Learning Permutation-invariant Macroscopic Dynamics

Mengyi Chen; Qianxiao Li; Zhichao Han

arxiv: 2605.30812 · v1 · pith:4VRKA4C3new · submitted 2026-05-29 · 💻 cs.LG · physics.comp-ph

Learning Permutation-invariant Macroscopic Dynamics

Zhichao Han , Mengyi Chen , Qianxiao Li This is my paper

Pith reviewed 2026-06-28 23:39 UTC · model grok-4.3

classification 💻 cs.LG physics.comp-ph

keywords permutation-invariantautoencodermacroscopic dynamicsparticle systemslatent representationsmass distributionfluid mixingpolymer dynamics

0 comments

The pith

A permutation-invariant autoencoder learns macroscopic dynamics from unordered microscopic states.

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

The paper establishes that standard autoencoders fail for systems like particle collections because they assume fixed input ordering. It introduces a permutation-invariant encoder paired with a decoder that reconstructs the overall mass distribution around observed points, then jointly optimizes latent states and the dynamics of macroscopic observables. A sympathetic reader would care because this removes the need to impose arbitrary labels on indistinguishable components. Demonstrations cover energy evolution in particle interactions, mixing in fluids, and stretching extracted from polymer videos. If correct, the approach makes data-driven reduced-order modeling viable for any collection where order carries no meaning.

Core claim

We adopt a permutation-invariant encoder and design the decoder to reconstruct the mass distribution centered at the observed points rather than per-sample reconstruction. We then jointly learn the macroscopic dynamics of the observables together with the latent states. The resulting framework is shown to be effective and robust across interacting particle systems, Lennard-Jones fluids, and video observations of polymers under elongational force.

What carries the argument

Permutation-invariant encoder together with mass-distribution decoder, which produces order-independent latent representations usable for dynamics prediction.

If this is right

The method reproduces energy dynamics in interacting particle systems.
It predicts mixing dynamics in Lennard-Jones fluids.
It recovers stretching dynamics from video of polymers in an elongational force field.
Performance holds across multiple microscopic settings without requiring fixed input order.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same encoder-decoder pattern could be tested on other unordered data structures such as point clouds or sets in general machine-learning tasks.
One could check whether the learned latent states transfer across simulations that differ only in how particles are labeled.
A direct comparison on the same datasets with and without the mass-distribution reconstruction step would isolate its contribution to dynamics accuracy.

Load-bearing premise

Reconstructing the mass distribution rather than individual points supplies a latent representation sufficient to capture the dynamics of the macroscopic observables.

What would settle it

Running the trained model on inputs whose particle ordering has been randomly shuffled and checking whether the predicted macroscopic observables remain accurate to the same degree as on the original ordering.

Figures

Figures reproduced from arXiv: 2605.30812 by Mengyi Chen, Qianxiao Li, Zhichao Han.

**Figure 1.** Figure 1: Existing and desired autoencoder for learning closure variables. (a) Standard autoencoder to learn the latent zˆ. Microstate X is represented by a vector based on the ordering. The encoder φˆ and decoder ψ are typically implemented as MLPs. (b) The desired autoencoder to learn zˆ. The model should preduce a permutation-invariant latent variable zˆ for inputs that lack the canonical ordering. Recent data-d… view at source ↗

**Figure 2.** Figure 2: Overview of our distribution-aware autoencoder for closure modeling. The encoder φˆ maps unordered microstate X to a permutation-invariant latent variable zˆ. The microstate X induces a target distribution qX. Conditioned on zˆ, the decoder ψ generates a density pθ(x|zˆ) to approximate the target density qX. In general, φˆ is a permutation-invariant set function and ψ is a conditional density function. We … view at source ↗

**Figure 3.** Figure 3: Workflow of learning macroscopic dynamics. We are given high-dimensional observations at the microscopic scale that evolve over time. The deterministic encoder φ¯ extracts the macroscopic feature of interest z¯ (for example, the system energy). A learned encoder φˆ extracts the closure macroscopic variables zˆ. A dynamic model is trained to predict the macroscopic dynamics of z¯ and zˆ together. sure vari… view at source ↗

**Figure 4.** Figure 4: Distributional reconstruction with different ϵ and zˆdim. The gray dotted curve denotes the Gaussian mixture distribution from which particles are sampled, and the blue rug marks indicate sample locations. The blue solid curve shows the target density. The orange dashed curve is the density learned by our model. does not necessarily require a reconstruction objective, we also consider the strategy of direc… view at source ↗

**Figure 5.** Figure 5: Experiment on the interacting particle system. (a) Microstate snapshots examples. (b) z¯ predicted by our model versus the ground truth. (c) Permutation-invariant verification. component GMM with independently resampled component means and isotropic covariances. This is for in-distribution evaluation, and we denote it as in-dst. Secondly, we generate 1k trajectories with the same number of particles as tr… view at source ↗

**Figure 6.** Figure 6: Experiment on the binary particle mixing. (a) An example microstate trajectory for illustrating the mixing process. (b) The mean of RAB and RBA under three initial separation boundaries: x = 12 (left), x = 16 (middle), and x = 20 (right). under different permutations, even with permutation augmentation during training. In contrast, our method is exactly permutation-invariant: the three prediction curves o… view at source ↗

**Figure 7.** Figure 7: Polymer extension prediction. (a) The snapshots of the first trajectory in these three test datasets at t=0 (initial configuration), t=100, t=200 and t=300. (b) The mean and standard deviation of our predicted stretching length and the ground-truth. defines the microstate X, and we compute the stretching length z¯ from X. Note that the number of non-white pixels varies across time steps, and our model ca… view at source ↗

**Figure 8.** Figure 8: Exemplary trajectory in the polymer video dataset. We test the state-of-the-art image models, CNN and ViT, for learning the closure variables. We train CNN and ViT autoencoders on image reconstruction with MSE loss and use the latent vector as the closure variables. The macroscopic dynamics learned with their closure variables are shown in [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗

**Figure 9.** Figure 9: Polymer extension prediction with the closure variables learned by (a) CNN and (b) ViT. B.3. The Induced Density The qX(x) constructed in Sec. 3.2 can be viewed as a kernel density estimate of the empirical measure associated with X. As ϵ goes close to 0 (with |X| fixed), qX concentrates and converges in the weak sense to the empirical measure 1 |X| P j δxj (i.e., a sum of point masses) (Silverman, 2018). … view at source ↗

read the original abstract

Accurately modeling the macroscopic dynamics of high-dimensional microscopic systems is of broad interest across the sciences. Many data-driven approaches learn a low-dimensional latent state through an autoencoder trained for pointwise input reconstruction. These methods typically assume a fixed ordering of microscopic degrees of freedom in the input. However, in many settings, such as particle systems, the microscopic state is inherently unordered. This motivates an autoencoder framework that learns permutation-invariant latent representations. To this end, we adopt a permutation-invariant encoder and design the decoder to reconstruct the mass distribution centered at the observed points rather than per-sample reconstruction. We then jointly learn the macroscopic dynamics of the observables together with the latent states. We demonstrate the effectiveness and robustness of the proposed method across a range of microscopic settings, including learning the energy dynamics in interacting particle systems, predicting mixing dynamics in Lennard-Jones fluids, and modeling the stretching dynamics from video data of polymers moving in an elongational force field.

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.

Desk Editor's Note private letter to a colleague

The permutation-invariant encoder plus mass-distribution decoder is a direct fix for unordered particle inputs, but whether it preserves the info needed for accurate dynamics is the part that needs checking.

read the letter

The main thing here is that the authors replace standard pointwise reconstruction in an autoencoder with a decoder that targets the mass distribution around the observed points. This lets them keep a permutation-invariant encoder while still learning a latent state that they then evolve jointly with the macroscopic observables. They test it on interacting particles, Lennard-Jones fluids, and polymer stretching from video.

That architectural choice is the concrete contribution. It directly addresses the ordering problem that comes up whenever the microscopic degrees of freedom have no natural sequence, and the joint training objective is a reasonable way to tie the latent dynamics to the quantities people actually care about.

The soft spot is the one the stress-test note flags. Reconstructing only the distribution can erase higher-order configuration details (relative positions, connectivity, specific interaction geometries) that still matter for how the macro variables evolve. The abstract gives no derivation showing why those quantities are preserved, and without the actual loss functions, ablations, or quantitative error tables it is impossible to tell whether the reported robustness comes from the method or from the test cases being forgiving. If the full paper has solid controls on that point, the concern shrinks; if not, it stays central.

The work is aimed at people already building latent dynamical models from particle or fluid simulation data. A reader who has run into the ordering issue before will see a usable variant to try or adapt. It is worth sending to referees because the problem is real, the proposed fix is specific, and the experiments span different physical regimes; the review process can sort out whether the distribution decoder actually delivers on the dynamics claim.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes a permutation-invariant autoencoder for modeling macroscopic dynamics of unordered microscopic systems (e.g., particle systems). A permutation-invariant encoder is paired with a decoder that reconstructs the mass distribution centered at observed points rather than per-sample reconstruction; macroscopic dynamics of observables are then learned jointly with the latent states. Effectiveness is demonstrated on energy dynamics in interacting particles, mixing in Lennard-Jones fluids, and stretching dynamics from polymer video data.

Significance. If the central claim holds, the framework offers a practical route to data-driven macroscopic modeling without assuming fixed ordering of degrees of freedom, which is common in physical systems. The joint training and cross-domain demonstrations could strengthen data-driven approaches in physics-informed machine learning.

major comments (2)

[Decoder objective] Decoder objective (method description): the design reconstructs the mass distribution rather than individual points. This choice is load-bearing for the claim that the resulting latent state suffices to learn accurate macroscopic dynamics, yet the manuscript provides no derivation or controlled test showing that higher-order statistics or specific configurations (e.g., relative particle positions or chain connectivity) are preserved when they affect the observables.
[Experiments] Joint training procedure (experiments section): while results are reported across three settings, there is no ablation that isolates whether the mass-distribution decoder (versus a pointwise or permutation-equivariant alternative) is necessary for the reported dynamics accuracy; without this, it is unclear whether the permutation-invariance alone or the specific reconstruction objective drives the performance.

minor comments (2)

[Method] Notation for the mass distribution and latent-state evolution equations could be introduced earlier and used consistently when describing the joint loss.
[Figures] Figure captions should explicitly state the quantitative metric (e.g., prediction error on observables) used to assess dynamics learning in each panel.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments. We address each major point below and describe the revisions we will make.

read point-by-point responses

Referee: [Decoder objective] Decoder objective (method description): the design reconstructs the mass distribution rather than individual points. This choice is load-bearing for the claim that the resulting latent state suffices to learn accurate macroscopic dynamics, yet the manuscript provides no derivation or controlled test showing that higher-order statistics or specific configurations (e.g., relative particle positions or chain connectivity) are preserved when they affect the observables.

Authors: The mass-distribution decoder reconstructs the empirical measure, which encodes the full set of moments and statistics of the point cloud. This is sufficient for the permutation-invariant macroscopic observables considered in the work. We agree that an explicit derivation and controlled test would strengthen the justification. In the revision we will add a short derivation linking the decoder objective to preservation of distribution moments relevant to the dynamics, together with a controlled synthetic experiment that verifies retention of higher-order statistics when they influence the target observable. revision: yes
Referee: [Experiments] Joint training procedure (experiments section): while results are reported across three settings, there is no ablation that isolates whether the mass-distribution decoder (versus a pointwise or permutation-equivariant alternative) is necessary for the reported dynamics accuracy; without this, it is unclear whether the permutation-invariance alone or the specific reconstruction objective drives the performance.

Authors: We acknowledge that the current experiments do not isolate the contribution of the mass-distribution decoder from permutation invariance alone. In the revised manuscript we will add ablation studies that replace the mass-distribution decoder with a pointwise reconstruction baseline (adapted for unordered inputs) and with a permutation-equivariant decoder, reporting the resulting change in macroscopic dynamics accuracy on all three experimental domains. revision: yes

Circularity Check

0 steps flagged

No circularity; derivation is a methodological proposal without self-referential reductions

full rationale

The paper proposes a permutation-invariant autoencoder with mass-distribution decoder and joint dynamics learning for unordered microscopic systems. No equations, fitted parameters, or self-citations are present that reduce any claimed prediction or result to an input by construction. The central claims rest on design choices and empirical demonstrations across particle systems, fluids, and polymers, with no load-bearing self-definition, uniqueness theorems, or renamed known results. The derivation chain is self-contained.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review; no explicit free parameters, axioms, or invented entities are stated or can be extracted.

pith-pipeline@v0.9.1-grok · 5688 in / 1002 out tokens · 25852 ms · 2026-06-28T23:39:56.131877+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

2 extracted references · 1 canonical work pages

[1]

PMLR, 2019. Lee, K. and Carlberg, K. T. Model reduction of dynamical systems on nonlinear manifolds using deep convolutional autoencoders.Journal of Computational Physics, 404: 108973, 2020a. Lee, K. and Carlberg, K. T. Model reduction of dynamical systems on nonlinear manifolds using deep convolutional autoencoders.Journal of Computational Physics, 404: ...

work page doi:10.1007/978-3-540-78841-6 2019
[2]

test fast

ISBN 978-1-6654-2812-5. Yang, Y ., Feng, C., Shen, Y ., and Tian, D. Foldingnet: Point cloud auto-encoder via deep grid deformation. In2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 206–215. IEEE, 2018. ISBN 978-1-5386-6420-9. 11 Learning Permutation-invariant Macroscopic Dynamics Yu, H., Tian, X., Weinan, E., and Li, Q. O...

2018

[1] [1]

PMLR, 2019. Lee, K. and Carlberg, K. T. Model reduction of dynamical systems on nonlinear manifolds using deep convolutional autoencoders.Journal of Computational Physics, 404: 108973, 2020a. Lee, K. and Carlberg, K. T. Model reduction of dynamical systems on nonlinear manifolds using deep convolutional autoencoders.Journal of Computational Physics, 404: ...

work page doi:10.1007/978-3-540-78841-6 2019

[2] [2]

test fast

ISBN 978-1-6654-2812-5. Yang, Y ., Feng, C., Shen, Y ., and Tian, D. Foldingnet: Point cloud auto-encoder via deep grid deformation. In2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 206–215. IEEE, 2018. ISBN 978-1-5386-6420-9. 11 Learning Permutation-invariant Macroscopic Dynamics Yu, H., Tian, X., Weinan, E., and Li, Q. O...

2018