Introduction

Artificial Intelligence (AI) is one of the most influential fields of the 21st century (Zhang et al., 2021). Rich, E (2019) candidly described it as “the study of how to make computers do things which, at the moment, people do better”, today AI often surpasses human ability in tasks like large-scale data mining and pattern recognition - its true strength. AI’s subfields - Machine Learning (ML) and deep learning (DL), play a critical role in expanding the usage to a vast variety of fields like planetary science, astronomy,  earth observations, and remote sensing, just to name a few. There is an expected inclination towards incorporating AI more frequently in the studies of planetary science given the vast and complex nature of planetary data. In fact, AI has already been instrumental in extracting meaningful insights and advancing research in both interplanetary and astronomical studies.

In planetary sciences, several AI techniques have been employed in order to bridge gaps in our understanding of the varied patterns and occurrences for studying the natural features observable from the data returned by scientific payloads. For example, PCA and cluster analysis can help in detecting patterns of compositional variation from multi and hyper-spectral imagery (Moussaoui et al., 2008; D’Amore & Padovan, 2022). Furthermore, to study specific features and patterns in their occurrences, correlations with neighbouring features; unsupervised algorithms and more complex -supervised techniques can be helpful depending on the scale of the task. 

From simple methods of unsupervised learning like clustering used to study the spectral signatures of Jezero crater on Mars (Pletl et al., 2023) to applying large language models to track asteroids affected by gravitational effects which alter the asteroid’s orbit (Carruba et al., 2025), such applications highlight the prospects of AI in the field of planetary science. Henceforth, to develop a deeper understanding of the potential and applications of ML, below is a typical AI workflow.

Typical AI workflow

A typical workflow for an AI model involves an initial step of selecting a model  suitable for your goals (Figure 2). Data format, quality of the data, static or dynamic features of interest, etc can influence the choice of AI model or techniques. Data preparation steps, like normalizing the data i.e. scaling the data from [-1,1] values, prevents dominance of any one feature in the data and stabilizes the model training process. 

Furthermore, parameters or hyperparameters are selected depending on the complexity of the model. While more complex models; deep neural networks or Vision Transformers will need hyperparameter adjustment to maximize performance, simpler models mainly rely on predefined weights or fixed rules. Likewise, a model architecture shall be established as per the data and targets. One example is the usage of the Faster-R-CNN - a robust and high accuracy yielding model which can be employed to train on high-resolution labelled images to perform object-detection tasks like identifying craters. 

In scientific use cases, the workflow often encompasses actual data that must be separated into a training set, validation set (to optimize the hyper-parameters) and a test set, completely independent from the training. To evaluate “how well” the model has learnt from the training dataset, accuracy, precision, recall, F1-score, and intersection-over-union (IoU) are the most popular statistics. Subsequently, model predictions can help in developing an understanding of the potential areas for fine-tuning and refining the model for the use case. Henceforth, fine-tuning the model is another crucial step.

Figure 2: A typical AI workflow

Potential of AI

A successful application of ML in planetary science can be driven by a collaboration of scientists and ML experts. Scientists (astronomers, geologists, planetary scientists etc.) are arguably more equipped to answer science-based questions like what can be called a crater and conversely, ML experts may be more adept at assessing data preparation techniques to eradicate noise. The field of planetary sciences encompasses themes like anomaly detection, simulation and surface modeling, atmospheric studies, gravitational behavior and its effects on planets and smaller bodies, instrumentation and spacecraft design etc. which necessitates such collaborations for the optimum result. In recent years, Large Language Models (LLMs) have had a significant paradigm shift in AI applications due to their understanding of patterns acquired through their vast pre-training phase. For time series analysis, image classification, and pattern identification tasks common in planetary sciences, LLMs can significantly streamline workflows by reducing the need for specialized preprocessing steps.

Given the enormous data from missions and observational surveys, and the numerous applications of planetary sciences, it is the need of the hour to produce workflows that not only automates but helps in an objective/standard decision making for problem statements of planetary sciences. The Europlanet Machine Learning Working Group does exactly this by sharing the latest techniques, tools, and applications and opens doors for people who want to apply these robust techniques.

References

• Rich, E. (2019). Artificial Intelligence 3E (Sie) (Vol. 63, No. 4). Tata McGraw-Hill Education.
• Moussaoui, S. et.al. (2008) On the decomposition of Mars hyperspectral data by ICA and Bayesian positive source separation Neurocomputing for Vision Research; Advances in Blind Signal Processing, 71, 2194-2208, http://dx.doi.org/10.1016/j.neucom.2007.07.034 
• Carruba, V. et al. (2025). Vision Transformers for identifying asteroids interacting with secular resonances. Icarus, 425, 116346. https://doi.org/10.1016/j.icarus.2024.116346
• D’Amore, M., & Padovan, S. (2022). Chapter 7 Automated surface mapping via unsupervised learning and classification of Mercury Visible–Near-Infrared reflectance spectra. In J. Helbert, M. D’Amore, M. Aye, & H. Kerner (Eds.), Machine Learning for Planetary Science (pp. 131–149). Elsevier. https://doi.org/10.1016/B978-0-12-818721-0.00016-1
• Pletl, A. et.al. (2023). Spectral Clustering of CRISM Datasets in Jezero Crater Using UMAP and k-Means. Remote Sensing, 15(4), Article 4. /10.3390/rs15040939
• Zhang, D. et al. (2021). The AI Index 2021 Annual Report (No. arXiv:2103.06312). arXiv. /10.48550/arXiv.2103.06312

How to cite: Kacholia, D., Verma, N., D’Amore, M., Angrisani, M., Frigeri, A., Schmidt, F., Carruba, V., Hatipoğlu, Y. G., Roos-Serote, M., Smirnov, E., Sassarini, N. A. V., Solmaz, A., Oszkiewicz, D., and Ivanovski, S.: Artificial Intelligence in Planetary Science and Astronomy: Applications and Research Potential, EPSC-DPS Joint Meeting 2025, Helsinki, Finland, 7–13 Sep 2025, EPSC-DPS2025-1467, https://doi.org/10.5194/epsc-dps2025-1467, 2025.

How to cite: Smirnov, E.: Open-source large language models in astronomical data classification: applications and benchmarking, EPSC-DPS Joint Meeting 2025, Helsinki, Finland, 7–13 Sep 2025, EPSC-DPS2025-226, https://doi.org/10.5194/epsc-dps2025-226, 2025.

Asteroid families are groups of asteroids formed by a collision or fission event.  Some asteroid families interact with secular resonances. Because of planetary perturbations, the pericenter and nodes of planets and asteroids precess with frequencies g and s. When the precession frequency of the asteroid (g) is close to that of Saturn (g₆), or g-g₆≈0, the ν₆ resonance occurs.

Figure (1): The location of main secular resonances in the (a, sin(i)) domain.

Contrary to mean-motion resonances, secular resonances cannot be easily identified in a proper elements' 2-D domains. To identify if an asteroid is in a secular resonance, we need to investigate the time behavior of its resonant argument. With over 8 million asteroids predicted to be discovered by the Vera C. Rubin Observatory, traditional visual analysis of arguments will no longer be feasible.

Figure (2): Examples of resonant arguments for asteroids circulating, alternating phases of circulation and libration, and in libration states.

The first deep learning approach for identifying asteroids interacting with secular resonance was introduced in Carruba et al. (2021), with a multi-layer perceptron model. This is a five-step process:

1. We integrate the asteroid orbits under the gravitational influences of all planets.

2. We compute the time series of the resonant argument.

3. Images of these time series are obtained for each asteroid.

4. The model is trained on a set of labeled image data.

5. The model predicts the labels for a set of test images.

Carruba et al. (2022) applied Convolutional Neural Networks (CNN) for the classification of large databases of images and regularization techniques for correcting overfitting. In Carruba et al. (2024), digitally filtered images of resonant arguments were used to enhance the performance of CNNs. Finally, in Carruba et al. (2025), there was the first application of Vision Transformers.

Convolutional neural networks, or CNNs, are a neural network model originally designed to work with two-dimensional picture data. Their name derives from the convolutional layer. Convolution is a linear procedure involving the multiplication of a two-dimensional array of weights with an input array. The result of applying the filter is a two-dimensional array: the feature map.  Three of the most commonly used CNN models are the VGG (Simonyan & Zisserman 2014), Inception (Szegedy et al. 2015), and the Residual Network, or ResNet (He et al. 2015). 

Figure (3): An example of the application of Vision Transformers to the analysis of an image.

The Vision Transformer architecture for classifying images of resonant arguments was first applied in Carruba et al. (2025). The ViT model is based on the Transformer architecture (Vaswani et al. 2017), and it applies the Transformer architecture directly to image data, without the need for CNNs. In the ViT approach, an input image is split into fixed-size patches, usually 1/10 of the image size, which are then linearly embedded and fed into the Transformer encoder. The Transformer encoder consists of a series of Transformer blocks, which are made of two parts:

a. Self-Attention Mechanism: This allows the model to weigh the importance of different images for ViT, in a sequence relative to each other, enabling it to capture contextual relationships regardless of their distance in the input sequence.

b. Feed-Forward Neural Network: After the self-attention step, the output is passed through a feed-forward network, which applies transformations to the data independently for each position in the sequence.

Multiple transformer blocks can be stacked to form a complete Transformer model, allowing it to capture long-range dependencies and global information within the image.  Two key hyperparameters in our model are:

1. num_layers: The number of Transformer blocks.

2. num_heads: The number of attention heads in the Multi-Head Attention layer.

We applied CNNs and ViT to three publicly available databases of images of resonant arguments for the ν₆ (Carruba et al. 2022), g − 2g₆ + g₅ (Carruba et al. 2024a), and s − s₆ − g₅ + g₆ (Carruba et al. 2024b).

Figure (4): Semi-logarithmic plot of the computational time for applying different methods of image classification.

The models' performance was superior when applied to images of filtered resonant arguments. ViT models outperformed CNNs in terms of running times (10 times faster!) and evaluation metrics, and their results are comparable to those of models produced by the new LLM approach of Smirnov (2024).

References

Carruba V., Aljbaae S., Domingos R. C., Barletta W., 2021, Artificial Neural Network classification of asteroids in the M1:2 mean-motion resonance with Mars, MNRAS, 504, 692.

V. Carruba, S. Aljbaae , G. Carita, R. C. Domingos, B. Martins, 2022, Optimization of Artificial Neural Networks models applied to the identification of images of asteroids' resonant arguments, CMDA, 134, A59.

V. Carruba, S. Aljbaae, R. C. Domingos, G. Carita, A. Alves, E. M. D. S. Delfino, Digitally filtered resonant arguments for deep learning classification of asteroids in secular resonances, 2024, MNRAS, 531, 4432-4443.

V. Carruba, S. Aljbaae, E. Smirnov, G. Carita, 2025, Vision Transformers for identifying asteroids interacting with secular resonances, Icarus, 425C 116346.

E. Smirnov, 2024, Fast, Simple, and Accurate Time Series Analysis with Large Language Models: An Example of Mean-motion Resonances Identication. ApJ , 966(2), 220.to asteroid resonant dynamics.

K. Simonyan and A. Zisserman. Very Deep Convolutional Networks for Large-Scale Image Recognition. ArXiv e-prints , page arXiv:1409.1556, September 2014.

C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE conference on computervision and pattern recognition , pages 1#9, 2015.

K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition, 2015.

A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N Gomez, L. Kaiser, and I. Polosukhin. Attention is all you need. arXiv preprint arXiv:1706.03762, 2017.

How to cite: Carruba, V., Aljbaae, S., Smirnov, E., and Caritá, G.: Vision Transformers for identifying asteroids interacting with secular resonances., EPSC-DPS Joint Meeting 2025, Helsinki, Finland, 7–13 Sep 2025, EPSC-DPS2025-58, https://doi.org/10.5194/epsc-dps2025-58, 2025.

The solar system is home to a diverse population of small celestial objects, including asteroids, comets, and meteoroids. Most small bodies in the solar system are found in two distinct regions, known as the Main Belt and the Kuiper Belt. However, certain small bodies, such as Near-Earth Objects (NEOs), have orbits that bring them into close proximity with the Earth and in some cases even collide with our planet. The goal of Impact Monitoring (IM)
is to assess the risk of collision of a small body with Earth. Understanding the potential risk posed by an asteroid and monitoring objects with a higher risk of collision is crucial for developing strategies for planetary defense. Since in the following years, vast amounts of data from astronomical surveys will become available, it is essential to implement a preliminary filter to determine which objects should be prioritized for follow-up using traditional IM methods.

We present a novel method for estimating the Minimum Orbit Intersection Distance (MOID) of a NEO based on artificial Neural Networks (NNs). The MOID is defined as the minimum distance between the two osculating Keplerian orbits of the Earth and the NEO as curves in the three-dimensional space; it is usually used as an indicator of the possibilities of a collision between the asteroid and the Earth, at least for the period during which the Keplerian orbit of the asteroid provides a reliable approximation of the actual orbit. Since Machine Learning (ML) has gained enormous popularity in the last few years and has been applied also to some Celestial Mechanics problems, we decided to try to estimate the MOID with a multilayer feedforward NN, which takes as input the coordinates of the asteroid at a specified epoch. After being trained on an artificial dataset of about 800,000 NEOs generated with NEOPOP, the NN has been tested on the currently known population of Near-Earth Asteroids. The network exhibits near-instantaneous predictions of the MOID and achieves a mean absolute error of approximately 10−3 on the test set. Fig. 1 shows the histogram of the actual and predicted values. The overestimation of the number of asteroids with a MOID value of 0 is due to the activation function used in the final layer of the NN, namely ReLU, which, by definition, outputs 0 for any negative input. By selecting a threshold value of 0.05, we transformed the regression problem into a classification problem. In particular, we consider the positive class the one formed by all asteroids with a predicted MOID exceeding the threshold. The resulting accuracy and false positive rate (FPR) are approximately 96.61% and 2.56%, respectively. To reduce false positives, we propose to prioritize testing with classical IM methods every object with a predicted MOID of 0.10 or less. In fact, we believe that ML should serve as an initial screening tool, enabling us to prioritize follow-up assessments using traditional IM methods when managing large volumes of data.

Figure 1: Histogram of the actual and predicted values

As a follow-up, we are testing the possibility of developing a NN capable of predicting the MOID starting from computable quantities derived directly from the observations. This would eliminate the need to calculate a preliminary orbit and apply the differential corrections procedure.
Specifically, we intend to use as input vector for the NN an attributable (α, δ, ˙ α, ˙δ ), together with the second derivatives of right ascension and declination. In fact, given m ≥ 3 optical observations (αi, δi) at times ti, it is easy to compute, with a quadratic fit of both angular variables separately, the quantities α, ˙ α, ¨α and δ, ˙δ, ¨δ. Although this task is more difficult, both in terms of data acquisition and NN training, the preliminary findings are promising.

In conclusion, this research represents a step forward in addressing the urgent need for effective IM techniques, partially answering the question of whether ML can serve as a preliminary filter for some orbit determination problems.

[1] Vichi, V., Tommei, G. Exploring the potential of neural networks in early detection of
potentially hazardous near-earth objects, Celest Mech Dyn Astron 137, 17 (2025).

How to cite: Vichi, V. and Tommei, G.: Exploring the potential of neural networks in early detection of potentially hazardous Near-Earth Objects, EPSC-DPS Joint Meeting 2025, Helsinki, Finland, 7–13 Sep 2025, EPSC-DPS2025-149, https://doi.org/10.5194/epsc-dps2025-149, 2025.

Understanding the surface and subsurface temperature distributions of small bodies in the Solar System is fundamental to thermophysical studies, which provide insight into their composition, evolution, and dynamical behavior [1,2]. Thermophysical models are essential tools for this purpose, but conventional numerical treatments are often computationally expensive. This limitation presents significant challenges, particularly for studies requiring high-resolution simulations or large-scale, repeated calculations across parameter spaces.

To overcome these computational bottlenecks, we developed ThermoONet -- a deep learning-based neural network designed to efficiently and accurately predict temperature distributions for small Solar System bodies [3,4]. ThermoONet is trained on results from traditional thermophysical simulations and is capable of replicating their accuracy with dramatically reduced computational cost. We apply ThermoONet to two representative cases: modeling the surface temperature of asteroids and the subsurface temperature of comets. Evaluation against numerical benchmarks shows that ThermoONet achieves mean relative errors of approximately 1% for asteroids and 2% for comets, while reducing computation time by over five orders of magnitude.

We test the ability of ThermoONet with two scientifically compelling yet computationally heavy tasks. We model the long-term orbit evolution of asteroids (3200) Phaethon and (89433) 2001 WM41 using N-body simulations augmented by instantaneous Yarkovsky accelerations derived from ThermoONet-driven thermophysical modelling [3]. Results show that by applying ThermoONet, it is possible to employ actual shapes of asteroids for high-fidelity modelling of the Yarkovsky effect. Furthermore, we employ ThermoONet to simulate water ice activity of comets [4]. By fitting the water production rate curves of comets 67P/Churyumov-Gerasimenko and 21P/Giacobini-Zinner, we show that ThermoONet could be of use for the inversion of physical properties of comets that are difficult to achieve with traditional methods.

[1] Delbo, M., Mueller, M., Emery, J.P., Rozitis, B. and Capria, M.T., 2015. Asteroid thermophysical modeling. Asteroids iv, 1, pp.107-128.

[2] Prialnik, D., Benkhoff, J. and Podolak, M., 2004. Modeling the structure and activity of comet nuclei. Comets II, 1, pp.359-387.

[3] Zhao, S., Lei, H. and Shi, X., 2024. Deep operator neural network applied to efficient computation of asteroid surface temperature and the Yarkovsky effect. Astronomy & Astrophysics, 691, p.A224.

[4] Zhao, S., Shi, X. and Lei, H., 2025. ThermoONet: Deep learning-based small-body thermophysical network: Applications to modeling the water activity of comets. Astronomy & Astrophysics, in press.

How to cite: Zhao, S., Shi, X., and Lei, H.: ThermoONet -- Deep Learning-based Small Body Thermophysical Network, EPSC-DPS Joint Meeting 2025, Helsinki, Finland, 7–13 Sep 2025, EPSC-DPS2025-709, https://doi.org/10.5194/epsc-dps2025-709, 2025.

Asteroid families are groups of small bodies originating from a common progenitor asteroid that fragmented through collisional, tidal or rotational disruptions. Identifying these groups holds the benefit of furthering our understanding of the solar system’s dynamical and compositional evolution. Traditional ways to classify these families through the Hierarchal Clustering Method (HCM) [1] in proper orbital parameter space have a unique sensitivity to added data points. This is due to the increased possibility of chaining effects [2] that merge families without shared collisional histories into single ones, especially in the cases of complex or overlapping distributions. This sensitivity trade-off results in so-called 'halo' asteroids being excluded from core family membership.

The upcoming Vera C. Rubin Observatory’s Legacy Survey of Space and Time (LSST) is expected to detect more than 5 million new asteroids[GU1]  [3]. Entering this era of Big Data for asteroids means an increasing likelihood of chaining effects. Introducing additional parameters, such as magnitude and albedo, can add new constraints to avoid this. Artificial Neural Networks (ANNs) offer a promising approach to identifying families [4] by enabling such constraints. Training models on both real and synthetic data transforms large volumes into an asset rather than a limitation.

As these ANNs are only recently being developed, we must consider methods of confirming their family classifications, both in the realm of synthetic training data and for real results. Such methods require dedicated tools to render large, sparse datasets while also enabling productive interaction with them. A robust human interface for both data and results is thus needed that can handle multiparameter data.

The immersive Data Visualisation Interactive Explorer (iDaVIE) [5] software suite is an open-source, Unity-based virtual reality (VR) tool developed alongside the H I astronomy community for quality control of automated source finding of galaxies in SKA pathfinder projects. With iDaVIE, researchers in the field have successfully used VR both to identify real H I emission that was missed and to remove spurious detections [6-8]. Although iDaVIE’s primary scientific use has been through its tools for volumetric data analysis, its secondary particle rendering and interaction capabilities have remained underutilised. Asteroid family classification with ANNs presents a well-suited application for this flavour of iDaVIE.

In this contribution, we present early results on integrating iDaVIE to assist with ANN model development, particularly in the verification of asteroid family classification and the labelling of training data. The VR environment enables researchers to explore and validate families, whether derived from traditional techniques or modern synthetic methods. We have enhanced iDaVIE’s particle rendering capabilities and interactive tools to support this application, allowing users to inspect model classification results in an immersive 3D space. New features, such as a basic GUI interface, conditional subsets, navigation aids, and fine-tuning adjustments for rendering parameters, are described in how they were implemented and used. Many of these were either adapted from the volumetric context or developed from scratch based on our specific use cases.

Figure 1 Using iDaVIE to investigate main belt family halos in proper orbital parameter space with data from the Asteroid Family Portal [9]

We also review these use-cases, specifically how we verified results from an ANN model developed to classify asteroid halo members of the (163) Erigone family. Here we took the output probabilities of family membership and compared it to previous study results that relied on the HCM. By mapping proper orbital and physical properties to various spatial and rendering parameters in the 3D virtual space (see Figure 1), we were able to identify which halo asteroids were classified appropriately. The new interaction tools allowed us to isolate these asteroids and iteratively adjust parameters to refine and confirm results.

Our contribution demonstrates how immersive data visualisation can play a key role in advancing both methodological development and scientific discovery in planetary science, particularly when working alongside Artificial Intelligence. This also serves to open up the discussion to the broader planetary science community of how such interactive tools can be introduced to modern research workflows. A critical next step is collecting input on new features and tools that could support other applications, particularly with aiding in the development of machine learning models.

Acknowledgements:

Astronomical research at the Armagh Observatory & Planetarium is grant-aided by the Northern Ireland Department for Communities (DfC). Work by the authors was supported by the Leverhulme Trust. The visualisation work was supported by the iDaVIE development team hosted at the Inter-University Institute of Data Intensive Astronomy.

Bibliography:

• Zappala, V., et al., Asteroid Families. I. Identification by Hierarchical Clustering and Reliability Assessment. The Astronomical Journal, 1990. 100: p. 2030.
• Nesvorný, D., M. Brož, and V. Carruba, Identification and Dynamical Properties of Asteroid Families, in Asteroids IV, P. Michel, F.E. DeMeo, and W.F. Bottke, Editors. 2015. p. 297-321.
• Abell, P.A., et al., Lsst science book, version 2.0. 2009.
• Carruba, V., et al., Machine learning classification of new asteroid families members. Monthly Notices of the Royal Astronomical Society, 2020. 496(1): p. 540-549.
• Jarrett, T., et al., iDaVIE: Immersive Data Visualisation Interactive Explorer, in Zenodo Software. 2024, Zenodo: Zenodo. p. 4614115.
• Deg, N., et al., WALLABY pilot survey: the potential polar ring galaxies NGC 4632 and NGC 6156. Monthly Notices of the Royal Astronomical Society, 2023. 525: p. 4663-4684.
• Glowacki, M., et al., A serendipitous discovery of H I-rich galaxy groups with MeerKAT. Monthly Notices of the Royal Astronomical Society, 2024. 529: p. 3469-3483.
• Kleiner, D., et al., A MeerKAT view of pre-processing in the Fornax A group. Astronomy and Astrophysics, 2021. 648: p. A32.
• Novaković, B., et al., Asteroid families: properties, recent advances, and future opportunities. Celestial Mechanics and Dynamical Astronomy, 2022. 134(4): p. 34.

How to cite: Sivitilli, A., Marshall-Lee, A., and Christou, A.: Supporting Machine Learning-Based Asteroid Family Classification with Immersive Visualisation, EPSC-DPS Joint Meeting 2025, Helsinki, Finland, 7–13 Sep 2025, EPSC-DPS2025-1601, https://doi.org/10.5194/epsc-dps2025-1601, 2025.

Asteroid exploration, sample return, and defense missions critically require accurate three-dimensional (3D) reconstructions of target celestial bodies, including global shape models and high-fidelity local topographic details (Richardson et al. 2022; Thomas et al. 2023). Traditional methods, such as photogrammetry, stereophotoclinometry, and shape-from-shading have been extensively used to derive global and local 3D models from images (Gaskell et al. 2008; Palmer et al. 2022)). However, achieving reliable geometric accuracy and capturing fine-scale terrain features from limited observational data—such as sparse viewpoints and constrained illumination—remains a significant challenge.

This paper focuses on local topographic refinement, a process crucial for various advanced applications. It supports tasks such as high-precision engineering and scientific site reconstruction, provides high-detail maplets enabling more accurate global 3D reconstructions, and offers high-fidelity topographic features essential for precision Terrain Relative Navigation (TRN) (Gaskell et al. 2008, 2023; Olds et al. 2015).

The core innovation of the proposed method is the ability to achieve detailed topographic reconstruction of a local surface region using only a limited number of images—potentially even a single image—under constrained viewing and illumination conditions. To address the ill-posed problem of recovering high-frequency topographic details from sparse or limited observational data, we propose leveraging powerful diffusion priors from generative models pre-trained on low-resolution images and digital elevation models (DEMs). These priors encode geometric knowledge that can guide the reconstruction process, thereby significantly reducing the stringent requirement for extensive multi-view, multi-illumination image sets. Furthermore, the proposed generative framework is designed to inherently incorporate principles similar to GeoWizard’s geometry switcher and cross-domain attention mechanisms (Fu et al. 2024). This design facilitates a joint estimation of both elevation and implied surface orientation (normals). This joint estimation aims to ensure high geometric consistency between the derived elevation map and its corresponding surface normals, ultimately yielding a more reliable and geometrically accurate high-resolution topography.

Experimental validation was conducted on real data of NASA's OSIRIS-REx Mission to Asteroid Bennu, including MapCam, PolyCam, and NavCam images. Starting with 75 cm resolution local DEMs (99×99 grids), high-resolution DEMs were iteratively generated at 25 cm, 18 cm, and 10 cm resolutions. This process involved refining the DEM from the preceding resolution level using just 1-3 corresponding high-resolution images for each successive target resolution. The DEM with spatial resolution of 10 cm achieves less than 1 cm average root mean square error (RMSE) compared to NASA published 5 cm ground truth. Additionally, to evaluate the effectiveness of the refined DEM, we rendered it using the same illumination and observation conditions as the captured image. Scale-Invariant Feature Transform (SIFT) was then applied to match the rendered image against the captured image. The high matching success rate indicates the reconstructed terrain captures effective image textures, which further validates the reliability of our method in capturing terrain details. The refined DEM also exhibit enhanced compatibility with TRN task. The landmark matching success rate exceeds 95%, which is significantly higher than the rates achieved without terrain detail.

Reference

Fu, X., Yin, W., Hu, M., et al. (2024). Geowizard: Unleashing the diffusion priors for 3d geometry estimation from a single image. In European Conference on Computer Vision Conference (pp. 241-258)

Gaskell, R. W., Barnouin‐Jha, O. S., Scheeres, D. J., et al. (2008). Characterizing and navigating small bodies with imaging data. Meteoritics & Planetary Science, 43(6), 1049-1061.

Gaskell, R. W., Barnouin, O. S., Daly, et al. (2023). Stereophotoclinometry on the OSIRIS-REx mission: mathematics and methods. The Planetary Science Journal, 4(4), 63.

Palmer, E. E., Gaskell, R., Daly, M. G., et al. (2022). Practical stereo-photoclinometry for modeling shape and topography on planetary missions. The Planetary Science Journal, 3(5), 102.

Richardson, D. C., Agrusa, H. F., Barbee, B., et al. (2022). Predictions for the dynamical states of the Didymos system before and after the planned DART impact. The planetary science journal, 3(7), 157.

Olds, R., May, A., Mario, C., et al. (2015). The application of optical based feature tracking to OSIRIS-REx asteroid sample collection. In AAS Guidance, Navigation, & Control Conference, AAS pp. 15-124.

Thomas, C. A., Naidu, S. P., Scheirich, P., et al. (2023). Orbital period change of Dimorphos due to the DART kinetic impact. Nature, 616(7957), 448-451.

How to cite: Wang, Y., Xie, H., and Hestroffer, D.: High-Fidelity Local 3D Terrain Reconstruction for Asteroids via Generative Modeling, EPSC-DPS Joint Meeting 2025, Helsinki, Finland, 7–13 Sep 2025, EPSC-DPS2025-496, https://doi.org/10.5194/epsc-dps2025-496, 2025.

Introduction

Fluvial networks provide key insights into surface processes, underlying lithology, and the tectonic and climatic history of planetary bodies. On Earth, these drainage systems have been studied extensively, but classification still relies heavily on visual interpretation. Such manual methods are not scalable, especially when applied to large datasets or remote planetary terrains where direct observation is limited. This study presents a fully automated, unsupervised machine learning framework designed to classify fluvial patterns objectively, initially tested on terrestrial rivers but developed with a clear orientation toward planetary applications.

Methods

The classification pipeline begins with HydroRIVERS v1.0, a global hydrographic dataset, from which river segments are extracted. For each segment, a suite of morphometric descriptors is calculated, encompassing parameters related to geometry, orientation, curvature, and network topology. These features are used as inputs for unsupervised clustering. Three clustering techniques were evaluated: K-Means, Gaussian Mixture Models (GMM), and CLARANS. Among them, K-Means consistently delivered the highest internal validation scores across silhouette, Davies-Bouldin, and Dunn indices. CLARANS, while computationally more intensive, offered greater interpretability by selecting real river segments as cluster centers, which is especially valuable in a geomorphological context. Classification at the river-system level was then achieved by exploring multiple strategies, including majority voting among segments, and aggregation of morphometric features by their mean or median; they all produced reliable and coherent drainage typologies.

Results

The optimal solution identified six distinct morphometric clusters. These clusters were subsequently interpreted and labeled based on canonical fluvial patterns, such as dendritic, sub-dendritic, radial, trellis, and rectangular. This interpretation followed geomorphological typologies standardized by Donadio et al. (2021) (Figure 1), enabling the translation of purely numerical groupings into geological meaning.

Figure 1. Different classes of drainage patterns following the scheme proposed by Donadio et al. 2021: a) dendritic; b) sub-dendritic; c) pinnate; d) parallel; e) radial; f) rectangular; g) trellis; h) angular; i) annular; j) contorted.

Validation against a reference dataset containing thousands of manually classified segments showed strong agreement. The K-Means algorithm achieved high consistency with expert labels, while CLARANS proved useful in highlighting key reference cases. To visualize the effectiveness of clustering, a principal component analysis (PCA) was performed, projecting the high-dimensional feature space into three dimensions (Figure 2).

Figure 2. Images of the resulting distribution, where the six clusters form distinct, compact groupings: 2a shows the clustering results with k-means algorithm; 2b shows the clustering results with CLARANS algorithm.

Planetary Application and Discussion

The versatility of the proposed framework makes it well suited for planetary science. Many bodies in the Solar System, such as Mars and Titan, show evidence of ancient or current fluvial activity. These features, visible in high-resolution orbital imagery, share morphometric properties with terrestrial rivers. The automated and objective nature of our method is ideal for application to planetary surfaces, where field validation is not possible and manual classification is impractical. On Mars, dendritic valley networks such as Warrego Valles (Figure 3) could be objectively identified and differentiated from structurally controlled systems like those near Valles Marineris. On Titan, channel networks via Cassini RADAR data present complex patterns possibly influenced by tectonics or cryovolcanism. The ability to distinguish between morphometric types without supervision may support hypotheses on climate evolution and crustal processes in these environments.

Figure 3. A visual comparison between a dendritic river system on Earth and a Martian network: 3A shows the original image of Po River (Italy) before pre-processing; 3B shows the manually extracted image of Po River. 3B shows the original image of Warrego Valles (Mars) before pre-processing; 3B shows the manually extracted image of Warrego Valles (Mars).

In the future, this approach may also be extended to Venus, where missions like ESA’s EnVision could reveal previously hidden fluvial features through synthetic aperture radar. Additionally, expanding the feature set to include elevation data and terrain roughness could further enhance classification capability and geological interpretation.

Conclusions

This study introduces a robust, unsupervised framework for the classification of fluvial networks that moves beyond the limitations of subjective interpretation. While trained and tested on terrestrial data, the method is explicitly designed with planetary applications in mind. It demonstrates strong potential for analyzing fluvial systems on Mars, Titan, and other planetary bodies, offering new perspectives on landscape evolution and hydrologic history. Future work will focus on applying the pipeline to specific planetary case studies, enhancing feature sets, and integrating data from multiple remote sensing platforms.

How to cite: D'Aniello, M., Velotti, C., Esposito Mocerino, G., and Donadio, C.: A Novel Machine Learning Approach for Objective Fluvial Network Classification: Earth & Beyond, EPSC-DPS Joint Meeting 2025, Helsinki, Finland, 7–13 Sep 2025, EPSC-DPS2025-1703, https://doi.org/10.5194/epsc-dps2025-1703, 2025.

We present an overview of the performance and improvements of the Occultation Detection neural Network framework: ODNet, a convolutional neural network for asteroid occultation detection (Cazeneuve+ 2023). ODNet is the backbone of the occultation detection program of the Unistellar network; an expanding network of 15,000+ smart telescopes distributed around the world that allows citizen scientists to join various observational campaigns and provide invaluable data for multitude of scientific projects.

The original ODNet was built using the TensorFlow architecture and was trained on mostly synthetic data. The purpose of ODNet is to provide an independent and reproducible measure of the likelihood of the occultation event in the observed data set (a set of individual image frames), remove any influence of human analysis, and to allow a fully autonomous detection of occultations in all Unistellar observations in the occultations campaign.

The original results were promising. ODNet was able to analyze raw data from occultation events in minutes with very high precision (91%) and recall (87%). This enabled a completely automated pipeline that is able to handle multiple occultation events per day from multiple locations and observers. However, as more observations were made and different observing configurations, conditions, and telescope models appeared the efficiency and precision of the original model was not satisfactory.

Here, we present improvements to the original ODNet framework by increasing its precision and recall to >95%, ability to handle very short occultation events (<4 frames / 1 second), deal with more diverse observation conditions, and able to deal with datasets from various telescopes. We improved identification of false positives and were able to recover a number of short occultation events in the archival data. We also added a fully automated light curve analysis that provides additional information about the nature of the observation and significance of the event.

Ultimately, we show our latest advancements in our search for a lightweight solution that is based on a YOLOv8 framework (You-only-look-once) that would allow the observer to analyze both predicted and serendipitous occultation events in real time while using processing units on modern mobile devices.

How to cite: Pokorny, P., Hanus, J., Marchis, F., and Esposito, T. M.: Detecting stellar occultations using machine-learning techniques and smart telescopes powered by citizen scientists, EPSC-DPS Joint Meeting 2025, Helsinki, Finland, 7–13 Sep 2025, EPSC-DPS2025-982, https://doi.org/10.5194/epsc-dps2025-982, 2025.

Introduction

Impact cratering shapes the surfaces of airless bodies such as the Moon and Mercury, providing both a record of impacts history and a chronostratigraphic tool for dating geological units. Traditional crater catalogs, compiled manually or via semi-automated algorithms, remain incomplete below ~1 km diameter on the Moon[1] and ~5 km on Mercury[2], limiting studies of small-scale resurfacing, secondary-crater populations, and polar processes.

This study has compiled the first impact-crater dataset for Mercury by a multimodal deep-learning pipeline. Our approach adapts YOLOLens single-stage object detector model initially fine-tuned on lunar imagery to 3-bands inputs and through domain-adaptive transfer learning is applied on Mercury.

The key innovation is the construction of a three-band cube for each orthographic tile:

• WAC/MDIS visible spectrum mosaics

• Digital Terrain Model (DTM) 

• Hillshaded DTM version

Methods

Each band is co-registered at 100m/px resolution and normalized to [0,1] to ensure balanced feature learning . Tiles overlap by 128 px to mitigate edge effects, and a triple-scale subsampling captures craters whose diameters exceed individual tile dimensions .

To achieve acceptable generalization model capabilities, a balancing function is applied across latitude ranges and FoVs to ensure a robust model, finally, the entire workflow is shown in Fig.1.

                                                                                                            Fig. 1: Workflow.

The YOLOLens model[3], leverages on super-resolution techniques to improve impact crater detection on planetary surfaces. By addressing challenges such as low spatial resolution and variable solar illuminance in satellite imagery, YOLOLens demonstrated enhanced precision compared to state-of-the-art methods.

The parameters are modified at the first/last layers such that the model creates super-resolution (SR) images from a single-channel image. Then, the SR output is used by an object-detection model in an end-to-end model.

Results

The model has been trained across the lunar dataset (Robbins catalogue[1]) to demonstrate the effectiviness of the multimodal approach respect to the catalog derived in[4], thus, we applied a transfer learning step rapidly adapts to Mercury’s distinct albedo contrasts while retaining learned shape priors from the Moon.

Validation on multiple quadrangles yields recall rates of 58% – 95.7% across primary/secondary craters, confirming robust cross-body generalization (Tab.1 and Tab.2).

TABLE 1: Comparison of Crater Diameter Counts for Mercury.

The Herrick catalog reports initial insights derived from the database, while also addressing several limitations that constrain future analyses. Our validation set was compiled through different craters manual counts[5,6,7] see Tab.2 .

TABLE 2: Summary of crater detection performance by YOLOLens across different datasets.

In the investigated regions, high recall values are obtained, suggesting robust performances of the YOLOLens model to identify impact craters. Since the areas used for comparison are representative of all the latitudes, such robustness could be extended to all the Hermean surfaces. Fig. 2 shows the size-frequency distributions.

Fig. 2: Comparison of size-frequency distributions (SFDs) between the ME6M300TGT and [6], [7].

The development of a comprehensive global catalog allows us to investigate the distribution of craters and their morphometric characteristics, and from them observe clues about the evolution of the hermean surface. The crater density information provides us with important insights about the age of the different regions and the rates of degradation or remodeling (Fig.3).

Fig. 3: (a) Extract from the Geological Map of the Hokusai quadrangle (H05)[8] and (b) the same region extracted from the ME6M300TGT to derive the density map by diameter range.

An important parameter for the surface characterization is the depth-to-diameter (d/D) ratio, which is useful to define the post-impact degradation and modification of the crater. The d/D was measured for the entire catalog. Fig.4 shows the distribution of the d/D of the craters.

Fig. 4: The map shows the d/D distribution of Mercury’s craters. A) Diameter > 15km over Mercury, B) The North Pole (> 5km) and C) The South Pole (> 15km).

This study presents advancements in planetary surface analysis through the development of a comprehensive crater catalog for Mercury, which exploit images and topographical information simultaneously. The YOLOLens model has demonstrated its capability to detect craters with great accuracy and scale. The datasets also integrate morphometric information, enabling novel insights into crater distributions and the geological evolution of these celestial bodies. In addition, the catalog can function as an important tool in cartographic endeavors currently compiled as a multi-mapper regional-scale project[9].

References

[1] Robbins (2019) J. Geophys. Res. Planets, 124(4), 871-892.

[2] Herrick et al. (2011) Icarus 215(1), 452-454.

[3] La Grassa et al. (2023) Remote Sens. 15(5), 1171.

[4] La Grassa et al. (2025) ISPRS J. Photogramm. 220, 75-84.

[5] Giacomini et al. (2020) Geosci. Front. 11(3), 855-870.

[6] Galluzzi et al. (2016) J. Maps 12(sup1), 227-238.

[7] Martellato et al. (2023) Bull. AAS, 55(8), #116.06.

[8] Wright et al. (2019) J. Maps 15(2), 509-520.

[9] Galluzzi (2019) In: Hargitai (eds) Planetary Cartography and GIS, Springer.

Acknowledgements: we gratefully acknowledge funding from the Italian Space Agency (ASI) under ASI-INAF agreement 2024-40.HH.0

How to cite: La Grassa, R., Re, C., Martellato, E., Tullo, A., Bertoli, S., Galluzzi, V., Giacomini, L., Vergara Sassarini, N. A., Cremonese, G., and faletti, M.: Mercury Global Crater Catalog using Multimodal Deep Learning for Crater Detection and Morphometric Analysis, EPSC-DPS Joint Meeting 2025, Helsinki, Finland, 7–13 Sep 2025, EPSC-DPS2025-1625, https://doi.org/10.5194/epsc-dps2025-1625, 2025.

Introduction: The large amount of data from orbital imagery requires time-consuming analyses, especially for high-resolution products. On Mars, the High Resolution Imaging Science Experiment (HiRISE, McEwen et al., 2007) gathers several thousand images (McEwen et al., 2024). Automated algorithms are needed to extract data from such a large number of images. Small impact craters (with a diameter below 50 meters) are geological features that provide key information about the subsurface structure or the landscape changes through time. For now, a global mapping effort has been completed for craters larger than 1 km (Robbins et al., 2012). The challenge to tackle is now to pursue this task for lower diameters by using automated algorithms. 

Methods: We trained the well-known You Only Look Once (YOLO, Redmon et al., 2016) v11 algorithm to segment small craters over HiRISE images (McEwen et al., 2007). YOLO's architecture is a Convolutional Neural Network (CNN, e.g. Ronneberger et al., 2015) used for the segmentation and classification of images, and has been successfully applied on Martian data (e.g. Benedix et al., 2020). We used an exhaustive count (Millot et al., 2024) of 83820 craters located in Candor Chasma, with diameters ranging from ~2 m to 270 m over the HiRISE image ESP_024255_1750, for the training of the algorithm. We cut the original HiRISE image and the target file into tiles of 256 by 256 pixels. Then, we split the resulting dataset into training, validation, and test sets following a 70%, 20%, and 10% ratio. We trained the YOLO medium algorithm ('yolo11m-seg') over 100 epochs, following the set of hyperparameters defined in Table 1. Finally, we tested the algorithm over a new HiRISE image, ESP_026477_1615, located in the magmatic terrains of Tyrrhenus Mons.

Table 1: The set of hyperparameters for the ‘yolo11m-seg’ algorithm training.

Results and perspectives: The training process reached a mean Average Precision (mAP) of around mAP = 65% for the best models we trained. Precision P and recall R were respectively P = 68% and R = 58% for the best models. Most of the medium-sized craters, between 5 and 20 m in diameter, were segmented correctly by the algorithm after the training. Improvements can still be made on small (< 5 m) and large (over > 50 m) crater detections and segmentations. Figure 1 displays some examples of four tiles from the HiRISE image ESP_026477_1615. Panel 1a shows that the largest craters from a cluster are correctly mapped (blue masks), while some of the smallest in diameter are missing. Panel 1b presents one accurate result over a fresh and large crater (top left corner of the tile) along with a wrong detection (lowest part of the tile). Subplot 1c highlights the two detections of a large crater and a small crater set on a lava terrain. Finally, panel 1d underlines that one large crater (more than 100 m in diameter) is not mapped by the algorithm. Results are still variable but remain encouraging, as crater mapping has been made over a new HiRISE image, which has not been involved in the training process.

The potential to automatically map the small-sized craters (< 1 km) on planets with an atmosphere will bring new insights into the active crater modification processes (erosion or aeolian infilling).

Figure 1: Four tiles from the HiRISE image ESP_026477_1615 summarizing predictions (blue masks) from our trained YOLO model.

References

• K. Benedix, et al., “Deriving surface ages on Mars using automated crater counting.” Earth and Space Science, 7, e2019EA001005. https://doi.org/ 10.1029/2019EA001005. 2020.
• A McEwen, et al., “Mars reconnaissance orbiter's high resolution imaging science experiment (HiRISE).”Journal of Geophysical Research: Planets 112.E5 (2007).
• A.S. McEwen, et al., “The high-resolution imaging science experiment (HiRISE) in the MRO extended science phases (2009–2023)”, Icarus, 419, https://doi.org/10.1016/j.icarus.2023.115795, 2024.
• Millot, “Dataset from Millot et al., 2024”. Zenodo. doi: 10.5281/zenodo.14035904. 2024.
• Redmon, et al., “You Only Look Once: Unified, Real-Time Object Detection,”  IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA, pp. 779-788, doi: 10.1109/CVPR.2016.91, 2016.
• J. Robbins, & B.M. Hynek, “A new global database of mars impact craters > 1 km: 1. database creation, properties, and parameters”. Journal of Geophysical Research: Planets, 117 (E5). Retrieved from https://agupubs.onlinelibrary.wiley.com/doi/abs/10.1029/2011JE003966 doi:https://doi.org/10.1029/2011JE003966, 2012a.
• Ronneberger, et al., “U-net: Convolutional networks for biomedical image segmentation.” Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18. Springer international publishing, 2015.

How to cite: Millot, C., Quantin-Nataf, C., and Volat, M.: Supervised Deep Learning for Mapping Small Craters on Mars, EPSC-DPS Joint Meeting 2025, Helsinki, Finland, 7–13 Sep 2025, EPSC-DPS2025-1547, https://doi.org/10.5194/epsc-dps2025-1547, 2025.

Introduction:  Determining the internal mass distribution of planetary bodies, such as Ganymede, remains a challenging problem due to observational degeneracies. In the 2030s the JUICE mission with its several instruments will orbit Ganymede and provide information on parameters that depend on the interior structure of the moon, including estimates of the polar moment of inertia, the radial and gravitational Love numbers and associated phase-lags, the longitudinal libration amplitudes, as well as the phase and amplitude of the induced magnetic field due to the presence of a subsurface ocean. In order to impose the constraints on the interior structure in the most effective way a joint inversion of all available parameters is ideally  necessary. In this work, we use a machine learning approach  to predict the thicknesses and densities of Ganymede's internal layers and ocean conductivity using these parameters. To achieve this,  a synthetic dataset of  plausible internal structure models of Ganymede is generated via Monte Carlo sampling. For each of these internal structures we compute the corresponding observable parameters (Love numbers, libration amplitude, polar moment of inertia, etc.) using existing  models. We then train a Neural Network on the synthetic dataset to learn the intricate relationships between these parameters and the internal structure model. 

Our model is able to retrieve the internal structure parameters with varying levels of accuracy across different layers, with promising  performance in the prediction the icy shell and ocean thickness and density, ocean conductivity and  thickness of the high-pressure ice layer. The Monte Carlo dropout method is utilized to estimate the uncertainties in the predicted parameters. These results highlight the potential of machine learning as a preliminary and fast tool to detect families of interior structures compatible with the observed parameters. 

Interior Structure Model: The interior of the icy satellites is modeled as several spherically symmetric, uniform shells and is completely specified by the values of radius R_i, density ρ_i, rigidity µ_i, and viscosity η_i of each layer. The five layers are: an icy shell, a liquid subsurface ocean, an High-Pressure-ice (HP-ice) layer, a silicate mantle and a solid inner core.  For the solid layers we adopt an Andrade rheology, while the ocean and liquid core are treated are treated as inviscid fluids. 

Figure 1: schematic representation of the an internal structure model for Ganymede with five layers.

Dataset and training:  we trained the Neural Network on a synthetic dataset consisting of 10⁷ interior structures, that we generated by performing a Monte-Carlo sampling of the internal structure parameters y, namely: thicknesses and densities of each layer,   icy shell viscosity and ocean conductivity, subject to the total radius and mass constraints. For each interior structure we then computed a set of observables x, including the polar moment of inertia, the radial and gravitational Love numbers h₂ and k₂ using the ALMA³ code [1], the libration amplitude L_s at the orbital period [2], as well as the amplitude A and phase φ_A of the induced magnetic field at the orbital period [3]. 

Neural Network architecture. A schematic representation of the Neural Network is shown in Fig. 2.  The minimization of the loss function is performed using Adam optimizer [5]. In order to prevent any issues with model overfitting we adopt early stopping [6]. We train the neural network with 80% of the data set and use the remaining 20% for validation.

Figure 2: schematic representation of the neural network

Montecarlo dropout. In order to capture the uncertainty in the predicted parameters it would be desirable to have a posterior distribution of the interior structure parameters, rather than deterministic values. To do this we use the Montecarlo dropout approach [7], which allows to obtain an approximate Bayesian inference through dropout training. 

Results and Conclusions:  Our model demonstrates significant predictive accuracy in estimating the thickness and density distributions of Ganymede like icy satellites across its five-layer interior. In Fig. 3, we show a comparison between the actual values of the interior parameters y and those predicted by our trained neural network for the validation dataset y. A perfect prediction would fall on the green dashed line. From the validation dataset, the neural network effectively captured the characteristics of the icy shell and ocean, with excellent agreement between predicted and actual values. The HP-ice layer’s thickness was predicted with moderate accuracy, while its density estimates showed higher variability. The model performs poorly in the task of inferring the thickness and density of the core and mantle, suggesting limited sensitivity of the selected observables to these parameters. However, this had to be expected, and is in line with previous results present in the literature. 

 In Fig. 4 we show the posterior distributions obtained with the Montecarlo dropout method corresponding to the set of parameters x* drawn from the synthetic dataset. We observe that the true values of the internal structure parameters y*, shown as dashed vertical lines, fall within the posterior distributions, close to the mean values in the case of the icy shell, the ocean, and HP-ice layer. However, the methodology presented here come short in the task of assessing the uncertainty in the deeper interior, as the intrinsic degeneracy in the inverse problem allows for a broader range of deep interior structures than those predicted by the neural network.

Figure 3.  

Figure 4.  

References:  [1] Melini D., et al., 2022, Geophysical Journal International, 231, 1502 [2] Baland R.-M., Van Hoolst T., 2010, Icarus, 209, 651 [3] Vance S. D., et al., 2021, Journal of Geophysical Research: Planets, 126, [4] Srivastava N., et al., 2014, Journal of Machine Learning Research, 15, 1929  [5] Kingma D. P., Ba J., 2014, CoRR, abs/1412.6980 [6] Prechelt L., 1998, Neural Networks, 11, 761 [7] Gal Y., Ghahramani Z., 2016, in Balcan M. F., Weinberger K. Q., eds, Proceedings of Machine Learning Research Vol. 48.

How to cite: Macrì, G. and Casotto, S.: Enforcing multiple constraints on the interior structure of Ganymede: a machine learning approach, EPSC-DPS Joint Meeting 2025, Helsinki, Finland, 7–13 Sep 2025, EPSC-DPS2025-691, https://doi.org/10.5194/epsc-dps2025-691, 2025.

Simulating exoplanetary atmospheres is essential for describing them, estimating their composition, investigating the presence of haze and clouds, and identifying their relationship with observational signatures. With the emergence of JWST (James Webb Space Telescope) and the upcoming ARIEL (Atmospheric Remotesensing Infrared Exoplanet Large-survey) mission (Tinetti et al. 2018; Pascale, Bezawada, and al. 2018), the need for a fast implementation of the classical forward models is more important than ever. Machine learning (ML) can offer a revolutionary solution to this challenge by providing efficient surrogate models or emulators that approximate the behavior of specific components or the whole system, significantly accelerating the
simulation pipeline.
In our research in general, we are using a 1D self-consistent model including haze/cloud microphysics, disequilibrium chemistry, and radiative transfer interactions to simulate the atmospheric structure of temperate exoplanets from the deep (10³ bar) to the upper thermosphere ( 10⁻¹⁰ bar). This forward model was used in exoplanet studies, like in Arfaux and Lavvas (2022) and many other studies (Lavvas et al. 2019; Arfaux and Lavvas 2023; Lavvas, Paraskevaidou, and Arfaux 2023), offering a more detailed correspondence of the atmospheric composition with the transit observations. In this work we will try to develop a supervised neural network-based surrogate model, with inspiration from Hendrix, Louca, and Miguel (2023), trained on the outputs of the forward model, enabling fast approximation of atmospheric responses for a range of exoplanetary parameters (like planet mass, stellar radius, temperature-pressure profile, stellar flux, metallicity, etc.) without repeated execution of the full model.
We are currently replacing the microphysics of the forward model, which simulates the photochemical haze particle size distribution over a grid of particles radii, with a neural network trained on a given temperature (isothermal), pressure profile, viscosity (which correlates with metallicity), eddy mixing, and gravity. Our goal is to replace the entire forward model (or at least the most time-consuming parts) and improve atmospheric characterization speed and accuracy. This work has the potential to greatly benefit the research community by making comparative studies across planetary systems more accessible to a wider range of groups. It can also be used in hybrid frameworks in which ML handles expensive subcomponents (e.g., radiative transfer) and traditional models handle dynamics, preserving physical interpretability while increasing efficiency.

References
Arfaux, Anthony and Panayotis Lavvas (June 2022). “A large range of haziness conditions in hot-Jupiter atmospheres”. In: Monthly Notices of the Royal Astronomical Society 515.4, pp. 4753–4779. issn: 0035-8711. doi: 10.1093 /mnras /stac1772. eprint: https : / / academic.oup .com /mnras /article - pdf/515/4/4753/45475436/stac1772.pdf. url: https://doi.org/10.1093/mnras/stac1772.
Arfaux, Anthony and Panayotis Lavvas (Apr. 2023). “A physically derived eddy parametrization for giant planet atmospheres with application on hot-Jupiters”. In: Monthly Notices of the Royal Astronomical Society 522.2, pp. 2525–2542. issn: 0035-8711. doi: 10 . 1093 / mnras / stad1135. eprint: https : / /
academic . oup . com / mnras / article - pdf / 522 / 2 / 2525 / 50113176 / stad1135. pdf. url: https ://doi.org/10.1093/mnras/stad1135.
Hendrix, Julius L A M, Amy J Louca, and Yamila Miguel (June 2023). “Using a neural network approach to accelerate disequilibrium chemistry calculations in exoplanet atmospheres”. In: Monthly Notices of the Royal Astronomical Society 524.1, pp. 643–655. issn: 0035-8711. doi: 10 . 1093 / mnras / stad1763.eprint: https://academic.oup.com/mnras/article-pdf/524/1/643/54758485/stad1763.pdf. url: https://doi.org/10.1093/mnras/stad1763.
Lavvas, Panayotis, Sofia Paraskevaidou, and Anthony Arfaux (Oct. 2023). “Photochemical hazes clouds in the atmosphere of GJ 1214 b in view of recent JWST observations”. In: 55th Annual Meeting of the Division for Planetary Sciences, id. 223.08. Bulletin of the American Astronomical Society e-id 2023n8i223p08 55.8. url: https://ui.adsabs.harvard.edu/abs/2023DPS....5522308L/abstract.
Lavvas, Panayotis et al. (June 2019). “Photochemical Hazes in Sub-Neptunian Atmospheres with a Focus on GJ 1214b”. In: 878.2, 118, p. 118. doi: 10.3847/1538-4357/ab204e. arXiv: 1905.02976 [astro-ph.EP].
Pascale, Enzo, Naidu Bezawada, and et al. (July 2018). “The ARIEL space mission”. In: Space Telescopes and Instrumentation 2018: Optical, Infrared, and Millimeter Wave. Ed. by Makenzie Lystrup et al. Vol. 10698. Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, 106980H, 106980H. doi: 10.1117/12.2311838.
Tinetti, Giovanna et al. (Nov. 2018). “A chemical survey of exoplanets with ARIEL”. In: Experimental Astronomy 46.1, pp. 135–209. doi: 10.1007/s10686-018-9598-x.

How to cite: Paraskevaidou, S.: Implementing a Neural Network on Forward Models:A Case study for Exoplanet Atmospheres, EPSC-DPS Joint Meeting 2025, Helsinki, Finland, 7–13 Sep 2025, EPSC-DPS2025-272, https://doi.org/10.5194/epsc-dps2025-272, 2025.

 Introduction

Launched in 2006 aboard NASA/Jet Propulsion Laboratory's Mars Reconnaissance Orbiter spacecraft, the HiRISE (High-Resolution Imaging Science Experiment)[1] has provided the highest resolution (up to 25cm/pixel) of orbital imagery of Mars for more than twenty years. It features a set of 3 different color filtered CCDs: the RED array has a range of 570-830nm, while the blue-green (BG) and near-infrared (IR) filters covers respectively less than 580nm and greater than 790nm. Color imagery provide valuable information to create geological maps, such as was done by Mandon et al[2].

The RED sensor provides a wide angle of view, but the "color" sensors (BG, IR) have a narrower coverage of 20% of the center of the RED swath. This limitation, aggravated by the already low global coverage of the HiRISE dataset, means planetary scientists must work with very limited data.

The overlap of the RED and colored datasets mean we have a strong target set to train a deep neural network (DNN) to extrapolate the NIR and BG channels and provide an extended colored dataset. While there have been various work on using such algorithm to colorize imagery, we are not aware of previous work targeting martian orbital data.

Architecture

We use a straightforward U-Net model, which are based on autoencoders. An autoencoder is the result of a function that convert its input into a different representation (coded) by a function that transform coded data into the original representation. When applied on imagery, the encoding function often take the form of layers of convolution filters, and the decoder of transpose convolutional filters. The U-Net architecture concatenate each encoding layer to the result of the corresponding decoding layer to steer the process, as defined by Ronneberger et Al[3].

We define the images size to 256x256 pixels to run the model on individual data that provide a rich feature set. We define 5 levels of encoding and decoding to account for the input data size.

One signifiant difference with previous works is that given the different nature of HiRISE channels compared to "normal" imagery, we do not convert the RGB data into a perceptual colorspace as it would be usually done, as our channels do not have the same relationship as RGB channels have. We set the model target to be the NIR and BG channels of an HiRISE image.

To train this model, we evaluate the validation results using L1 norm, mean square error (MSE), binary cross-entropy (BCE) function as loss function. When evaluating the structural similarity (SSIM) of validation step, we observe very close results with the MSE providing slightly better results.

Training dataset

The “color” dataset of HiRISE provide us a needed training target. In theory, we could randomly select datasets across the complete catalog, but we found that operational condition randomness leaves us with varying levels of image quality.

A first step in our training was to establish a criteria to find “good” references images. We decided to approach this issue based on how the RED and BG are correlated: both those channels are first normalized and 256 buckets histograms are computed. The histograms values are also normalized by the number of input pixels to that image size do not impact the values. We smooth those histograms with a gaussian filter and compute the product of both histograms and sum the resulting values.

For a first attempt, we selected and scored 228 images over the Mawrth Vallis region using the MarsSI[4] portal to select and process data, both due to the large amount of observation over this area and the diversity of terrain of this area. We selected the 35 best-scored images and split them in 256x256 tiles, our objective being to train the model to run at 1:1 scale. This resulted in a training dataset of 27,555 images, of whom 23,421 where used for loss function computation and 4,134 for validation during the training. With this dataset, we find that the model will converge over 25 epochs.

 Results

As a first step, we expect to validate the results of our colorizer on images that would have highly correlated R/BG histograms as this is the kind of data we selected for our training. At this step of our study, we only evaluate the BG channel generation as the “natural colors” images evaluation comes easier to the human eye (analytic evaluation is already done during the training).

First results are encouraging revealing that color contrasts can be well reproduced forming image products directly usable for mapping. The red vs blue balance of some units is not always perfectly reproduced like the dark capping unit of the image below, redder in the training set while bluish in the results. Our perspective is to play with the training set and to apply the pipeline to RB color.

Closeup sample of recolorized (left) and COLOR (right) version of image ESP_016196_2050

References

[1] McEwen, Alfred S., et al. "Mars reconnaissance orbiter's high resolution imaging science experiment (HiRISE)." Journal of Geophysical Research: Planets 112.E5 (2007).

[2] Mandon, Lucia, et al. "Morphological and spectral diversity of the clay-bearing unit at the ExoMars landing site Oxia Planum." Astrobiology 21.4 (2021): 464-480.

[3] Ronneberger, Olaf, Philipp Fischer, and Thomas Brox. "U-net: Convolutional networks for biomedical image segmentation." Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18. Springer international publishing, 2015.

[4] Quantin-Nataf, C., et al. "MarsSI: Martian surface data processing information system." Planetary and Space Science 150 (2018): 157-170.

How to cite: Volat, M., Quantin-Nataf, C., and Millot, C.: HiRISE image colorization with a U-Net deep neural network, EPSC-DPS Joint Meeting 2025, Helsinki, Finland, 7–13 Sep 2025, EPSC-DPS2025-969, https://doi.org/10.5194/epsc-dps2025-969, 2025.