Evo-MCTS: Evolutionary Monte Carlo Tree Search for Gravitational Wave Detection

🧬 Overview

Evo-MCTS represents a breakthrough in automated scientific algorithm discovery, introducing the first integration of Large Language Model (LLM) guidance with domain-aware physical constraints for gravitational wave detection. This groundbreaking framework systematically explores algorithmic solution spaces through tree-structured search enhanced by evolutionary optimization, addressing fundamental limitations of existing approaches like matched filtering and deep neural networks.

Developed by researchers from the International Centre for Theoretical Physics Asia-Pacific (ICTP-AP), University of Chinese Academy of Sciences (UCAS), and Tsinghua University, this work demonstrates that LLMs can discover interpretable, high-performance algorithms while respecting strict physical constraints inherent in gravitational wave astronomy.

As highlighted on the project website: “Give Evo-MCTS a few iterations, and get a state-of-the-art gravitational wave detection algorithm in return!”

🎯 Key Achievements

Performance Metrics

• 20.2% improvement over state-of-the-art domain-specific methods
• 59.1% improvement over competing LLM-based optimization frameworks
• Consistently exceeds detection thresholds on MLGWSC-1 benchmark dataset
• Discovers novel algorithmic combinations not found in traditional approaches

Scientific Impact

• First framework to combine LLM guidance with physical constraints for scientific computing
• Establishes transferable methodology applicable across computational science domains
• Generates interpretable algorithmic pathways suitable for scientific validation
• Minimizes hidden biases common in black-box deep learning approaches

🚀 Key Innovations

Core Technical Advances

1. Reflective Code Synthesis

   • LLMs leverage domain expertise to generate executable algorithm components
   • Incorporates physical constraints directly into code generation process
   • Produces human-readable, scientifically interpretable solutions

2. Multi-Scale Evolutionary Operations

   • Operates on structured code representations at multiple abstraction levels
   • Combines local optimizations with global algorithmic restructuring
   • Balances exploration of novel solutions with exploitation of known patterns

3. Tree-Guided Exploration

   • Monte Carlo Tree Search systematically navigates algorithm design space
   • Produces interpretable pathways showing algorithmic evolution
   • Reveals distinct performance patterns and optimization trajectories

4. Domain-Aware Constraints

   • Integrates gravitational wave physics directly into search process
   • Ensures generated algorithms respect physical principles
   • Validates solutions against LIGO/Virgo/KAGRA detector characteristics

5. Multi-LLM Ecosystem

   • Compatible with OpenAI (GPT-3.5-turbo, GPT-4o, GPT-4-turbo)
   • Supports DeepSeek, Zhipu AI, and LiteLLM-compatible services
   • Flexible model selection based on task requirements and budgets

🔬 Technical Architecture

Three-Layer Framework

Layer 1: LLM-Guided Code Synthesis

• Domain knowledge injection through carefully designed prompts
• Reflective reasoning to evaluate and refine generated code
• Integration of physical constraints into generation process
• Validation against detector noise characteristics and signal morphology

Layer 2: Evolutionary Optimization

• Population Management: Maintains diverse algorithm candidates
• Fitness Evaluation: Tests algorithms on MLGWSC-1 benchmark
• Selection Mechanisms: Fitness-based tournament selection
• Genetic Operations: Crossover and mutation tailored for algorithmic structures
• Elitism: Preserves best-performing solutions across generations

Layer 3: Monte Carlo Tree Search

• Tree Construction: Builds search tree of algorithmic combinations
• UCB1 Selection: Balances exploration and exploitation
• Simulation: Evaluates promising branches through fitness testing
• Backpropagation: Updates node values based on descendant performance
• Best Path Extraction: Identifies optimal algorithmic pathways

Algorithm Discovery Pipeline

1. Initialization: Seed population with domain-inspired algorithms and LLM-generated variants
2. MCTS Expansion: Explore tree of algorithmic modifications and combinations
3. LLM Refinement: Use language models to propose promising variations
4. Evolutionary Selection: Fitness-based ranking and tournament selection
5. Validation: Multi-episode testing on MLGWSC-1 benchmark dataset
6. Iteration: Repeat until convergence or budget exhaustion
7. Analysis: Extract interpretable pathways and performance insights

📈 Performance Results

MLGWSC-1 Benchmark

The Machine Learning Gravitational-Wave Search Challenge 1 (MLGWSC-1) provides a rigorous testing ground with:

• Real LIGO detector noise characteristics
• Diverse signal morphologies (binary black hole coalescences)
• Multiple signal-to-noise ratio (SNR) regimes
• Continuous parameter spaces (masses, spins, sky locations)

Quantitative Improvements

Qualitative Advantages

• Interpretability: Generated algorithms consist of readable code with clear logic flow
• Physical Validity: All solutions respect gravitational wave physics constraints
• Robustness: Performs consistently across different noise realizations
• Generalization: Transfers to signal parameters not seen during training
• Computational Efficiency: Optimized algorithms reduce detection latency

Discovered Algorithmic Patterns

Evo-MCTS reveals distinct successful strategies:

1. Multi-resolution preprocessing: Wavelet-based time-frequency representations
2. Adaptive filtering: SNR-dependent feature extraction pipelines
3. Ensemble approaches: Combining complementary detection statistics
4. Physics-informed features: Leveraging chirp mass relationships and spin effects

💻 Getting Started

Installation

# Clone the repository
git clone https://github.com/iphysresearch/evo-mcts.git
cd evo-mcts

# Create virtual environment (recommended)
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

Key Dependencies

• scipy 1.12.0 - Scientific computing
• numpy - Array operations
• joblib - Parallel processing
• h5py - HDF5 data handling
• gwpy - Gravitational wave data analysis
• pandas - Data manipulation
• lalsuite - LIGO Algorithm Library
• hydra-core - Configuration management

Configuration

Create a .env file with your LLM API credentials:

# OpenAI
OPENAI_API_KEY=your_openai_key

# DeepSeek
DEEPSEEK_API_KEY=your_deepseek_key

# Zhipu AI
ZHIPUAI_API_KEY=your_zhipuai_key

Configure experiment parameters via Hydra config files in config/:

• model.yaml - LLM provider and model selection
• mcts.yaml - Tree search parameters
• evolution.yaml - Genetic algorithm settings
• dataset.yaml - MLGWSC-1 data paths

Running Experiments

# Basic run with default configuration
python main.py

# Specify LLM provider
python main.py model=openai model.name=gpt-4o

# Adjust MCTS parameters
python main.py mcts.num_iterations=100 mcts.exploration_weight=1.5

# Custom evolution settings
python main.py evolution.population_size=50 evolution.num_generations=20

# Run with specific random seed for reproducibility
python main.py seed=42

Analyzing Results

# Visualize search tree evolution
python scripts/plot_tree.py --experiment_dir outputs/YYYY-MM-DD/HH-MM-SS

# Compare algorithm performance
python scripts/compare_algorithms.py --results_path outputs/

# Extract best algorithm code
python scripts/extract_best.py --checkpoint path/to/checkpoint.pkl

📊 Application to Gravitational Wave Astronomy

Scientific Context

Gravitational waves are ripples in spacetime predicted by Einstein’s general relativity. LIGO, Virgo, and KAGRA detectors have observed dozens of signals from:

• Binary black hole mergers
• Binary neutron star mergers
• Neutron star-black hole mergers

Detection requires sophisticated algorithms to identify weak signals buried in detector noise.

Traditional Approaches

1. Matched Filtering

   • Correlates data with theoretical waveform templates
   • Computationally expensive (millions of templates)
   • Assumes accurate waveform models
   • Struggles with glitches and non-Gaussian noise

2. Deep Neural Networks

   • Learn patterns directly from training data
   • Fast inference once trained
   • Black-box nature hinders scientific validation
   • May encode hidden biases from training data

Evo-MCTS Advantages

• Interpretability: Scientists can validate algorithmic logic
• Flexibility: Adapts to detector characteristics without retraining
• Efficiency: Discovers compact algorithms with low computational cost
• Novelty: Finds combinations not conceived by human experts
• Transferability: Methodology applies to future detector generations (Einstein Telescope, Cosmic Explorer, LISA)

🌐 Broader Impact

Transferable to Other Domains

The Evo-MCTS framework is domain-agnostic and applicable to:

• Astronomy: Exoplanet detection, transient classification, pulsar timing
• Particle Physics: Event reconstruction, trigger algorithms
• Materials Science: Crystal structure prediction, property optimization
• Climate Science: Pattern recognition in climate models
• Bioinformatics: Sequence analysis, protein folding
• Drug Discovery: Molecular optimization, binding affinity prediction

Methodological Contributions

1. LLM-Scientific Computing Integration: Demonstrates effective combination of language models with rigorous scientific constraints
2. Interpretable AI: Shows that automated discovery need not sacrifice explainability
3. Hybrid Search: Combines tree search, evolution, and LLM guidance synergistically
4. Evaluation Framework: Establishes principles for validating discovered algorithms

🔗 Resources

• Paper: arXiv:2508.03661 (76 pages, submitted August 2025)
• Code: GitHub Repository (GPL-3.0 License)
• Website: Project Homepage
• Dataset: MLGWSC-1 Benchmark (available via LIGO Open Science Center)

📝 Citation

If you use Evo-MCTS in your research, please cite:

@article{wang2025evomcts,
  title={Evo-MCTS: Evolutionary Monte Carlo Tree Search for Gravitational Wave Detection},
  author={Wang, He and others},
  journal={arXiv preprint arXiv:2508.03661},
  year={2025},
  note={76 pages}
}

🤝 Collaboration

This project represents collaborative research between:

• International Centre for Theoretical Physics Asia-Pacific (ICTP-AP)
• University of Chinese Academy of Sciences (UCAS)
• Tsinghua University

🌟 Future Directions

Short-term Goals

• Extend to multi-messenger astronomy (joint gravitational wave + electromagnetic observations)
• Incorporate additional detector data (KAGRA, future Einstein Telescope simulations)
• Optimize computational efficiency for real-time detection pipelines
• Develop interactive visualization tools for algorithm evolution

Long-term Vision

• Generalize framework to other scientific computing domains
• Investigate transfer learning between related detection problems
• Develop theoretical understanding of LLM-guided search convergence
• Create benchmark suite for automated algorithm discovery
• Establish best practices for interpretable AI in scientific research

Open Challenges

• Scaling to extremely large algorithm search spaces
• Incorporating multi-objective optimization (detection rate vs. false alarm rate)
• Handling non-differentiable performance metrics
• Ensuring reproducibility across different LLM providers
• Balancing computational cost of fitness evaluation with search efficiency

Ready to revolutionize gravitational wave detection? Explore the project website, dive into the GitHub repository, or read the full paper to learn more!

Join the scientific computing revolution powered by LLM-guided automated algorithm discovery.