AI4S-Agent Co-building Initiative | Open Source Collaboration: ABACUS-agent-tools

Posted on In Word count in article: 1.1k Reading time ≈ 4 mins.

Background

ABACUS is a density functional theory (DFT) software initiated by the University of Science and Technology of China, and open-sourced and co-constructed by multiple domestic teams including Peking University, Institute of Physics of Chinese Academy of Sciences, Beijing Academy of Science and Intelligence, and Hefei Comprehensive Artificial Intelligence Research Institute. Since adopting the LGPL3.0 open-source license in 2021 and further embracing the open-source sharing concept together with the DeepModeling community, it has successively released 70 iterative versions. Both the software functions and ecosystem have been significantly developed thanks to the selfless contributions of the open-source community.

Currently, ABACUS supports approximately 30 major categories of functions and has 449 adjustable parameters (statistics from the online user documentation on GitHub as of June 13, 2025), providing users with rich functional choices. Meanwhile, ABACUS supports various high-performance computing methods such as CPU thread/process/thread-process hybrid parallelism and CPU+GPU/DCU heterogeneous parallelism. It has been pre-deployed on mainstream scientific computing platforms including domestic supercomputers, the Supercomputer Internet, and the Bohrium cloud computing platform of Deep Potential Technology, providing convenient and efficient computing support for the majority of users.

In terms of user tutorials and documentation, developers and users of the ABACUS open-source community, adhering to the open-source spirit, have released diverse tutorials and user success cases, including online documentation updated with versions, ABACUS Chinese user tutorial series, Bohirum Notebook tutorials, etc. However, with the rapid iteration of ABACUS in terms of functional diversity, computing efficiency, and development friendliness, the content of some tutorials inevitably shows minor inconsistencies with the features of the latest software version, making it difficult to meet the rapidly growing diverse computing needs of users.

With the widespread application of large language models (LLM) and related technologies in specific scenarios, the concept of agents, centered around LLM and equipped with mature tools such as knowledge bases and highly automated workflows, has emerged. The exploration of its technology stack, product design, product deployment, and rapid iteration have become the forefront of leading productivity transformation. The DeepModeling community has launched the AI4S-agent-tools project, covering various scenarios such as literature reading, material intelligent simulation, and analysis. First-principles material calculation, as the core function of material intelligent simulation and a data production engine, has a wide coverage and self-evident importance. However, there are many optional methods and software for first-principles material calculation, which require a lot of development and iteration to have relevant tool interfaces that support diverse and reliable scenarios.

In view of this, the ABACUS development team hereby launches a new open-source initiative: jointly building the best first-principles material calculation agent tool interface collection (tentatively named ABACUS-agent-tools). The goal is to achieve efficient connection between the physical problems of material calculation and the technical problems of first-principles calculation, integrating ABACUS pre-processing scripts, post-processing scripts, material property calculation workflows, ABACUS+other open-source software collaborative calculation workflows, ABACUS successfully completed calculation verification cases, and parameter combinations, etc. Whether ordinary users, developers, or other teams hoping to integrate ABACUS-agent-tools into agent products, they can become contributors through open-source collaboration to achieve mutual benefit and win-win results!

Goals of the ABACUS-agent-tools Co-construction Plan

  1. Promote the Open-Source Spirit and Break Down Knowledge Barriers

The technical experience of first-principles calculation simulation adapted to a specific type of scientific problem often requires a long time of experimentation, testing, and accumulation, which means it has a strong dependence on specific operators. Fortunately, some successful application cases and skills of simulation have been open-sourced as data attached to academic articles, written on web pages, or encoded in post-processing scripts or software. However, after the information volume begins to grow rapidly, the difficulty of information retrieval significantly increases, while the speed of information dissemination remains slow.

LLM/Agent-related technologies have brought new technical paths for the rapid dissemination of experience: encoding successful cases, skills, and data into the knowledge base connected by the Agent, and indexing and retrieving the data in the form of natural language description. We plan to effectively integrate experience into the Agent through direct or indirect encoding in ABACUS-agent-tools, making it a basic tool that can be migrated, repeated, and used as a basis for further inference, effectively improving the accuracy and effectiveness of the entire agent in performing related calculations.

  1. AI+DFT

In the wave of AI for Science, the rapid development of machine learning technology has also brought forth an endless stream of AI-integrated DFT multi-scale simulation methods, enriching the technical routes available for users to study materials, expanding the diversity of material objects that users can study, and the scope of material-related phenomena. However, it has also increased the threshold for users to learn new methods: compared with a few years ago, users now need to master relevant knowledge in multiple fields, be familiar with the use of multiple software from different fields, and have strong script development capabilities.

ABACUS-agent-tools is committed to reverting the complicated technical problems faced by users to physical problems as much as possible. Based on a well-organized program architecture that supports multi-terminal collaboration and asynchronous development, the technical details of various programs are encapsulated into functions callable by LLM, effectively helping users to focus on physical problems to the greatest extent and interact with LLM in natural language. In this process, users can also effectively combine the learning and actual use process, and quickly apply the most cutting-edge material simulation calculation methods involving machine learning-related technologies to their own research work.

  1. Flexible Customization to Deconstruct Complex Material Simulation Calculation Problems

ABACUS-agent-tools is committed to solving the problem of "lack of reliable open-source tool sets for agents to call for first-principles calculations" and is not an agent product itself. ABACUS-agent-tools can be customized, integrated, and secondarily developed into a powerful agent for solving complex material simulation calculation problems. We welcome other material simulation agent teams to use or participate in the development of ABACUS-agent-tools to provide a better experience for the majority of material simulation agent users.

Ways to Contribute and Experience

Repository address: https://github.com/deepmodeling/ABACUS-agent-tools/tree/develop

We encourage users to carry out testing, verification, and functional development based on the ABACUS long-term support version LTSv3.10.0 1 and the pseudopotential orbital set APNS-v1.0 2 to unify the behavior and experience of each interface of the agent.

[1]https://github.com/deepmodeling/abacus-develop/releases/tag/LTSv3.10.0
[2]https://aissquare.com/datasets/detail?pageType=datasets&name=ABACUS-APNS-PPORBs-v1%253Apre-release&id=326