Zoobot: Galaxy Morphology Models

• Zoobot is a family of transfer-learning deep models that use Galaxy Zoo labels to produce probabilistic vote fractions mimicking citizen-science classifications.
• It employs diverse architectures—from EfficientNet to ConvNeXt and U-net—to support tasks like classification, regression, ranking, and segmentation across multiple surveys.
• The framework demonstrates high sample-efficient transfer and adaptability, though its performance can be affected by domain shifts and detection limitations.

Searching arXiv for Zoobot foundation and recent application papers to ground the article. arXiv Search Query: Zoobot galaxy morphology Walmsley 2023 arXiv Zoobot is the Galaxy Zoo team’s family of galaxy foundation models designed to generalize visual morphology judgments across surveys and tasks. In its canonical form, it is a probabilistic deep-learning model for galaxy morphology that reproduces the Galaxy Zoo decision tree, outputting calibrated probabilities or vote fractions corresponding to the fraction of citizen-science votes for each morphological question; in later work, the same pretrained representation has also been adapted to binary and multiclass classifiers, regressors, ranking systems, and segmentation networks (Walmsley et al., 2021, Collaboration et al., 19 Mar 2025). Across the literature, “Zoobot” therefore denotes both a specific pretrained morphology engine and a broader transfer-learning framework for astronomical imaging, spanning DECaLS, HST, HSC-SSP, DESI Legacy Surveys, Pan-STARRS, Euclid, JWST, and UNIONS (O'Ryan et al., 2023, Pearson et al., 26 Sep 2025).

1. Origins and model family

Zoobot emerged from deep supervised representation learning on Galaxy Zoo morphology tasks. The foundational formulation used an EfficientNet-B0 backbone trained to answer every Galaxy Zoo DECaLS question, then removed the final classification layer and treated the 1280-dimensional global-average-pooled vector as a transferable morphology embedding (Walmsley et al., 2021). In that setting, the pretrained DECaLS CNN weights, embedding extraction code, and fine-tuning utilities were released as a public package intended for astronomers with little or no deep learning experience (Walmsley et al., 2021).

Subsequent work established that Zoobot is not a single fixed architecture but a model family. Older multitask implementations used EfficientNet-B0 and Dirichlet–Multinomial outputs for Galaxy Zoo vote counts (Butterworth et al., 27 Apr 2026, Collaboration et al., 2024). Euclid Q1 morphology work used a pretrained encoder that produces 512-dimensional image embeddings and trains a new linear head with one unit per morphology answer while freezing the base encoder (Collaboration et al., 19 Mar 2025). Other studies used ConvNeXt-Nano as the Euclid Q1 encoder, with a 640-dimensional penultimate embedding (Wu et al., 27 Oct 2025), ConvNeXt-Base for JWST AGN-fraction regression (Margalef-Bentabol et al., 2024), ResNet50 as a Faster R-CNN backbone for clump detection (Popp et al., 2023), and a U-net-style residual encoder-decoder dubbed ZooBot:3D for per-pixel segmentation of spiral arms and bars (Spindler et al., 15 Jun 2026).

This architectural variability reflects a consistent design principle rather than a single network definition: broad pretraining on Galaxy Zoo labels supplies a morphology-rich representation, and downstream studies then attach survey-specific or task-specific heads. The common empirical claim is that this representation transfers efficiently to new domains with limited custom labels (Walmsley et al., 2021, Omori et al., 2023).

2. Prediction formalism and output heads

The classical Zoobot output is a set of predicted vote fractions that approximate how Galaxy Zoo volunteers would answer each node of a hierarchical decision tree. In Euclid preparation work, a fine-tuned model predicted 13 morphology questions with 40 answers, using outputs treated as Dirichlet parameters for each question’s answer probabilities; the loss exploited the conjugacy of Dirichlet and Multinomial and handled missing answers by applying no gradient when a question had no volunteer responses (Collaboration et al., 2024). In Euclid Q1, the catalogue reports both per-answer “fraction” estimates and Dirichlet parameters, with

$E(X_i) = \alpha_i / \sum \alpha$

and

$$Var(X_i) = \alpha_i (1 - \alpha_i) / (1 + \sum \alpha),$$

and predictions are set to NaN when the leaf probability along the decision-tree path falls below 0.5 (Collaboration et al., 19 Mar 2025).

This probabilistic morphology formalism enables explicit scientific selection rules. In Euclid Q1 bar-fraction work, barred discs are selected by the conjunction

$$p_\mathrm{feature} > 0.5,\quad p_\mathrm{edge\text{-}on} < 0.5,\quad p_\mathrm{bar} > 0.5,$$

with

$f_\mathrm{bar} = N_\mathrm{bar}/N_\mathrm{gal}$

and 68% intervals from a beta–binomial posterior (Collaboration et al., 19 Mar 2025). In the 4MOST CHANCES cluster-filament study, Zoobot vote fractions for “Merger (M-M)” and “Major Disturbance (M-D)” were thresholded as

$FMM \ge 0.5 \ \text{or}\ FMD \ge 0.5$

to define a high-purity merger sample (Dulcien et al., 6 Mar 2026).

Later applications repurposed the pretrained representation with different output heads. Warp detection in Pan-STARRS and Euclid used a custom binary head on top of a Zoobot ConvNeXT-Nano encoder, with global average pooling, a fully connected layer with 128 units, ReLU, dropout 0.2, and a final linear layer with two logits (Suguna et al., 12 Jun 2026). Merger-stage classification on mock JWST images used one-stage and two-stage softmax heads for non-merger, pre-merger, and post-merger labels (Graaff et al., 10 Feb 2025). AGN–host decomposition reformulated the problem as regression, attaching a sigmoid regression head that predicts the aperture-defined nuclear light fraction

$$f_{\rm AGN} = \frac{F_{\rm AGN}}{F_{\rm AGN} + F_{\rm host}}$$

in JWST F150W (Margalef-Bentabol et al., 2024). ZooBot:3D moved from global labels to dense outputs, predicting two-channel soft segmentation maps for spiral arms and bars by MAE regression to volunteer pixel fractions (Spindler et al., 15 Jun 2026).

The result is a heterogeneous but internally coherent formalism: one pretrained morphology backbone, multiple probabilistic heads, and outputs that remain closely tied either to volunteer consensus or to physically defined target quantities.

3. Transfer learning and survey adaptation

Transfer learning is the central operational mechanism of Zoobot. The original DECaLS work showed that models pretrained on all Galaxy Zoo DECaLS tasks outperform ImageNet-pretrained or from-scratch models when fine-tuned to new morphology problems with limited labels, and that multi-task pretraining is materially better than pretraining on any single DECaLS task (Walmsley et al., 2021). That result became the template for later survey adaptation.

Euclid Q1 provides the clearest large-scale example. The morphology catalogue for 378,000 bright or extended galaxies was created by finetuning Zoobot galaxy foundation models on annotations from a one-month Galaxy Zoo campaign in which 9,976 volunteers contributed 2.9 million annotations (Collaboration et al., 19 Mar 2025). The Euclid adaptation used three cutout presentations—an RGB composite with $R = Y_E$, $B = I_E$, $G = \mathrm{mean}(R,B)$, a greyscale $I_E$ image at maximum resolution, and a low-surface-brightness-enhanced greyscale $$Var(X_i) = \alpha_i (1 - \alpha_i) / (1 + \sum \alpha),$$0 rendering—to align volunteer judgement and model inputs with Euclid imaging characteristics (Collaboration et al., 19 Mar 2025). The separate Euclid Q1 bar-fraction analysis states explicitly that the model was fine-tuned on new citizen-science labels for Euclid galaxies collected between Aug–Sep 2024 to address domain shift from prior Galaxy Zoo datasets to Euclid VIS imaging (Collaboration et al., 19 Mar 2025).

Other domains required different adaptation strategies. HSC-SSP merger identification froze a Zoobot backbone pretrained on Galaxy Zoo DECaLS and fine-tuned a compact binary head using roughly 1,200 HSC-realistic synthetic images from TNG50, achieving a sample-efficient transfer from DECaLS to deeper HSC imaging (Omori et al., 2023). HST archival interaction classification fine-tuned Zoobot from DECaLS to single-band HST F814W images by freezing 4,048,989 feature-extraction parameters and retraining a smaller head of 86,209 parameters, followed by optional low-learning-rate unfreezing (O'Ryan et al., 2023). Warp identification fine-tuned Zoobot’s ConvNeXT-Nano encoder on 1,000 Pan-STARRS i-band FITS cutouts and then applied the same pipeline to Euclid Q1 VIS cutouts without explicit domain adaptation (Suguna et al., 12 Jun 2026).

Several studies adapted the method to highly specialized survey configurations. EGIDE trained three separate band-specific classifiers for gri, grz, and gi subsets because DESI DR10 does not uniformly provide all griz bands, and then used an ensemble of nine independently fine-tuned models for each subset (Marchuk et al., 15 Jun 2026). GOBLIN fine-tuned a ConvNeXt-Nano Zoobot on aligned UNIONS gri cutouts using soft expert labels and then averaged an ensemble of 10 fold-trained models (Heesters et al., 23 May 2025). Strong-lens discovery in Euclid Q1 fine-tuned a ConvNeXt-Nano Zoobot by partially freezing the pretrained backbone and found that fine-tuning the last three layers optimized early-rank retrieval (Collaboration et al., 19 Mar 2025).

These applications collectively indicate that Zoobot is best understood as a transfer-learning substrate for morphology rather than as a survey-specific classifier.

4. Scientific deployments

Zoobot has been used to generate survey-scale morphology catalogues. The Euclid Q1 “First visual morphology catalogue” measured bars, spiral arms, mergers, bulges, clumps, and artefacts for approximately 378,000 bright or extended galaxies, described as the first 0.4% of the approximately 100 million galaxies where Euclid will ultimately resolve detailed morphology (Collaboration et al., 19 Mar 2025). A companion Euclid Q1 bar-fraction study used Zoobot to identify 7,711 barred galaxies with $$Var(X_i) = \alpha_i (1 - \alpha_i) / (1 + \sum \alpha),$$1 over 63.1 deg$$Var(X_i) = \alpha_i (1 - \alpha_i) / (1 + \sum \alpha),$$2, finding a mean bar fraction of $$Var(X_i) = \alpha_i (1 - \alpha_i) / (1 + \sum \alpha),$$3 and a steeper decline with redshift in lower-mass systems (Collaboration et al., 19 Mar 2025). In Galaxy Zoo: Cosmic Dawn, Zoobot contributed probabilistic morphology outputs and uncertainties for 45,742 subjects, including more than 41,000 galaxies in ultra-deep HSC imaging (Pearson et al., 26 Sep 2025).

The framework has also supported highly specialized morphology catalogues. EGIDE built a catalogue of 149,215 edge-on galaxy candidates in DESI DR10 using a Zoobot classifier fine-tuned specifically for the binary question “edge-on or not edge-on,” followed by manual supervision (Marchuk et al., 15 Jun 2026). GOBLIN used a fine-tuned Zoobot to classify $$Var(X_i) = \alpha_i (1 - \alpha_i) / (1 + \sum \alpha),$$4 million UNIONS low-surface-brightness candidates and identified 42,965 dwarf candidates with probability $$Var(X_i) = \alpha_i (1 - \alpha_i) / (1 + \sum \alpha),$$5, including 23,072 with probability $$Var(X_i) = \alpha_i (1 - \alpha_i) / (1 + \sum \alpha),$$6 (Heesters et al., 23 May 2025). A large HST archival search applied Zoobot to 126 million Hubble Source Catalogue extended sources in ACS/WFC F814W and, after representation-space filtering and visual vetting, produced a clean catalogue of 21,926 interacting systems (O'Ryan et al., 2023).

Beyond catalogue production, Zoobot has been adapted to targeted scientific measurements. In “Caught in the web: galaxy mergers along cosmic filaments,” Zoobot merger-related vote fractions identified 698 mergers among 43,922 galaxies in the 4MOST CHANCES low-$$Var(X_i) = \alpha_i (1 - \alpha_i) / (1 + \sum \alpha),$$7 sample, enabling the result that mergers lie significantly closer to filaments than non-mergers, especially outside $$Var(X_i) = \alpha_i (1 - \alpha_i) / (1 + \sum \alpha),$$8 (Dulcien et al., 6 Mar 2026). In Subaru HSC-SSP, a Zoobot-based merger score $$Var(X_i) = \alpha_i (1 - \alpha_i) / (1 + \sum \alpha),$$9 was used to study merger incidence versus stellar-mass overdensity, with higher merger scores favoring lower-density environments on scales of $$p_\mathrm{feature} > 0.5,\quad p_\mathrm{edge\text{-}on} < 0.5,\quad p_\mathrm{bar} > 0.5,$$0 to $$p_\mathrm{feature} > 0.5,\quad p_\mathrm{edge\text{-}on} < 0.5,\quad p_\mathrm{bar} > 0.5,$$1 in both simulations and observations (Omori et al., 2023). For merger stages in mock JWST images, one-stage classification achieved $$p_\mathrm{feature} > 0.5,\quad p_\mathrm{edge\text{-}on} < 0.5,\quad p_\mathrm{bar} > 0.5,$$2 accuracy, moderately outperforming a two-stage setup at $$p_\mathrm{feature} > 0.5,\quad p_\mathrm{edge\text{-}on} < 0.5,\quad p_\mathrm{bar} > 0.5,$$3 (Graaff et al., 10 Feb 2025).

Other deployments moved outside classical morphology classification. Fine-tuned Zoobot identified 2,088 warped and 1,398 non-warped Pan-STARRS galaxies at $$p_\mathrm{feature} > 0.5,\quad p_\mathrm{edge\text{-}on} < 0.5,\quad p_\mathrm{bar} > 0.5,$$4, and 1,209 warped plus 1,025 non-warped galaxies in Euclid Q1 VIS at the same threshold (Suguna et al., 12 Jun 2026). In AGN–host decomposition, a Zoobot regressor recovered injected JWST F150W AGN fractions with mean difference $$p_\mathrm{feature} > 0.5,\quad p_\mathrm{edge\text{-}on} < 0.5,\quad p_\mathrm{bar} > 0.5,$$5, RMSE $$p_\mathrm{feature} > 0.5,\quad p_\mathrm{edge\text{-}on} < 0.5,\quad p_\mathrm{bar} > 0.5,$$6, RAE $$p_\mathrm{feature} > 0.5,\quad p_\mathrm{edge\text{-}on} < 0.5,\quad p_\mathrm{bar} > 0.5,$$7, and outlier fraction 6.5%, substantially outperforming GALFIT on the same dataset (Margalef-Bentabol et al., 2024). ZooBot:3D produced soft segmentation maps for spiral arms and bars for 639,636 DESI Legacy Survey galaxies, and also generated cross-matched segmentation products for 29,006 MaNGA and 5,019 SAMI targets (Spindler et al., 15 Jun 2026). In Euclid strong-lens discovery, a fine-tuned Zoobot ranked one million Q1 objects and placed 122 grade A and 41 grade B lenses in its top 1,000 candidates (Collaboration et al., 19 Mar 2025). In clump science, a Zoobot-ResNet50 backbone inside Faster R-CNN achieved GSFC detections with completeness and purity of $$p_\mathrm{feature} > 0.5,\quad p_\mathrm{edge\text{-}on} < 0.5,\quad p_\mathrm{bar} > 0.5,$$8 while trained on $$p_\mathrm{feature} > 0.5,\quad p_\mathrm{edge\text{-}on} < 0.5,\quad p_\mathrm{bar} > 0.5,$$9 galaxy images (Popp et al., 2023).

The range of these deployments shows that Zoobot has evolved from a Galaxy Zoo vote-fraction emulator into a general infrastructure for morphology-conditioned inference at scale.

5. Scaling, validation, and interpretability

A recurrent theme in the Zoobot literature is that operational performance depends not only on architecture but on uncertainty handling, data curation, and inspection budget. In Galaxy Zoo: Cosmic Dawn, Zoobot predicted Dirichlet distributions over vote fractions and used the lower 90 per cent confidence bound on $f_\mathrm{bar} = N_\mathrm{bar}/N_\mathrm{gal}$0 to retire simple subjects early; 77.8 per cent of galaxies were retired before reaching the nominal volunteer depth, the mean volunteer votes per subject dropped from 40 to 14, and the workflow accelerated by a factor $f_\mathrm{bar} = N_\mathrm{bar}/N_\mathrm{gal}$1 (Pearson et al., 26 Sep 2025). In Euclid Q1 morphology, high-confidence volunteer labels on the validation subset were matched with accuracies above 99% for 7 of 13 questions and never below 95%, while regression-style vote-fraction deviations were typically $f_\mathrm{bar} = N_\mathrm{bar}/N_\mathrm{gal}$2 (Collaboration et al., 19 Mar 2025).

Studies that examined training protocol found that performance gains are not unlimited. The augmentation analysis on Galaxy Zoo DECaLS trained 24 EfficientNet-B0 Zoobot variants and concluded that augmentations improve performance most when data are scarce, but the improvement is significantly diminished as training dataset size increases; for a fixed-capacity model of approximately 5.3 million parameters, a saturation point appears by 50–100% training size for several questions, and complex augmentations increase total training time without robust gains at large scale (Butterworth et al., 27 Apr 2026). This is important because later large-survey applications often inherit the assumption that more augmentation is always beneficial; the DECaLS study does not support that generalization (Butterworth et al., 27 Apr 2026).

Interpretability has also become a substantive topic. Warp classification used LayerCAM and found that warped galaxies show enhanced activation in outer-disc regions where the mid-plane deviates from linearity, while non-warped galaxies exhibit symmetric activation distributed along the disc plane (Suguna et al., 12 Jun 2026). The Euclid Q1 sparse-autoencoder study treated Zoobot as a supervised baseline and showed that SAEs trained on its 640-dimensional embeddings retain stronger average alignment with Galaxy Zoo labels than PCA among the top 64 features—$f_\mathrm{bar} = N_\mathrm{bar}/N_\mathrm{gal}$3 versus $f_\mathrm{bar} = N_\mathrm{bar}/N_\mathrm{gal}$4 for all classes—and also surface interpretable features outside the Galaxy Zoo decision tree, including dust lanes in edge-on disks and elliptical galaxies with bluer companions (Wu et al., 27 Oct 2025).

This suggests that Zoobot’s internal representation is not exhausted by the original human-defined taxonomy. The model space remains strongly aligned with Galaxy Zoo labels, but post-hoc analysis indicates additional structure that can support discovery-oriented workflows.

6. Limitations and outlook

The principal limitations reported across the literature concern detectability, domain shift, and the distinction between apparent and intrinsic morphology. A Euclid-like mock study of TNG50 bars found that, at $f_\mathrm{bar} = N_\mathrm{bar}/N_\mathrm{gal}$5, Zoobot recovers only 31 of 141 barred galaxies in VIS $f_\mathrm{bar} = N_\mathrm{bar}/N_\mathrm{gal}$6 mocks and fails outside VIS in a representative borderline case; when non-detections are counted as unbarred, the apparent bar fraction drops from 44 percent in idealized mass maps to 10 percent for Zoobot and 34 percent for the combined priority method (Gonçalves et al., 22 Jan 2026). The implication is explicit in that work: theory–observation comparisons must be performed on mocks with realistic instrumental conditions and, ideally, with the same classification pipeline used on survey data (Gonçalves et al., 22 Jan 2026).

Merger work reaches similar conclusions by different routes. In the CHANCES filament study, conservative thresholds produce high purity but only 45% completeness for clear mergers, low-surface-brightness tidal features may not exceed the thresholds, and projection plus photometric-redshift uncertainties tend to dilute merger–filament associations (Dulcien et al., 6 Mar 2026). JWST merger-stage classification likewise found that precision and recall depend strongly on merger time relative to coalescence and that post-mergers are frequently confused with non-mergers (Graaff et al., 10 Feb 2025). HSC merger identification reported that minor and mini mergers often occupy intermediate probabilities and that no explicit probability calibration was applied beyond thresholding (Omori et al., 2023).

Other tasks expose additional failure modes. Warp detection is limited by the faint outskirts of edge-on discs, mild overfitting, and the absence of explicit domain adaptation between Pan-STARRS and deeper Euclid images (Suguna et al., 12 Jun 2026). GOBLIN notes that, without distance information, unresolved background spirals can appear dwarf-like, and that structural parameters from MTObjects are approximate rather than precision photometry (Heesters et al., 23 May 2025). AGN decomposition is explicitly band- and aperture-defined—F150W within a fixed 2″ aperture—so application to other bands or apertures requires retraining or careful recalibration (Margalef-Bentabol et al., 2024). Strong-lens discovery in Euclid Q1 emphasizes that Zoobot scores are not calibrated probabilities and that the ranker remains biased toward larger-radius, higher-S/N systems unless retrained with more representative real lenses and targeted simulations (Collaboration et al., 19 Mar 2025).

The current record therefore supports a precise characterization. Zoobot is not a universal morphology oracle; it is a pretrained morphology representation that can be adapted efficiently, often with unusually strong sample efficiency, but whose scientific validity depends on task-specific thresholding, survey-matched preprocessing, realistic mocks, and careful accounting of observational selection. A plausible implication is that future Zoobot development will continue in three coupled directions: instrument-aware fine-tuning, richer output structures such as dense segmentation and physically defined regression targets, and post-hoc interpretability methods that expose morphology-relevant features beyond the original Galaxy Zoo decision tree (Spindler et al., 15 Jun 2026, Wu et al., 27 Oct 2025).