Agentic Exploration of Physics Models

This paper introduces SciExplorer, a large language model-based agent capable of autonomously discovering the laws of unknown physical systems through iterative experimentation and code execution without requiring domain-specific tuning or instructions.

The Gist

Imagine you are handed a mysterious, locked box. You can't see inside, and you don't know what's making the noises coming from it. Your job is to figure out exactly how the machine inside works just by shaking it, listening to the sounds, and watching how it reacts when you push it.

This is essentially what scientists do when they study the universe: they observe nature, guess the rules (hypotheses), test those guesses, and refine them until they find the "law" that explains everything.

This paper introduces SciExplorer, a new kind of AI scientist designed to do this detective work entirely on its own.

The Problem: The "Specialist" vs. The "Generalist"

In the past, we've built AI tools that are like specialized apprentices.

• If you want to predict the weather, you use a weather AI.
• If you want to design a new drug, you use a chemistry AI.
• If you want to solve a specific math equation, you use a math AI.

These tools are great at their one job, but if you ask a weather AI to design a drug, it's completely lost. They need to be taught exactly what to do for every single new task.

SciExplorer is different. It's like a curious, super-smart intern who has read every physics textbook ever written but has never seen the specific machine in front of them. You don't give it a manual or a step-by-step guide. You just say, "Figure out how this thing works," and it has to figure out the rest.

How SciExplorer Works: The "Try, Fail, Learn" Loop

The paper describes SciExplorer as an "agentic" system. Think of it as a robot with a brain (a Large Language Model, or LLM) and a set of hands (computer tools).

Here is the cycle it goes through, using a simple analogy: The Detective's Notebook.

1. The Plan (Hypothesis Generation):
   The AI looks at the mystery system. It says, "Okay, I think this might be a swinging pendulum. Or maybe it's a wave. Let's test that."

   • Analogy: The detective looks at a crime scene and says, "It looks like a robbery. Let's check for fingerprints."

2. The Experiment (Tool Use):
   The AI writes a computer program to simulate an experiment. It might say, "I'm going to push the system hard from the left and see what happens." It runs this code.

   • Analogy: The detective sets up a trap or interviews a witness to see if their theory holds up.

3. The Analysis (Observation):
   The AI looks at the results. Did the system swing? Did it stop? Did it explode? It draws graphs and plots to see patterns.

   • Analogy: The detective looks at the evidence. "Hmm, the witness said it was raining, but the ground is dry. My theory about the robbery is wrong."

4. The Pivot (Self-Correction):
   If the theory was wrong, the AI doesn't give up. It says, "Okay, it's not a pendulum. Maybe it's a spring? Let's try a different experiment." It writes new code, runs new tests, and updates its notebook.

   • Analogy: The detective realizes the suspect is actually a spy, not a robber, and starts looking for different clues.

5. The Solution (Discovery):
   Eventually, the AI finds a set of rules (an equation) that perfectly predicts what the system will do in any situation. It writes down the final formula.

What Did It Actually Do?

The researchers tested SciExplorer on three very different types of "mystery boxes":

• Mechanical Systems (The Swinging Things): They gave it data from things like double pendulums (two swings attached) and particles moving in 2D. The AI had to figure out the exact equations of motion (like Newton's laws) just by watching the movement.

  • Result: It successfully "rediscovered" the laws of physics for these systems, often getting a perfect score.

• Wave Systems (The Ripples): They gave it data on how waves move through a grid of connected points (like ripples in a pond or light waves). The AI had to guess the complex equations that govern these waves.

  • Result: It figured out complex wave equations, including ones with "non-linear" effects (where the wave changes its own shape).

• Quantum Systems (The Tiny Spins): This is the hardest part. They gave it data on quantum particles (spins) and asked it to find the "Hamiltonian" (the master rulebook that dictates how these tiny particles interact).

  • Result: Even though quantum physics is notoriously difficult and counter-intuitive, the AI successfully identified the correct interaction rules for these systems.

Why Is This a Big Deal?

1. No "Cheating": The AI wasn't told what kind of system it was looking at. It didn't know "This is a pendulum." It just knew "Here is some data, find the rule."
2. It's a Generalist: It didn't need to be retrained for each new task. The same AI solved the pendulum, the wave, and the quantum problems using the same "brain."
3. It Handles Noise: Real-world data is messy. The researchers added "noise" (random errors) to the data, and SciExplorer could still figure out the correct laws, just like a human scientist would.
4. It's Fast (for an AI): While it takes the AI a few minutes to an hour to solve a problem, the researchers estimate that a human expert would take much longer to do the same "open-ended" exploration from scratch.

The Limitations

The paper is honest about where the AI struggles. Sometimes it gets "stubborn." If it guesses a model early on, it might stick to it even when the data says it's wrong. It also sometimes misses subtle visual clues in the graphs that a human might spot immediately. It's not perfect yet, but it's a massive leap forward.

The Bottom Line

SciExplorer is a proof-of-concept that we are moving toward a future where AI can act as a true scientific partner. Instead of just crunching numbers for us, it can look at a mystery, design its own experiments, analyze the results, and discover new laws of physics—all without needing a human to hold its hand and tell it exactly what to do next.

It's like giving a robot a library of all human knowledge and a set of tools, then saying, "Go explore the universe and tell us what you find." And for the first time, the robot is actually starting to find things on its own.

Technical Summary

1. Problem Statement

Scientific discovery traditionally relies on an iterative loop of observation, analysis, and hypothesis generation. While machine learning (ML) has been applied to specific sub-tasks (e.g., predicting experimental outcomes or fitting known equations), fully automating the heuristic, iterative loop required to discover the laws of an unknown system remains an open challenge. Existing approaches often require task-specific fine-tuning, predefined blueprints, or specialized tools for specific domains. The authors aim to bridge this gap by creating a general AI agent capable of exploring physical systems without domain-specific prior knowledge, mimicking the open-ended nature of human scientific inquiry.

2. Methodology: SciExplorer

The authors introduce SciExplorer, an autonomous AI agent built upon a Large Language Model (LLM) with tool-use capabilities.

• Core Architecture:

  • LLM Backbone: The agent uses state-of-the-art LLMs (primarily GPT-5) as the reasoning engine.
  • Tool Integration: The agent does not have hardcoded physics knowledge but accesses a set of generic tools:
    • Code Execution: A Python interpreter (using numpy, scipy, jax) to perform arbitrary data analysis, numerical integration, and model simulation.
    • Visualization: A plotting tool to generate images from data, which the LLM analyzes via its multimodal capabilities to extract qualitative insights (e.g., symmetry, decay, oscillation).
    • External Memory: A persistent storage system for experimental results and intermediate analysis data.
    • Simulators: Domain-specific simulators (e.g., ODE solvers for mechanics, split-step integrators for fields, spin-chain solvers for quantum systems) that the agent can query.
  • Autonomous Loop: The agent operates in a cycle:
    1. Reasoning: Formulates a hypothesis or plan based on current data.
    2. Action: Calls tools to run numerical experiments (setting initial conditions) or analyze existing data.
    3. Observation: Receives results (text, arrays, or images).
    4. Iteration: Updates its hypothesis, refines parameters, or designs new experiments until a "testable final answer" is reached.

• Minimalist Prompting: Crucially, the system prompt is generic and domain-agnostic. The agent is told to act as a "cautious scientist" but is not given specific instructions about the type of system (e.g., it is not told that mechanical systems follow ODEs or that quantum systems follow Hamiltonians). It must infer the governing laws from the data alone.

3. Key Contributions

• Generalist AI Physicist: Demonstrates that a single, non-finetuned LLM agent can successfully discover models across disparate physical domains (mechanics, wave dynamics, and quantum many-body physics) using only generic coding and visualization tools.
• Active Learning in Science: The agent does not just fit data; it actively designs experiments (selecting initial conditions, varying parameters) to maximize information gain and distinguish between competing hypotheses.
• Program Discovery: Unlike traditional symbolic regression which finds mathematical expressions, SciExplorer often discovers the entire simulation code (the "program") that reproduces the system's dynamics, effectively solving a program discovery problem.
• Robustness to Noise: The framework is tested under significant measurement noise (Gaussian noise for classical systems, shot noise for quantum systems), showing the agent can still recover accurate models.

4. Results

The authors evaluated SciExplorer on a broad set of benchmark tasks:

• Mechanical Systems:

  • Task: Discover equations of motion (ODEs) for systems like damped double pendulums, coupled oscillators, and particles in 2D potentials.
  • Performance: The agent successfully recovered governing equations with high accuracy ($R^2 \approx 1$) for a large subset of systems. It correctly identified non-linearities, damping terms, and external potentials.
  • Challenges: Performance degraded for "arbitrary" potentials with no known analytical form, though the agent often found approximate numerical solutions.

• Wave and Field Dynamics:

  • Task: Discover partial differential equations (PDEs) governing complex fields (e.g., Nonlinear Schrödinger Equation, Complex Ginzburg-Landau Equation).
  • Performance: The agent successfully identified the correct PDE structure, including kinetic terms, potential terms, and non-linearities. It could distinguish between conservative (Schrödinger) and dissipative (Ginzburg-Landau) dynamics.
  • Failure Mode: The agent struggled with an artificial model involving a sinusoidal relaxation term ($\sin(0.1|\phi|^2)\phi$), which is not a standard physical term.

• Quantum Many-Body Physics:

  • Task: Infer the Hamiltonian of spin systems (e.g., Heisenberg, Transverse Field Ising, Cluster Ising) from either time-evolution data or ground-state expectation values.
  • Performance: The agent demonstrated an active working knowledge of quantum models, correctly identifying standard Hamiltonians and even discovering tunable parameter families. It successfully handled scenarios with partial observability (only observing a subset of spins).
  • Precision: In ground-state tasks, the agent achieved high fidelity ($|\langle \psi_{agent} | \psi_{true} \rangle|^2 \approx 1$) in discovering the correct Hamiltonian structure.

• Ablation Studies:

  • Tools: Access to both coding and plotting tools was found to be critical. Without tools, the LLM failed to recover models even with visualizations provided.
  • Model Dependency: GPT-5 significantly outperformed other models (Gemini 2.5 Pro, GPT-5 Nano), highlighting the importance of advanced reasoning capabilities.
  • Runtime: A single exploration takes minutes to ~1.5 hours, which is comparable to or faster than an expert human physicist for open-ended discovery.

5. Failure Modes and Limitations

The authors identified common failure modes:

• Premature Commitment: The agent sometimes locks onto an incorrect model structure early in the process and fails to reconsider despite poor fits.
• Overlooking Qualitative Cues: The agent occasionally misses visual patterns in plots (e.g., periodic oscillations in acceleration) that would hint at the correct model.
• Parameter Precision: While the qualitative structure is often correct, the agent sometimes struggles with precise numerical parameters (e.g., missing a factor of 2 or sign errors in quantum Hamiltonians).
• Hallucination: Like all LLMs, the agent can hallucinate facts or code, though the tool-use loop provides a mechanism for self-correction.

6. Significance and Future Outlook

• Paradigm Shift: This work moves AI in science from "assisting with specific tasks" to "autonomously conducting research." It demonstrates that general-purpose LLMs, when equipped with code execution, can act as generalist scientists.
• Scalability: The framework is applicable to any domain where experiments can be simulated or controlled via code (e.g., chemistry, biology, materials science).
• Cost-Effectiveness: While the computational cost per run is higher than specialized algorithms (like SINDy or AIFeynman), it is lower than the human labor required for open-ended discovery.
• Future Directions: The authors suggest that future iterations could integrate more sophisticated prompting to reduce sign errors, improve visual reasoning, and apply the framework to real-world experimental hardware (e.g., cold atom labs or quantum processors).

In conclusion, SciExplorer represents a significant step toward fully automated scientific discovery, proving that an AI agent can navigate the heuristic search space of physical laws without human intervention or domain-specific training.