acceptodds
Under review as a conference paper at ICLR 2027

Representation Stability in Domain Space: Cross-Domain Validation for Unsupervised Scientific Discovery

Abstract

Unsupervised learning is increasingly used for scientific discovery in settings where labeled data are unavailable or ill-defined. Validation is critical but challenging, and internal cluster-quality metrics encode geometric or statistical assumptions that scientific data routinely violate, and external ground truth is typically absent. This paper takes the ensemble-based NEMI workflow (Sonnewald, 2026) and develops it into a formal validation principle for unsupervised scientific discovery, evaluated across four Earth science domains. The principle is representation stability in domain space. If a discovered structure reflects a real property of the underlying system, the partition it implies in geographic, sequence, or property space should persist under perturbations of the learned latent representation. The same ensemble that NEMI uses for clustering yields uncertainty estimates in two coordinate systems: 1) the embedding space where clustering operates, and 2) the domain space in which scientific meaning is found. We argue that the domain-space entropy field is the scientifically meaningful diagnostic, since domain-space output from any single fixed embedding is a deterministic projection of within-embedding uncertainty alone. Across ocean ecology, ocean biogeochemistry, ocean dynamics, sea-ice, and atmospheric dynamics, low-entropy regions identify regimes that support stable scientific interpretation, while the geometry of where uncertainty concentrates generates hypotheses that can be tested. We find a recurring pattern, where low-entropy structure within regions that domain intuition identifies as transitional, suggesting candidate refinements to established regime boundaries. We also analyze sensitivity to cluster alignment strategy across domains and show that the spatial structure of the entropy field is preserved under both Greedy and Hungarian alignment, with quantifiable differences tied to split/merge events. NEMI underlies domain-science publications and manuscripts under peer review across these applications, and is contributing to model development toolkits aimed at improving sea-level predictability.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
unsupervised learning, clustering, ensemble methods, uncertainty quantification, representation stability, validation, scientific machine learning, AI for Earth science

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_RwjNOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper developes a novel workflow for validating unsupervised scientific discovery, centered around the principle of representational stability. Primarily the paper addresses the notion of clustering similar physical phenomena based on stochastic embeddings and it asks whether perturbations in latent space affect the type of clusters found. The central premise is to build on an earlier framework, NEMI, which proposes to overcome any such instabilities by an ensemble-based approach where the same features are embedded multiple times, then clustered and returned to physical coordinates. These clusters are then aligned using one of two matching techniques and a consensus cluster is established using geographic entropy. The paper evaluates this approach on four datasets ranging from Ocean ecology over Atmospheric dynamics to Ocean biogeochemistry.

Critical strengths

Overall, I think the approach in the paper is sensible and contains an interesting approach to stability, that is very much used in other areas of scientific computing as well: ensemble-based approaches are widely used to enhance stability and robustness of forecasting and classification algorithms, so using them in the present context is a worthwhile endeavour. I think the authors provide sufficient mathematical details to describe their methodology (including clustering, alignment and consensus extraction) and the datasets used in the numerical experiments appear thorough. Presentation-wise I would like to highlight that the paper appears human written.

Critical weaknesses
  1. I found the paper to be quite difficult to read. I actually want to emphasise that this is not because it is overloaded with information (I would definitely not consider this as unedited AI output) but rather because the paper is formulated really abstractly and misses some details in crucial places. For example:

(a) In the abstract the authors refer to "the same ensemble that NEMI uses for clustering". This assumes some familiarity with the NEMI technique which until this point is only introduced as an "ensemble-based workflow".

(b) On p. 2 the authors mention "In general terms, the space where the interpretation should take place is the coordinate system in which scientific entities have natural relationships, constraints, or causal structure.". It would be great if they could be more explicit here as well, this is a very broad statement that does not really convey a practical application in my opinion.

(c) At the beginning of section 2 the authors mention that "The NEMI workflow comprises four stages: 1) stochastic embedding, 2) deterministic clustering in latent space, 3) ensemble construction, and 4) projection back to domain space" there is no mention what type of input data or indeed computation the principle is supposed to make. Are the authors intending to use this to make predictions of physical fields (e.g. weather forecasting), or is this intended to classify certain events? Or to characterise similar phenomena. In my opinion this is not clear at least by this point in the paper.

  1. While the datasets are overall a good choice for validation, the paper does not place this novel approach sufficiently well in the broader literature. In particular, the introduction cites several alternative clustering stability and consensus clusterin methods (Ben-Hur; Lange; von Luxburg; Minti et al.) but does not benchmark against any of them in the numerical experiments section. It is thus very difficult to ascertain the practical benefit of the novel approach.

  2. One might argue that the novelty of the contribution is somewhat limited with the main differences to prior work highlighted to be "1) we perturb the learned representation before applying deterministic clustering, and 2) we evaluate agreement for each location in domain space, rather than as a global score." This suggests the current approach is incremental and together with the lack in baselines, the algorithmic contribution might not be enough to merit publication in ICLR.

  3. The labels in Figures 2-6 are unusually small making it impossible to read and understand any of the numerical results in a printed version of the manuscript, and even on a screen the results are not easy to read.

Questions for authors
  1. I wonder if the authors have explored the effect of and possibility for adverserial perturbations rather than just random perturbations on their models?

  2. I wonder if the authors can explain more clearly what the role of random perturbation in latent space might correspond to in real scenarios? It seems to me more likely that observational data might be perturbed through noise in the state space and that the effects of such perturbations on downstream clustering should be examined. Is the NEMI approach able to cope with this?

Reviewer_C1n9OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper extends NEMI and presents representation stability in domain space as a validation principle for unsupervised scientific discovery. It repeatedly runs UMAP, clusters each embedding, aligns cluster labels, and computes per-location entropy after projecting the assignments back to physical space. Experiments cover ocean ecology, ocean dynamics, ocean biogeochemistry, and atmospheric dynamics.

Critical strengths

I. Mapping representation instability back to domain coordinates produces entropy maps that domain experts can inspect.

II. The comparison between Greedy and Hungarian alignment identifies failures caused by split/merge events and small clusters.

Critical weaknesses

I. The technical novelty is limited. Sonnewald (2026) already uses stochastic UMAP ensembles, geographic-overlap alignment, majority voting, and per-location geographic entropy. The main additions are a new interpretation, more case studies, and a Greedy/Hungarian comparison, but these additions do not clearly establish a new validation method.

II. The main validation claim is not supported. Low entropy only shows stability across UMAP random seeds under the chosen pipeline and does not rule out stable artifacts. The paper provides no known-ground-truth experiment, downstream evaluation, independent physical validation, or matched stability baseline showing that low entropy corresponds to correct or useful scientific structure. The same entropy is also used to select UMAP hyperparameters and then interpret robustness.

Questions for authors

None

Reviewer_HEskOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This work studies on the unsupervised learning for scientific discovery. To address the lack of ground-truth labels and the uncertainty introduced by representation learning, it introduces a validation principle based on representation stability in domain space. Building on the NEMI framework, this work introduces a domain-space stability diagnostic, which constructs an ensemble of stochastic UMAP embeddings, clusters each embedding deterministically, and maps the aligned labels back to physical coordinates. With the introduction of consensus labels and per-location normalized entropy, this work summarizes the resulting partitions and corresponding sensitives to represent variability. Experiments across four Earth science applications provide several domain-specific case studies, illustrate how this diagnostic can assess partition robustness and generate hypotheses about regime boundaries and observational gaps.

Critical strengths
  1. Well-motivated problem. This paper focuses on scientific discovery, an important and frontier research question, addresses an issue under the unsupervised setting.
  2. An informative perspective. This paper reports ensemble disagreement at individual physical locations, which maybe helpful in identifying localized instability and highlight regions that warrant further investigation.
  3. Clear case studies and sensitivity analysis. This work gives clear case studies on the relationships between the resulting partitions and physical and ecological variables, and provide context for interpreting the clustering outputs and discussing localized uncertainty.
Critical weaknesses
  1. Overclaimed paper contributions. This work studies scientific discovery limited in earth science scenarios, which is actually a limited domain in scientific discovery research. All the cases introduced in this work concern spatial regime partitioning with an explicit mapping to physical coordinates, which are not sufficiently tested in other domains with different data structures or discovery objectives.

  2. Insufficient evidence for the central validation claim. The evaluation of this work is lack of a controlled benchmark with ground truth. Also, there is no evidence that proposed stability diagnostic can distinguish scientifically meaningful partitions from reproducible but misleading ones, all makes the validation of proposed method unstable and unreliable

  3. Insufficient clarification. This work emphases the distinction between embedding-space and domain-space uncertainty as a core contribution. However, the detailed additional inferential information is obtained in domain space is still lack of clarification, which makes this contribution unstable.

Questions for authors
  1. For the proposed refinements to ecological boundaries, what evidence beyond the clustering results supports these interpretations?
  2. For the diagnostic provided in this work, can authors identify the additional information beyond simpler alternatives?
  3. For the relationships between lower ensemble disagreements and more scientifically partitions, can authors further provide analyses on it?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.