Local Consistency Does Not Guarantee Global Conservation: Auditing Zero-Shot Composition of Airway Flow Operators

arXiv:2610.02056 · eess.SY, cs.SY · Submitted 2026-10-01 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: Today's paper: "Local Consistency Does Not Guarantee Global Conservation".

Dev: Neural operators approximate Partial Differential Equation (PDE) solutions, but independently learned local operators need not form a consistent global simulator.

Rosa: First, who's behind it and why it matters.

Title and authors: Rosa: We've talked about how local operators don't guarantee global conservation in the paper "Local Consistency Does Not Guarantee Global Conservation: Auditing Zero-Shot Composition of Airway Flow Operators." Now, let's look at who put this out there.

Dev: The authors are Nichula Sathmith Wasalathilaka, Navodya Heshan Samarasinghe, Dhanujaya Suraweera, Kevin Dawson, Chinthaka Jacob Mervyn Parakrama Bandara Ekanayake, and Roshan Godaliyadda from the Department of Electrical and Electronic Engineering and Mechanical Engineering at the University of Peradeniya.

Taro: It’s interesting to see a team with both electrical and mechanical engineering backgrounds tackling this flow problem, suggesting they are thinking about the implementation side as well as the underlying physics.

Rosa: That makes sense, especially since this paper is deeply rooted in fluid dynamics simulation, but having expertise in electrical systems might suggest they're looking at how these operators interface with hardware later on.

Dev: They are certainly looking at that interface aspect because the whole point of using neural operators is to potentially speed up those high-fidelity simulations, and you need robust methods for integrating those outputs into a larger control loop.

Taro: I hope their work on autonomous systems means they have a solid grasp on how these flow fields translate into actionable data for decision-making in complex, real-world scenarios.

Rosa: I certainly hope so; the implications of this work could be significant for any field that relies on fast, accurate flow analysis, even if the simulation itself isn't perfectly converged globally yet.

Dev: If we can use these frozen operators for rapid inference, it opens up possibilities for things that need to react quickly to changing conditions in a physical setup.

The paper's summary: Rosa: To get into the actual substance of "Local Consistency Does Not Guarantee Global Conservation: Auditing Zero-Shot Composition of Airway Flow Operators," the paper explains that they study steady two-dimensional airway flow in idealized two-dimensional airway trees organized as Tube, bifurcation (Y2), and trifurcation (Y3) components.

Dev: They then mention that each component family gets an independent DeepONet, called a DeepONet Gfc, which maps geometry, inlet-speed scale, outlet resistances, and normalized local coordinates to two-dimensional velocity and pressure fields.

Taro: It sounds like they are treating the different parts of the airway tree as if they are completely separate entities initially before trying to link them up.

Rosa: That’s exactly right; they train these operators on four thousand eight hundred seventy-two primitive CFD cases using field supervision and auxiliary constraints like "divergence," "port-flux," "component-balance," and "port-pressure penalties."

Dev: Those auxiliary constraints are what help tune the local models to be good at their specific job, but they are still just local checks; they don't enforce a global rule across the whole system.

Taro: So, the paper is showing that even with those fine-tuned local constraints, you still end up with components that might not connect perfectly in a larger structure.

Rosa: That’s the central tension: how to use these powerful local tools without having to retrain them on the entire system every time you change the tree topology.

Dev: The paper sets up a validation process where they freeze their selected deployment before inspecting whole-tree CFD fields, without doing any tree training or iterative coupling, flux correction, or CFD-informed adjustments.

Taro: That freezing step is critical because it tests the hypothesis that local consistency can hold up on its own when put together in a larger structure without further training.

The paper's improvements: Rosa: Now, let's talk about what the authors suggest as improvements to this situation, beyond just trying to get better local diagnostics with those auxiliary objectives they mentioned earlier.

Dev: The key improvement they propose is moving towards methods that actively enforce conservative interface coupling, global projection, or iterative correction instead of relying only on local regularization.

Taro: It’s clear they are suggesting that the architecture needs a mechanism to manage the flow mismatch at the junctions, which is where most of these errors seem to cluster.

Rosa: They detail an algorithm called COMPOSETREE, which systematically audits the tree structure by calculating component-level metrics like "component-residual" and "interface-mismatch RMS."

Dev: This algorithm calculates a mismatch term 'me' for every edge in the tree, representing the parent–child flux mismatch, and then tries to find interface pressure offsets γc to minimize that mismatch.

Taro: So they are trying to solve the problem mathematically by quantifying exactly how much flow is leaking or misbehaving at each connection point before attempting a final assembly.

Rosa: The final step involves assembling the geometry using these calculated offsets and component data to get the global metrics like "external residual Rext = Pcrc − Pe me".

Dev: And they also calculate a normalized imbalance metric, ε ref mass which is defined as one hundred times the absolute value of Rext divided by Qref in. This gives them a clear way to see if their assembled structure is actually performing well globally.

Conclusion: Rosa: So, wrapping up what we've discussed about "Local Consistency Does Not Guarantee Global Conservation: Auditing Zero-Shot Composition of Airway Flow Operators," the main message is that while auxiliary objectives improve local audit scores, they don't guarantee improved global fields.

Dev: The paper concludes that when you compose these operators without conservative interface coupling or global projection, the mean tree velocity error increases by seven point five percent, and the external residual goes up from seventeen point four nine ± four point three eight percent to twenty-six point seven seven±three point six five percent.

Taro: That means we need to be careful not to mistake a locally accurate piece for a globally correct solution when you're trying to build something bigger and more complex than just summing up the parts.

Rosa: It really hammers home that local field accuracy doesn't automatically translate into a good global flow simulation; the paper "Local Consistency Does Not Guarantee Global Conservation: Auditing Zero-Shot Composition of Airway Flow Operators" shows us that for reliable composition, you need more structure than just local regularization.

Dev: For our listeners, the implications are that if we want to deploy these types of AI models in a real-time environment, we need to bake in those conservative coupling mechanisms or iterative correction steps into the design from the start.

Taro: My final thought is that this paper emphasizes that for any complex system, ensuring local accuracy doesn't guarantee global conservation; you have to actively manage the interfaces if you want reliable results.

Rosa: That’s a solid summary of how these operators behave when composed in isolation, and I think we're ready to take a quick break before we move on to what else is out there.

Nichula Sathmith Wasalathilaka, Navodya Heshan Samarasinghe, Dhanujaya Suraweera, Kevin Dawson, Chinthaka Jacob, Mervyn Parakrama Bandara Ekanayake, Roshan Godaliyadda

Department of Electrical and Electronic Engineering, University of Peradeniya · Department of Mechanical Engineering, University of Peradeniya

eess.SY, cs.SY

Submitted: 2026-10-01

Updated: 2026-10-01

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 70/100

The gist: Neural operators approximate Partial Differential Equation (PDE) solutions, but independently learned local operators need not form a consistent global simulator.

Key concepts

DeepONet
A type of neural operator used here to map complex physical inputs (like geometry and flow scales) directly to the resulting fluid fields (velocity and pressure). These models are trained independently for different parts of the airway tree.
COMPOSE algorithm
A systematic auditing protocol used to check the assembled airway structure. It calculates component-level errors, measures mismatches between connected parts, and attempts to adjust interface pressures before calculating global metrics like external residual.
External Residual (Rext)
A diagnostic metric calculated during the assembly audit that measures only the external boundary flux budget. The paper found this value is condition-invariant across different flow types and does not signal internal errors within the tree structure.

Terminology

Summary

Neural operators approximate Partial Differential Equation (PDE) solutions, but independently learned local operators need not form a consistent global simulator. This study investigates whether frozen, single-pass composition of these local operators into an unseen larger system can guarantee global conservation or accuracy without full-domain retraining. The core finding is that while auxiliary objectives improve primitive and assembled local diagnostics, they do not ensure improved whole-tree field error or external conservation when composition occurs without iterative exchange or global projection.

The gist

Frozen primitive operators are useful diagnostics, but reliable composition requires conservative interface coupling, global projection, or iterative correction rather than local regularization alone.

How it works

  1. The study uses steady two-dimensional airway flow in idealized two-dimensional airway trees organized as Tube, bifurcation (Y2), and trifurcation (Y3) components.

  2. Each component family receives an independent DeepONet, denoted as a DeepONet Gfc, which maps geometry, inlet-speed scale, outlet resistances, and normalized local coordinates to two-dimensional velocity and pressure fields. These operators are trained on 4,872 primitive CFD cases using field supervision and auxiliary constraints like divergence, port-flux, component-balance, and port-pressure penalties.

  3. The validation process involves a crucial step: the selected deployment is frozen before whole-tree CFD fields are inspected, without tree training or iterative coupling.

  4. A post-hoc sensitivity protocol repeats the Data, Div, and Full models across three seeds while holding the Tube fixed to audit primitive consistency versus assembled component and interface discrepancies.

Composition Audit Protocol

The COMPOSE algorithm systematically audits the assembled tree structure:

  1. It first calculates component-level metrics for all components by evaluating them independently using their respective operators, yielding quantities like component-residual and interface-mismatch RMS.

  2. For every edge in the tree, it calculates a mismatch term, denoted as me, representing the parent–child flux mismatch.

  3. It then determines interface pressure offsets, denoted as γc, by minimizing a function involving these mismatches:

  4. The final assembly step involves calculating the global metrics: (ubg, pbg) ← Assemblegeom for all components in the tree using the calculated offsets and component data.

  5. Diagnostic metrics derived from this assembly include external residual Rext = Pcrc − Pe me and a normalized imbalance metric, ε ref mass = 100Rext/Qref in.

Key Findings on Consistency

The results demonstrate that improvements in the primitive audit score and divergence-MSE are achieved by training Full models over Data or Div models. However, when these improved operators are composed into the tree, mean tree velocity error increases by 7.5% and the external residual increases from 17.49 ± 4.38% to 26.77±3.65%. This indicates that local regularization alone does not guarantee accurate global fields or conservation; local field accuracy ≠ globally consistent composition.

Conservation Decomposition

A secondary post-hoc audit, focusing on the Fixed-Tube and branching-only conservation attribution, reveals how errors decompose:

  1. The external boundary residual is condition-invariant across normal, stenosis, and dilation conditions.

  2. The Component RMS diagnoses component-level residual magnitude (e.g., Tube component RMS).

  3. The Branch–branch edge RMS diagnoses internal connection mismatch between components (e.g., Branch–branch interface RMS).

  4. The external boundary residual (Rext) diagnoses only the external boundary-flux budget; it does not localize internal error, and it contains no pathology signal because the external-port vectors do not change across conditions.

Conclusion on Composition

The study concludes that auxiliary objectives substantially improve primitive audit scores and local branching conservation metrics, but Full reduces assembled component-residual and interface-mismatch RMS by 7.2% and 16.6%, while simultaneously increasing the external residual from 17.49% to 26.77%. Therefore, reliable composition requires conservative interface coupling, global projection, or iterative correction rather than local regularization alone. The final audit shows that improved primitive and assembled-local diagnostics do not guarantee improved global fields or conservation.

Limitations

The non-clinical study treats steady laminar flow in idealized 2D rigid-wall airways. The external residual telescopes to the external-port budget and cannot localize internal error. Furthermore, the assembler enforces neither interface-flux continuity nor global balance. Unmatched hardware precludes a normalized speedup claim, and the results are descriptive rather than inferential due to limited statistical replicates. Pathology conditions are not independent replicates.

Improvements for AI systems

Here are the specific improvements and capabilities an AI system could gain by implementing the concepts from this research:

The core improvement is shifting from training monolithic, full-domain surrogate models to a robust, auditable framework of independently learned, frozen local operators that are composed through a controlled audit mechanism.

The improved AI system will possess the capability to perform high-fidelity flow simulations on complex airway trees with verifiable guarantees regarding local accuracy and component conservation, even when the global composition is not yet converged.

Here are the specific improvements:

  1. A system for creating and deploying reusable, family-specific neural operators (DeepONets) trained on primitive CFD cases (Tube, Y2 bifurcation, Y3 trifurcation).

  2. A mechanism to freeze these local operators after training and use them in a zero-shot composition pipeline to simulate large airway trees without retraining or iterative coupling.

  3. A diagnostic suite capable of auditing the frozen composition by calculating specific metrics:

  4. Individual component residuals (Branch-component RMS).

  5. Interface mismatch errors (Branch–branch edge RMS).

  6. External boundary flux budget consistency (External boundary residual, which diagnoses only external port balance and does not localize internal error).

The improved AI system can do the following:

  1. Perform rapid, near-real-time flow simulations on complex airway geometries (like those in respiratory systems) with inference speeds of 0.204–0.215 seconds per component/segment, which is suitable for embedded or real-time applications.

  2. Provide a frozen prediction that retains major flow patterns and controlled pathology responses (e.g., changes in stenosis velocity or dilation pressure) based on the primitive validation data, even if the global conservation law is not perfectly satisfied across all junctions.

  3. Act as a powerful diagnostic tool: When deployed, it can instantly report whether errors are localized to a specific component (like a single branch operator) or distributed across interfaces and external boundaries, preventing false confidence in the global solution before iterative correction methods are applied.

  4. Serve as a robust initialization or pre-processing step for other solvers by providing high-fidelity local estimates of velocity and pressure fields based on geometry and local inlet/outlet conditions, without requiring a full-domain training run.

Abstract

Neural operators approximate PDE solutions within a geometry family, but independently learned local operators need not form a consistent global simulator. We study frozen, single-pass composition for steady incompressible flow in idealized two-dimensional airway trees. Separate Tube, bifurcation, and trifurcation DeepONets are trained on 4,872 primitive CFD cases using field supervision and auxiliary divergence, port-flux, component-balance, and port-pressure penalties. The validation-selected deployment is frozen before whole-tree CFD fields are inspected and assembled without tree training, iterative coupling, flux correction, or CFD-informed adjustment. It retains major flow patterns and controlled pathology responses with 0.204-0.215 s CPU inference, but has a 22.68% prescribed-inlet-normalized external residual. A post-hoc sensitivity protocol, frozen before new training and evaluation, repeats Data, Div, and Full models across three seeds with Tube fixed. Relative to Data, Full reduces primitive composite scores by 26.7% for Y2 and 26.4% for Y3 and reduces assembled component-residual and interface-mismatch RMS by 7.2% and 16.6%, respectively. Nevertheless, mean tree velocity error increases by 7.5%, while external residual increases from 17.49 +/- 4.38% to 26.77 +/- 3.65%. Local regularization can therefore improve primitive and assembled local diagnostics without ensuring accurate global fields or conservation.

Sources

Related papers