Agentic AI Explores PDE Spaces for Scientific Discovery

The integration of multi-agent Large Language Models (LLMs) with Latent Foundation Models (LFMs) marks a significant advancement in automated scientific discovery, particularly for systems governed by Partial Differential Equations (PDEs). This methodology addresses the inherent challenges of exploring continuous, high-dimensional, and often chaotic physical phenomena, which have historically been constrained by the computational expense of numerical simulations or the practical limitations of laboratory experiments. By creating explicit, compact, and disentangled latent representations of flow fields, the LFM acts as an on-demand surrogate simulator, allowing AI agents to query vast parameter configurations at negligible cost. This capability fundamentally changes the economics and speed of scientific exploration, moving beyond discrete, tokenizable representations typically associated with LLM applications in fields like drug discovery.

The hierarchical agent architecture orchestrates a closed-loop process of hypothesis generation, experimentation, analysis, and verification, operating with a tool-modular interface that requires no human intervention. This autonomous exploration was demonstrated in a complex fluid dynamics problem involving flow past tandem cylinders at a Reynolds number of 500. The system successfully evaluated over 1,600 parameter-location pairs, leading to the discovery of previously unknown divergent scaling laws. Specifically, it identified a regime-dependent two-mode structure for minimum displacement thickness and a robust linear scaling for maximum momentum thickness, both exhibiting dual-extrema structures emerging at critical flow transitions. These empirical discoveries underscore the framework's capacity to not only simulate but also to derive novel physical insights from complex systems.

The coupling of learned physical representations with agentic reasoning establishes a generalizable paradigm for automated scientific inquiry across various PDE-governed domains. This could accelerate breakthroughs in engineering design, climate modeling, and fundamental physics by enabling machines to independently formulate and test hypotheses at scales impossible for human researchers. The implications extend to democratizing access to advanced simulation and discovery tools, potentially fostering innovation in fields where computational resources or specialized expertise are currently bottlenecks. However, the validation and interpretability of such autonomously discovered laws will be critical, ensuring that the AI's findings are robust and align with established physical principles, or genuinely represent new, verifiable phenomena.