Self-Evolving Scientific Agent Discovers Generalizable Physically-Reasoned Fluid Control

Abstract

While data-intensive deep reinforcement learning can optimize complex control policies, scientific control design in physical systems fundamentally requires an interpretable chain of reasoning that connects physical evidence to structured control architectures. Here, we present a self-evolving scientific agent workflow, driven by large language models and iterative code generation, that automates controller construction while preserving strict interpretability and rigorous physical reasoning. Instead of adjusting weights, the agent deploys candidate whitebox controllers into physical simulations, actively diagnoses dynamic behaviors from multimodal evidence, and translates these observations into progressive source-code refinements. We demonstrate this framework on a highly non-linear fluid-structure interaction problem: an underactuated, two-joint dogfish swimmer tasked with spatial target reaching in an unsteady flow using only joint angular accelerations. Starting from a target-blind propulsive seed, the agent autonomously designs and refines a unified controller that reaches a target embedded in an unsteady four-cylinder wake. Without retraining, retuning or case-specific branching, the retained controller achieves target capture across the full generalization test matrix, spanning variations in target position, rear-row geometry, cylinder count and inflow speed. The auditable evolution log reveals an emergent control architecture built upon travelling-wave propulsion, body-frame bearing guidance, phase-selective steering, corrective burst and adaptive relief. Our results show that an autonomous scientific agent can successfully transform accumulated physical evidence into a robust, mathematically readable control policy, while maintaining a fully traceable process of scientific control design.