Programa AIA 2026

Human centric machine learning as a design choice in transient astronomy

Autor: Emille Ishida (emille.ishida@clermont.in2p3.fr)

Afiliação: CNRS/Laboratoire de Physique de Clermont Auvergne, Université Clermont-Auvergne (UCA), France

The recent increase in volume and complexity of available astronomical data, coupled with widespread use of Al tools, has led to an indiscriminate use of machine learning techniques. Despite the undeniable opportunity open by these new tools, their applicability is highly dependent on the design choice and scientific context in which they are used. In this talk, I will briefly describe the historical path taken by astronomy until the arrival of the era of big data and discuss how we can design machine learning tools which are specific for the context of transient detection. I will give a few examples of how this can be done and results found so far using ZTF and Rubin data and discuss the challenges ahead.

Computational Kinetic Analysis of Sulfur Chemical Networks in Low-Temperature Radiative Environments

Autor: Nicolas da Silva Jonker (prof.nicolas.jonker@gmail.com)

Afiliação: UNIVAP

Recent JWST observations of K2-18b have raised questions about sulfur chemistry and its astrobiological significance, given sulfur's role in prebiotic pathways and sulfur-bearing amino acid synthesis (Mifsud et al., 2021). Understanding the radiation-driven chemical networks that could give rise to such species remains a central challenge in astrochemistry and exoplanet science. Using PROCODA v7c benchmarked against Bonfim et al. (2017), we simulated a reaction network of 31 sulfur-bearing species to characterize the chemical pathways triggered by soft X-ray irradiation at 12 K. The simulations revealed the efficient formation of SO₃, transient sulfur radicals, and short-lived intermediates undetectable by conventional FTIR spectroscopy. Notably, polymerization pathways leading to elemental sulfur (S₈) were identified, suggesting mechanisms for building sulfur reservoirs in icy grain mantles and planetary environments with direct implications for the well-known sulfur depletion problem in the interstellar medium (Mifsud et al., 2021). These results provide a quantitative framework for interpreting radiation-driven oxidation pathways in sulfur-rich ices, offering new insights into the chemical reservoirs that may ultimately seed the organosulfur chemistry observed in exoplanetary atmospheres such as K2-18b's. This study offers new chemical routes connecting SO₂ photochemistry to the formation of complex organosulfur molecules such as DMS and DMDS. By linking kinetic modeling to observed atmospheric signatures, the work contributes to the broader effort of characterizing the molecular complexity and potential habitability of exoplanetary environments.

ML-FAIR: An Automated Machine Learning Framework for the Fast Identification of Eccentricity-Type Mean-Motion Resonances

Autor: Mansur M. Bala (m.bala@unesp.br)

Afiliação: UNESP Guaratinguetá

Identifying asteroids in mean-motion resonances (MMR) with the planets is central to understanding the dynamical evolution of small-body populations, but classical resonant-argument analysis becomes computationally prohibitive for the very large catalogues expected from next-generation surveys such as the Vera C. Rubin Observatory (LSST). We present ML-FAIR, an automated framework that extends the FAst Identification of mean-motion Resonances (FAIR) method by combining its geometric approach, based on counting stripe intersections in angle-versus-mean-anomaly space, with unsupervised machine learning. Circulating orbits are first separated from resonant candidates using density-coverage thresholds; the remaining objects are analysed with density-based clustering (OPTICS) and circular kernel density estimation to locate the peaks that define the resonance order and degree, followed by a consistency check against the asteroid's semi-major axis. We applied ML-FAIR to the Atira and Aten near-Earth asteroid populations, which cross numerous weak eccentricity-type MMR with the terrestrial planets, automatically classifying more than 85% of 2882 analysed cases and leaving only a small fraction for visual inspection. Results were compared with a state-of-the-art benchmark based on long-term numerical integration (Smirnov, 2023). The two approaches show good qualitative agreement for the great majority of resonances identified, including strong consistency for Venus (2V:3, 3V:4, 4V:5) and substantial overlap for Earth and Mars; discrepancies are limited to a few resonances, notably 3M:7 and 2M:5. ML-FAIR additionally revealed two candidate objects librating around the unusual 270° equilibrium point in resonances with Mercury, potentially the first such detections. These results show that ML-FAIR provides a scalable, efficient alternative to manual resonant-argument analysis, well suited to future large-scale dynamical surveys of near-Earth asteroid populations.

Identifying local Little Red Dots in S-PLUS with Machine Learning

Autor: Micheli Trindade Moura (micheli.t.moura@gmail.com)

Afiliação: IAG-USP

Little Red Dots (LRDs) are a population of compact, red sources identified by JWST at high redshift, whose nature and physical properties remain under debate. Identifying their low-redshift analogs can provide insights into their physical nature and evolutionary connection to other galaxy populations. We aim to develop a scalable photometric strategy for identifying local LRD analogs over wide areas of the sky and to assess the potential of machine learning for morphological classification of candidate sources. We developed a multi-stage selection pipeline combining optical photometry from S-PLUS, ultraviolet data from GALEX, and imaging from the DESI Legacy Survey. Candidates are refined through photometric quality and color cuts, a structural compactness criterion, and cross-matching with Gaia DR3 and the Milliquas v8 catalogue to remove Galactic stars and background quasars. For morphological classification, we employ CLIP (Contrastive Language-Image Pre-training) embeddings based on the Vision Transformer architecture (ViT-L/14) to distinguish compact, LRD-like sources from extended galaxies and stars. The resulting catalogue is designed for direct comparison with independently selected LRD samples from Euclid, providing an external reference for evaluating the selection strategy.

Classificação Fotométrica de Quasares e Estimativa de Redshift com Aprendizado de Máquina

Autor: Gabriel Silva Costa (gabriel.scosta@gmail.com)

Afiliação: INPE

Este artigo apresenta a construção de um catálogo fotométrico de quasares a partir de dados do Dark Energy Survey (DES) DR2, utilizando técnicas de aprendizado de máquina para enfrentar dois problemas fundamentais: distinguir quasares de estrelas e estimar seus redshifts fotométricos (photo-z). Essa classificação é importante porque quasares são excelentes traçadores da estrutura em larga escala do Universo, mas sua aparência pontual e suas cores podem ser semelhantes às de estrelas, especialmente em altos redshifts. Para construir a amostra de treinamento, os autores realizaram um cross-match entre objetos pontuais do DES DR2 e objetos espectroscopicamente classificados no SDSS DR16. Para a classificação entre quasares/galáxias e estrelas, foi utilizado o algoritmo K-Nearest Neighbors (KNN), tendo como principais características as magnitudes PSF nas bandas g, r, i e z. Após otimização, o modelo apresentou precisão de 99% e recall de 77%, com k igual a 11, permitindo aplicá-lo ao catálogo completo do DES. A estimativa de photo-z foi realizada por uma abordagem híbrida, combinando um Boosted Decision Tree (BDT) implementado no ANNz e um Decision Tree Regressor (DTR) do scikit-learn, com ponderação dinâmica baseada em KNN. Essa combinação apresentou melhor desempenho do que os modelos individuais, produzindo uma concentração principal de objetos com redshifts entre aproximadamente 0.5 e 3 e recuperando uma população de quasares em torno de z próximo de 4. Para reduzir erros catastróficos, definidos por diferenças absolutas entre redshift fotométrico e espectroscópico maiores ou iguais a 0.4, foi desenvolvido ainda um classificador stacked combinando KNN, Gradient Boosting e Random Forest. O modelo alcançou AUC-PR igual a 0,92, reduzindo significativamente a presença de outliers. Aplicando todo o pipeline aos dados do DES, os autores obtiveram 872.373 candidatos a quasares/galáxias, dos quais 675.683 permaneceram após a remoção de outliers. O catálogo apresenta uma ampla distribuição em redshift e permite construir mapas tomográficos que acompanham a estrutura em larga escala do Universo. Para redshifts maiores que aproximadamente 3,5, a amostra permanece particularmente confiável, pois os erros catastróficos praticamente não se estendem para essa região. Os resultados demonstram que técnicas de aprendizado de máquina podem produzir um catálogo fotométrico de quasares suficientemente robusto para estudos cosmológicos. O catálogo complementa levantamentos baseados em galáxias e pode ser utilizado em análises conjuntas de estrutura em larga escala, incluindo estudos que combinem dados ópticos e levantamentos de rádio.

PoincAlré: Machine Learning for the Classification of Spin-Orbit Dynamical Regimes

Autor: Tiago Francisco Lins Leal Pinheiro (franciscopinheiro@on.br)

Afiliação: ON

Poincaré surfaces of section (PPS) constitute a well-established method for visualizing the structure and dynamical behavior of trajectories in dynamical systems. By recording the intersections of trajectories with two-dimensional cross sections in phase space, they enable the identification of periodic, quasi-periodic, and chaotic motion. However, their construction and analysis generally require the numerical integration of a large number of trajectories over sufficiently long timescales, making this approach computationally demanding and potentially time-consuming, particularly when exploring extensive parameter spaces. PoincAlré is an ensemble of open-source machine learning models developed to accelerate the analysis of dynamical systems using the PPS approach for the spin-orbit of an elongated satellite under the gravitational influence of a spherical central body. The first model employs Convolutional Neural Network (CNN) and Vision Transformer (ViT) architectures trained to automatically classify three distinct dynamical regimes from PSS images: resonance, circulation, and chaos, achieving an accuracy of around 99%. The second model is based on the XGBoost algorithm and uses the system's initial conditions as input to predict the probability of the trajectory belonging to each dynamical class. This approach significantly accelerates parameter-space exploration, enabling high-resolution dynamical maps to be generated without explicitly integrating every trajectory. This unprecedented model can classify 10^5 different initial conditions in a matter of seconds, with an accuracy of approximately 97%. Furthermore, combining the predicted class probabilities with normalized entropy quantity provides a performance measure of classification confidence.

Supervised Gradient-Domain Learning for Galaxy Morphology and Transient Classification in the Rubin LSST Era

Autor: Reinaldo Roberto Rosa (reinaldo.rosa@inpe.br)

Afiliação: COPDT-INPE

The scale and diversity of imaging and time-domain data from the Vera C. Rubin Observatory LSST demand scalable and interpretable Al methods. We present a supervised gradient-domain learning framework based on Gradient Pattern Analysis (GPA). Rather than relying exclusively on statistical descriptions of amplitude fluctuations, GPA derives Gradient Moments that act as structural and temporal fingerprints. These quantities can serve both as interpretable features and as domain-informed targets for Machine and Deep Learning, providing algorithmic supervision directly derived from the data. Applications to galaxy morphology and transient classification are presented, while SHAP and LIME are used to investigate which information drives model predictions and to relate learned representations to interpretable gradient-domain properties.

SURF-T: A Standardized Interpolation and Uniform Resampling Method for Modeling and Classifying Sparse Astronomical Transients

Autor: Mariana Rubet da Costa (rubetmariana@gmail.com)

Afiliação: ON

We present SURF-T (Standardized Uniform Resampling Framework for Transients), a Python-based computational framework designed to improve the photometric classification of sparsely sampled optical transients. The method enforces a standardized interpolation using shape-preserving polynomials followed by uniform temporal resampling, ensuring that all light curves share identical dimensionality prior to feature extraction. From the resampled curves, we construct a three-dimensional parameter space defined by (i) the Detrended Fluctuation Analysis scaling exponent (α), (ii) a structural variability metric derived from Gradient Pattern Analysis (G₂), and (iii) a morphology-based asymmetry parameter (Ψ). Using a balanced pilot sample of 10 sources across 15 transient classes, we evaluate separability through the distributions of the extracted parameters. Preliminary results based solely on α reveal a systematic displacement of supernovae relative to the other subclasses, despite substantial overlap among the remaining populations. These results demonstrate that the DFA scaling exponent, still underexplored in transient classification, captures class-dependent temporal structure and provides a complementary source of information for more robust photometric classification. The inclusion of G₂ and Ψ will provide complementary morphological information and enable a more comprehensive assessment of class separability in the combined parameter space.

Interpretable Galaxy Morphology with Gradient Pattern Analysis: From Structural Fingerprints to Explainable Al

Autor: Carlos Eduardo Falandes (carlos.falandes@inpe.br)

Afiliação: INPE

The growing scale and complexity of astronomical imaging surveys are making Machine Learning (ML) increasingly important for automated galaxy characterization, while also raising a fundamental question: which morphological information is actually driving model predictions? In this work, we investigate Gradient Pattern Analysis (GPA) [1] as an interpretable gradient-domain representation for Al-assisted galaxy morphology. Rather than relying exclusively on statistical properties of image amplitudes, GPA represents an astronomical image through its gradient lattice and its decomposition according to bilateral symmetry. The resulting asymmetric gradient lattice is characterized by established Gradient Moments, which act as structural fingerprints encoding the organization and departures from bilateral symmetry. GPA has already been incorporated into galaxy morphometry frameworks such as CyMorph [2], where gradient-based information is combined with complementary descriptors from the CAS and EGG systems [3]. We use these multidimensional morphological representations as inputs to ML models and apply Explainable Artificial Intelligence (XAI), particularly SHAP (SHapley Additive exPlanations) [4], to quantify how individual morphological descriptors contribute to model predictions and to investigate whether classification decisions can be related to physically meaningful structural properties. We also present methodological refinements and an implemented computational tool that, beyond the Gradient Moments, enables direct mapping of the asymmetric gradient lattice, retaining spatial information about the magnitude, orientation, and location of the gradient vectors responsible for departures from bilateral symmetry. This opens the possibility of connecting feature-level XAI with spatially resolved gradient-domain information. While established applications have focused primarily on classical galaxy morphological types and subtypes, we extend the methodology to the more challenging case of interacting and merging galaxies and present promising preliminary results, including objects already observed by the Vera C. Rubin Observatory Legacy Survey of Space and Time (LSST). We further discuss the potential integration of GPA with SAGUI [5], a framework for spatially resolved and SED-based segmentation of multiband galaxy images, as a route toward combining gradient-domain structural fingerprints, spatially resolved photometric information, ML, and explainability. This framework illustrates how domain-informed representations and XAI can complement data-driven models toward more transparent and scientifically interpretable Al for galaxy morphology.

Referências:
[1] Rosa et al. (2018), MNRAS 477, L101-L105. https://doi.org/10.1093/mnrasl/sly054
[2] Barchi et al. (2020), Astronomy and Computing 30, 100334. https://doi.org/10.1016/j.ascom.2019.100334
[3] Kolesnikov et al. (2024), MNRAS 528, 82-107. https://doi.org/10.1093/mnras/stad3934
[4] Aguilar-Argüello et al. (2025), MNRAS 537, 876. https://doi.org/10.1093/mnras/staf085
[5] de Souza et al. (2026), MNRAS 549, stag1062. https://doi.org/10.1093/mnras/stag1062

Uso de Aprendizado Profundo para Classificação de Trânsito Planetários em Curvas de Luz

Autor: Helena Buschermohle (buschermohleh@gmail.com)

Afiliação: INPE

Este trabalho tem como objetivo a criação e validação de um classificador de janelas de curvas de luz baseado em arquitetura MLP. Por meio de normalização Min-Max, regularização L2 e otimização bayesiana de hiperparâmetros, foram obtidas 5 arquiteturas distintas capazes de resolver o problema com acurácia k > 0,94. A contribuição individual de cada técnica de pré-processamento foi investigada, revelando que a normalização Min-Max é o fator determinante para o desempenho do modelo. Os resultados indicam que a detecção de trânsitos em janelas isoladas de curvas de luz é um problema tratável por arquiteturas simples, desde que o pré-processamento adequado seja aplicado.

Classificação da emissão nuclear de galáxias com aprendizado de máquina

Autor: Roberto Bertoldo Menezes (roberto.menezes@maua.br)

Afiliação: Instituto Mauá de Tecnologia

Nesse seminário, irei apresentar os resultados obtidos pelo meu grupo de pesquisa no Instituto Mauá de Tecnologia, envolvendo o uso de aprendizado de máquina para a classificação da emissão nuclear de galáxias. Parte da análise foi realizada com dados fotométricos, obtidos com o Southern Photometric Local Universe Survey (S-PLUS) e resultou em acurácias próximas de 90% na classificação binária da emissão nuclear de galáxias em regiões HII e LINERS/Seyferts. O restante da análise envolveu espectros obtidos com o Sloan Digital Sky Survey (SDSS), os quais foram utilizados como entrada para os algoritmos de aprendizado de máquina e resultaram em acurácias superiores a 90% na classificação da emissão nuclear de galáxias em regiões HII, Seyferts ou LINERS.

Estimativa de linhas de emissão em espectros de galáxias usando regressão simbólica

Autor: Eduardo Galvani Massino (eduardo.massino@usp.br)

Afiliação: IAG-USP

Modelos semi-analíticos de formação de galáxias, como o L-GALAXIES, são capazes de prever o contínuo espectral das galáxias, mas não incorporam métodos para estimar as larguras equivalentes (EWs) das linhas de emissão nebular. Esta dissertação propõe uma abordagem baseada em aprendizado de máquina para construir modelos analíticos interpretáveis que expressem explicitamente a dependência das EWs em relação aos parâmetros de síntese de populações estelares, idades e metalicidades ponderadas pelo fluxo e pela massa, massa estelar e extinção por poeira, obtidos pelo código STARLIGHT a partir de espectros do SDSS DR8 combinados com fotometria ultravioleta do GALEX GR6. As EWs-alvo são [NII]λ6584, [OIII]λ5007, Hα e Hβ. Compararam-se três algoritmos, Random Forest, Operon e PySR, sendo que a regressão simbólica, em especial o Operon, apresentou desempenho comparável ao Random Forest com a vantagem de produzir equações explícitas e legíveis, ou seja, que mostram relações analíticas entre as variáveis. Uma abordagem estatística baseada em amostragem de normais multivariadas foi desenvolvida, na qual o Operon modela as médias, os desvios-padrão e as covariâncias das EWs, capturando as correlações físicas entre as linhas. A linha [OIII]λ5007 revelou-se a mais difícil de modelar em todos os cenários avaliados. Os modelos obtidos representam um passo concreto para a incorporação de linhas de emissão em modelos semi-analíticos em função dos parâmetros de populações estelares.

Photometric Metallicity Mapping of the Inner Sagittarius System with JEDIS

Autor: Matheus Agenor (agenor.math@gmail.com)

Afiliação: INPE

The Sagittarius (Sgr) system is a prime example of ongoing accretion in the Milky Way (MW), with a well-defined core and extended tidal streams. A metallicity gradient along the leading and trailing arms links its dynamical evolution to its chemical structure. The central region is harder to observe due to Galactic plane contamination and extinction, making the determination of stellar metallicities particularly challenging in this region. In this work, we investigate the use of Machine Learning techniques to estimate photometric metallicities and map the chemical structure of the inner Sgr system. Observational data from the JEDIS (JoinEd narrow-band Data of Inner Sagittarius) survey, obtained with the T80-South telescope (exclusively for this project) in 12 photometric bands, provide a novel basis for this analysis. The survey covers Λ ∈ [0, 15°] and B ∈ [-3.5°, 6.5°], enabling the construction of metallicity maps in a region that remains difficult to observe and hard to explore. To estimate stellar metallicities from these data, we construct a supervised Machine Learning training sample using 12-band synthetic photometry derived from Gaia XP spectra and spectroscopic [Fe/H] measurements from GALAH DR4 as reference labels. The photometric information is represented by the 12 individual magnitudes and all possible combinations of colors, resulting in 78 initial features, from which the most informative features are selected through Recursive Feature Elimination. Random Forest and Histogram-based Gradient Boosting regression models are then trained and compared on an independent test sample to determine the most accurate photometric metallicity estimator. From these models, we aim to derive metallicities for the JEDIS stellar sample, measure the amplitude and spatial variation of the metallicity gradient, and test whether the leading arm is systematically more metal-poor than the trailing arm in the inner regions. As a complement, we performed initial hydrodynamic simulations of the MW-Sgr system using the Arepo code, with orbits derived from dynamical integrations in fixed potentials from 6 Gyr in lookback time to the present. These experiments reproduce the present-day phase space of Sgr and support the existence of a metallicity gradient consistent with observations, with the leading arm being -0.19 dex more metal-poor. The next steps include (i) validating the Machine Learning metallicity estimates against independent spectroscopic samples, (ii) applying the trained model to the reduced and calibrated JEDIS photometric catalog, and (iii) constructing a spatially resolved metallicity map of the inner Sagittarius system to directly compare the observed chemical structure with the hydrodynamic simulations.

Aplicação de Redes Neurais Profundas no Ajuste de Modelos de Distribuição Espectral de Energia em Estrelas T Tauri (POSTER)

Autor: Gustavo Martin Garcia Souza (gustavo.martin@aluno.ufabc.edu.br)

Afiliação: UFABC

Estrelas T Tauri constituem objetos jovens em estágio de pré-sequência principal, cujas propriedades físicas estão diretamente relacionadas à interação entre a estrela central e seu ambiente circunstelar. A determinação desses parâmetros a partir de Distribuições Espectrais de Energia (SEDs) constitui um problema inverso de elevada complexidade, devido à degenerescência entre diferentes propriedades estelares e circunstelares. Este trabalho apresenta uma abordagem computacional baseada em Redes Neurais Profundas com Regularização Física para a inferência de parâmetros físicos de estrelas T Tauri a partir de dados fotométricos. Para isso, foi desenvolvido um universo sintético de estrelas virtuais, gerado por amostragem estatística do espaço paramétrico e por um modelo analítico de SED que considera as contribuições da fotosfera, do disco circunstelar e do envoltório de poeira. O conjunto sintético incorpora diferentes condições observacionais, incluindo ruído fotométrico e limites instrumentais, permitindo o treinamento de modelos capazes de atuar em cenários próximos aos dados reais. A arquitetura proposta utiliza penalidades físicas associadas às restrições evolutivas e geométricas do sistema, além de uma Matriz Híbrida de Especialistas, construída a partir da seleção de redes especializadas em diferentes grupos de parâmetros. O método foi aplicado a uma população sintética de 200 mil sistemas e posteriormente utilizado na análise de objetos do catálogo Pico dos Dias Survey. Os resultados demonstram a capacidade da abordagem em realizar inferência multivariada e reproduzir propriedades termodinâmicas e morfológicas dos sistemas estudados, além de possibilitar a análise automatizada da classificação entre diferentes estágios evolutivos. A metodologia desenvolvida estabelece uma estratégia computacional para a interpretação de grandes conjuntos de dados fotométricos e para a investigação automatizada das propriedades de estrelas jovens.