🤖 AI

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

This paper introduces a framework that couples multi-agent large language models with latent foundation models to enable automated, large-scale exploration of continuous partial differential equation solution spaces, successfully discovering novel scaling laws in fluid dynamics without human intervention.

Original authors: Abhijeet Vishwasrao, Francisco Giral, Mahmoud Golestanian, Federica Tonti, Andrea Arroyo Ramo, Adrian Lozano-Duran, Steven L. Brunton, Sergio Hoyas, Soledad Le Clainche, Hector Gomez, Ricardo Vinuesa

Published 2026-04-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Abhijeet Vishwasrao, Francisco Giral, Mahmoud Golestanian, Federica Tonti, Andrea Arroyo Ramo, Adrian Lozano-Duran, Steven L. Brunton, Sergio Hoyas, Soledad Le Clainche, Hector Gomez, Ricardo Vinuesa

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 you are trying to understand the weather patterns of a new, mysterious planet. Traditionally, scientists would have to build a massive, expensive supercomputer for every single day they wanted to predict the weather. If they wanted to know what happens if the wind blows 1% faster, or if a mountain is 1 meter taller, they'd have to build a new computer simulation from scratch. It's slow, expensive, and limits how much they can explore.

This paper introduces a revolutionary new way to do science, acting like a team of AI detectives equipped with a super-powered crystal ball.

Here is the breakdown of how it works, using simple analogies:

1. The Problem: The "Expensive Lab" Trap

In the real world, studying fluids (like water or air) involves solving complex math equations called PDEs (Partial Differential Equations).

The Old Way: To study a fluid, you run a simulation. It takes hours or days. If you want to test 1,000 different scenarios, you have to wait 1,000 times that long. It's like trying to taste every flavor of ice cream by baking a whole new batch for every single spoonful.
The Limitation: Because it's so slow, humans can only test a few flavors. They miss the weird, interesting ones in between.

2. The Solution: The "Crystal Ball" (Latent Foundation Model)

The authors built a Latent Foundation Model (LFM). Think of this as a super-smart crystal ball that has "dreamed" about millions of fluid simulations.

Instead of memorizing every single simulation, it learned the essence or the "DNA" of how fluids move.
The Magic: Once trained, this crystal ball can instantly "dream up" (generate) a realistic fluid simulation for any scenario you ask for, even ones it has never seen before. It's like asking a chef who has tasted every dish in the world to instantly describe what a "spicy-sweet-savory" dish would taste like, without actually cooking it.
Cost: Asking the crystal ball takes a fraction of a second and costs almost nothing, compared to the hours needed for a real simulation.

3. The Team: The "AI Detective Squad" (Multi-Agent System)

The paper doesn't just use the crystal ball; it puts it in the hands of a team of AI agents, each with a specific job, working together like a research lab:

The Planner (The Strategist): This agent looks at the map and says, "We haven't checked the area where the wind is very fast yet. Let's go there." It decides where to look next.
The Analyst (The Explorer): This agent takes the Planner's orders, asks the Crystal Ball for the data, and measures the results. It's the one actually "touching" the fluid.
The Critic (The Quality Control): This agent double-checks the work. "Wait, does this simulation look physically possible? Did the AI hallucinate a floating cylinder?" If it looks fake, the Critic sends it back for a redo.
The Writer (The Reporter): Once the team is done, this agent writes the final report, explaining what was discovered.

4. The Experiment: Two Cylinders in a Stream

To test this, the team set up a classic physics problem: Two cylinders (like pipes) standing in a river.

The Question: How does the water flow between them change as you move the cylinders closer or further apart?
The Old Way: A human would pick 10 or 20 distances, run simulations, and guess the pattern.
The AI Way: The AI team autonomously tested 1,600 different distances and locations in a fraction of the time it would take a human.

5. The Discovery: Two Different Rules

The AI found something humans might have missed because they didn't test enough points. They discovered that the water behaves in two completely different ways depending on the distance:

The "Hugging" Rule (Close together): When the cylinders are close, the water gets stuck between them. The "thickest" part of the water flow stays right next to the first cylinder.
The "Dancing" Rule (Far apart): When they are far apart, the water starts to swirl independently around each cylinder. The "thickest" part of the flow suddenly jumps to a new spot, following the second cylinder.

The AI noticed a sharp "jump" or transition point where the rules change. It found that one measurement (how much water is missing) follows a complex, two-step rule, while another measurement (how much momentum is lost) follows a simple, straight-line rule.

Why This Matters

This paper proves that we can now automate scientific discovery.

No more manual guessing: The AI doesn't need a human to tell it what to look for. It figures out the strategy itself.
Infinite exploration: Because the "Crystal Ball" is so cheap to use, we can explore every corner of the parameter space, finding hidden patterns and "regime transitions" that are invisible to traditional methods.
The Future: This isn't just for water pipes. This framework could be used to design better airplane wings, discover new drugs, or optimize energy grids, all by letting AI agents explore the physics of the world on their own.

In short: The authors built a team of AI detectives with an instant-simulation crystal ball. They sent them to explore a fluid physics problem, and the team found complex, hidden patterns that would have taken humans years to uncover, proving that AI can now be a true partner in scientific discovery.

1. Problem Statement

Scientific discovery in fluid dynamics and other fields governed by Partial Differential Equations (PDEs) traditionally relies on systematic parameter studies involving expensive numerical simulations and manual interpretation. This approach faces two major bottlenecks:

Computational Cost: Exhaustive exploration of high-dimensional, continuous parameter spaces is prohibitively expensive.
Representation Mismatch: Unlike discrete domains (e.g., drug discovery with SMILES strings or materials science with crystal graphs), physical systems governed by PDEs produce continuous, high-dimensional, and chaotic spatiotemporal data. This makes them incompatible with Large Language Models (LLMs), which typically operate on tokenized, discrete inputs.
Passive Tools: Existing data-driven reduced-order models (ROMs) act as passive predictors within human-driven workflows, lacking the autonomy to decide what to simulate or how to synthesize findings.

The paper addresses the challenge of creating an autonomous, automated scientific discovery system capable of exploring continuous PDE parameter spaces without human intervention.

2. Methodology

The authors propose a framework coupling Multi-Agent LLMs with a Latent Foundation Model (LFM). This system creates a closed loop of hypothesis, experimentation, analysis, and verification.

A. Latent Foundation Model (LFM)

The LFM serves as an "on-demand surrogate simulator," replacing costly numerical solvers with a generative model that can sample flow fields for arbitrary parameters. It consists of two components:

Total-Correlation Variational Autoencoder (TC-VAE):
- Learns a compact, disentangled latent representation ( $z$ ) of high-dimensional flow fields ( $x$ ).
- Uses a $\beta$ -TC-VAE loss function to penalize statistical dependence between latent dimensions, ensuring that each dimension captures independent factors of variation from the parameter space.
Parameter-Conditioned Latent Diffusion Model:
- Operates in the latent space learned by the VAE.
- Takes input parameters $P$ (e.g., cylinder spacing) and noise levels to generate latent trajectories.
- Uses a transformer architecture with factorized spatiotemporal attention and rotary positional embeddings (RoPE) for temporal coherence.
- Capability: Enables continuous interpolation and moderate extrapolation across the parameter space, generating physically plausible flow fields at negligible computational cost.

B. Multi-Agent Exploration Framework

Built on LangGraph, the system employs a hierarchical agent architecture to orchestrate the exploration:

Planner: Formulates investigation strategies (e.g., coverage-based sampling or targeted refinement) based on the current knowledge state and dataset. It decides which regions of the parameter space to explore next.
Analyst: A deterministic execution node that interfaces with the LFM and physics-aware tools. It queries the LFM to generate flow fields, computes interpretable statistics (e.g., displacement thickness, momentum thickness), and compiles structured data.
Critic: Validates outputs for geometric fidelity (e.g., checking cylinder positions) and data consistency. It operates via a ReAct loop to assess coverage and smoothness of trends, flagging unreliable runs.
Writer: Synthesizes the final findings into a structured report, visualizations, and documentation.
Tool Interface: Agents access physics-based analysis tools (e.g., flow_generator, scan, evaluate_pair) as callable functions, allowing them to perform domain-specific analysis without explicit physical knowledge in their prompts.

3. Key Contributions

Bridging Continuous Physics and LLMs: The paper introduces the LFM as a critical intermediate representation, translating continuous, chaotic PDE solutions into a discrete, tokenizable latent space accessible to LLM reasoning.
Autonomous Scientific Discovery: Demonstrates a fully autonomous loop where agents can formulate hypotheses, execute "experiments" (via the LFM), analyze results, and refine strategies without human guidance.
Surrogate Simulator for Exploration: Establishes the LFM not just as a predictor, but as a generative engine that allows agents to query arbitrary parameter configurations at near-zero marginal cost, enabling dense sampling impossible with traditional solvers.
Tool-Modular Architecture: The framework decouples the agent reasoning from the analysis tools, allowing the system to be adapted to different physical quantities (e.g., vortex shedding frequencies, aerodynamic coefficients) simply by swapping tools.

4. Results

The framework was applied to incompressible flow past tandem circular cylinders at Reynolds number $Re = 500$. The agent autonomously explored cylinder spacing ( $S$ ) from $3.5D$ to $10D$ .

Scale of Exploration: The agent evaluated 1,600 parameter-location pairs across 41 spacing configurations, a scale impractical for manual workflows.
Discovery of Divergent Scaling Laws:
- Maximum Momentum Thickness ( $\theta$ ): The location of peak momentum deficit ( $x^*_\theta$ ) followed a robust single linear law ( $x^*_\theta \approx S - 1.2D$ ) across all spacings, tracking the downstream cylinder.
- Minimum Displacement Thickness ( $\delta^*$ ): The location of minimum mass deficit ( $x^*_{\delta^*}$ $x_{δ^{*}}^{*}$ ) exhibited a regime-dependent two-mode structure:
  - Near-wake regime ( $S \lesssim 6\text{--}7D$ ): $x^*_{\delta^*}$ remained fixed near $0.5D$ (downstream of the upstream cylinder).
  - Co-shedding regime ( $S \gtrsim 7D$ ): $x^*_{\delta^*}$ shifted downstream following a new linear trend ( $0.71S - 2.2D$ ).
Regime Transition Detection: The agent identified a sharp transition near $S \approx 7D$ . The divergence between the two optimal locations ( $|\Delta x^*|$ ) grew steadily in the near-wake regime, dropped sharply at the transition, and plateaued in the co-shedding regime. This non-monotonic signature served as a quantitative marker for the physical regime change.
Validation: The LFM-derived results (filled markers) showed high agreement with simulation ground truth (open markers), validating the surrogate's accuracy.

5. Significance

Paradigm Shift in PDE Exploration: This work moves beyond passive data analysis to active, agentic exploration. It proves that LLMs can drive scientific discovery in continuous physical systems when paired with learned physical representations.
Efficiency and Scalability: By using the LFM as a surrogate, the system bypasses the computational bottleneck of running millions of CFD simulations, making large-scale parameter sweeps feasible.
Generalizability: The framework is not limited to fluid dynamics; it offers a general paradigm for any PDE-governed system where an LFM can be trained. Future work could extend this to 3D flows, multi-parameter spaces, and causal mechanism discovery.
Engineering Impact: The autonomous discovery of regime-dependent scaling laws provides immediate engineering insights for designing flow control systems (e.g., drag reduction vs. flow blockage reduction) that require different strategies depending on the flow regime.