PDE-Agent: A toolchain-augmented multi-agent framework for PDE solving
📌 Contents
🆕 News
- [2025-12-18]: 🎉 PDE-Agent is Now Available in This Repository!
- [2025-11]: The PDE-Agent project will available on GitHub.
⌚️ Overview
In this work, we frame PDE solving as tool invocation via LLM-driven agents and introduce PDE-Agent, the first toolchain-augmented multiagent collaboration framework, inheriting the reasoning capacity of LLMs and the controllability of external tools and enabling automated PDE solving from natural language descriptions.
PDE-Agent leverages the strengths of multi-agent and multi-tool collaboration through two key innovations: A Prog-Act framework with graph memory and A Resource-Pool integrated with a tool-parameter separation mechanism.
Key Features:
- PDE Tools: The first LLM-driven multiagent framework equipped with modular PDE-toolkits rigorously encapsulated for automated PDEs solving.
- Prog-Act: A Prog-Act framework with graph memory for multi-agent
collaboration, which enables effective dynamic planning and
error correction via dual-loop mechanisms (localized fixes
and global revisions).
- Resource-Pool: A Resource-Pool integrated with a tool-parameter separation mechanism for multi-tool collaboration. This centralizes the management of runtime artifacts and resolves inter-tool dependency gaps in existing frameworks.
📦 Project Framework
/SciToolAgent
├── data # Data storage directory
├── scripts # Scripts for running PDE-Agent
├── test # Testing scripts
├── SciToolEval # SciToolEval related files
└── HiveMinds # PDE-Agent main files
├── minds # Multi-Agent
├── toolkits # PDE tools
├── memory # Memory and resources pool
├── utils # Common utility functions and toolchain
├── hive_mind.py # The core logic of PDE-Agent
└── .env # YOUR API KEY
⚡ Getting Started
🔧️ Installation
-
Clone the repository
First, clone the project to your local machine:
git clone https://github.com/ScienceOne-AI/PDE-Agent.git
cd PDE-Agent
-
Create and activate a virtual environment
Set up a new virtual environment using Conda and activate it:
conda create -n PDEA python=3.9
conda activate PDEA
-
Install project dependencies
Install the necessary dependencies for the project:
pip install -r ./pyenv/requirements.txt
🚀Quick Start
You need to modify the `.env` files to set your `API_KEY` and `API_BASE`.
*
```
DEEPSEEK_API_KEY=YOUR_DEEPSEEK_API_KEY
DEEPSEEK_BASE_URL=DEEPSEEK_BASE_URL
```
-
Load dependence.
-
Run the PDE-Agent
cd ../scripts
bash run.sh
You can also give the case_file or edit run.sh:
bash run.sh ./data/yaml/inputs/Allen-Cahn equation/eval1.yaml
The results will be saved in ./scripts/logs
📊 Data Preparation
🔄 Data is in Progress!
We are currently performing data cleaning, standardization, and validation for the dataset of this project:
- Eliminating redundant, incomplete, or invalid data entries
- Unifying data formats to ensure consistency across the dataset
- Verifying data accuracy to support reliable subsequent use
The data preparation work is ongoing, and we will update the processed dataset, along with relevant supporting materials, in this repository once this process is completed.
📝️ Cite
If you find this work useful, please consider citing:
@misc{liu2025pdeagenttoolchainaugmentedmultiagentframework,
title={PDE-Agent: A toolchain-augmented multi-agent framework for PDE solving},
author={Jianming Liu and Ren Zhu and Jian Xu and Kun Ding and Xu-Yao Zhang and Gaofeng Meng and Cheng-Lin Liu},
year={2025},
eprint={2512.16214},
archivePrefix={arXiv},
primaryClass={cs.AI},
url={https://arxiv.org/abs/2512.16214},
}