A Sampling Framework for Machine Learning–Based Retrieval of Surface Composition in Small Bodies ​I. TNOs case​

The era inaugurated by the James Webb Space Telescope revealed the presence of carbon dioxide, carbon monoxide, water ice in amorphous and crystalline phases, methanol, and a forest of complex organic materials on the surface of small and mid-sized trans-Neptunian objects (TNOs) [1-3]. The TNOs can be seen as a population of small bodies in the solar system that have undergone little surface alteration since their formation to the present day. Due to that, constraining the properties (e.g., abundances, grain sizes, distribution) of these materials on their surface can help us to understand the protoplanetary disk conditions at the place where they formed and how the migration process in the late stages of the outer solar system formation affected them. 

Retrieving the surface compositional properties of TNOs through modelling techniques (e.g., Hapke, Shkuratov) using a large spectral range is not a straightforward task. Depending on the chosen modelling technique, the number of free parameters to estimate and the degenerated solutions can overwhelm the process. In order to obtain a quantification of the abundances and grain sizes of the materials present on TNOs’ surface, we are training a machine learning (ML) algorithm with a library of synthetic spectra containing materials that are expected to be found there. Our goal is to obtain a tool able to find a match for a real JWST/NIRSpec spectra of a TNO that belongs to the prominent water ice or prominent organic groups [4] in the dataset we create, and, in this way, constraint their surface compositional properties.  

In this work, we present a preliminary database containing 1000 synthetic spectra created with the Hapke model [5] and optical constants of ice, red, and dark materials for training purposes of the ML match algorithm. To avoid repetitions and clustering in the training dataset, we use the Latin Hypercube sampling statistical tool [6] combined with a Dirichlet distribution for generating a near-random sample of parameter values for abundance and grain size parameters (Fig. 1). As the abundances must sum to unity, we implemented a weight system to avoid an oversample of abundances at 1/N, where N is the number of materials used in the mixtures. The next step is to evaluate the physical properties of the synthetic models and train the ML algorithm to recognize the compositional groups we are interested in.

Figure 1. Abundance (left side) and grain size (right side) distribution of an experimental dataset with 1000 mixtures containing five materials each. The grain size selected range is from 7 to 150 microns.

References:

[1] De Pra et al. 2025. Widespread CO2 and CO ices in the trans-Neptunian population revealed by JWST/DiSCo-TNOs. Nature Astronomy, Volume 9, p. 252-261.

[2] Pinilla-Alonso et al. 2025. A JWST/DiSCo-TNOs portrait of the primordial Solar System through its trans-Neptunian objects. Nature Astronomy, Volume 9, p. 230-244

[3] Fernandez-Valenzuela et al. 2021. Compositional Study of Trans-Neptunian Objects at λ > 2.2 μm. Planet. Sci. J. 2 10.

[4] Holler et al. 2025. A Descriptive Taxonomic Nomenclature for Intermediate-sized Trans-Neptunian Object Spectra. Res. Notes AAS 9 241.

[5] Hapke 2012. Theory of Reflectance and Emittance Spectroscopy, by Bruce Hapke. Cambridge: Cambridge University Press. OCLC: 775869853. eISBN: 9781139025683.

[6] Songa & Kawai, 2023. “Monte Carlo and variance reduction methods for structural reliability analysis: A comprehensive review”. Probabilistic Engineering Mechanics 73, 103479.