LeoLai689/SimulCost-baseline-sim_res

SimulCost-Bench

Project Page | Paper | GitHub | Tools

SimulCost is a cost-aware benchmark and toolkit for evaluating how well LLM agents tune simulation parameters under realistic computational budgets. Unlike prior evaluations that focus on correctness while implicitly treating tool usage as "free," SimulCost explicitly measures both accuracy and simulation compute consumption (defined in a platform-independent way via analytical cost models or FLOPs).

The benchmark spans 2,916 single-round (initial guess) and 1,900 multi-round (trial-and-error with feedback) tasks across 12 simulators.

Simulators Included

The benchmark covers 12 physics simulators across fluid dynamics, solid mechanics, and plasma physics:

• 1D Heat Transfer
• 2D Steady Heat Transfer
• 1D Burgers Equation
• 1D Euler Equations
• 2D Navier-Stokes (Channel and Transient)
• 1D EPOCH Particle-in-Cell (PIC)
• 2D Material Point Method (MPM)
• 1D Diffusion-Reaction
• 2D Euler Gas Dynamics
• Hasegawa-Mima (Linear and Nonlinear)
• 2D Finite Element Method (FEM)

This repository stores simulation result archives (.zip/.tar.gz) used as a cache to facilitate research without needing to re-run expensive physics simulations.

Sample Usage

To run inference on the benchmark using the toolkit provided in the GitHub repository:

# Example for 1D Heat Transfer task using OpenAI
python inference/langchain_LLM.py -p openai -m gpt-5-2025-08-07 -d heat_1d -t cfl -l medium -z

Parameters:

• -p: LLM provider (openai, gemini, bedrock, custom_model)
• -m: Model name/identifier
• -d: Dataset name
• -t: Problem task type
• -l: Precision level
• -z: Enable zero-shot mode

Citation

@article{cao2026simulcost,
  title={SimulCost: A Cost-Aware Benchmark and Toolkit for Automating Physics Simulations with LLMs},
  author={Cao, Yadi and Lai, Sicheng and Huang, Jiahe and Zhang, Yang and Lawrence, Zach and Bhakta, Rohan and Thomas, Izzy F. and Cao, Mingyun and Tsai, Chung-Hao and Zhou, Zihao and Zhao, Yidong and Liu, Hao and Marinoni, Alessandro and Arefiev, Alexey and Yu, Rose},
  journal={arXiv preprint arXiv:2603.20253},
  year={2026}
}