arxiv: 2604.09584 · v1 · submitted 2026-02-26 · 💻 cs.AI · cs.CV

Recognition: 2 theorem links

· Lean Theorem

Agentic Exploration of PDE Spaces using Latent Foundation Models for Parameterized Simulations

Abhijeet Vishwasrao , Francisco Giral , Mahmoud Golestanian , Federica Tonti , Andrea Arroyo Ramo , Adrian Lozano-Duran , Steven L. Brunton , Sergio Hoyas

show 3 more authors

Soledad Le Clainche Hector Gomez Ricardo Vinuesa

Authors on Pith no claims yet

Pith reviewed 2026-05-15 18:34 UTC · model grok-4.3

classification 💻 cs.AI cs.CV

keywords agentic LLMslatent foundation modelsPDE parameter explorationfluid dynamicsscaling lawstandem cylinderssurrogate modelingautomated discovery

0 comments

The pith

Multi-agent LLMs coupled with latent foundation models autonomously discover scaling laws in tandem cylinder flows by querying over 1600 parameter pairs.

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

The paper establishes that coupling multi-agent large language models with latent foundation models enables efficient automated exploration of continuous, high-dimensional PDE solution spaces that were previously limited by expensive simulations. The latent models learn compact, disentangled representations of flow fields and act as fast surrogate simulators for arbitrary parameter queries. Applied to viscous flow past tandem cylinders at Reynolds number 500, the hierarchical agent loop evaluates more than 1600 parameter-location combinations and identifies divergent scaling behaviors: a regime-dependent two-mode structure in minimum displacement thickness alongside robust linear scaling in maximum momentum thickness. Both quantities show a dual-extrema landscape that appears precisely at the transition between near-wake and co-shedding regimes. This combination of learned physical representations and agentic reasoning supplies a reusable template for discovery across other PDE-governed systems.

Core claim

The framework autonomously evaluates over 1,600 parameter-location pairs and discovers divergent scaling laws: a regime-dependent two-mode structure for minimum displacement thickness and a robust linear scaling for maximum momentum thickness, with both landscapes exhibiting a dual-extrema structure that emerges at the near-wake to co-shedding regime transition.

What carries the argument

Latent foundation models that learn explicit, compact, and disentangled latent representations of flow fields, used as on-demand surrogate simulators inside a closed-loop hierarchical multi-agent architecture.

If this is right

Automated exploration of high-dimensional PDE parameter spaces becomes feasible without repeated expensive simulations.
Regime transitions in fluid flows can be located automatically through emergent structures in displacement and momentum thickness landscapes.
Divergent scaling laws (two-mode versus linear) can be extracted directly from surrogate queries in wake flows.
The same agentic loop supplies a general template for closed-loop hypothesis generation and verification in other PDE-governed physical systems.

Where Pith is reading between the lines

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

The same surrogate-plus-agent architecture could be applied to other chaotic or high-dimensional systems such as turbulence or reacting flows once suitable latent models exist.
Integration with real-time experimental sensor data would allow the agents to close the loop between simulation surrogates and physical measurements.
The discovered dual-extrema structure might serve as a diagnostic signature for identifying wake-regime boundaries in related bluff-body configurations.
If the latent representations remain faithful at higher Reynolds numbers, the framework could accelerate the mapping of drag or lift scaling across wider parameter ranges.

Load-bearing premise

The latent foundation model supplies sufficiently accurate approximations of the true PDE solutions across the full parameter space so that the discovered scaling laws are not artifacts of the surrogate.

What would settle it

High-fidelity direct numerical simulations of a random subset of the 1600 configurations, followed by direct comparison of the measured minimum displacement thickness and maximum momentum thickness scaling curves against the agent-reported laws.

Figures

Figures reproduced from arXiv: 2604.09584 by Abhijeet Vishwasrao, Adrian Lozano-Duran, Andrea Arroyo Ramo, Federica Tonti, Francisco Giral, Hector Gomez, Mahmoud Golestanian, Ricardo Vinuesa, Sergio Hoyas, Soledad Le Clainche, Steven L. Brunton.

**Figure 2.** Figure 2: Extremal locations x ∗ δ ∗ (blue circles, minimum δ ∗ ) and x ∗ θ (orange squares, maximum θ) as functions of cylinder spacing S. Filled markers: LFM-derived values; open markers: simulation ground truth. The minimum δ ∗ location transitions from a near-constant value (≈ 0.5D) in the nearwake regime to a linearly growing trend (0.71S − 2.2D) in the co-shedding regime. The maximum θ location follows a sing… view at source ↗

**Figure 3.** Figure 3: Optimization landscape of (a) displacement thickness δ ∗ (S, xp) and (b) momentum thickness θ(S, xp) across the inter-cylinder gap. Stars: LFM-derived optima; red diamonds: simulation ground truth; dashed lines indicate the downstream cylinder center (xp = S) and surface (xp = S − 0.5D). (a) For small spacings a single minimum zone appears near xp ≲ 1D; for large spacings a second deeper minimum emerges … view at source ↗

**Figure 4.** Figure 4: Time-averaged streamwise velocity u¯ (rows 1–2) and streamwise Reynolds stress Ruu (rows 3–4): simulation (rows 1, 3) versus LFM (rows 2, 4) for S/D = 4, 8, 9.5 (from test dataset). The LFM faithfully reproduces the mean wake topology, recirculation zones and wake-recovery structure, as well as the turbulent stress distribution, across the full spacing range. captures the qualitative landscape structure an… view at source ↗

read the original abstract

Flow physics and more broadly physical phenomena governed by partial differential equations (PDEs), are inherently continuous, high-dimensional and often chaotic in nature. Traditionally, researchers have explored these rich spatiotemporal PDE solution spaces using laboratory experiments and/or computationally expensive numerical simulations. This severely limits automated and large-scale exploration, unlike domains such as drug discovery or materials science, where discrete, tokenizable representations naturally interface with large language models. We address this by coupling multi-agent LLMs with latent foundation models (LFMs), a generative model over parametrised simulations, that learns explicit, compact and disentangled latent representations of flow fields, enabling continuous exploration across governing PDE parameters and boundary conditions. The LFM serves as an on-demand surrogate simulator, allowing agents to query arbitrary parameter configurations at negligible cost. A hierarchical agent architecture orchestrates exploration through a closed loop of hypothesis, experimentation, analysis and verification, with a tool-modular interface requiring no user support. Applied to flow past tandem cylinders at Re = 500, the framework autonomously evaluates over 1,600 parameter-location pairs and discovers divergent scaling laws: a regime-dependent two-mode structure for minimum displacement thickness and a robust linear scaling for maximum momentum thickness, with both landscapes exhibiting a dual-extrema structure that emerges at the near-wake to co-shedding regime transition. The coupling of the learned physical representations with agentic reasoning establishes a general paradigm for automated scientific discovery in PDE-governed 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 agentic LFM framework automates PDE parameter sweeps at scale and reports concrete scaling laws in tandem cylinder wakes, but the surrogate's accuracy against CFD ground truth is not shown.

read the letter

The paper pairs a hierarchical multi-agent LLM system with a latent foundation model that serves as a cheap surrogate for parameterized flow simulations. Applied to tandem cylinders at Re=500, the agents run over 1600 evaluations and surface two specific scaling behaviors: a regime-dependent two-mode pattern for minimum displacement thickness and a linear trend for maximum momentum thickness, both showing dual extrema near the wake-to-co-shedding transition. That closed-loop setup, where agents handle hypothesis, query, analysis, and verification without manual intervention, is the clearest new piece. It moves beyond one-off surrogate use and lets the system traverse continuous parameter spaces that direct CFD cannot reach at the same volume. The modular tool interface also looks practical for extension to other PDE problems. The soft spot is validation. No error metrics, side-by-side field comparisons, or bounds on how well the LFM reproduces integral quantities in the transition regimes appear in the description. The scaling laws therefore rest on the assumption that the generative model stays faithful enough not to invent structure. If the full manuscript contains those checks, the claim strengthens; otherwise the discoveries stay preliminary. This work is aimed at computational fluid dynamicists and researchers building AI surrogates for physics. A reader already working on agentic or latent-model pipelines will see a concrete integration worth testing. It deserves peer review so referees can examine the validation data and any ablation on the agent hierarchy. The idea is grounded enough to justify the time.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces a framework coupling multi-agent LLMs with latent foundation models (LFMs) as on-demand surrogates for parameterized PDE simulations. Applied to flow past tandem cylinders at Re=500, the system autonomously evaluates over 1,600 parameter-location pairs via a closed-loop hypothesis-experimentation-analysis-verification process and reports the discovery of a regime-dependent two-mode structure for minimum displacement thickness, robust linear scaling for maximum momentum thickness, and dual-extrema structures emerging at the near-wake to co-shedding transition.

Significance. If the LFM surrogate fidelity is established, the work offers a potentially valuable paradigm for automated exploration of high-dimensional, continuous PDE spaces where traditional simulation is prohibitive. The demonstration of extracting specific, falsifiable scaling laws from large-scale parameter sweeps without manual intervention could accelerate discovery in fluid dynamics, provided the agentic loop produces insights that survive direct comparison to ground-truth solvers.

major comments (2)

[Abstract] Abstract and results: The central claim that the framework discovers genuine regime-dependent scaling laws rests on the unverified assumption that LFM-generated flow fields yield integral quantities (displacement thickness, momentum thickness) sufficiently close to high-fidelity CFD solutions. No quantitative error metrics, validation plots, or side-by-side comparisons are reported for the 1,600+ evaluations, especially across the near-wake to co-shedding transition where the dual-extrema structure is claimed to emerge.
[Methods] Methods and results: The hierarchical agent architecture is described at a high level, but the manuscript provides no concrete account of how verification steps guard against surrogate-induced artifacts or post-hoc interpretation when extracting scaling laws. A worked example showing the full loop (hypothesis, LFM query, analysis, verification against an independent check) for one reported law would be required to support the claim of scientifically valid insights.

minor comments (2)

[Abstract] The term 'divergent scaling laws' is introduced without a precise definition; clarify whether it refers to qualitatively different functional forms across regimes or simply to the coexistence of two-mode and linear behaviors.
Notation for the latent space, parameter vector, and extracted integral quantities should be introduced with a compact table or diagram to improve readability for readers outside the immediate subfield.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their thorough review and valuable suggestions. We have carefully considered the comments and revised the manuscript to address the concerns regarding surrogate validation and the agentic loop details.

read point-by-point responses

Referee: [Abstract] Abstract and results: The central claim that the framework discovers genuine regime-dependent scaling laws rests on the unverified assumption that LFM-generated flow fields yield integral quantities (displacement thickness, momentum thickness) sufficiently close to high-fidelity CFD solutions. No quantitative error metrics, validation plots, or side-by-side comparisons are reported for the 1,600+ evaluations, especially across the near-wake to co-shedding transition where the dual-extrema structure is claimed to emerge.

Authors: We agree that explicit validation of the LFM surrogate is crucial for supporting the claims. While the LFM training and validation details are provided in the Methods section, we did not include aggregate error metrics for the specific integral quantities across the explored parameter space. In the revised manuscript, we have added a dedicated validation subsection with quantitative error metrics (MAE and relative errors for displacement and momentum thickness), validation plots comparing LFM predictions to CFD at representative points including the transition regime, and a table summarizing errors for the 1,600 evaluations. This establishes that the surrogate errors are sufficiently low to support the discovered scaling laws. revision: yes
Referee: [Methods] Methods and results: The hierarchical agent architecture is described at a high level, but the manuscript provides no concrete account of how verification steps guard against surrogate-induced artifacts or post-hoc interpretation when extracting scaling laws. A worked example showing the full loop (hypothesis, LFM query, analysis, verification against an independent check) for one reported law would be required to support the claim of scientifically valid insights.

Authors: We acknowledge the need for a more concrete illustration of the verification process. The manuscript describes the closed-loop process at a high level, but to address this, we have included a new worked example in the revised Methods section. This example traces the discovery of the regime-dependent two-mode structure for minimum displacement thickness, detailing the agent's hypothesis generation, specific LFM queries for parameter sweeps, the analysis steps to identify the scaling, and the independent verification using additional high-fidelity CFD simulations at selected points to confirm the dual-extrema structure and rule out surrogate artifacts. revision: yes

Circularity Check

0 steps flagged

No circularity in derivation chain; scaling laws emerge from surrogate queries

full rationale

The paper trains an LFM surrogate on parameterized flow simulations to enable cheap queries, then deploys a hierarchical agent loop that autonomously samples over 1600 parameter-location pairs and extracts scaling laws (two-mode structure for min displacement thickness, linear scaling for max momentum thickness, dual-extrema at regime transition) directly from those surrogate outputs. No equation or step reduces the reported scaling laws to fitted parameters by construction, nor renames a known result, nor imports a uniqueness theorem via self-citation. The derivation chain is therefore self-contained: the LFM provides an independent generative approximation whose integral quantities are queried and analyzed, without the target structures being presupposed in the model definition or training objective.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Based on abstract only; no specific free parameters, axioms, or invented entities detailed beyond the implicit assumption that the LFM serves as a faithful surrogate. The main unstated premise is model fidelity for arbitrary parameters.

axioms (1)

domain assumption The latent foundation model accurately represents the underlying PDE solutions for arbitrary queried parameters and boundary conditions.
This is required for the LFM to function as an on-demand surrogate simulator in the agentic loop.

pith-pipeline@v0.9.0 · 5602 in / 1368 out tokens · 38667 ms · 2026-05-15T18:34:27.527720+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.

The LFM consists of two components: a total-correlation variational autoencoder (TC-VAE) that learns a compact, disentangled latent representation of the flow-fields and a parameter-conditioned diffusion model that enables generative sampling across the parameter space P.
IndisputableMonolith/Foundation/AlphaCoordinateFixation.lean J_uniquely_calibrated_via_higher_derivative unclear

?

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

the agent discovers qualitatively divergent scaling laws: minimum displacement thickness follows a regime-dependent two-mode structure... maximum momentum thickness follows a single linear law

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.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

AeroJEPA: Learning Semantic Latent Representations for Scalable 3D Aerodynamic Field Modeling
cs.LG 2026-05 unverdicted novelty 6.0

AeroJEPA applies joint-embedding predictive learning to produce scalable, semantically organized latent representations for 3D aerodynamic fields that support both field reconstruction and downstream design tasks.

Reference graph

Works this paper leans on

8 extracted references · 8 canonical work pages · cited by 1 Pith paper

[1]

URLhttps://doi.org/10.1038/s41586-023-06924-6

doi: 10.1038/s41586-023-06924-6. URLhttps://www.nature.com/articles/ s41586-023-06924-6. Timo Schick, Jane Dwivedi-Yu, Roberto Dessi, Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language Models Can Teach Themselves to Use Tools. InAdvances in Neural Information Processing Systems, Novem- ...

work page doi:10.1038/s41586-023-06924-6 2023
[2]

URLhttps://www.nature.com/articles/ s43588-022-00264-7

doi: 10.1038/s43588-022-00264-7. URLhttps://www.nature.com/articles/ s43588-022-00264-7. Ricardo Vinuesa, Steven L. Brunton, and Gianmarco Mengaldo. Explainable AI: Learning from the Learners, January 2026. URLhttp://arxiv.org/abs/2601.05525. arXiv:2601.05525. Alexius Wadell, Anoushka Bhutani, Victor Azumah, Austin R. Ellis-Mohr, Celia Kelly, Hancheng Zha...

work page doi:10.1038/s43588-022-00264-7 2026
[3]

all reported best values(S ∗, x∗ p)are computed from the CSV via tool calls,

work page
[4]

termination decisions are made only after verifying convergence thresholds on the primary metric,

work page
[5]

This design ensures that the Critic cannot “accept” fabricated values produced by Planner reasoning

refinement suggestions are justified by gaps in sampling density or detected discontinuities inx ∗ p(S). This design ensures that the Critic cannot “accept” fabricated values produced by Planner reasoning. If evidence is absent from the CSV , the Critic must returnneeds refinement, forcing further evaluation. B.7 WRITER ANDEVIDENCEPACKAGING The Writer age...

work page
[6]

ingestion of the final CSV and extraction of the bestx ∗ p(S)per spacing,

work page
[7]

generation of all discovery figures (scaling laws, minimum-thickness curves and(S, x p) heatmaps),

work page
[8]

By generating tables and plots programmatically from the stored CSV , the Writer is prevented from introducing numerical values that are not supported by evidence

creation of a Markdown report containing the figures and CSV-derived tables. By generating tables and plots programmatically from the stored CSV , the Writer is prevented from introducing numerical values that are not supported by evidence. This final report step functions as an automated supplementary-material generator, producing a complete trace from t...

work page