PaperlayerView exact revision on arXiv

Source-linked AI summary

Self-Evolving Scientific Agent Designs Physically Reasoned White-Box Fluid Control

Boai Sun, Wenjin Guo, Zongmin Yu, Liu Yang

arXiv:2606.08405v4cs.AIphysics.flu-dyn

TL;DR

The work addresses designing inspectable white-box controllers for fluid navigation as an alternative to black-box neural-network control. Self-evolving agents refine controller code using simulation evidence, and the resulting fixed controllers generalize across the tested flow variations while 2D control priors accelerate 3D adaptation.

  • Problem

    The work addresses designing inspectable white-box controllers for fluid navigation beyond black-box neural-network control.

    Source: Discussion · pp. 11–12

  • Method

    Self-evolving agents revise explicit controller code using rollout evidence and physical priors, while a fixed CFD evaluator scores each candidate.

    Source: Self-evolving agents for white-box… · p. 3

  • Results

    The retained champion reached targets across the full test matrix without retuning, and 2D-controller initialization reached a high-score regime earlier than minimal-seed initialization.

    Source: Discussion · pp. 12–13

  • Takeaways & Limitations

    Design-time reasoning is separated from standalone controller execution, which requires no language-model or neural-network inference.

    Source: Discussion · p. 12

    The present results are limited to simulation, and controller evaluation relies entirely on high-fidelity CFD simulations.

    Source: Discussion · p. 13

Abstract

from arXiv · show

While neural networks excel in autonomous control, their black-box nature makes control decisions difficult to interpret and diagnose in dynamic fluids. Here, we show how self-evolving scientific agents can design explicit, neural-network-free white-box controllers by iteratively interpreting simulation evidence, accumulating control knowledge and refining controller code. We demonstrate this approach on an underactuated two-joint swimmer navigating unsteady flows via joint angular accelerations. Starting from a target-blind propulsive controller, the agent gradually constructs key mechanisms, including travelling-wave propulsion, body-frame guidance, phase-selective steering, redirect bursts and adaptive relief. The resulting controllers reach targets and generalize across changes in target position, wake geometry, cylinder count, and inflow speed without revision. Moreover, 2D control priors transfer successfully to accelerate 3D adaptation. Our work demonstrates that self-evolving agents can autonomously design physically reasoned and generalizable white-box fluid control, showing a promising paradigm beyond traditional reinforcement learning and black-box neural network control.

Results

Self-evolving agents iteratively revise executable white-box controllers using simulation evidence and physical priors, while the fixed CFD evaluator scores each candidate. The resulting controllers navigate unsteady wakes, generalize across tested flow changes, and provide a basis for continued evolution in three dimensions.

Self-evolving agents for white-box fluid control

  • Agents co-evolve design agents and executable controller code, testing each candidate through a fixed CFD evaluator that returns scores and multimodal evidence.The controller runs as explicit code; the design-time agent is not required during deployment.Source: Self-evolving agents for white-box… · p. 3

Autonomous design and generalization in two-dimensional flow

  • The fixed champion reached the target across the full tested matrix of target, wake-geometry, cylinder-count and inflow-speed changes without revision; frozen DRL policies performed poorly under the same perturbations.Other retained policies reached the target in most tests, and the champion's successful responses were closed-loop rather than case-specific.Source: Autonomous design and generalization… · p. 6, Figure 3 · p. 7

Evolution of the white-box control strategy

  • The controller retained the Seed’s travelling body wave and added target-bearing and body-motion feedback, later incorporating phase-selective steering, redirect bursts and load-conditioned relief.By Iteration 19, source lines, parameters and feedback signals grew to 7.3, 4.2 and 3.5 times their seed values, without adding actuation degrees of freedom.Source: Evolution of the white-box control… · p. 8

Physical mechanisms underlying adaptive wake navigation

  • During redirection, the travelling deformation persisted as wake- and posture-dependent pressure loads acted at different body moment arms, linking target-guided commands to turning through fluid loading.The three-wake-state tests also showed target capture without phase-specific retuning.Source: Physical mechanisms underlying… · p. 9
  • Matched-state counterfactuals identify bearing guidance, half-cycle steering asymmetry and redirect bursts as a hierarchy that turns bearing error into phase-selective deformation and yaw.The isolated half-cycle addition made capture roughly one-fifth faster (19.6%); the burst further accelerated an already reliable turn.Source: Physical mechanisms underlying… · p. 9

Transferring two-dimensional control priors to three dimensions

  • Across three independent experiments per condition, initializing 3D evolution from a retained 2D policy reached a high-score regime earlier, produced more consistent outcomes and attained a higher observed score ceiling than starting from the minimal seed.The 3D task added a finite span, spanwise flow structures and a deforming three-dimensional fluid–body interface.Source: Transferring two-dimensional control… · p. 11

Discussion

The study demonstrates self-evolving agents designing physically reasoned, inspectable white-box fluid controllers, with explicit code that runs independently of the design-time agent. Frozen controllers generalize across tested flow changes, and retained 2D control structures support faster, more consistent 3D adaptation; real-world performance remains unestablished.

Discussion

  • Self-evolving agents designed physically reasoned, inspectable white-box controllers, with design-time reasoning separated from standalone run-time execution.Source: Discussion · pp. 11–12
  • The retained champion reached targets across the full test matrix without retuning, while frozen DRL policies showed poor target-capture performance under the same perturbations.Source: Discussion · p. 12
  • The physical roles of individual control structures are testable, but which structures account for robustness across different environments remains an open question.Source: Discussion · p. 12
  • Across three experiments per condition, 2D-policy initialization reached a high-score regime earlier, with more consistent outcomes and a higher observed score ceiling than minimal-seed initialization.Source: Figure 5 · p. 12, Discussion · p. 13
  • The study relies entirely on high-fidelity CFD simulations; real-world performance and simulation-to-real transfer remain to be established.Source: Discussion · p. 13

Methods

EvE evolves explicit controller code through population-based design and fixed CFD evaluation, retaining simulation evidence and design records. The methods specify the underactuated swimmer and simulation, a behavior-cloned PPO comparator, and the use of generative AI in manuscript preparation.

Implementation of self-evolving agents for fluid control

  • EvE combines editable controller code, a fixed CFD evaluator, rollout evidence and a Knowledge Shelf to evolve standalone white-box controllers.Policies, scores, logs and agent design records are retained, connecting controller revisions to physical outcomes.Source: Implementation of self-evolving agents… · pp. 13, 15

Fluid control simulation

  • The swimmer’s only control inputs are two joint angular accelerations; propulsion and steering emerge from body deformation and hydrodynamic loads.The search specifies no gait or waveform, and policies may use subsets of a rich state input including target geometry, loads, velocity, joint state and local flow.Source: Fluid control simulation · pp. 16, 18, 19

Behavior-cloned PPO comparator

  • BC–PPO was trained in the nominal two-dimensional CFD–FSI task after behavior cloning a target-blind seed controller that supplied propulsion without steering.The comparator used the shared observation and actuation limits; high-scoring policies were tested without retraining or case-specific tuning.Source: Behavior-cloned PPO comparator · pp. 19–20

AI use declaration

  • The methods describe the model, execution settings and instructions for language-model agents in Supplementary Section B.Generative AI also assisted manuscript revision; Codex assisted with figure-script modifications under author direction, and the authors retain responsibility for interpretation and verification.Source: AI use declaration · p. 20

Funding

The study was supported by the National Research Foundation, Singapore, through an NRF Fellowship to L.Y.

Funding

  • The National Research Foundation, Singapore, supported the work through L.Y.’s NRF Fellowship.Project No. NRF-NRFF17-2025-0006.Source: Funding · p. 20

a Two-joint deformable swimmer b CFD–body integration

Extended Data Fig. 1 depicts the deformable swimmer and its CFD–body integration loop.

a Two-joint deformable swimmer b CFD–body integration

  • The figure shows a fixed thickness profile around a smooth two-joint spine, with joint commands updating the moving boundary and fluid loads driving translation and yaw.The updated pose closes the coupling loop.Source: a Two-joint deformable swimmer b… · p. 25

A CFD verification and validation

The numerical components were assessed through external moving-boundary validations in two and three dimensions and a three-dimensional moving-window implementation check.

A CFD verification and validation

  • The CFD components were evaluated using two-dimensional moving-boundary validation, three-dimensional moving-boundary and power-integration validation, and a check of the 3D moving-window algorithm.Source: A CFD verification and validation · p. 26

A.1 Two-dimensional wake-transition validation

The two-dimensional solver reproduced a reported wake-transition boundary within 10% at Nc = 64, while thrust and propulsive efficiency converged as resolution increased.

A.1 Two-dimensional wake-transition validation

  • At Nc = 64, the predicted BvK-to-reversed-BvK transition across seven Strouhal numbers lay within 10% of the reported wake-neutral boundary.The comparison used the pure-pitch, leading-edge-pivot NACA 0016 benchmark at ReC = 1173.Source: §A.1 · p. 26
  • For the foil at Sr = 0.25 and AD = 0.874, cycle-mean thrust coefficient and propulsive efficiency converged as resolution increased from Nc = 16 to 128.Nc = 64 was selected for two-dimensional calculations to balance numerical accuracy and computational cost.Source: §A.1 · p. 26
  • Supplementary Figure 1 compares the predicted transition boundary at Nc = 64 with the reported boundary and shows mean-thrust and propulsive-efficiency convergence with resolution.The transition plot shows AD ± 0.05 bars.Source: §A.1 · p. 27

A.2 Three-dimensional oscillating-cylinder validation

The three-dimensional solver’s mean signed power approached published and experimental references as resolution increased; deviations from both were below 10% at n = 64.

A.2 Three-dimensional oscillating-cylinder validation

  • The validation compared mean signed power across nine resolutions with the reported numerical value −4.39 and experimental value −4.525.The oscillating-cylinder benchmark used ReD = 7620, Vr = 5.4, Ax/D = 0.4, Ay/D = 1.6 and θ = π/6.Source: §A.2 · p. 27
  • At n = 64, mean signed power deviations from both the published and experimental references were below 10%.With increasing resolution, mean signed power approached both values and remained close across finer grids.Source: §A.2 · p. 27
  • The production calculations used n = 64, matching the nominal resolution parameter used in the two-dimensional study.The choice followed the reported convergence and accuracy-cost assessment.Source: §A.2 · pp. 27–28

A.3 Three-dimensional moving-window implementation verification

Across five matched three-dimensional releases, moving-window simulations preserved controlled swimming dynamics while reducing active cells and per-step wall time substantially.

A.3 Three-dimensional moving-window implementation verification

  • The moving window contained approximately 4.72 × 10^6 cells, 3.13% of the fixed full field’s approximately 1.51×10^8 cells, reducing active cell count by 96.9%.The window measured 4L×3L×1.5L within a fixed 24L×16L×1.5L domain at L = 64.Source: §A.3 · p. 28
  • The five-release comparisons matched initial conditions and frozen policies, varied only domain treatment, and ran for 10T per pair.Moving-window cases underwent 69–86 integer-cell translations.Source: §A.3 · p. 28
  • Across five releases, mean RMS differences were 0.0183L for centre position, 9.95×10−3 rad (0.570°) for heading, and 0.0100U for velocity.Maximum centre-position and heading differences were 0.0985L and 5.06 × 10−2 rad (2.90°), respectively; the fixed full field used 20.8 times the wall time per solver step on average.Source: §A.3 · p. 29
  • The authors conclude that closed-loop state feedback makes the trajectory separation compatible with robust feedback, and that the moving window preserves relevant swimming dynamics without materially affecting control-capability assessment.The reported validation attributes the reduced cost to 32-fold fewer active cells and approximately 21-fold less wall time per step.Source: §A.3 · p. 29

B Self-evolving agents: implementation, prompts and instructions

The implementation instructions require agents to consult mechanism references at specified points, translate transferable invariants into the controller contract, and make evidence-backed policy changes.

B Self-evolving agents: implementation, prompts and instructions

  • Before the first architecture proposal, workers must consult the fish-control shelf; on later iterations, consultation is required after three completed iterations without a new mechanism or semantic improvement.Adopting a shelf primitive remains optional, and it cannot justify scalar-only gain tuning.Source: B Self-evolving agents:… · p. 30
  • Workers must produce one improved fixed-morphology policy for the four-cylinder wake without editing the morphology, solver, episode, geometry or scoring code.Each worker produces exactly one candidate and preserves the public target-policy contract.Source: B Self-evolving agents:… · p. 30
  • The shelf’s transferable object is an explanatory invariant translated into normalized body-frame observations, two-joint actuation and physical constraints, not a published gain or task-specific route.The instructions treat references as hypotheses and require current keyframes and metrics to be read first.Source: B Self-evolving agents:… · pp. 30–31

C Design lineage of the evolved controller

The controller evolved from a target-blind travelling-wave oscillator into a swimmer that steers using body-frame guidance and phase-selective authority. Later iterations added redirect bursts and response-based relief, culminating in force-confirmed wake assistance.

Related papers

Loading 2606.08405v4…