Generative modeling of granular flow on inclined planes using conditional flow matching

Rui Li; Teng Man; Xuyang Li; Yimin Lu

arxiv: 2604.04453 · v2 · pith:ERTGNUO3new · submitted 2026-04-06 · 💻 cs.CE · cs.LG

Generative modeling of granular flow on inclined planes using conditional flow matching

Xuyang Li , Rui Li , Teng Man , Yimin Lu This is my paper

Pith reviewed 2026-05-10 20:11 UTC · model grok-4.3

classification 💻 cs.CE cs.LG

keywords granular flowconditional flow matchinginverse reconstructiondiscrete element methodgenerative modelingsparse observationsphysics-informed inference

0 comments

The pith

A conditional flow matching model reconstructs interior granular velocities and stresses from sparse boundary data by training on particle simulations and enforcing consistency via a differentiable forward operator.

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

The paper develops a generative framework that learns to map sparse boundary measurements to full interior flow fields in granular materials sliding down inclined planes. It trains exclusively on high-fidelity discrete-element simulations and then uses gradient guidance from a physics-based forward model at inference time to keep predictions consistent with the available observations. The approach succeeds even when only 11 to 16 percent of the informative data window is supplied and yields not only mean fields but also spatially resolved uncertainty estimates and derived quantities such as stress and granular temperature. A deterministic convolutional baseline fails to maintain physical consistency in the same ill-posed regimes. The result is a practical, non-invasive route to infer hidden bulk mechanics that conventional direct observation or expensive forward simulation cannot provide.

Core claim

Trained on particle-resolved discrete-element simulations of granular flow on inclined planes, a conditional flow matching model, guided at inference by a differentiable forward operator with sparsity-aware gradient guidance, recovers physically consistent interior velocity fields, mean and deviatoric stresses, and granular temperature from boundary observations that are reduced to 16 percent or even 11 percent of the informative window, while outperforming a deterministic CNN baseline and supplying ensemble-based uncertainty maps.

What carries the argument

Conditional flow matching generative model guided by a differentiable forward operator that enforces measurement consistency without hyperparameter tuning.

If this is right

Interior velocity fields remain accurate down to 16 percent of the informative observation window.
The same reconstructions stay reliable at strongly diluted spatial resolutions using only 11 percent of the data.
A physics decoder converts the velocity output into mean stress, deviatoric stress, and granular temperature without additional training.
Ensemble sampling from the generative model supplies spatially resolved uncertainty estimates.
The method outperforms a deterministic CNN baseline in the most under-determined reconstruction settings.

Where Pith is reading between the lines

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

The same conditional guidance mechanism could be applied to other particulate or multiphase inverse problems where only boundary data are accessible.
Real-time deployment would require replacing the current differentiable operator with a faster surrogate that still preserves measurement consistency.
Uncertainty maps could guide adaptive sensor placement in future experimental setups.

Load-bearing premise

A model trained only on simulated particle data will generalize to real sparse boundary observations without producing unphysical artifacts or requiring extra tuning.

What would settle it

Perform a controlled experiment with transparent side walls and interior particle tracking on the same inclined-plane geometry, then compare the model-reconstructed velocity and stress fields against the directly measured interior data at the same sparse boundary sampling density.

Figures

Figures reproduced from arXiv: 2604.04453 by Rui Li, Teng Man, Xuyang Li, Yimin Lu.

**Figure 2.** Figure 2: Overview of the proposed framework. The architecture integrates (a) a continuous-time proba [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗

**Figure 3.** Figure 3: Empirical cumulative distribution functions (eCDFs) of the three velocity components ( [PITH_FULL_IMAGE:figures/full_fig_p019_3.png] view at source ↗

**Figure 4.** Figure 4: Reconstruction of internal velocity fields from full boundary observations at [PITH_FULL_IMAGE:figures/full_fig_p020_4.png] view at source ↗

**Figure 5.** Figure 5: Reconstruction of internal streamwise velocity fields ( [PITH_FULL_IMAGE:figures/full_fig_p022_5.png] view at source ↗

**Figure 6.** Figure 6: End-to-end inference of physical fields from partial boundary observations at [PITH_FULL_IMAGE:figures/full_fig_p023_6.png] view at source ↗

**Figure 7.** Figure 7: Comparison of the reconstructed streamwise velocity ( [PITH_FULL_IMAGE:figures/full_fig_p026_7.png] view at source ↗

**Figure 8.** Figure 8: Uncertainty analysis of the internal flow reconstruction under partial observation at [PITH_FULL_IMAGE:figures/full_fig_p028_8.png] view at source ↗

**Figure 9.** Figure 9: Ablation study comparing the proposed sparsity-aware guidance (top row) and the baseline [PITH_FULL_IMAGE:figures/full_fig_p030_9.png] view at source ↗

**Figure 10.** Figure 10: Sensitivity analysis of the reconstruction fidelity under varying observation constraints (evaluated [PITH_FULL_IMAGE:figures/full_fig_p032_10.png] view at source ↗

read the original abstract

Granular flows govern many natural and industrial processes, yet their interior kinematics and mechanics remain largely unobservable, as experiments access only boundaries or free surfaces. Conventional numerical simulations are computationally expensive for fast inverse reconstruction, and deterministic models tend to collapse to over-smoothed mean predictions in ill-posed settings. This study, to the best of the authors' knowledge, presents the first conditional flow matching (CFM) framework for granular-flow reconstruction from sparse boundary observations. Trained on high-fidelity particle-resolved discrete element simulations, the generative model is guided at inference by a differentiable forward operator and a novel sparsity-aware gradient guidance mechanism. This mechanism avoids the gradient dilution inherent to standard mean-squared-error approaches, preserves the absolute physical scale of observation errors, enforces measurement consistency without hyperparameter tuning, and prevents unphysical velocity predictions in non-material regions. A physics decoder maps the reconstructed velocity fields to stress states and energy fluctuation quantities, including mean stress, deviatoric stress, and granular temperature. The framework accurately recovers interior flow fields from full observation to only 16\% of the informative window, and it remains effective under strongly diluted spatial resolution with only 11% of data. It also outperforms a deterministic CNN baseline in the most ill-posed reconstruction regime and provides spatially resolved uncertainty estimates through ensemble generation. These results demonstrate that conditional generative modeling offers a practical route for non-invasive inference of hidden bulk mechanics in granular media, and it suggests potential applicability for inverse problems in particulate and multiphase systems.

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 paper applies conditional flow matching to reconstruct interior granular flows from sparse boundary data using DEM training and a physics decoder, but all quantitative claims come from simulation-only tests.

read the letter

The core contribution is a conditional flow matching model trained on discrete element simulations of inclined granular flows. At inference it uses a differentiable forward operator with sparsity-aware gradient guidance to enforce consistency with limited boundary measurements, then decodes the velocity field into stresses and granular temperature. This generative route lets them output ensembles and uncertainty maps, which is a step beyond deterministic CNN baselines in the most under-determined regimes (down to 11-16% of the data window). The combination of CFM with the physics decoder and guidance mechanism appears new for this inverse problem in granular mechanics. It handles the ill-posedness by sampling from a learned distribution rather than collapsing to a mean field, and the abstract shows it recovers fields more faithfully than the baseline when observations are heavily diluted. That part is useful for anyone thinking about non-invasive inference in particulate systems. The main limitation is that every reported recovery accuracy, outperformance, and uncertainty estimate comes from sparsifying observations drawn from the same DEM distribution used for training. No physical experiments, sensor noise, boundary irregularities, or independent interior validation are described, so the claim that the model will stay physically consistent on real granular flows remains untested. The abstract also gives no numbers, error bars, or ablation details, which makes it hard to judge how robust the gains actually are. This work is aimed at researchers in granular mechanics and ML-for-physics inverse problems who already work with simulation data and want a generative alternative to deterministic reconstruction. A reader focused on method novelty would find the guidance mechanism and decoder worth looking at; someone needing demonstrated real-world performance would not get much yet. It deserves a serious referee because the technical setup is coherent and the problem is well-motivated, even though the validation needs expansion before the practical-route claim can be taken at face value. I would send it for review with a request for real-data tests or at least clearer quantitative reporting.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces the first conditional flow matching (CFM) generative framework for reconstructing interior velocity fields in granular flows on inclined planes from sparse boundary observations. Trained exclusively on high-fidelity particle-resolved DEM simulations, the model is guided at inference by a differentiable forward operator incorporating sparsity-aware gradient guidance to enforce measurement consistency and avoid unphysical predictions. A physics decoder then maps the reconstructed velocities to mean stress, deviatoric stress, and granular temperature. The work reports accurate recovery down to 16% of the informative window and effectiveness at 11% data dilution, outperformance versus a deterministic CNN baseline in ill-posed regimes, and ensemble-based uncertainty quantification.

Significance. If the performance claims hold, the approach would represent a meaningful advance for inverse problems in granular and particulate systems by providing a generative, uncertainty-aware route to infer hidden bulk mechanics from limited boundary data. Strengths include the integration of CFM with a physics-informed guidance mechanism that requires no hyperparameter tuning and the addition of a decoder for stress and fluctuation quantities. These elements address the ill-posedness of sparse reconstruction more flexibly than deterministic baselines.

major comments (2)

[Abstract] Abstract: the statements that the framework 'accurately recovers interior flow fields' from full observation to only 16% of the informative window and 'remains effective' at 11% data are presented without any numerical metrics, error bars, validation splits, or ablation results. These quantitative details are load-bearing for the central performance claims and must be supplied explicitly in the results section.
[Results] Results section: all reported quantitative evaluations (recovery accuracy, CNN comparison, stress/temperature decoding) are generated by sparsifying observations drawn from the identical DEM simulation distribution used for training. This leaves untested the claim of a 'practical route for non-invasive inference' on real-world sparse boundary observations that may include sensor noise, boundary irregularities, and material variability; a concrete test or discussion of distribution shift is required to support the broader applicability assertion.

minor comments (2)

Figure captions and axis labels for velocity field reconstructions and uncertainty maps should be expanded to indicate the exact sparsity level and ensemble size used in each panel.
[Methods] The definition and implementation of the 'sparsity-aware gradient guidance' term in the inference procedure should be given a dedicated equation or pseudocode block for reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. The comments have helped us strengthen the presentation of quantitative results and clarify the scope of applicability. We address each major comment below and indicate the revisions made.

read point-by-point responses

Referee: [Abstract] Abstract: the statements that the framework 'accurately recovers interior flow fields' from full observation to only 16% of the informative window and 'remains effective' at 11% data are presented without any numerical metrics, error bars, validation splits, or ablation results. These quantitative details are load-bearing for the central performance claims and must be supplied explicitly in the results section.

Authors: We agree that the abstract claims require explicit quantitative support. In the revised manuscript we have expanded the Results section with detailed numerical metrics, including mean squared error and relative L2 errors (with standard deviations) for velocity reconstruction at the 16% and 11% sparsity levels, obtained from multiple validation splits and ablation studies. The CNN baseline comparison and stress/temperature decoding errors are also reported with the same rigor. The abstract has been lightly edited to reference these concrete results while preserving its length. revision: yes
Referee: [Results] Results section: all reported quantitative evaluations (recovery accuracy, CNN comparison, stress/temperature decoding) are generated by sparsifying observations drawn from the identical DEM simulation distribution used for training. This leaves untested the claim of a 'practical route for non-invasive inference' on real-world sparse boundary observations that may include sensor noise, boundary irregularities, and material variability; a concrete test or discussion of distribution shift is required to support the broader applicability assertion.

Authors: The referee is correct that all quantitative results are obtained by sparsifying data from the same DEM distribution. To address this, the revised manuscript adds a new subsection discussing distribution shift. We report additional experiments in which Gaussian noise is superimposed on the boundary observations to emulate sensor noise; the model retains useful reconstruction accuracy under moderate noise levels. We also discuss boundary irregularities and material variability as open challenges and note how the generative, ensemble-based nature of CFM may offer advantages over deterministic models in such settings. A full evaluation on experimental (non-simulated) data is not feasible in the present study because training relies on high-fidelity DEM simulations; we now explicitly list this as a limitation and identify it as a key direction for future work. revision: partial

Circularity Check

0 steps flagged

No significant circularity; framework applies existing CFM to external DEM data with independent guidance

full rationale

The paper introduces a conditional flow matching model trained exclusively on high-fidelity DEM simulations and guided at inference by a separate differentiable forward operator plus sparsity-aware gradients. No self-definitional steps, fitted inputs renamed as predictions, or load-bearing self-citations appear in the abstract or described chain. Performance claims rest on empirical recovery from sparsified simulation data (standard held-out evaluation) rather than any reduction of outputs to inputs by construction. The derivation is therefore self-contained as a domain application of generative modeling.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review prevents identification of specific free parameters or axioms; the approach implicitly assumes that DEM simulation data sufficiently represents real granular behavior and that the differentiable forward operator accurately encodes measurement physics without additional fitted constants.

pith-pipeline@v0.9.0 · 5544 in / 1204 out tokens · 68145 ms · 2026-05-10T20:11:11.040351+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

conditional flow matching (CFM) framework... differentiable forward operator with a sparsity-aware gradient guidance mechanism... physics decoder maps the reconstructed velocity fields to stress states and energy fluctuation quantities
IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Trained on high-fidelity particle-resolved discrete element simulations... outperforms a deterministic CNN baseline

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.