Performance of neural networks after training with first training + validation splits with regard to the spectrum contained in the validation subset. The plot shows predicted and true labeling values for molecular fractions and temperature. Left panel: Molecular fractions of the ice spectrum. Each ice spectrum (specific composition and temperature) within the validation subset corresponds to six circles in the plot. One is per target molecule (shown in color). The size of the circle indicates the temperature of the ice. The lower panel reports differences between predicted molecular fractions and experimental label values. Right panel: Temperature of the ice spectrum, represented by each circle. Again, the bottom panel reports the difference between the predicted and indicator values. – Astro-Ph.im
Current observations taken by James Webb Space Telescope (JWST) allow us to observe the absorption function of the ice mantle, which covers interstellar dust particles, mainly composed of H.2o, co, and co2along with other small species.
Thanks to its sensitivity and spectral resolution, JWST could observe ice features towards hundreds of sources at various stages along the star formation process. However, identifying spectral features of different species and quantifying ice composition is not trivial and requires complex spectroscopic analysis.
We present an automatic ice composition estimator (AICE), a new tool based on artificial neural networks. Based on infrared (IR) ice absorption spectra of 2.5-10 microns, AICE predicts fractional composition of ice from the standpoint of H2O, Co, co2ch3Ah, NH3and ch4. To train the model, we used hundreds of laboratory experiments of ice mixtures from various databases reprocessed with baseline subtraction and normalization.
Trained AICE takes less than 1 second on a traditional computer, predicting ice composition related to the observed IR absorption spectrum, with a typical error in speciation rate of approximately 3%. The performance was tested on two spectra reported to the background stars of NIR38 and J110621 observed within the JWST ICE AGE program, showing good agreement with previous estimates of ice composition.
The rapid and accurate performance of AICE allows for systematic analysis of hundreds of different ICE spectra, allowing for a modest time investment. Furthermore, this model can be enhanced and retrained with more experimental data, improving prediction accuracy and expanding the list of predictive species.
Andrés Megías, Izaskun Jiménez-Serra, François Dulieu, Julie Vitorino, Belén Maté, David Ciudad, Will RM Rocha, Marcos Martínez Jiménez, Jacobo Aguirre
Comment: 24 pages, 20 digits. I accepted that it will be featured in Astronomy and Astrophysics
Subjects: Astrophysics of Galaxies (Astro-PH.GA); Astrophysics of Earth and Planets (Astro-Ph.ep); Astrophysics Instruments and Methods (Astro-PH.IM); Astrophysics of Sun and Stars (Astro-PH.SR)
Quote: arxiv: 2509.04331 [astro-ph.GA] (Or arxiv: 2509.04331v1 [astro-ph.GA] For this version)
https://doi.org/10.48550/arxiv.2509.04331
Focus to learn more
Submission history
From: Andrés Megías
[v1] Thu, September 4, 2025, 15:49:28 UTC (6,652 kb)
https://arxiv.org/abs/2509.04331
Astrology, astronomy,
