Agentic Exploration of PDE Spaces using Latent Foundation Models for Parameterized Simulations
TLDR
Couples multi-agent LLMs with latent foundation models to autonomously explore PDE parameter spaces, discovering scaling laws in fluid dynamics.
Reasoning
The paper presents a novel integration of LLM agents with generative surrogate models for efficient exploration of PDE spaces, demonstrating real-world application with flow past tandem cylinders. However, the approach is limited to a specific case and scalability to broader PDE problems is not fully addressed.
Read-first score
Read-first score 64, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 74.
Field roles
Rank sensitivity
Stability: volatile; rank range: 28.
Keyword Scores
Deep Analysis
Innovations
- Coupling multi-agent LLMs with latent foundation models (LFMs) for continuous exploration of PDE parameter spaces
- LFM as a generative model learning explicit, compact, disentangled latent representations of flow fields, serving as an on-demand surrogate simulator
- Hierarchical agent architecture with closed-loop hypothesis, experimentation, analysis, and verification, using a tool-modular interface
- Discovery of divergent scaling laws (two-mode structure for minimum displacement thickness, linear scaling for maximum momentum thickness) and dual-extrema structure at regime transition
Methodology
A latent foundation model (LFM) is trained on parametrized flow simulations to provide a low-cost surrogate. Multi-agent LLMs in a hierarchical architecture iteratively propose hypotheses, query the LFM for flow fields at chosen parameter-location pairs, analyze the results, and verify findings, enabling autonomous exploration of the PDE solution space.
Key Results
Autonomous evaluation of over 1,600 parameter-location pairs for flow past tandem cylinders at Re=500 revealed a regime-dependent two-mode scaling for minimum displacement thickness and a robust linear scaling for maximum momentum thickness, with both landscapes showing a dual-extrema structure emerging at the near-wake to co-shedding transition.
Limitations
- Demonstration limited to a single flow configuration (tandem cylinders at Re=500); generalizability to other PDE systems not shown.