GWAgent: Building Interpretable Scientific Simulation Surrogate Models with Agent AI

GWAgent is an agent workflow system based on large language models, designed to automatically construct physically interpretable analytical surrogate models from simulation data. It resolves the contradiction in traditional surrogate models where machine learning black boxes lack interpretability and symbolic regression has insufficient precision. In gravitational wave signal modeling tasks, it achieves high precision (median waveform mismatch of 6.9×10⁻⁴) and an 8.4x speedup, outperforming traditional symbolic regression and machine learning methods.

In many scientific fields such as gravitational wave astronomy, climate simulation, and nuclear physics, high-precision numerical simulations often consume enormous computational resources. Taking the gravitational wave signal generated by binary black hole mergers as an example, a complete numerical relativity simulation may take weeks or even months. To accelerate scientific reasoning and parameter estimation, researchers have developed various "surrogate models"—using fast-computing approximate models to replace expensive simulations.

However, traditional surrogate models face a fundamental contradiction: machine learning-based black-box models are fast but lack interpretability, making it difficult for scientists to understand the internal physical mechanisms; while methods like symbolic regression can produce explicit formulas, they often lack precision in high-dimensional complex problems. This dilemma gives rise to a core question: Can we build surrogate models that are both fast and accurate, as well as physically interpretable?

GWAgent (Gravitational Wave Agent) is an agent workflow system based on large language models proposed by researchers. Unlike traditional one-time reasoning, GWAgent adopts an iterative model construction strategy:

1. Candidate Model Generation: LLM generates candidate mathematical expressions based on current data and domain knowledge
2. Quantitative Validation: Each candidate model is numerically compared with ground-truth simulation results
3. Feedback Iteration: Based on validation results, LLM adjusts the model structure and generates an improved version
4. Convergent Output: When the model precision reaches the threshold, the final interpretable analytical expression is output

This "validation-constrained iterative optimization" paradigm is the core innovation of GWAgent. Surrogate models are particularly suitable for this agent workflow because each candidate model can be objectively evaluated through numerical simulation, providing a clear optimization signal for LLM.

GWAgent's key design allows researchers to inject physically inspired domain ansatz into the system. In gravitational wave modeling tasks, researchers provided the agent with physical intuition about the waveform structure of binary black hole mergers—including the characteristic forms of the inspiral phase, merger phase, and ringdown phase.

Experimental results show that injecting such physical knowledge significantly improves the accuracy of the output model. Compared with purely data-driven methods, GWAgent combined with physical priors converges faster to expressions conforming to physical laws while maintaining model simplicity and interpretability. This reveals an important insight: In scientific discovery tasks, LLM should not be a black box replacing human domain knowledge, but a collaborative tool integrating human intuition and computational capabilities.

Researchers conducted a comprehensive evaluation of GWAgent on the gravitational waveform modeling task of eccentric binary black hole mergers. The results show:

• Precision Metric: The median waveform mismatch of the generated analytical surrogate model on the Advanced LIGO detector is only 6.9×10⁻⁴, reaching a level almost indistinguishable from high-precision numerical simulations
• Speedup Ratio: Compared to direct numerical simulations, the surrogate model's evaluation speed is increased by about 8.4 times
• Comparative Advantage: On the same task, GWAgent outperforms both traditional symbolic regression methods and conventional machine learning baseline models

More importantly, GWAgent-generated models are not just black-box predictors but analytical expressions with clear physical meanings. Researchers can identify compact physical structures from learned representations, providing a new perspective for understanding complex gravitational wave emission mechanisms.

To demonstrate GWAgent's practical scientific value, the team applied it to the analysis of the real astrophysical event GW200129—a gravitational wave event detected by LIGO/Virgo from two black hole mergers.

Using GWAgent's surrogate model, researchers accurately inferred the system's orbital eccentricity. The analysis shows that at a 20 Hz reference frequency, the binary black hole system's eccentricity is e₂₀ₕ₂ = 0.099⁺⁰·⁰⁶³₋₀.₀₄₄. This result not only constrains the astrophysical system's formation channel but also demonstrates the practicality of agent-assisted surrogate models in real scientific reasoning tasks.

GWAgent's success provides important methodological insights for AI applications in scientific computing:

First, it proves the effectiveness of the "validation-constrained iterative optimization" agent paradigm in scientific discovery tasks. Unlike open-ended text generation, scientific modeling tasks have objective evaluation criteria (e.g., waveform mismatch), providing clear feedback signals for LLM to continuously improve through iterations.

Second, GWAgent demonstrates how to effectively combine human domain knowledge (physical assumptions) with LLM's reasoning capabilities. Pure end-to-end learning often struggles to capture deep physical laws, while human-guided domain assumptions help the agent quickly locate promising areas in the vast hypothesis space.

Finally, this work emphasizes the importance of interpretability in scientific AI. GWAgent's analytical models are not just for prediction but for understanding—scientists can read, analyze, and modify these expressions to deepen their understanding of physical systems.

The GWAgent framework has wide applicability. Although the current demonstration focuses on gravitational wave astronomy, its core method—iterative model discovery based on validation feedback—can be extended to any scientific simulation task with quantifiable evaluation criteria. Fields like climate modeling, material simulation, and fluid dynamics computing may all benefit from this agent-driven surrogate model construction method.

This work also marks a shift in the role of large language models in scientific research: from simple text processing tools to intelligent collaborators that actively participate in the scientific discovery process. As LLM capabilities continue to improve and agent workflows mature, we can expect AI to play an increasingly important role in accelerating scientific discovery and revealing natural laws.