acceptodds
Withdrawn from ICLR 2027

EvolveODE: Emulating Underlying Dynamics through Sorting and Coupling

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Observational-level data arises from processes beyond the scope of direct observation. These processes naturally couple across the scales they natively occupy. In real-world settings, they may include global weather subsystems and time-delayed traffic flows, among other such systems. Through self-attention, Graph Neural Networks (GNNs) offer a natural way to untangle these confounding factors, namely by leveraging self-organization to sort variables that strongly interact with one another, forming a sub-process. We propose EvolveODE, a multiscale GNN that solves attraction-repulsion equations to identify groupings, thereby learning the coupling between the physical processes underlying observation-level phenomena. EvolveODE then emulates these relationships with a dynamical system. We show that EvolveODE outperforms the current state-of-the-art baselines across various tasks and settings. Code is available at https://anonymous.4open.science/r/EvolveODE/.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
NeuralODE, Graph, Discovery, Coupling, Dynamics, Forecasting, Classification

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

The text proposes EvolveODE, a higher order Graph Neural Network that aims to resolve underlying dynamics directly via attractive and repulsive node interaction terms. Significant subprocesses of the latent dynamics are linked to a complete subgraph topology. The method augments the energy surface implied by the interaction terms by corresponding subprocess energy wells. Supplied theory is empricially evaluated on short- and longterm forecasting tasks with varying degree of data sparsity and node homophily.

Critical strengths

1. Interdependence of coarser and more local structure. Conceptually, I find the paper's recognition of the interdependence between scales important. Moreover, the idea to use attraction-repulsion terms in combination with Graph Neural Networks is sound and reasonable as, e.g., already verified by the cited ACMP.

Critical weaknesses

1. Lack of Clarity and Consistency. Overall, the text does not sufficiently describe how the proposed method is actually implemented. For instance, Definition 1 introduces sub-processes as the interaction between nodes that form a complete subgraph but it never defines what is meant with "an interaction between nodes". The same definition further introduces a "noisy" graph, without alluding to a "noisy" graph later on in the text. Furthermore, in what sense, is time relevant for evaluated datasets like Cora and Squirrel that have no inherent time axis? In general, the text does not clarify over what dimensionality the coarsening takes place. That is, whether links are established across spatial, temporal or other scales.

2. Lack of Reproducibility. The text describes their experiment setup only partially. While dropout and learning rate are given, optimizer the number of epochs etc., but most importantly a comprehensive description of the EvolveODEs GNN architecture is missing. Furthermore, the supplied code does not provide a requirements, environment or similar file that would facilitate reproducibility and portability. More importantly, the code does not run anyhow as it is incomplete. Experiment scripts import modules (e.g., baselines or profiling_utils) that are not part of the supplied code. A smaller faux pas is the README describing the code as part of a NeurIPS submission.

Minor:

3. Citation style. The authors use a number citation style despite the author guidelines prescribing an author-year format. The used number style leads to significant more available space for this submission compared to other ones that adhered to the guidelines.

4. Writing.

  • l. 071-072: "... in EvolveODE breaks monotonically reduces error." is probably missing a word or two
  • l. 172-173: "challange" should be "challenge"
  • l. 216-217: "...with the form from Equation 3 for is well established" probably got an extra "for"
  • l. 366: "... variables, whereas in weither forecasting." ends abruptly
Questions for authors
  1. Could the authors substantiate their definition 1 (see Critical Weaknesses)?

  2. Could the authors explain over what feature dimension a coarsening is achieved conceptually (see Critical Weaknesses)?

  3. Could the authors provide a full description of their experiment setup (see Critical Weaknesses)?

  4. What is meant with the observational level of dynamics? (e.g. l. 032-033) In what way does the method allow for / encode hidden states?

  5. In what sense can GNNs capture spatio-temporal interactions via physical networks? (l. 093-094)

  6. In what sense is EvolveODE "enabling the first marriage of ideas between NeuralODEs and multiscale learning"?(l. 106-107) A quick web search shows already various papers on the topic (e.g. Xu et al., "Multi-scale Neural ODEs for 3D Medical Image Registration" in MICCAI 2021).

  7. What is an edge index variable (l. 291)? Is here meant the actual edge index or the edge features? How do individual coupling matrixes Qu↦vQ_{u \mapsto v} look like?

  8. The authors, at points, state phrases that are rather grand or general but are not further explained:

    • "Behind all naturally occurring observations are underlying coarse-scale processes governing the evolution of these local phenomena." (l. 133-135)
    • "Graph topology describes intricate real-world relationships, yet confounding factors entangle their features." (l. 164-165)
    • "From an information-theoretic perspective, such feature revitalization injects previously invisible contextual cues into the disentanglement process" (l. 313-315) Could the authours explain what exactly is meant or, at least, have a reference provided?
  9. Experiment variance: Experiment tables mostly do not report a metric for spread. Why is that and could it be added, e.g., over training seeds? Conversely, why does Table 1 which would have a natural spread over input buffer sizes not report it? In contrast, over what is the spread reported in Table 3?

  10. What do the line colors in Figure 4 imply? Most of the panels (all but "Actor") do device two lines, orange and blue. The figure does not indicate what the different colors imply (and why there is none for "Actor").

  11. Can the authors provide information on limitations and future works? On the one hand, this section is missing, on the other the authors refer to it in lines 944-945, where they would want to discuss the "finite split problem" further.

1 comment

We thank the reviewer for their time and feedback. We address each point below.

Weaknesses

W1. Please see our responses to the questions below.

W2 (Reproducibility). We apologize for the difficulty in running our code. The issue can be resolved by removing baselines and profile_utils from the dependencies, which is a straightforward change. We have tried to make our work as reproducible as possible: we have released almost our entire codebase, the .sh files used to run our experiments, and instructions in the README on how to start new experiments. Given this, we would kindly ask the reviewer to reconsider the reproducibility score.

W3 (Writing). Thank you for pointing out these issues. They have been fixed in the updated paper.

Questions

Q1. The second paragraph of the introduction explains why variables need to be sorted accurately into sub-processes. As stated just before the definition, we formalize the mismatch between variables and sub-processes discussed in the introduction to make it concrete. EvolveODE aims to avoid such mismatches, so the definition serves to motivate the subsequent examples by making the intuition from the introduction precise.

Q2. We believe there may be a small misunderstanding here. The kernel does not apply to a single dimension; it applies to a given input graph. All main-paper experiments have a known graph. The weather experiments (in the Appendix) learn an N×TN \times T graph, but this graph still does not occupy a single dimension.

Q3. Yes. The exact hyperparameter setups, including learning rate and weight decay, are available at the code link, and the README contains the commands we used to run our experiments.

Q4. The observational scale refers to the level of granularity of the observations. For example, for the Max Planck weather dataset (Appendix), the scale is the sampling frequency. As is standard, the observations are embedded into a latent space via a learnable embedding map.

Q5. The physical network represents spatial characteristics of the system dynamics, which allows the GNN to learn the temporal evolution of the observations. For example, in a road network, traffic bottlenecks in the graph will be reflected in the measurements at certain times, such as peak hours. Extracting spatiotemporal trends via a physical network is well established in the spatiotemporal GNN literature.

Q6. In context, the phrase refers to GNN research not having explored NeuralODEs specifically for this purpose. The citation at the end of our statement (reference [35], which was not included in the quoted text) supports this. Regarding the paper suggested by the reviewer, we believe it differs from our setting in several respects: it is (a) not a NeuralODE paper, (b) not a GNN paper, and (c) applied to medical imaging. It solves known ODEs across scales by coarsening target variables, but it does not find couplings in multiscale systems, which our work does.

On the other points: "edge index" refers to the edge_index variable in PyTorch Geometric, the most widely used library for GNN research and implementation. We define QQ as assignment matrices with entries in {0,1}\{0, 1\}.

Q7. The first two of these three phrases build on earlier context in the paper:

  • First statement. The explanation is given in the first paragraph of the introduction, along with several supporting citations. Rather than repeat the full argument, this paragraph summarizes its core idea, which we hoped would be familiar by this point in the paper and to readers with a background in multiscale GNNs or forecasting.
  • Second statement. Both parts of this statement are well established. Features become noisy because real-world measurements are unreliable, which obscures their true meaning. In addition, intra-variable differences are hard to discern because the features are low-dimensional and lie close together in observational space. This motivates the use of deep attention as the backbone of our method, which must sort variables into sub-processes.
  • Third statement. This statement is self-contained within the paragraph in which it appears.

Q8. Reporting results across multiple random seeds is not standard practice in forecasting. The variance is also very small, being one standard deviation less than the mean.

Q10. Thank you for the suggestion; we have added this section. It discusses how performance cannot be improved indefinitely by adding equations, since the number of such relationships is finite. Future work should therefore focus on the operations performed on the uncovered dynamical system. This is a promising direction, as it may allow performance to scale, an approximation of the underlying system having been learnt.

Reviewer_18HfOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes EvolveODE, a graph neural ODE that discovers latent “sub-process” graphs. It integrates an attraction–repulsion ODE with self-attention kernels, adds an m-well potential whose minima sit at points evenly spaced on a hypersphere, and then coarsens the nodes into clusters with Graclus. Repeating this K times gives a hierarchy of graphs. All levels are then coupled through products of the cluster-assignment matrices and solved jointly as one dynamical system, followed by a task-specific output head.

Critical strengths
  1. Interesting idea. Using attraction–repulsion dynamics with a multi-well potential to drive node features into several clusters, and turning those clusters into a learned hierarchy of coupled graphs, is a creative link between graph neural ODEs and hierarchical pooling.
Critical weaknesses
  1. The central claim of physical understanding is not supported by the experiments. The paper repeatedly claims that EvolveODE learns and emulates the physical processes underlying observations, and the conclusion says it “approximates, from a bedrock level, the governing equations”. Yet the evidence is almost entirely predictive accuracy: (a) Traffic forecasting and node classification measure only MAE or accuracy. Node classification has no physical interpretation at all. (b) The only test against known latent structure is the synthetic Kuramoto experiment. It reports graph-similarity scores (“Dist.” and “Sim.”, which are not defined) without any baseline. It also checks only whether the cluster graphs match. It does not check whether the learned dynamics recover the true couplings, natural frequencies or governing equations.
  2. Baselines. Accuracy comparisons have no standard spatio-temporal traffic GNNs and hierarchical GNN baselines, even though the latter are the paper’s main motivation.
Questions for authors
  1. Validating physical understanding. Can you show that EvolveODE recovers physically meaningful structure rather than only improving accuracy? For example: (a)In the Kuramoto setting, do the learned dynamics recover the true coupling strengths or natural frequencies, and how do the recovered graphs compare against simple baselines (correlation-based spectral clustering, Graclus or DiffPool without the ODE)? (b) On PEMS, do the discovered sub-processes align with known road-network structure? (c) Please also define “Dist.” and “Sim.” and clarify whether the latent oscillators are part of the observed signal Y(t).

  2. Baselines. Please consider to compare your method with standard spatiio-temporal GNN and hierarchical GNN baselines.

1 comment

We thank the reviewer for their feedback. Our responses to the weaknesses are covered in our answers to Q1 and Q2 below.

Questions

Q1. We address points (a)–(c) below.

(a) Physical reasoning and coupling analysis. We understand the reviewer's concern about evidence for physical reasoning. We would also like to emphasize the importance of finding inter- and intra-scale couplings in the first place. Learning these connections is a challenging task, and recovering them with high fidelity directly supports the main claims of the paper, since finding such couplings is one of its central motivations. Figure 2 shows that EvolveODE accurately uncovers the inter-scale couplings of the system. Intra-scale couplings are much harder to visualize, but the table in the same figure evaluates both intra- and inter-scale topology.

The reviewer also notes a missing comparison with "correlation-based spectral clustering, Graclus or DiffPool without the ODE". These comparisons are included in the Appendix (Tables 6 and 7 and the accompanying text), and we apologize if they were easy to miss. These experiments show that the mm-well potential ODE is needed: spectral clustering and 2-well ODEs are not sufficient to uncover the system couplings.

We agree that the coupling strengths are an interesting question. However, the Kuramoto oscillator and the ODE we use are not directly comparable, as they have different governing equations. Our primary focus is on first finding the couplings, and our work does not extend to equation discovery beyond this point, although we see this as a very promising direction for future work. We also believe the suggested frequency analysis falls outside the scope of this paper, as it is typically carried out by Fourier-based methods that train on specific modes.

(b) Latent structure in traffic networks. We appreciate this suggestion. However, there is no known latent structure for a traffic network that could serve as ground truth. It is unclear which coarse-scale topology is most responsible for the observations, even if that topology reflects spatial characteristics. For this reason, we believe such an experiment would not be a well-posed test of uncovering topology. Instead, we constructed a synthetic experiment, which allows a direct comparison with the ground truth. We have also conducted a real-world case study (Figure 8) that shows recovery of known physical processes.

(c) Definitions. Thank you for pointing this out. We have added definitions of Dist and Sim in Appendix B. Could the reviewer please clarify what is meant by whether the latent oscillators are part of the observed signal Y(t)\mathbf{Y}(t)? We define Y(t)\mathbf{Y}(t) in terms of x(t)\mathbf{x}(t). As stated in Definition 1, xi(t)∈RD\mathbf{x}_i(t) \in \mathbb{R}^D is a latent vector describing transient properties of particle motion.

Q2 (Baselines). Thank you for this suggestion. We would like to note that Tables 1 and 2 already include multiscale GNN and spatiotemporal GNN baselines, which are separated from the other baselines by a horizontal line. In particular, we compare against MAGE and CITRUS, which are state-of-the-art spatiotemporal (continuous-time) baselines. CITRUS connects multiple graphs into a unified view through tensor products, which makes it comparable to our method, since EvolveODE also constructs an underlying dynamical system. MAGE is another strong comparison because it builds multiple views of the same graph, similar to how EvolveODE learns such views, though EvolveODE does so at multiple scales. These comparisons allow us to assess the key learning assumptions. If the reviewer has suggestions for further baselines, we would be happy to include them in the paper.

Reviewer_M4UpOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

EvolveODE constructs a multiscale graph model by evolving features under attraction–repulsion dynamics with an m-well potential, coarsening the resulting features with Graclus, and jointly evolving the graphs through cross-scale coupling. The intended interpretation is that sorting reveals latent subprocesses and their interactions. The paper gives clustering/stability claims and evaluates traffic forecasting, graph node classification, a synthetic coupled-oscillator system, and long-horizon weather forecasting.

Critical strengths
  1. The combination of multi-well attraction–repulsion dynamics, explicit graph coarsening, and cross-scale coupling is an interesting architectural idea. It provides a concrete alternative to conventional hierarchical aggregation.
  2. The model shows promising results under limited training data and at longer forecast horizons. It also uses substantially fewer FLOPs than DCRNN and STGCN in the reported comparison, while GPU memory changes little as hierarchy depth increases.
  3. The synthetic reconstruction experiment and hierarchy-depth ablations offer useful evidence for EvolveODE within the tested settings.
Critical weaknesses
  1. The interpretation of the learned graphs as underlying physical subprocesses remains only partially validated. Direct structural recovery is evaluated on a constructed Kuramoto system, while the real-data experiments primarily measure predictive accuracy. Those improvements do not establish that the learned hierarchy corresponds to genuine physical processes.

  2. The paper doesn't really say how to pick the hierarchy and dynamical parameters in practice. It leans on choices like the number of scales and wells, the repulsion strength, the integration settings, and the coupling topology. The appendix does include sensitivity studies, but these offer little guidance on how to set these values for a new dataset.

  3. Table 4 shows EvolveODE-1 taking 27.46 seconds per epoch and EvolveODE-5 taking 120.61 seconds, next to 26.26 seconds for STGCN. Coupled with the modest—and sometimes negative—gains in some short-horizon settings, this leaves the accuracy–runtime trade-off pretty poorly characterized.

  4. The paper does not follow the ICLR template. As submitted, the manuscript departs from the official ICLR paper template.

Questions for authors
  1. The paper's central claim is that EvolveODE recovers the underlying physical sub-processes, yet on real data (traffic, weather, node classification) there is no ground truth for sub-processes. The only direct check—Figure 2—uses a synthetic Kuramoto system that the authors themselves construct, with coupling rules and noise they specify in Appendix B. How do we know the model is recovering true sub-processes rather than merely fitting the generative structure the authors built in? Could the performance gains in Tables 1–2 instead be explained by stronger regularization or better fitting, independent of any "physical" recovery?

  2. Eq. (12) bounds the deviation by C(τ−t)C(\tau-t), and the proof then takes τ−t\tau-t small—a one-step bound. But the theorem is titled "in the Limit" and Lemma 3.5 claims convergence as t→∞t\to\infty. Monotone energy decrease does not by itself guarantee finite-time entry into Bξ(αq)B_\xi(\alpha_q). How does the one-step bound compose across steps, and is the accumulation point necessarily some αq\alpha_q?

  3. For G(y)=(y2−1)2\mathbf{G}(\mathbf{y})=(\mathbf{y}^2-1)^2, if all node states are zero, both ∇G(0)\nabla\mathbf{G}(0) and the pairwise terms vanish—an equilibrium away from either well. What excludes this? Lemma 3.5 claims convergence for all initializations; if a non-degenerate initialization is intended, it should be stated explicitly.

2 comments

We thank the reviewer for their detailed feedback. We address each point below.

Weaknesses

W1. Please see our response to Q1 in the reply below.

W2 (Hyperparameters). Thank you for raising this. We would like to clarify that not every non-learnable parameter is a hyperparameter that needs tuning. Tuning is usually needed when the relationship between a parameter and the experimental outcome is unclear. For almost all of the listed parameters, we believe this is not the case for EvolveODE. We address each parameter below.

  • Repulsion strength ε\varepsilon. For forecasting tasks, we fix ε=0.2\varepsilon = 0.2, since these tasks tend to be numerically stable. Performance is invariant to ε\varepsilon within [0.1,0.5][0.1, 0.5], which suggests that the attention mechanism corrects for this quantity and still uncovers the same sub-processes. It only begins to affect outcomes once the attention values (bounded above by 1) can no longer compensate. For node classification, Figure 4 shows that the repulsion strength should be high for heterophilic graphs and low for homophilic graphs. This gives a clear rule for setting it.
  • Integration time TT and step size. Figure 7 shows a clear relationship between integration time and performance. As expected, more accurate settings of the forward Euler method give stronger performance, and integrating for longer lets the dynamics converge. For example, T=5,10,15T = 5, 10, 15 have all settled and are roughly invariant to step size. The deviations that remain come from the numerical instability of the forward Euler method. T=20T = 20 reduces error at step length 25, but only because it improves numerical precision. In short, we do not believe this parameter requires tuning. Adaptive step-size solvers, which take larger steps when x(t)\mathbf{x}(t) changes slowly, would also remove the need to choose TT at all.
  • Coupling topology. This parameter also does not appear to need tuning. The difference in performance is small (at most 0.6%) and only appears for a system with 5 equation sets. In addition, the principle behind EvolveODE does not depend on a specific coupling architecture.
  • Coarsening factor. We agree that this is the one genuinely tunable parameter in the list. What matters is that the similarity between learnable graph structures overlaps only to a limited extent. A factor of 0.5 was enough to obtain strong performance across our experiments, so we did not find extensive tuning necessary.

W3 (Accuracy–runtime trade-off). We respectfully disagree that EvolveODE shows a poor accuracy–runtime trade-off, whether for short-term forecasting or for the other tasks. On average, EvolveODE improves margins by 1.3%–4.1% across all short-term forecasting datasets, horizons, buffer sizes and baselines, and it loses on only 4 of 36 individual settings. Gains on the other tasks in the main paper are larger: 2.8%–10% on long-term traffic forecasting and up to 7.2% on node classification. Across these tasks, the model loses on only 1 of 13 individual experiments, by a margin of 0.7%. For weather forecasting at long lead times, gains reach 17.8%. We believe these improvements are substantial and represent a favourable trade-off.

W4 (Citations). Thank you for pointing this out. We have corrected the citation formatting, which required a few minor changes.

Questions

Q1 (Source of gains / regularization). We appreciate this question. It is reasonable to ask whether the gains could come from a source other than the one described in the main paper. However, we have not found evidence that they can be explained by "better regularization", and attraction–repulsion ODEs of this kind are not known to have under-constrained dynamics. ACMP, for instance, constrains post-convergence features to lie in one of two feature groups. The mm-well potential of EvolveODE allows convergence to up to mm different groups. If anything, this has the opposite effect to regularization, because it allows for greater feature diversity.

We have also investigated this question in detail in the Appendix, through ablation studies and a real-world case study:

  • Ablations. Tables 6 and 7 show that the mm-well attraction–repulsion ODE is needed to construct the hierarchy. A 2-well ODE performs reasonably at single-scale forecasting but fails at multiscale forecasting, which indicates that it does not recover the underlying structure.
  • Real-world case study. In Figure 8, we run EvolveODE on real-world weather observations and show that it discovers plausible governing dynamics. The only input to EvolveODE is the raw observations, so this structure cannot have been inherited from any other input signal. If features were grouped randomly, the clustering operation would produce random groups of variables with no apparent relationship. Table 9 reports the long-term weather forecasting results from this experiment. EvolveODE remains stable across prediction horizons while the error of other methods increases, which suggests that EvolveODE captures first-order underlying dynamics when it generates predictions.

Q2 (Proof). Thank you for the careful reading of the proof. We would like to clarify a few points.

First, the proof considers the possibility of a particle leaving Bξ(αq)B_\xi(\alpha_q) at some time step. The limit t→∞t \to \infty ensures that the free-energy term converges, which removes any edge conditions. We do not claim to show accumulated particle motion. Instead, we show that if a particle is close to ∂Bξ(αq)\partial B_\xi(\alpha_q) (the boundary of the region), its directional derivative points away from the boundary, roughly inwards towards αq\alpha_q. For this reason, arguments about accumulated particle motion do not affect the outcome of the proof.

Second, we do not show that αq\alpha_q is necessarily a fixed point of a given particle. This depends on the relative interactions between particles adjacent to it in the graph, i.e. the set N(i)\mathcal{N}(i). This does not affect the result, however. The features lie in a relatively small neighbourhood, and the clustering kernel K\mathcal{K} collects them into the same sub-process, which recovers the underlying topology.

We also agree that a particle may leave Bξ(αq)B_\xi(\alpha_q) for sufficiently large step sizes. For any directional derivative pointing away from αq\alpha_q, such a step size always exists, however small the derivative. This is why we assume that the step size is sufficiently small.

We would also like to respectfully clarify the statement "Monotone energy decrease does not by itself guarantee finite-time entry into Bξ(αq)B_\xi(\alpha_q)". As discussed in Appendix B, the free energy is a Lyapunov function, so it decreases strictly with time and is always positive. Let Φ\Phi be the free energy of the unconstrained diffusion process (Equation 1). Then Φ+G\Phi + G is the free energy of the constrained diffusion process (Equation 3), where GG is the polynomial potential energy and both Φ\Phi and GG are always positive. Since αq\alpha_q is a minimum of GG, it is a minimum of Φ+G\Phi + G. Therefore, for some qq, monotonic energy decrease does guarantee entry into Bξ(αq)B_\xi(\alpha_q). Some αq\alpha_q will also be minima of Φ\Phi, i.e. fixed points of dxdt\frac{d\mathbf{x}}{dt}, although this does not apply to all αq\alpha_q. The mm-well potential creates diversity among the set of possible minima.

Q3 (Instability at the origin). We believe there may be a small sign discrepancy here. The origin of this potential energy function is an unstable fixed point, i.e. particles move away from it. The review writes the term as +G(y)=(y2−1)2+G(y) = (y^2 - 1)^2, whereas Equation 3 writes dxdt\frac{d\mathbf{x}}{dt} in terms of −G(y)-G(y), so the effective influence of G(y)G(y) on particle dynamics is −G(y)-G(y). For particles to remain at this fixed point, all features would have to be initialized to exactly zero: not close to zero, but numerically zero. For 32-bit floats, this would require initial values smaller than roughly the 45th decimal place. In practice, the only way to initialize at zero would be an embedding matrix whose entries are all zero, and such weights would receive no gradients.

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

Related papers

Discussion (0)

Sign in to comment.