pith. sign in

arxiv: 2509.04331 · v3 · submitted 2025-09-04 · 🌌 astro-ph.GA · astro-ph.EP· astro-ph.IM· astro-ph.SR

A fast machine learning tool to predict the composition of astronomical ices from infrared absorption spectra

Andr\'es Meg\'ias , Izaskun Jim\'enez-Serra , Fran\c{c}ois Dulieu , Julie Vitorino , Bel\'en Mat\'e , David Ciudad , Will R. M. Rocha , Marcos Mart\'inez Jim\'enez
show 1 more author
Jacobo Aguirre
This is my paper

Pith reviewed 2026-05-18 19:03 UTC · model grok-4.3

classification 🌌 astro-ph.GA astro-ph.EPastro-ph.IMastro-ph.SR
keywords astronomical icesinfrared spectroscopymachine learningneural networksJWSTice compositionastrochemistryspectral analysis
0
0 comments X

The pith

A neural network predicts the fractional composition of astronomical ices from infrared spectra in under a second with typical errors of 3%.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces AICE, an artificial neural network that estimates the molecular makeup of icy mantles on interstellar dust grains. It takes the infrared absorption spectrum from 2.5 to 10 microns as input and returns the relative amounts of H2O, CO, CO2, CH3OH, NH3, and CH4. Training draws on hundreds of laboratory ice mixture experiments that were first baseline-subtracted and normalized. Once trained, the network delivers predictions in less than one second on ordinary computers while keeping errors around 3 percent. Tests on two JWST spectra from the Ice Age program match earlier manual estimates, showing the approach works on real astronomical data and can handle large numbers of sources efficiently.

Core claim

AICE is an artificial neural network that, after training on reprocessed laboratory infrared spectra of ice mixtures, predicts the fractional abundances of H2O, CO, CO2, CH3OH, NH3, and CH4 in astronomical ices from observed spectra between 2.5 and 10 microns, achieving typical errors of around 3 percent in under one second.

What carries the argument

The artificial neural network trained on baseline-subtracted and normalized laboratory ice absorption spectra to map spectral input directly to species fractional outputs.

Load-bearing premise

Laboratory ice mixtures used for training adequately sample the range of compositions, temperatures, and processing histories present in real astronomical ices so the network generalizes without large systematic bias.

What would settle it

Direct comparison of AICE predictions to independently derived compositions from traditional spectroscopic fitting on a new set of JWST spectra where the true values are known but were withheld from training.

Figures

Figures reproduced from arXiv: 2509.04331 by Andr\'es Meg\'ias, Bel\'en Mat\'e, David Ciudad, Fran\c{c}ois Dulieu, Izaskun Jim\'enez-Serra, Jacobo Aguirre, Julie Vitorino, Marcos Mart\'inez Jim\'enez, Will R. M. Rocha.

Figure 1
Figure 1. Figure 1: Diagram of the model behind AICE: artificial neural networks. For each target molecule (H2O, CO, CO2, CH3OH, NH3, and CH4), a multilayer perceptron is used to predict the corresponding molecular fraction. The IR ice spectrum in absorbance is fed to the model, which transforms the input through a series of steps or layers. In each one, a linear combination of the previous values (xi) is followed by the appl… view at source ↗
Figure 2
Figure 2. Figure 2: Diagram of the architecture of the neural networks of AICE: Multi-layer perceptrons whose layer size (in parenthesis) decreases from the input spectrum to the output prediction. For the dropout step on the first dense layer, the probability of dropout is stated in parenthe￾ses. for each neural network (i.e. the weights and biases). Dropout was applied only after the first hidden layer with a probability of… view at source ↗
Figure 3
Figure 3. Figure 3: Example of the baseline reduction for experiment #35 from LIDA, performed with the interactive Python interface developed by us and included within AICE. The plot shows the spectra corresponding to this experiment (H2O : CO2 : CH3OH, 9:2:1), for several temperatures. In black and blue, we display the original and reduced spectrum from the experiment at 10 K. In light grey and light green+yellow, we indicat… view at source ↗
Figure 4
Figure 4. Figure 4: Example of synthetic generated spectrum (#21), normalised by its mean absorbance value. The colours represent the different molecular contributions. This spectrum represents a layered mixture of H2O, CO, CH3OH, NH3, CO2, and CH4, in decreasing order molecular fraction, all of them at a temperature of 11.7 K. The pure ice spectra used to construct this curve come from different experiments from LIDA, OCdb a… view at source ↗
Figure 5
Figure 5. Figure 5: Performance of our neural networks after the training with the first training+validation split, with respect to the spectra included in the validation subset. The plots show the predicted versus true labeled values for the molecular fractions and temperature. Left panels: Molecular fractions of ice spectra. Each ice spectrum in the validation subset (with a certain composition and temperature) corresponds … view at source ↗
Figure 6
Figure 6. Figure 6: Absorbance spectra obtained by JWST towards the background stars NIR38 (top) and J110621 (bottom) in the Chamaeleon I molecular cloud, and predictions obtained by AICE for the corresponding ice composition and temperature. Left panels: Absorbance spectra (in black) obtained after the reduction of the original JWST spectra in flux (McClure et al., 2023). Right panels: Direct numerical predictions given by A… view at source ↗
read the original abstract

Current observations taken by James Webb Space Telescope (JWST) allow us to observe the absorption features of icy mantles that cover interstellar dust grains, which are mainly composed of $\mathrm{H_2O}$, $\mathrm{CO}$, and $\mathrm{CO_2}$, along with other minor species. Thanks to its sensitivity and spectral resolution, JWST has the potential to observe ice features towards hundreds of sources at different stages along the process of star formation. However, identifying the spectral features of the different species and quantifying the ice composition is not trivial and requires complex spectroscopic analysis. We present Automatic Ice Composition Estimator (AICE), a new tool based on artificial neural networks. Based on the infrared (IR) ice absorption spectrum between 2.5 and 10 microns, AICE predicts the ice fractional composition in terms of $\mathrm{H_2O}$, $\mathrm{CO}$, $\mathrm{CO_2}$, $\mathrm{CH_3OH}$, $\mathrm{NH_3}$, and $\mathrm{CH_4}$. To train the model, we used hundreds of laboratory experiments of ice mixtures from different databases, which were reprocessed with baseline subtraction and normalisation. Once trained, AICE takes less than one second on a conventional computer to predict the ice composition associated with the observed IR absorption spectrum, with typical errors of $\sim$3 $\%$ in the species fraction. We tested its performance on two spectra reported towards the NIR38 and J110621 background stars observed within the JWST Ice Age program, demonstrating a good agreement with previous estimations of the ice composition. The fast and accurate performance of AICE enables the systematic analysis of hundreds of different ice spectra with a modest time investment. In addition, this model can be enhanced and re-trained with more laboratory data, improving the precision of the predictions and expanding the list of predicted species.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript presents Automatic Ice Composition Estimator (AICE), an artificial neural network tool that predicts the fractional composition of astronomical ices (primarily H₂O, CO, CO₂, CH₃OH, NH₃, and CH₄) from 2.5–10 μm infrared absorption spectra. Trained on hundreds of reprocessed laboratory ice mixture spectra with baseline subtraction and normalization, the model is claimed to deliver predictions in under one second on standard hardware with typical errors of ~3% in species fractions. It is tested on two JWST spectra (NIR38 and J110621 from the Ice Age program), reporting good agreement with prior manual estimates, and is intended to enable rapid analysis of large numbers of sources.

Significance. If the generalization performance holds, AICE would provide a practical, high-throughput method for extracting ice compositions from the growing archive of JWST spectra, addressing a clear bottleneck in astrochemistry. The reliance on existing laboratory databases and the explicit provision for retraining with new data are constructive features that could support iterative improvement.

major comments (3)
  1. [Abstract] Abstract: the central performance claim of '~3% in the species fraction' and sub-second inference is stated without any description of network architecture, training/validation protocol, cross-validation method, baseline handling details, or error estimation procedure. These omissions prevent independent assessment of whether the quoted accuracy applies to the reported laboratory test cases, let alone to astronomical spectra.
  2. [Results / JWST testing] Testing on JWST spectra: only two background-star spectra (NIR38 and J110621) are examined, with agreement described qualitatively. No quantitative error distribution, mean absolute deviation, or comparison against a larger independent JWST sample is provided. Because the headline applicability is to 'hundreds of different ice spectra,' this limited validation is load-bearing for the generalization claim.
  3. [Methods / Training dataset] Training data coverage: the manuscript does not quantify how the laboratory mixtures span the temperature, irradiation, and minor-species ratios expected in interstellar ices. Without such coverage metrics or domain-shift experiments, the risk that spectral features outside the training distribution produce systematic biases remains unaddressed.
minor comments (2)
  1. [Methods] Add a table or supplementary section listing the exact number of laboratory spectra per mixture type, temperature range, and processing history to allow readers to judge representativeness.
  2. [Model description] Clarify whether the network outputs are post-processed to enforce summation to unity or whether the raw predictions can sum to values other than 100%.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive and detailed review. We address each major comment below and indicate where revisions will be incorporated in the next version of the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central performance claim of '~3% in the species fraction' and sub-second inference is stated without any description of network architecture, training/validation protocol, cross-validation method, baseline handling details, or error estimation procedure. These omissions prevent independent assessment of whether the quoted accuracy applies to the reported laboratory test cases, let alone to astronomical spectra.

    Authors: We agree that the abstract, owing to length constraints, does not detail these elements. The full manuscript describes the neural-network architecture, training and cross-validation protocol, baseline subtraction and normalization steps, and error estimation in the Methods and Results sections. We will revise the abstract to include a concise statement of the training approach and validation method while retaining its brevity and directing readers to the relevant sections. revision: yes

  2. Referee: [Results / JWST testing] Testing on JWST spectra: only two background-star spectra (NIR38 and J110621) are examined, with agreement described qualitatively. No quantitative error distribution, mean absolute deviation, or comparison against a larger independent JWST sample is provided. Because the headline applicability is to 'hundreds of different ice spectra,' this limited validation is load-bearing for the generalization claim.

    Authors: We acknowledge that astronomical validation is restricted to the two sources for which independent manual composition estimates were available. We will add quantitative metrics, including per-species absolute deviations and mean errors for these two cases, to the revised Results section. We will also note the limited size of the current JWST test set as a limitation and the reliance on laboratory cross-validation for the primary generalization assessment. revision: partial

  3. Referee: [Methods / Training dataset] Training data coverage: the manuscript does not quantify how the laboratory mixtures span the temperature, irradiation, and minor-species ratios expected in interstellar ices. Without such coverage metrics or domain-shift experiments, the risk that spectral features outside the training distribution produce systematic biases remains unaddressed.

    Authors: The training set comprises hundreds of reprocessed laboratory spectra drawn from multiple databases and includes a range of temperatures, irradiation conditions, and mixture ratios representative of interstellar ices. We did not provide explicit coverage statistics or domain-shift tests in the submitted version. We will add a summary table or figure in the Methods section quantifying the spanned parameter space and include a short discussion of how cross-validation performance and the diversity of the training data address potential extrapolation risks. revision: yes

Circularity Check

0 steps flagged

No significant circularity in the AICE model's derivation chain

full rationale

The paper trains an artificial neural network on reprocessed laboratory ice mixture spectra drawn from external databases to predict fractional compositions of H2O, CO, CO2, CH3OH, NH3, and CH4. Predictions are generated via standard forward inference on the trained model, with performance evaluated on two independent JWST astronomical spectra (NIR38 and J110621) that were not part of training. No equations, derivations, or self-citations reduce the output predictions to fitted parameters defined from the target data itself, nor is any uniqueness theorem or ansatz smuggled in via prior author work. The workflow is a conventional supervised ML pipeline relying on external training data and external validation, rendering the claimed performance self-contained against benchmarks outside the present manuscript.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that laboratory ice mixtures are representative of astronomical ices and that the neural network generalizes without large domain shift. No new physical entities or ad-hoc constants are introduced beyond standard neural-network training.

axioms (1)
  • domain assumption Laboratory ice spectra adequately represent the range of compositions and conditions in interstellar ices
    Invoked when claiming that training on lab data enables accurate predictions for JWST observations

pith-pipeline@v0.9.0 · 5928 in / 1280 out tokens · 28637 ms · 2026-05-18T19:03:05.456353+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

12 extracted references · 12 canonical work pages · 1 internal anchor

  1. [1]

    Very Deep Convolutional Networks for Large-Scale Image Recognition

    Accolla, M., Congiu, E., Dulieu, F., et al. 2011, Physical Chemistry Chemical Physics (Incorporating Faraday Transactions), 13, 8037 Altwegg, K., Balsiger, H., Bar-Nun, A., et al. 2016, Science Advances, 2, e1600285 Bisschop, S. E., Fuchs, G. W., Boogert, A. C. A., van Dishoeck, E. F., & Linnartz, H. 2007, A&A, 470, 749 Boogert, A. A., Gerakines, P. A., &...

    work page internal anchor Pith review Pith/arXiv arXiv 2011

  2. [2]

    A.2.Spectrum taken by JWST towards the background star NIR38 in the molecular cloud Cha I, and fitted continuum, used to compute the absorbance spectrum

    0 1 2 3 absorbance NIR38 Fig. A.2.Spectrum taken by JWST towards the background star NIR38 in the molecular cloud Cha I, and fitted continuum, used to compute the absorbance spectrum. The fit was performed by McClure et al. (2023) joining several polynomials fitted within the dashed regions and extrap- olating to the rest. when applied to the correspondin...

    work page 2023

  3. [3]

    0.00 0.25 0.50 0.75 1.00 1.25 1.50 corrected absorbance NIR38 Fig. A.3.Spectrum in absorbance derived from data by JWST towards the background star NIR38 in the molecular cloud Cha I, and fitted sili- cate contribution, which has to be subtracted to obtain the ice spectrum. The fit was performed by McClure et al. (2023) using the software Op- Tool (Domini...

    work page 2023

  4. [4]

    In case that both the input and the output are vectors, (x,y), we would use equation B.3 in a vectorial form, calculating the squared norm of the vectorial dif- ference betweenˆyk andy k, that is,||ˆyk −y k||2. In the training of a neural network, its parameters are initialised randomly and then modified in consecutive steps orepochs, to explore the param...

    work page 1986

  5. [5]

    15 K original spectrum windows Fig

    0.0 0.1 0.2 0.3 absorbance AICE Interactive Toolkit check terminal for instructions 12-HCOOH+H2O-9+91.csv - abs. 15 K original spectrum windows Fig. C.1.Spectra obtained in the experiment #12 of LIDA, from a mix- ture of H2O : HCOOH (91:9), as seen with our AICE Interactive Toolkit. The curve in black is the spectrum at 15 K, and the rest of coloured curv...

    work page 2007

  6. [6]

    15 K original spectrum baseline fit edited spectrum windows Fig

    0.02 0.00 0.02 0.04 0.06 0.08 0.10 absorbance AICE Interactive Toolkit check terminal for instructions 12-HCOOH+H2O-9+91.csv - abs. 15 K original spectrum baseline fit edited spectrum windows Fig. C.3.Same as Fig. C.2 in a zoomed region, after a smoothing the spectra with a size of 7 and interpolating two contaminated regions. The curve in black shows the...

    work page 1993

  7. [7]

    We can see how the losses decrease rapidly at the beginning of the training and slowly sta- bilise after a maximum amount of 160 epochs. In all cases, the loss in the validation subset is larger than the loss in the train- ing subset, since the optimisation of the model’s parameters is done with respect to the training subset. The noisy patterns of the cu...

    work page 2023

  8. [8]

    We used these three PDFs to represent values with asymmetric uncertainties using the same reasoning as in the work by Possolo et al

    are large enough and there is a low or medium asymmetry in the uncertainties (s1, s2), we can use alternative functions to model the PDF of the rich value: the PDFs of a split normal distribution, a log-normal distribution and a generative extreme value (GEV) distribution. We used these three PDFs to represent values with asymmetric uncertainties using th...

    work page 2019

  9. [9]

    In any case, it does not seem that the standard AICE models are underpredicting any molecular fraction due to band saturation

    In short, the results by both models are compatible within the uncertainties, although we see some systematic differences. In any case, it does not seem that the standard AICE models are underpredicting any molecular fraction due to band saturation. Appendix H: Retraining with a smaller spectral range As we commented in Section 3.5, we retrained our neura...

    work page 2000

  10. [10]

    Despite the reduction in the number of spectral points, we obtained a good correlation between the predicted and actual values in the validation, as shown in Fig. H.1, although 0.0 0.2 0.4 0.6 0.8 1.0predicted molecular fraction H2O CO CO2 CH3OH NH3 CH4 10 K 90 K 10 K 90 K 10 K 90 K 10 K 90 K 10 K 90 K 10 K 90 K 0.0 0.2 0.4 0.6 0.8 1.0 labeled molecular f...

    work page 2000

  11. [11]

    Table I.1.Column densities obtained by the automatic band integrator included within AICE. molecule column density(10 18 cm−2) NIR38 J110621 H2O(%) 7.0878±0.0011 9.0±0.3 CO(%) 2.007±0.008 2.86±0.10 CO2 (%) 1.238±0.020 1.6 +0.5 −0.2 CH3OH(%) 0.81 +0.07 −0.05 1.2+0.8 −0.3 NH3 (%) 0.34 +0.04 −0.03 0.31+0.08 −0.05 CH4 (%) 0.171 +0.008 −0.007 0.211+0.011 −0.01...

    work page 2015

  12. [12]

    I.1.Spectra taken by JWST towards NIR38 in the molecular cloud Cha I, converted to absorbance scale after fitting the stellar continuum and the silicate contribution

    0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5absorbance H2O CO CO2 CH3OH NH3CH4 NIR38-c-s.txt AICE fit integrated area observations Fig. I.1.Spectra taken by JWST towards NIR38 in the molecular cloud Cha I, converted to absorbance scale after fitting the stellar continuum and the silicate contribution. In pale red are the fitted Gaussians by the band integrator include...

    work page 2023