lie.