llm-autonomous-physics-discovery

name	llm-autonomous-physics-discovery
description	Autonomous LLM agent methodology for computational physics discovery using progressive local search, knowledge accumulation from successful/failed attempts, and interpretable exploration trajectories. Covers PhyNex framework for scorable scientific tasks with domain-specific tools enforcing physical consistency. Activation: LLM autonomous discovery, physics agent, progressive local search, computational physics agent, PhyNex, automated physics optimization, 大语言模型物理发现, 自主科学发现代理
metadata	{"arxiv_id":"2606.14266","published":"2026-06-12","authors":"Hang Lin, Chongwen Liu, Gang Yan","tags":["LLM","autonomous-discovery","computational-physics","agent","optimization"]}

LLM Autonomous Physics Discovery

Description

PhyNex: autonomous LLM agent framework for computational physics discovery. Combines LLM-guided progressive local search with domain-specific tools enforcing physical consistency, accumulating knowledge from both successful and failed attempts.

Activation Keywords

• LLM autonomous discovery
• physics agent
• progressive local search
• computational physics agent
• PhyNex
• automated physics optimization
• 大语言模型物理发现
• 自主科学发现代理
• scientific discovery agent
• LLM-guided exploration

Core Methodology

Problem

Scientific discovery in computational physics involves optimizing quantitatively evaluable objectives subject to physical constraints, but researchers spend substantial effort on iterative refinement of methods.

PhyNex Framework

1. Exploration: Systematically explore solution space of scorable scientific tasks
2. LLM-guided search: Use LLMs to propose method modifications and parameter adjustments
3. Domain tools: Enforce physical consistency through computational tools
4. Knowledge accumulation: Learn from both successful and failed attempts
5. Interpretable trajectories: Reveal which algorithmic components drive improvements

Validated Applications

Task	Improvement	Metric
Dielectric spectra prediction	+3.8%	Spectral similarity
Max-Cut heuristics	+15.0%	Normalized mean cut
Quantum battery charging	+5.9%	Ergotropy at 80k checkpoint

Usage Patterns

Pattern 1: Automated Algorithm Design

When optimizing computational physics algorithms:

1. Define scorable objective function with physical constraints
2. Set up LLM agent with domain-specific tool access
3. Run progressive local search with knowledge accumulation
4. Analyze exploration trajectories for interpretable insights

Pattern 2: Method Discovery

When discovering new approaches for scientific problems:

1. Frame problem as optimization with quantitative metrics
2. Provide LLM with prior knowledge and constraints
3. Allow agent to explore methodological search space
4. Extract reusable knowledge from successful attempts

Instructions for Agents

Step 1: Define Scientific Task

• Quantifiable objective function (e.g., spectral similarity, ergotropy)
• Physical constraints (e.g., conservation laws, boundary conditions)
• Evaluation protocol with reproducible metrics

Step 2: Set Up Agent Framework

• LLM with domain knowledge context
• Domain-specific computational tools
• Progressive local search mechanism
• Knowledge storage for attempt history

Step 3: Run Exploration

• Agent proposes method modifications
• Execute with physical consistency checks
• Score and record results
• Accumulate knowledge from successes and failures

Step 4: Analyze Results

• Extract interpretable exploration trajectories
• Identify which components drove improvements
• Document reusable knowledge for future tasks

Error Handling

Physical Constraint Violation

If agent proposes physically invalid solutions:

• Enforce constraints via domain tools
• Reject and record failed attempts as knowledge
• Guide agent toward valid search regions

Convergence Failure

If agent cannot improve beyond baseline:

• Expand search space exploration
• Provide additional domain priors
• Reset with different initial configurations

References

• arXiv: 2606.14266v1
• Validated on: dielectric spectra, Max-Cut heuristics, quantum batteries