arXiv Science⌕ Search

arXiv subjects

J. Gracia-Carpio

Publications and source records attributed to J. Gracia-Carpio.

At least 19 recordsLinked to original sources

Euclid: Quick Data Release (Q1) -- The evolution of the radio-loud AGN fraction-stellar mass relation to $z\sim3$ in the Euclid Deep Field North

The incidence of radio-loud active galactic nuclei (RLAGN) and their energetic output via AGN feedback are crucial in understanding the evolution of galaxies and the large-scale structure of the Universe. In the local Universe, the incidence of RLAGN has been shown to correlate strongly with host galaxy properties, particularly stellar mass. Using the large-area survey capabilities provided by Euclid and the LOw Frequency ARray (LOFAR), we present the first study of the RLAGN incidence--stellar mass relation to $z=3.0$. We perform a stringent analysis accounting for cross-matching reliability and stellar-mass incompleteness. In addition, we test the robustness of our results against AGN and host galaxy classification. This yields a sample of $3\,154$ RLAGN in the Euclid Deep Field North, comprising $2\,803$ low-excitation radio galaxies (LERGs). We find that the LERG incidence as a function of stellar mass and radio luminosity follows a tight correlation over more than two orders of magnitude in stellar mass to $z=3.0$. The slope of the power-law relation decreases from $γ\approx2.5$ in the local Universe to $γ\approx1.0$ at $z=3.0$, indicating a relative increase in low-mass LERGs at earlier epochs. Separating the LERG population into actively star-forming and quiescent galaxies reveals that quiescent LERGs show a much steeper relation between incidence and stellar mass, with $γ\approx2.5$ compared to star-forming LERGs that show $γ\approx1.5$. However, the relation seems to be independent of whether galaxies are star-forming or quiescent at $z\geq2$, where both populations show relatively flat relations with $γ\approx1.0$. Our results suggest an evolutionary picture in which the LERG population transitions from cold-mode, merger-dominated accretion at earlier epochs to a scenario in which systems require a massive hot halo to fuel the AGN in the local Universe.

astro-ph.GA↗

Euclid Data Release 1 (DR1). Double-source-plane lens candidates

Strong gravitational-lensing systems with multiple source planes are powerful tools for probing the stellar and dark matter content of galaxies. Lens modelling can measure the relative Einstein radii of the two source planes, which encode a ratio of angular-diameter distances determined by the lens and source redshifts and the dark energy equation of state, independently of the Hubble constant. Galaxy-scale double-source-plane lenses (DSPLs) are extremely rare. We present a sample of 84 galaxy-scale DSPL candidates in Euclid Data Release 1 (DR1), of which 80 are newly identified. These systems were selected using machine-learning lens-finding models followed by visual inspection by citizen scientists and experts. We then used lens modelling to test the DSPL interpretation. Among our lens candidates, 11 form the 'Gold' sample: each shows two sets of multiple images at different radii, and the observed configuration is reproduced by a DSPL lens model. The remaining 73 'Silver' candidates have at least two lensed arcs from visual inspection, but either lack multiple images for at least one arc, have a best-fitting DSPL distance ratio close to the single-source-plane limit, remain compatible with a single-plane model with multiple sources, or have not yet been modelled. Even if all 84 candidates are DSPLs, the sample falls below the approximately 90 systems expected from extrapolating the Euclid Quick Data Release 1 (Q1) yield and the approximately 200 obtained by scaling the LensPop forecast to the DR1 area, indicating that many DSPLs remain undetected in DR1. Spectroscopy or higher-resolution imaging is required before we can be confident that the Silver candidates are DSPLs.

astro-ph.GA↗

Euclid: Data Release 1 (DR1) -- Bayesian 3D shape inference of star-forming galaxy populations

Understanding the intrinsic 3D shapes of galaxies provides crucial insights into their formation and evolution, but determining these from projections remains challenging. We apply a Bayesian inference framework to optical images of galaxies from the Euclid Data Release 1, leveraging its large sample size to test the scalability and robustness of this methodology. Applying Hamiltonian Monte Carlo (HMC) sampling to Sérsic axis ratios and effective radii, we constrain the population-level mean 3D ellipticity, triaxiality, and size, together with their covariances for $40\,864\,304$ star-forming galaxies with stellar masses $M_\ast=10^{9}$--$10^{10.5}M_\odot$ at redshifts $0<z<1.2$. Our sample size exceeds previous shape inference studies by factors of $440$--$52\,800$ relative to early observations by JWST, and factors of $440$--$17\,600$ relative to legacy observations from HST in overlapping mass-redshift bins, enabling finer binning alongside improved statistical constraints on shape evolution. We additionally include over 9 million galaxies at $z<0.5$, a regime historically limited by sample size in previous shape inference studies. We recover key evolutionary trends identified in previous smaller-sample studies. We find high ellipticities ($μ_E \sim 0.600$--$0.821$) across all bins, indicating that the underlying 3D shape distributions are statistically dominated by flattened geometries. Most notably, we recover the decline of the prolate fraction towards $z=0$ at low masses ($M_\ast=10^{9}$--$10^{9.5}M_\odot$) reported by previous studies. This trend is seen consistently in comparable analyses of independent data sets spanning different depths and resolution regimes, showing that it is not an artefact of small-sample statistics, suggesting it may be a genuine feature of galaxy evolution. We also identify future methodological improvements for subsequent Euclid data releases.

astro-ph.GA↗

Euclid Data Release 1 (DR1): The MERge processing

The ESA satellite mission \Euclid was launched in July 2023 and started its regular observing mode in February 2024. The data from the \Euclid instruments Visible Camera (VIS) and the Near IR Spectrometer and Photometer (NISP) are complemented by ground-based imaging data in the $u$, $g$ $r$, $i$, and $z$ bands to enable the determination of photometric redshifts and other galaxy properties. This paper describes the MERge processing function (MER PF) that has been developed and run for the Euclid Data Release 1 (DR1). The aim of the MER PF is to detect objects in the \Euclid imaging data, measure their properties, and combine them into a single multi-wavelength catalogue. In this paper we illustrate the design of the MER PF and introduce the different photometric methods applied to the data, which are Kron total flux, aperture photometry on Point Spread Functions (PSF) -matched images, template fitting photometry, and Sérsic fitting photometry. We also present the various morphological measurements and quality flags that improve the scientific content of the output catalogues and better characterize the detected sources. We illustrate the processing and measurements that help to validate the MER PF results, which is an essential step to qualify the output for the public DR1 release. We discuss the global assessment of the photometric homogeneity and provide an internal comparison of our processing results. In addition, we give an extensive overview of the DR1 output data and the various processing scenarios

astro-ph.IM↗

Euclid Data Release 1 (DR1). AstroVink-TREX: Automated strong gravitational lens segmentation as input for mass modelling pipelines

Strong gravitational lens modelling requires the lens light, lensed source light, and contaminants to be separated in each image before mass modelling codes can fit the system. Within Euclid this separation has so far been done by hand, which does not scale to the more than 15,000 candidates in the Euclid Lens Identification Pipeline and Scientific Exploration (ECLIPSE) catalogue of Data Release 1 (DR1). We present AstroVink-TREX , a neural network that performs this separation automatically and predicts three flux maps, in a standard format directly usable by mass-modelling pipelines. The network uses a hybrid architecture, pairing a vision transformer encoder reused from the lens detection network with a parallel convolutional branch. A custom loss function penalises both flux and structural error, while allowing independent weighting per component. From the network's predictions we also derive the modelling centre, fitting region, conjugate points, and binary maps, providing all the inputs that a modelling pipeline needs to run without manual setup. We evaluate the efficiency of the network on DR1 systems against existing fits and through expert validation. Across approximately 3000 inspected systems, taking 75% agreement between inspectors, just 39 lens models fail because of a segmentation error. Following this validation, the network provides the component separation and modelling setup for every system modelled in DR1. With automated segmentation, lens modelling is no longer limited by manual preprocessing and can now operate at the same scale as lens discovery.

astro-ph.IM↗

Euclid Data Release 1 (DR1) : Performance of the redshift measurement and spectroscopic analysis

The European Space Agency's Euclid mission has issued its first major data release(Data Release 1, DR1), providing a transformative view of the extragalactic sky through deep photometry and near-infrared slitless spectroscopy. This paper presents the Euclid DR1 spectroscopic data set and details the evolution of the spectroscopic processing function (SPE PF) pipeline, which is responsible for the automated analysis of 1-dimensional spectra acquired by the Near-Infrared Spectro-Photometer (NISP). The SPE PF pipeline workflow comprises spectroscopic classification, redshift determination, and the systematic measurement of spectral features. Emission line fluxes are derived using two independent methodologies: direct integration (DI) and Gaussian fitting (GF). For sources exhibiting broad H$α$ profiles (velocity dispersion larger than $1500$ km s$^{-1}$), a dedicated `QSO mode' is implemented to robustly deconvolve nuclear and host galaxy components. The data set was validated by cross-matching Euclid results with external spectroscopic catalogues, specifically Dark Energy Spectroscopic Instrument (DESI) DR1, and through an extensive visual inspection campaign of approximately 4000 sources within the Euclid Deep Field North (EDF-N). For high-quality samples selected in the EDF-N, we find a 98.5\% agreement between Euclid and DESI spectroscopic redshifts. This success rate increases to 98.9\% when incorporating pipeline probability and reliability estimators. (abridged)

astro-ph.CO↗

Euclid Data Release 1 (DR1). Near-infrared processing and data products

The Euclid near-infrared processing function (NIR PF) converts the raw observations acquired by the Near-Infrared Spectrometer and Photometer (NISP) into calibrated images and source-level products. Following the first implementation delivered with the Euclid Quick Data Release 1 (Q1), the NIR PF has been revised for the first major Euclid Data Release (DR1), with improvements to detector-level processing, calibration, image reconstruction, and quality control. The DR1 processing includes observations covering more than $1900\,\rm{deg}^2$ and has produced $71\,659$ calibrated near-infrared (NIR) images. We assess the performance of the processing chain through extensive astrometric and photometric validation. The astrometric solution reaches an average root-mean-square (RMS) of 10 mas, remaining below 20 mas for 99% of the dithers in all three NISP bands. Temporal variations in the instrumental throughput are typically corrected to a level of 2 mmag RMS, while the relative photometric accuracy reaches 0.9% for sources brighter than 19 mag in $Y_\mathrm{E}$, $J_\mathrm{E}$, and $H_\mathrm{E}$, and the residual spatial variations of the photometric calibration across the focal plane are at the level of approximately 10 mmag for the entire sample. Despite some remaining limitations, the DR1 NIR imaging products are consistent with the required astrometric and photometric performance over the large area covered by this release.

astro-ph.IM↗

Euclid Data Release 1 (DR1). The PHZ processing function for the physical properties of galaxies

Euclid is a European Space Agency mission mapping the extragalactic sky, aiming to provide the most-accurate 3D view of the Universe at optical and near-infrared wavelengths, by measuring the photometric redshifts of billions of galaxies. Within the Euclid Consortium, this task is executed by the PHZ Organisation Unit. PHZ takes as input the multi-wavelength photometry of Euclid and ground-based data, and returns as output the source classification, photometric redshifts for the Euclid core science (i.e., cosmology), and photometric redshifts plus physical parameters and absolute magnitudes for non-cosmological science. As part of the first Euclid Data Release (DR1), we describe the details of the DR1 pipeline for non-cosmological science, along with the released data products, and their validation. The non-cosmological science data products are obtained with NNPZ, a nearest-neighbour algorithm developed within PHZ that infers the properties of each target galaxy from its closest matches, in colour-magnitude space, within a reference sample of galaxies with known redshifts and physical parameters. We built this reference sample from Euclid observations of COSMOS, one of the Euclid Auxiliary Fields, labelling each galaxy with redshifts and physical parameters derived from the COSMOS2020 multi-wavelength photometry, consistently with the Euclid core-science results. Combined with the adjustments and priors required to generalise the results across the entire extragalactic sky, this approach yields high-quality and internally consistent data products. We tested the performance using both internal metrics and comparisons to ancillary data sets (e.g., COSMOS2020, 3D-HST, and DESI) and scaling relations (e.g., the star-forming main sequence); the results are in line with expectations and pre-DR1 simulations for both the Euclid Wide Survey and Euclid Deep Fields.

astro-ph.GA↗

Euclid Data Release 1 (DR1). An overview of the Science Ground Segment architecture and operations

The Euclid mission is designed to address fundamental questions on the nature of dark energy and dark matter by mapping the extragalactic sky with a wide-area survey complemented by targeted deep observations. The required scientific performance places stringent demands on the Euclid Science Ground Segment (SGS), which transforms raw telemetry into calibrated, validated, and scientifically exploitable products. The SGS is a distributed system comprising the Euclid Science Operations Centre, the Science Data Centres, Instrument Operation Teams, and supporting Organisation Units, within a common architectural and procedural framework. This paper presents the SGS as implemented for Data Release 1 (DR1), describing its organisation, architecture, processing functions, and operational model. We outline the main components of the SGS common infrastructure. A key element is the Euclid Archive System (EAS), the central metadata repository based on a common data model. It comprises a Distributed Storage System, providing scalable storage of data files across the Science Data Centres; a Data Processing System used by the pipelines to register and access files and their lineage; and a Science Archive, distributing validated data products. Processing is coordinated by a common orchestration framework based on the Infrastructure Abstraction Layer, providing workflow management and resource abstraction of distributed compute clusters. Software development, integration, and deployment are supported by the Common Development Environment. We describe daily operations, calibration, and quality control, followed by the DR1 production campaign, including processing flows, validation strategy, and computational and data volumes. This infrastructure enabled Euclid DR1, involving 35 million jobs across nine Science Data Centres and 6.7 million core-hours, demonstrating robustness and scalability of the SGS.

astro-ph.IM↗

Euclid Data Release 1 (DR1). NIR astrometric calibration

This paper presents the methodology and results of the astrometric calibration applied to the Euclid Near-Infrared (NIR) images in the Euclid Data Release 1, focusing on the techniques used to model and correct instrumental distortions, as well as the assessment of calibration quality. This work details the calibration pipeline, including quality control tests, and the evaluation of astrometric residuals against the Gaia DR3 catalogue and internal consistency metrics based on data encompassing more than one year of Euclid operation: the achieved accuracy is 30 mas for 97% of the total area, thus exceeding the requirements for Data Release 1. For the remaining 3%, accuracy declines due to a persistence correction issue: a method to correct the issue is presented. Furthermore, we discuss the stability of the solution over time, its dependence on source properties and instrument configuration, and outline the improvements planned for future Euclid data releases to enhance the astrometric calibration accuracy and efficiency.

astro-ph.IM↗

Euclid Data Release 1 (DR1) - NISP-P point-spread function modelling pipelines

The images collected in photometric mode by the Near-Infrared Spectrometer and Photometer (NISP) instrument of the Euclid satellite are used to model the point-spread function (PSF) of a selection of sources detected in single and combined frames. We used Q1 and DR1 data to improve the development of the PSF module of the near-infrared processing function (NIR PF) and to refine the models currently provided. We explored a few methods normally used to evaluate the quality of the derived models and discuss possible variations related to the position of the sources in the focal plane and to the wavelength and time of the observations. The procedure of fitting the PSF profile, compared to other methods used for flux estimation, such as aperture photometry, indicates that updated models are better than those provided by Q1 images. The current results are in agreement with several key requirements defined for the mission, but there is still room for further improvements, which we plan to investigate and implement in future releases.

astro-ph.IM↗

Euclid: Data Release 1 (DR1) - First NISP emission-line maps: Methodology, validation, and applications to galaxy evolution

We demonstrate the capability of the slitless spectroscopic mode of the Euclid Near-Infrared Spectrometer and Photometer (NISP) to produce emission-line maps analogous to narrow-band images with user-defined central wavelengths. We present the first such maps of galaxies at $0.4 < z < 0.9$, constructed from blue-grism exposures obtained during the Euclid performance-verification observations of the Cosmic Evolution Survey (COSMOS) field in November 2023. We demonstrate the use of these maps for identifying star-forming clumps, measuring radial star-formation profiles, detecting projected internal velocity gradients, deblending the H$α$--[N II] complex, and assessing spectroscopic redshift measurements. As a first statistical application, we reconstruct the galaxy mass--metallicity relation for 885 galaxies observed with the blue grism at $0.4 < z < 0.9$ by deblending the H$α$--[N II] complex in the emission-line maps, and measure stacked H$α$ radial profiles for 2634 galaxies over the same redshift range. The mass--metallicity relation is consistent with that obtained from the Euclid 1D spectroscopic measurements, demonstrating overall agreement between the two deblending approaches. The recovered H$α$ radial profiles broadly follow the $Y_{\mathrm{E}}$ stellar continuum but remain modestly more extended after matching the $Y_{\mathrm{E}}$ images to the H$α$ PSF, consistent with previous results in the literature. Across the four $Y_{\mathrm{E}}$ size bins, we measure a mean $R_{50,\mathrm{H}α}/R_{50,Y_{\mathrm{E}}}=1.15$.

astro-ph.GA↗

Euclid Data Release 1 (DR1). Colouring Euclid with external ground-based imaging data

Accurate photometric-redshift estimation and calibration of the weak-lensing point-spread function (PSF) for Euclid require complementary ground-based optical imaging with high photometric precision and spatial homogeneity. The Organisation Unit for External Data (OU-EXT) integrates heterogeneous ground-based surveys into the Euclid Science Ground Segment (SGS), delivering uniformly calibrated, co-added images, and associated PSF models. To minimise systematics in spectral energy distribution (SED) reconstruction and photometric-redshift estimation, these data must achieve relative photometric stability at the $\sim$10 mmag level on Euclid-tile scales. We present the ground-based imaging component of Euclid Data Release 1 (DR1), comprising $\sim2500$ deg$^2$ across the wide, deep, and auxiliary fields. We summarise the OU-EXT production workflow and validation framework, and quantify the achieved data quality of the co-added products in terms of depth, photometric zero-point stability, and spatial colour homogeneity. In the Euclid Wide Survey (EWS), the point-source $10σ$ depths in the $griz$ bands reach 24.8/24.3/24.0/23.0 AB mag, with the deep fields extending 1--2 magnitudes deeper. Most EWS DR1 tiles meet the 10 mmag relative photometric requirement, with 1.1% and 1.5% showing offsets exceeding 50 mmag in the northern and southern hemispheres, respectively. Spatial colour variations remain typically below 5--10 mmag across the southern EWS footprint, while larger residual structures of 10--25 mmag appear in some northern regions. These results demonstrate that the EXT production chain provides a largely homogeneous framework for integrating ground-based datasets into Euclid, while identifying calibration and co-addition limitations that have to be addressed in DR2 to meet the mission's most stringent cosmological requirements. (Abridged)

astro-ph.CO↗

Euclid Data Release 1 (DR1): Calibration strategy and observations

This paper reviews how Euclid's top-level calibration requirements were translated into a calibration plan, how this plan is implemented and maintained during the survey, and how it performs in the current DR1 release. We show that DR1 already meets or exceeds most end-of-mission spectrophotometric calibration requirements. The construction of spectroscopically complete and pure samples relies on ongoing data acquisition for the Euclid Deep Survey, while point-spread-function modelling and shear calibration are still converging to their stringent accuracy requirements. With all DR1 data reprocessed with the DR1 pipeline configuration, more subtle effects have been detected and characterised, and correction strategies have been developed for DR2. These include improved absolute flux calibration, astrometry, charge-persistence correction, shutter effects, wavelength calibration, and spectrophotometric corrections for water-ice contamination. Finally, synthetic photometric transformations between Euclid and each of Roman, the James Webb Space Telescope, and the Chinese Space-station Survey Telescope, are presented.

astro-ph.IM↗

Euclid Data Release 1 (DR1). Persistence in NISP detectors

Detector persistence is a significant source of systematic error in infrared observations and affects the scientific performance of the \Euclid Near Infrared Spectrometer and Photometer (NISP). We characterise persistence using complementary ground calibration and in-flight data, develop empirical models of its temporal evolution, and assess its impact on science observations and the \Euclid data-processing pipeline. Persistence is highly non-uniform across the focal plane, with variations of up to two orders of magnitude between and within detectors. Its amplitude is typically below 1\% of the stimulus signal and depends nonlinearly on the preceding fluence. The decay rate is faster for brighter sources. During mission operations, the persistence signal has progressively decreased by roughly 10\%, while the decay rate has remained approximately stable. Persistence following over-saturating signals does not follow the empirical relations derived for under-saturating fluences and plausibly reduces persistence response over longer timescales. The impact of charge trapping and release on flat-field observations and on the background in science exposures is also discussed. Persistence measured in science observations is observed to differ systematically from that derived from dark calibration data. In addition, persistence of energetic particle hits is observed to be different from that of astronomical sources of comparable signal. Finally, we discuss the impact of persistence on \Euclid NISP observations and describe the masking strategy adopted for the Q1 abd DR1 releases, focusing on identifying which pixels are affected by persistence and for how long they should be excluded from science observations.

astro-ph.IM↗

Euclid Data Release 1 (DR1). Multi-band photometry validation on COSMOS extragalactic sources

The Euclid Data Release 1 (DR1) contains catalogues of sources covering 1900 deg2, including deep coverage of legacy fields that enable an extensive comparison between Euclid imaging and multiwavelength observations from other instruments. The Cosmic Evolution Survey (COSMOS) field, observed by Euclid and included in DR1, has been observed with the James Webb Space Telescope, resulting in the COSMOS2025 catalogue. We compare the Euclid DR1 photometry with the multiwavelength COSMOS2025 data to assess the accuracy of the photometric measurements produced by the different methods implemented in the OU-MER pipeline. We use both the Euclid Wide and Deep Surveys, focussing exclusively on extragalactic sources. We use the fluxes of COSMOS2025 to predict the expected fluxes in the Euclid bands via spectral energy distribution fitting, while the redshifts are either spectroscopic, for a sample of brighter sources, or photometric, to include fainter sources. Both the fluxes in the common Hyper Suprime-Cam bands and the SEDs from COSMOS2025 are used to validate the analysis and to characterise any biases or limitations of the method. The predicted fluxes and the OU-MER measurements are in good agreement, especially for bright and isolated sources. Median offsets depend on band, photometric method, survey depth, and redshift sample, most well within 10%, the largest offset seen in the NIR bands for the deep photo-z model-fitting measurements. However, complementary measurements on Gaia stars show an opposite trend, highlighting that photometric validation approaches may suffer from intrinsic uncertainties that prevent a thorough assessment of systematic effects. Among the three methods, template-fitting photometry generally shows the smallest offset from the SED-based predictions in the NISP bands and toward fainter magnitudes, while aperture photometry and model-fitting are comparable in IE.

astro-ph.CO↗

Euclid data release 1 (DR1) - SIR spectroscopic calibration

The Euclid space mission is dedicated to probing the nature of dark energy and dark matter, through a comprehensive mapping of the Universe's large-scale structure. Central to these cosmological goals are the spectra produced by the Near-Infrared Spectrometer and Photometer (NISP) and reduced by the scientific pipelines of the SIR (Spectral near InfraRed) processing function. Due to the intrinsic complexity of slitless spectroscopic data, scientific observations cannot be self-calibrated by the science pipelines. Consequently, the development of dedicated SIR calibration pipelines, utilising specific pointing sequences and observation fields, is essential. The resulting calibration products are fundamental to ensuring that reduced spectra are characterised within requirements in both wavelength and flux. This paper provides a detailed description of the SIR calibration pipelines responsible for generating the products utilised by the science pipelines for Data Release 1 (DR1). A significant challenge in these calibration tasks arises from the requirement to model the entire focal plane as a unified entity. Such a global modelling approach is critical for maintaining astrometric and spectroscopic continuity across the entire field of view. DR1 provides more than 170 millions validated spectra in both red- and blue-grism modes. The accuracy of the reduced spectra is strictly dependent on the performance of the calibration pipelines. Comprehensive analyses of the calibration products and the resulting science spectra demonstrate that current outputs successfully meet the mission requirements. Nevertheless, specific areas for optimisation have been identified; ongoing development is focused on addressing these critical points to further enhance data quality for future data releases.

astro-ph.IM↗

Euclid Data Release 1 (DR1). Enhancing deep-learning models: the impact of image processing and training data selection for strong-lens detection

Euclid is expected to discover ~$10^5$ galaxy-galaxy strong gravitational lenses across ~14,000deg$^2$, requiring automated searches with high completeness and low false-positive rates. Despite advances in deep learning, the effects of training-set composition, image representation, and preprocessing on survey-scale detection remain poorly understood. We present an ablation study of the convolutional neural network used in the ECLIPSE strong-lens detection pipeline, evaluating different training configurations on a fixed test set. Using simulated and real Euclid data, we compare VIS-only and pseudo-RGB representations, FITS and JPEG formats, training-set composition, hard-negative selection, and multi-stage training. We also assess ensembling on ~67 million DR1 cutouts. Training-set composition and strategy strongly affect high-Recall performance. Omitting simulation based pre-training reduces high-Recall weighted average Precision by 83.7% for VIS-only and 74.5% for pseudo-RGB models. Negative-sample selection also substantially affects performance, while FITS yields modest gains over JPEG at high Recall (~2-5%). On DR1, the logit-mean ensemble improves average Precision by 28.0% and 39.9% over the individual VIS-only and pseudo-RGB models. Among ~70,000 candidates selected for visual inspection, it recovers 11,256 unique grade A (secure) and B (probable) lenses, compared with 9650 and 10,395 for the individual models. The branches show complementary ranking preferences, particularly for lenses benefiting from near-infrared information. Careful training-set design, realistic negatives, and multi-band ensembling improve survey-scale searches. Our final ensemble achieves ~80% Recall at the operational selection threshold, outperforming ablation models with higher global classification metrics and highlighting the need to optimise performance for realistic survey conditions.[Abridged]

astro-ph.CO↗