跳转至

DREAMS：基于密度泛函理论的智能材料仿真研究引擎

文章背景与核心概要

大型语言模型（LLM）智能体能够为长周期科学工作流程提供强大的自动化支持，但由于上下文丢失、验证检查作弊以及产生看似合理却无效的结果，其数值输出常常缺乏可靠性。为了解决这一问题，研究人员推出了 DREAMS（基于密度泛函理论的智能材料仿真研究引擎，Density Functional Theory Based Research Engine for Agentic Materials Simulation）。这是一个分层的多智能体框架，构建有多层强大的安全防护机制。

DREAMS 针对每一个参数应用确定性检查和限定范围的 LLM 判定，将每个数值追溯到其注册来源，从而确保从工具调用阶段到报告生成的全过程绝对完整性。该系统在 Sol27LC 晶格常数基准测试中实现了低于 1% 的平均误差，成功复现了 CO/Pt(111) 难题中专家级的吸附能差异，并通过增强 L2（L2+）自动化水平展示了接近 L3 自动化的能力，为实现值得信赖、高通量的自主材料仿真开辟了道路。

DREAMS: Density Functional Theory Based Research Engine for Agentic Materials Simulation

Summary

Large language model (LLM) agents offer powerful automation capabilities for long-horizon scientific workflows, but their numerical outputs often lack reliability due to lost context, gamed verification checks, and plausible yet invalid results. To combat this, researchers introduce DREAMS (Density Functional Theory Based Research Engine for Agentic Materials Simulation), a hierarchical multi-agent framework built with a robust multi-tier safety guard.

DREAMS applies deterministic checks and scoped LLM judgments parameter-by-parameter, tracing every value back to its registered source to ensure absolute integrity from tool-call time to report generation.

Key Achievements & Features:

  • High Accuracy: Achieves average errors below 1% on the Sol27LC lattice-constant benchmark and successfully reproduces expert-level adsorption-energy differences on the CO/Pt(111) puzzle.
  • Uncertainty Quantification: Quantifies functional-driven uncertainty using Bayesian ensemble sampling (confirming FCC site preference at the GGA level).
  • Robust Verification: Verifies every essential step (at roughly 13× the input tokens compared to unguarded systems), with tunable verification layers and transferable judge rules across five different models.
  • Autonomous Milestone: Operates at an enhanced L2 (L2+) automation level, displaying capabilities that closely approach L3 automation for trustworthy, high-throughput materials simulations.

Article Metadata

Field Details
arXiv ID arXiv:2507.14267 [cs.AI]
Subjects Artificial Intelligence (cs.AI); Materials Science (cond-mat.mtrl-sci)
Authors Ziqi Wang, Hongshuo Huang, Hancheng Zhao, Changwen Xu, Shang Zhu, Jan Janssen, Venkatasubramanian Viswanathan
Submission Dates Submitted on 18 Jul 2025; Revised on 11 Aug 2026 (v2)
Full-Text Links View PDF | HTML Version | TeX Source
License Creative Commons Attribution 4.0 license icon

Abstract

大语言模型（LLM）智能体能够执行长周期的科学工作流程，但其数值输出很难让人信赖：智能体会丢失上下文、在验证检查中作弊，并可能产生大量看似合理实则无效的结果。我们引入了基于 DFT 的智能材料仿真研究引擎（DREAMS），这是一个针对密度泛函理论（DFT）的分层多智能体框架，其核心围绕多层安全防护网构建。该防护网在存在明确标准的地方应用确定性检查，在其他地方应用限定范围的 LLM 判定，每次评估一个参数并将每个数值追溯到其注册来源。验证过程从工具调用阶段一直延伸到报告生成阶段：在工具调用阶段，虚构的、伪造的或无来源的数值在进入工作流程之前就会被拒绝；在报告生成阶段，裁判会对每个声明背后的完整来源图进行审计；共享画布则在数百个步骤中保持了信息完整性。DREAMS 在 Sol27LC 晶格常数基准测试中实现了低于 1% 的平均误差，在 CO/Pt(111) 难题上复现了专家级的吸附能差异，并通过贝叶斯集成采样量化了泛函驱动的不确定性，确认了广义梯度近似（GGA）水平下的面心立方（FCC）格位偏好。与未加防护的同类系统（仅 81% 的关键步骤成功就得到了接近正确的答案）相比，受防护的系统以大约 13 倍的输入 Token 验证了每个关键步骤；验证层可以单独禁用，以平衡可信度与成本，并且经过调优的裁判规则可以在五个裁判模型之间迁移。DREAMS 在增强的 L2（L2+）自动化水平下运行，并展示了接近 L3 自动化的能力，为实现值得信赖、高通量的自主材料仿真提供了一条路径。

Large language model (LLM) agents can execute long-horizon scientific workflows, but their numerical outputs are difficult to trust: agents lose context, game verification checks, and can produce large volumes of plausible yet invalid results. We introduce the DFT-based Research Engine for Agentic Materials Simulation (DREAMS), a hierarchical multi-agent framework for density functional theory (DFT) built around a multi-tier safety guard. The guard applies deterministic checks wherever explicit criteria exist and scoped LLM judgment elsewhere, evaluating one parameter at a time and tracing every value to its registered source. Verification extends from tool-call time, where fabricated, laundered, or unsourced values are rejected before entering the workflow, to report time, where a judge audits the full provenance graph behind every claim; a shared canvas preserves information integrity across hundreds of steps. DREAMS achieves average errors below 1% on the Sol27LC lattice-constant benchmark, reproduces expert-level adsorption-energy differences on the CO/Pt(111) puzzle, and quantifies functional-driven uncertainty with Bayesian ensemble sampling, confirming the face-centered-cubic (FCC) site preference at the generalized gradient approximation (GGA) level. Compared with its unguarded counterpart, which reached a nearly correct answer while only 81% of its essential steps succeeded, the guarded system verifies every essential step at approximately 13 times the input tokens; verification layers can be disabled individually to balance trustworthiness against cost, and the tuned judge rules transfer across five judge models. DREAMS operates at an enhanced L2 (L2+) automation level and demonstrates capabilities approaching L3 automation, providing a path toward trustworthy, high-throughput autonomous materials simulation.

External Resources & Tools

外部资源与工具