TLDR: SciExplorer is an AI agent that uses large language models (LLMs) and code execution tools to autonomously explore and discover the laws of unknown physical systems. It has been successfully applied to mechanical dynamical systems, wave evolution, and quantum many-body physics, demonstrating impressive performance in tasks like recovering equations of motion and inferring Hamiltonians, all without domain-specific blueprints or fine-tuning.
Scientific discovery, a process traditionally driven by human ingenuity, observation, and hypothesis, is increasingly seeing the integration of machine learning. However, fully automating the iterative and open-ended nature of discovering the fundamental laws of unknown systems has remained a significant challenge. This is where a new agent called SciExplorer steps in, aiming to bridge this gap by leveraging the advanced tool-use capabilities of large language models (LLMs) to explore physical systems without requiring specific domain blueprints.
SciExplorer is designed to mimic the scientific process, allowing an LLM to autonomously select and analyze experiments. Instead of being given a rigid set of tools, the agent primarily uses a flexible code execution tool, allowing it to generate and run arbitrary Python code for experiments, data analysis, and simulations. It also has access to a code-based plotting tool, which, combined with the LLM’s ability to interpret visual information, helps it gain qualitative insights. An external memory stores experimental and analysis results, enabling the agent to build upon its findings. The system operates with minimal domain-specific input, relying on a general prompt that encourages systematic hypothesis formulation and iterative verification.
Exploring Mechanical Systems
One of the key areas where SciExplorer was tested is in mechanical systems, which offer a wide range of complexities from conservative to dissipative and driven regimes. The agent’s task was to recover the equations of motion from observed dynamics. For instance, in a scenario involving a single particle moving in two spatial dimensions, SciExplorer could infer that the system was a Kepler-like central-force problem with an off-origin center. It achieved this by observing multiple evolutions, analyzing diagnostics like speed and angular momentum, and using sparse regression to determine the force law. The agent demonstrated impressive performance, often achieving near-perfect fits (R² ≈ 1) for a large subset of systems, even when only observing a subset of particles and inferring hidden initial conditions. However, its performance could degrade for more complex or atypical systems, suggesting a reliance on its broad knowledge base for success in other scenarios.
Unraveling Waves and Fields
SciExplorer also ventured into the dynamics of waves and fields, a cornerstone of modern physics. Here, the agent was allowed to set up initial field configurations and observe their subsequent evolution, governed by an unknown classical equation. It could also simulate hypothesized field equations for comparison. The agent proved capable of discovering the true models for various 1D field equations, including the nonlinear Schrödinger equation and the time-dependent Ginzburg-Landau equation, even with variations like external potentials or longer-range couplings. The agent often began with a detailed exploration plan, starting with Gaussian wave packets and then systematically extracting dispersion relations by launching plane waves. It effectively translated between abstract evolution equations and their corresponding Python specifications for simulation.
Also Read:
- Beyond Brute Force: How a New AI Learns Game Rules and Strategy
- HeDA Agent System Maps Complex Heatwave Vulnerabilities
Deciphering Quantum Many-Body Physics
Perhaps one of the most challenging domains tackled by SciExplorer is quantum many-body physics, involving complex systems of interacting spins or qubits. The agent was tested in two settings: observing the temporal evolution of single-spin operators from product states, and obtaining expectation values of arbitrary observables in the ground state of an unknown Hamiltonian. SciExplorer demonstrated an active working knowledge of quantum many-body physics, recognizing standard models like transverse Ising and Heisenberg. It employed strategies such as analyzing short-time dynamics to check for conserved quantities, computing dominant frequency components, and fitting hypothesized families of Hamiltonians. Even in scenarios where it could only observe a subset of spins or control unknown parameters, the agent often made significant progress, showcasing its ability to discover entire families of Hamiltonians.
The effectiveness of SciExplorer highlights the potential of LLM-based agents in automating open-ended scientific discovery. By combining the heuristic and general knowledge of LLMs with the reliability of external tools, it can characterize complex physical systems in a single-shot manner, without extensive fine-tuning or task-specific instructions. While the runtime for complex tasks can be substantial (up to 1.5 hours), it is often estimated to be faster than what a human expert would typically require for such an open exploration. This framework opens doors for similar scientific exploration in other domains, such as chemistry and biology, where code-based interfaces are common for experimental control. For more in-depth information, you can read the full research paper here.
