🤖 machine learning

An Agentic AI Scientific Community for Automated Neural Operator Discovery

This paper introduces an agentic AI scientific community comprising a swarm of interacting virtual laboratories that autonomously discover high-accuracy, low-parameter neural operator architectures through a citation-based economy, demonstrating that LLM-driven agency is essential for maintaining architectural diversity and that no single universal neural operator architecture exists across diverse scientific problems.

Original authors: Luis Loo, Ulisses Braga-Neto

Published 2026-07-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Luis Loo, Ulisses Braga-Neto

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a bustling, futuristic city of science where every building is a Virtual Laboratory. But these aren't just quiet rooms with scientists; they are alive with three distinct characters working together: a Dreamer (a super-smart AI planner), a Builder (a tireless robot that actually does the math), and a Critic (a sharp-eyed AI judge).

This city isn't run by a mayor. Instead, it runs on a citation economy. Think of it like a popularity contest where the most "cited" (highly recommended) labs get to clone themselves and spawn new labs that follow their ideas. The labs that fail to impress? They get evicted and replaced. The goal? To automatically discover the perfect recipe for solving complex physics puzzles, without any human telling them what to do.

The Great Experiment: Finding the Perfect Recipe

The researchers set up this AI city to solve five different types of physics problems, ranging from simple curve-fitting to simulating how air flows over a wing (Navier-Stokes) or how water moves through soil (Darcy flow).

They ran the simulation three times for each problem, using 16 virtual labs per run, for 20 rounds of evolution. In total, the AI made 9,623 calls to its "Dreamer" and "Critic" brains.

What did they find?
The AI city was incredibly successful at discovering high-accuracy, low-parameter-count neural operator architectures. But here is the twist: There is no single "best" recipe.

Just like you wouldn't use a hammer to fix a watch or a screwdriver to drive a nail, the AI discovered that different problems need different architectural "recipes."

For problems involving smooth, wave-like patterns (like the Navier-Stokes and Burgers equations), the winning labs tended to stack Fourier blocks (think of them as musical notes that capture smooth waves).
For problems with sharp jumps or complex shapes (like Darcy flow or piecewise regression), the winners preferred Wavelet or DeepONet blocks (think of these as tools for handling jagged edges).

In fact, the AI found that a "one-size-fits-all" approach doesn't work. A recipe that won the Navier-Stokes race performed terribly on the Darcy problem, and vice versa. The paper suggests a "no-free-lunch" theorem: there is no universal winner that dominates every single problem.

The Secret Sauce: Why the AI Dreamers Matter

The researchers wanted to know: Is the "Dreamer" (the Large Language Model) actually necessary, or could a simple rule-based robot do the same job?

To find out, they ran a "what-if" scenario. They replaced the AI Dreamers and Critics with rule-based agents (simple, pre-programmed robots that follow strict instructions).

The Result?

With AI Dreamers: The labs kept mixing and matching different building blocks. In 99.8% of their decisions, the AI Dreamers chose to hybridize (mix different families of blocks together). They created unique, multi-family hybrids that were efficient and accurate.
With Rule-Based Robots: The community collapsed. The robots stopped mixing. They stuck to single-family stacks. For example, on the Navier-Stokes problem, they just built pure Fourier towers. On Darcy problems, they built pure Wavelet towers.

The paper suggests that the AI agency is needed to preserve diversity. Without the "Dreamer's" creative intuition, the community gets stuck in a rut, repeating the same old patterns instead of inventing new, hybrid solutions.

The Numbers Behind the Magic

The paper is very specific about the performance and the cost:

Accuracy: The AI-discovered hybrid for the Navier-Stokes problem reached a relative L2 error of 0.0005, which is on par with the best existing method (FNO) but used fewer parameters. The hybrid for Darcy flow reached an error of 0.0631, statistically tied with the best existing method (DeepONet).
Efficiency: The AI hybrids were leaner. For the Navier-Stokes problem, the AI hybrid used 2.2 million parameters, while the rule-based version used 5.6 million (more than double!) to get similar accuracy.
Cost: Using the AI Dreamers did cost more time. Because the AI had to "think" and "review" before building, each lab evaluation took longer. For 1D problems, it took about 19.4 seconds per lab with AI, compared to 4.9 seconds for the rule-based robots. For 2D problems, it jumped to 51.3 seconds vs 18.7 seconds. However, the researchers note this is a small price to pay for finding better, more diverse designs.

What This Means (and What It Doesn't)

The paper suggests that this "AI Scientific Community" is a powerful way to discover new architectures automatically. It rules out the idea that a single, universal neural operator exists for all problems. It also rules out the idea that simple, rule-based automation is enough to maintain the creative diversity needed for these discoveries.

However, the paper is careful to note that these results come from simulations. They haven't proven that this will work for every possible physics problem in the real world, only the five they tested. They also used a specific local AI model (gemma3:12b) and didn't test bigger models, so we don't know if a "super-smart" AI would change the results.

In short, the paper shows that if you give a swarm of AI labs the freedom to mix, match, and critique each other, they can invent better, more efficient tools for solving physics puzzles than if you just told them to follow a strict rulebook. But they still need a human to set up the city and check the work!

Technical Summary: An Agentic AI Scientific Community for Automated Neural Operator Discovery

Problem Statement

Operator learning aims to build fast surrogates for Partial Differential Equation (PDE) solvers. However, existing neural operator families (e.g., Fourier Neural Operators, DeepONets, Wavelet operators) encode fixed inductive biases suited to specific regimes. Furthermore, selecting architecture hyperparameters (depth, width, block composition) for new PDEs remains a manual process. The central question addressed is whether an automated, agentic simulation of a research community can discover the optimal neural operator architecture for a given problem without human prior knowledge or intervention.

Methodology

The AI Scientific Community Framework

The authors propose a decentralized swarm of $N=16$ virtual laboratories that explore the architecture search space over $T=20$ iterations. The system operates under a "citation-based economy of influence":

Swarm Dynamics: Well-cited labs (those with high fitness) gain influence and breeding priority, while uncited labs are replaced. This mimics real-world scientific communities where successful research directions attract followers.
Virtual Lab Structure: Each lab contains three distinct agents:
- LLM Planner: Proposes the next architecture genome based on its history, the global best, and a summary of peer labs.
- Numerical Worker: A non-LLM, gradient-based trainer that compiles the proposed architecture, trains it at "screening fidelity" (reduced data/epochs for speed), and measures the relative $L_2$ error.
- LLM Reviewer: Participates in cross-lab peer review, casting soft votes based on candidate performance and architectural novelty.
Search Space (Genome): Architectures are defined as sequences of blocks ( $L \in [2, 8]$ ) drawn from a vocabulary: $\{ \text{BRANCH\_TRUNK, FOURIER, ATTENTION, WAVELET, RESIDUAL\_CONV} \}$ . The genome also includes hyperparameters like hidden width, spectral modes, and learning rates.
Fitness Function: A composite score $F_i$ combines accuracy ( $A_i$ ), robustness to noise ( $G_i$ ), parameter efficiency ( $E_i$ ), and architectural novelty ( $\nu_i$ ). Accuracy is heavily weighted ( $w_a=0.70$ ), but novelty and efficiency prevent the collapse to trivial solutions.

LLM Ablation Study

To isolate the contribution of LLM agency, the authors conducted a controlled ablation study. In the "rule-based" condition, the LLM planner and reviewer were replaced by deterministic algorithms:

Rule-based Planner: Uses standard particle-swarm optimization (explore/exploit) updates.
Rule-based Reviewer: Uses accuracy-proportional voting.
Both conditions share identical swarm sizes, iteration counts, training budgets, and random seeds.

Agent Auditing

To ensure the system did not silently degrade to rule-based fallbacks (a known risk in agentic systems), every LLM call (prompt, response, parse status) was logged and audited. The system uses a local gemma3:12b model via Ollama to ensure reproducibility without paid APIs.

Key Results

1. Discovery of Multi-Family Hybrids

The LLM-driven community consistently discovered multi-family hybrid architectures.

Hybridization Rate: In 99.8% of planner decisions, the LLM chose to "HYBRIDIZE" (graft blocks from different families) rather than purely explore or exploit a single family.
Regime Appropriateness: The discovered architectures adapted to the problem physics. For spectral-friendly problems (Advection, Burgers, Navier-Stokes), Fourier blocks led the stacks. For piecewise regression and Darcy flow, Wavelet or DeepONet-style blocks were prominent.
Performance:
- Navier-Stokes (2D): The discovered hybrid achieved $0.0005 \pm 0.0000$ relative $L_2$ error, matching the state-of-the-art FNO ($0.0002$) but with fewer parameters (1.1M vs 1.2M for FNO).
- Darcy Flow (2D): The hybrid achieved $0.0631 \pm 0.0083$ error, statistically tied with DeepONet ($0.0649$) and significantly outperforming FNO ($0.3070$).

2. The "No-Free-Lunch" Theorem for Neural Operators

The results confirm that no single operator architecture dominates all problems.

The Navier-Stokes hybrid performed poorly on Darcy flow (error $\approx 0.71$ ), and the Darcy hybrid performed poorly on Navier-Stokes.
This validates the hypothesis that different PDE regimes require distinct inductive biases.

3. The Role of LLM Agency (Ablation Findings)

While the accuracy of the LLM and rule-based communities was statistically indistinguishable on the screening fidelity, the architectural outcomes differed significantly:

LLM Community: Maintained diversity, consistently producing parameter-lean multi-family hybrids.
Rule-Based Community: Tended to "collapse" into canonical, single-family stacks (e.g., pure Fourier for Navier-Stokes, pure Wavelet for Darcy).
Efficiency: On Navier-Stokes, the LLM hybrids achieved comparable accuracy to the rule-based pure-Fourier stacks but with 2.5× fewer parameters (2.2M vs 5.6M).
Conclusion: LLM agency is necessary to preserve architectural diversity and discover efficient hybrid solutions, whereas rule-based agents converge prematurely to standard families.

4. Computational Cost

The LLM-driven approach incurs an overhead of roughly 2–4× in wall-clock time compared to rule-based coordination (e.g., ~271 minutes vs ~103 minutes for Navier-Stokes). This cost is attributed to the LLM planning and review calls (two calls per lab per iteration). However, the use of "screening fidelity" (training in seconds rather than hours) makes the swarm approach feasible on a single GPU.

Significance and Claims

The paper claims three primary contributions:

A Novel Framework: An agentic AI scientific community that successfully automates neural operator discovery using a swarm of virtual labs with LLM planners and reviewers.
Evidence of LLM Necessity: An ablation study demonstrating that LLM agency is critical for maintaining architectural diversity and discovering efficient hybrid models, preventing the collapse to single-family solutions seen in rule-based systems.
Empirical Validation of No-Free-Lunch: A large-scale simulation confirming that no universal neural operator exists; the "best" architecture is problem-dependent, and automated communities can discover these problem-specific hybrids.

The authors explicitly state that the system does not claim to outperform human-designed baselines in raw accuracy (as the rule-based and LLM baselines were comparable in screening) but rather in architectural discovery and parameter efficiency. The work emphasizes that the LLM acts as a policy for decision-making, not as an autonomous tool executor, and that the "no-free-lunch" result is a rediscovery of known literature facts within a broader, automated context. All code, configurations, and complete LLM transcripts (9,623 calls) are released for reproducibility.