acceptodds
Withdrawn from ICLR 2027

Seeing Through Spectral Overlap: Recovering Mixture Composition as a Measure on a Learned Infrared Spectral Manifold

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Infrared mixture analysis seeks to recover component spectra and their abundances from a single observation, but spectral overlap allows different compositions to produce similar signals, making reconstruction alone insufficient for identification. We introduce SpecMUSE, which represents mixture composition as a finitely supported probability measure on a learned spectral manifold. Its support locations specify component spectra, while its masses encode their abundances. Masked reconstruction and coordinate alignment establish the manifold from pure spectra. A set estimator jointly predicts component coordinates, presence probabilities, and abundances from the observed mixture. A differentiable realization field and discrete projection connect coordinates to individual component spectra, while a straight-through surrogate carries spectral-loss gradients to the continuous selection queries. This formulation couples component recovery with spectral reconstruction while accommodating unknown component counts at inference. Experiments on mixtures constructed from simulated QM9S spectra and measured NIST spectra demonstrate improved recovery across mixture complexities and generalization to unseen mixtures. Anonymous code is available at https://anonymous.4open.science/r/SpecMUSE-9F68.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
molecular spectroscopy, infrared spectral decomposition, inverse problems

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_HB65OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a novel method to decompose a mixture of infrared spectra.

Critical strengths

The proposed method is interesting, addressing a difficult problem at the intersection of machine learning and chemistry. The paper is relatively well written, and the appendix provides several significant materials, including background material on infrared spectroscopy, allowing ML researchers to better understand the underlying problem.

Critical weaknesses

The proposed method is compared to 4 other methods. The challenging ones are the DeepCID for Raman spectra mixtures proposed by Fan et al. (2019) and the NNLS, which is the classical non-negative least squares (NNLS) recently explored on liquid-phase mixtures by Melle et al. (2026). It would have been relevant to explore related methods that are closely connected to the problem at hand. Comparing to methods that are designed to address Raman spectra mixtures or liquid-phase mixtures is not fair and misleading.

It is not clear which variant of the NNLS is used in the submitted paper, considering that several variants were studied by Melle et al. (2026) for liquid-phase mixtures. As given in (39) in Appendix B, it seems that the unregularized variant is used. However, Melle et al. (2026) have shown that the regularized variant outperforms other variants.

The proposed method relies on too many hyperparameters to be tuned by the user, including the many tradeoff parameters called λ\lambda such as comp, slot, mix, z, p, … The ablation studies nullify some of them, but it is not clear how all of them are fine-tuned.

At the end, it seems that the proposed method is fine-tuned to provide high performance, while compared methods are off-the-shelf. Moreover, compared methods often have fewer hyperparameters to tune.

Questions for authors

none

Reviewer_XGiUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a two-stage unmixing method: the first stage constructs a spectral manifold using a pure-component dataset, and the second performs manifold-constrained measure estimation from the observed signal. While this general idea is widely adopted in supervised signal separation, the implementation introduces certain novelties. The reported results look promising; however, the comparison scope is quite limited.

Critical strengths

The reported results are promising, and the method is applicable to certain real-world applications. The proposed method is beneficial to the study of supervised-learning-based spectral unmixing.

Critical weaknesses
  1. From the perspective of signal separation, the theoretical justification lacks sufficient rigor.
  2. Compared with existing spectral unmixing methods, the innovation of this work is limited, and the experimental comparisons are inadequate.
Questions for authors

How do you ensure that the projection from an observed signal to the abundances of pure spectral signatures within the learned manifold is orthogonal instead of overlapping? Likewise, how do you ensure that the pure spectral signatures can be distributed on the manifold with sufficient pairwise separation? n Table 1, the proposed method achieves considerably better performance than the other compared methods. Please provide a rigorous explanation for this observation.

Reviewer_qBVTOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper addresses the recovery of component spectra and their abundances from a single infrared mixture spectrum under an additive (Beer–Lambert) mixing model. The composition is represented as a finitely supported probability measure on a learned spectral manifold: support locations correspond to component spectra and masses to abundances.

The proposed method, SpecMUSE, has two stages:

Stage 1 pretrains on pure spectra. A masked-reconstruction encoder maps spectra to coordinates on a unit sphere, and a decoder (the “realization field”) maps coordinates back to spectra.

Stage 2 trains a slot-based set estimator that predicts, for each slot, a coordinate query, a presence probability and an abundance score. Queries are assigned to reference anchors through hard selection with a straight-through gradient. Training combines permutation-invariant matching, slot supervision and an additive reconstruction loss.

At inference, a refinement step re-selects each component among nearby anchors. Experiments use synthetic mixtures of two to four components, built from simulated QM9S spectra and from measured NIST gas-phase spectra. Baselines are NNLS, a fixed-library adaptation of EB-gMCR, a Transformer set predictor and a CNN retrieval model.

Critical strengths
  • Identifying and quantifying components from overlapping IR bands is a practically relevant problem for automated chemical analysis.
  • The method improves over the included baselines across mixture sizes.
  • The paper reports ablations, implementation details, and code.
Critical weaknesses
  • Motivation. The introduction relies heavily on specialized terminology, which makes the contribution hard to follow.

    • The repeated emphasis on the composition being “unordered” is not motivated. Library-based approaches such as NNLS index components by their reference spectra and do not face an ordering problem; nor do multivariate curve resolution methods appear to.
    • The introduction lists difficulties (compensation between similar spectra, weak components, unknown component count). These are generic. It does not explain why existing solvers of the linear model in Eq. (1) fail on them, i.e., what specific gap the method fills.
  • Positioning and novelty. There is a large body of work on spectral mixture analysis with deep learning, but the introduction does not state what is new relative to it.

    • The related-work section consists mostly of single-sentence descriptions of individual papers, without discussing how they relate to the proposed approach.
    • Closely related work is missing. In spectral unmixing, several methods already represent component (endmember) spectra through learned generative models and estimate them jointly with abundances under the same linear model. Examples:
      • Shi, Zhao, Zhang, et al., “Probabilistic generative model for hyperspectral unmixing accounting for endmember variability,” IEEE TGRS, vol. 60, 2022.
      • Deng, Qian, Nie, et al., “Diffusion-model-based hyperspectral unmixing using spectral prior distribution,” IEEE TGRS, vol. 62, 2024.
    • Recent work on infrared spectroscopy also pretrains on pure spectra and transfers to mixture component identification and quantification: Tan, Chen, Fang, et al., “Simulation-to-real transfer learning for infrared spectroscopic chemical sensing and analysis from molecules to complex samples,” arXiv:2608.13341, 2026.
    • The relationship of SpecMUSE to these approaches should be discussed.
  • Model formulation. Eq. (2) introduces a coordinate space on the unit sphere and a spherical distance without motivating why components should lie on a sphere or how this space relates to the mixture model of Eq. (1). Several symbols (e.g., ψψ, dzd_z, the Dirac δδ) are not defined where they first appear.

    • More importantly, the “probability measure” is not properly introduced. In Eqs. (2), (10) and (37) there is only a weighted sum of Dirac masses with nonnegative weights summing to one. This is the standard abundance vector.
    • The method seems deterministic; there is no probabilistic model, uncertainty or measure-theoretic tool.
    • At inference every support is snapped to a reference anchor, so the recovered measure is a sparse abundance vector over a fixed library.
    • The measure-on-a-manifold framing therefore overstates the conceptual method.
  • Section 4.1 (spectral manifold construction). The section introduces a new data representation (patch embeddings X_s, visible tokens, mask tokens, a restoration module) and an architecture without connecting them to the problem formulation.

    • Patches and the masking procedure are not defined in the main text.
    • The Stage 1 training data are specified only in Section 5.1.
    • As written, the section is difficult to follow.
  • Section 4.2 (measure estimation). The architecture (mixture encoder, slot decoder with cross- and self-attention, three prediction heads, soft kernel and hard selector) is introduced block by block, without an overall motivation for each design choice.

    • The two encoders are easy to confuse. The mixture encoder EθE_θ produces tokens CyC_y from patches of the mixture spectrum, while the pretrained encoder EψE_ψ maps the reference pool to anchors Z.
    • The text should state explicitly which inputs each module receives and which parameters are frozen during Stage 2.
  • Permutation-invariant matching. The matching in Eq. (11) is a consequence of the slot-based design. Since the final supports are reference anchors, a library-indexed output would avoid the ordering issue altogether, for example a presence probability and abundance per reference spectrum, as in sparse unmixing or multi-label classification. The choice of a slot-based set predictor over such a formulation should be motivated.

  • Baselines. The choice of baselines is not justified. NNLS, a fixed-library EB-gMCR adaptation, a Transformer and a CNN retrieval model do not cover recent multivariate curve resolution and spectral unmixing methods, including the generative unmixing approaches cited above and sparse unmixing.

  • Hyperparameters. The method has a large number of hyperparameters: loss weights λalignλ_{align}, λrawλ_{raw}, λcompλ_{comp}, λslotλ_{slot}, λmixλ_{mix}, λzλ_z, λpλ_p, λαλ_α; discrepancy weights ηaη_a, ηsη_s, ηdη_d; temperatures τcτ_c, τsτ_s, τzτ_z; the Huber threshold δHδ_H. Their values are not reported in Table 6, and a selection procedure is stated only for τzτ_z.

Questions for authors
  • Why is component ordering a problem that SpecMUSE needs to solve, compared with library-based NNLS or multivariate curve resolution, where components are indexed by their reference spectra?
  • What specific limitation of existing solvers of Eq. (1) does the method address?
  • How does SpecMUSE differ, conceptually and empirically, from generative-model-based unmixing methods such as Shi et al. (2022) and Deng et al. (2024)?
  • Does any part of the method use its probabilistic nature, for example for uncertainty estimation? In what sense is there a probability measure?
  • Given that supports are restricted to reference anchors at inference, why use a slot-based set predictor with permutation-invariant matching rather than per-reference presence and abundance outputs?
  • What motivated the choice of baselines, and how does the method compare with recent multivariate curve resolution and unmixing methods?
  • How were the hyperparameters selected?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.