智能体程序:计算材料科学中一种新兴的科学软件形态
文章背景与核心概要
传统计算材料科学长期遵循“计算机处理纯算法任务、人类负责所有科学判断”的协作模式。然而,随着大语言模型(LLM)智能体框架的飞速发展,一种被称为“智能体程序(agentic programs)”的新型科学软件正在崭露头角。这类程序有机地结合了确定性算法与受限的LLM推理能力、任务特定的验证机制、情境演化成熟机制,以及生产环境中的完全执行委托。
本文以 DeMARS 为例,生动展示了这一全新概念。DeMARS 是一种专为根据实验测得的无序晶体结构构建原子模型而设计的智能体程序,标志着科学计算向高度自主化迈出了重要一步。
Agentic Programs: An Emerging Form of Scientific Software in Computational Materials Science
Summary
This paper explores the paradigm shift in computational materials science, moving away from traditional models where computers handle purely algorithmic tasks while humans supply all scientific judgment. The authors introduce agentic programs—a new breed of scientific software enabled by recent Large Language Model (LLM) agent frameworks. These programs seamlessly integrate deterministic algorithms with bounded LLM-based reasoning, task-specific verification, episodic maturation, and complete execution delegation in production. The concept is demonstrated through DeMARS, a practical agentic program designed to construct atomistic models from experimentally measured disordered crystal structures.
本文探讨了计算材料科学中的范式转变,摆脱了传统模式(即计算机处理纯算法任务,而人类提供所有科学判断)。作者引入了智能体程序(agentic programs)——这是由近期大语言模型(LLM)智能体框架赋能的一种新型科学软件。这些程序将确定性算法与有界的基于LLM的推理、任务特定验证、情境演化成熟机制以及生产中的完全执行委托无缝集成。这一概念通过 DeMARS 得到了展示,DeMARS 是一种实用的智能体程序,旨在根据实验测得的无序晶体结构构建原子模型。
Publication Details
- arXiv Identifier: arXiv:2609.00795 [cond-mat.mtrl-sci]
- Submission Date: September 1, 2026
- Authors:
- Yunsung Lim
- Haekwan Jeon
- Jaesun Kim
- Jisu Kim
- Seungwu Han
- Primary Subject: Materials Science (
cond-mat.mtrl-sci) - Secondary Subject: Artificial Intelligence (
cs.AI) - Citation:
arXiv:2609.00795 [cond-mat.mtrl-sci]
出版详情
- arXiv 标识符: arXiv:2609.00795 [cond-mat.mtrl-sci]
- 提交日期: 2026年9月1日
- 作者:
- Yunsung Lim
- Haekwan Jeon
- Jaesun Kim
- Jisu Kim
- Seungwu Han
- 主学科: 材料科学 (
cond-mat.mtrl-sci)- 辅学科: 人工智能 (
cs.AI)- 引用格式:
arXiv:2609.00795 [cond-mat.mtrl-sci]
Abstract
Computational materials science has traditionally delegated algorithmic tasks to computers while leaving scientific judgments to humans. We argue that recent LLM-based agent harnesses enable an emerging form of scientific software, agentic programs, that combine deterministic algorithms with bounded LLM-based judgment, task-specific verification, episodic maturation, and complete delegation in production. We illustrate this concept with DeMARS, an agentic program for constructing atomistic models from experimentally measured disordered crystal structures.
摘要
传统计算材料科学一直将算法任务委托给计算机,而将科学判断留给人类。我们认为,近期基于 LLM 的智能体框架催生了一种新兴的科学软件形态——智能体程序,它将确定性算法与有界的基于 LLM 的判断、任务特定验证、情境演化成熟以及生产环境中的完全委托结合在一起。我们以 DeMARS 为例阐述了这一概念,DeMARS 是一种用于从实验测得的无序晶体结构构建原子模型的智能体程序。
Access Links & Resources
- Full-Text Options:
- View PDF
- HTML Version (Experimental)
- TeX Source
- External Indices & Citations:
- NASA ADS
- Google Scholar
- Semantic Scholar
- CrossRef DOI
访问链接与资源
- 全文选项:
- 查看 PDF
- HTML 版本(实验性)
- TeX 源码
- 外部索引与引用:
- NASA ADS
- Google Scholar
- Semantic Scholar
- CrossRef DOI