文章背景与核心概要
从观测数据中发现底层的控制偏微分方程(PDE)是各个科学领域中的一项核心挑战。传统的科学方法(如稀疏回归、符号回归以及早期的纯大语言模型技术)往往受制于诸多限制,例如刚性的预定义函数库、对噪声的敏感性、模型幻觉以及缺乏迭代修正机制。
为了克服这些痛点,本文作者推出了 MAGE(Multimodal Agentic Governing Equation Discovery,多模态智能体控制方程发现框架)。受经典科学方法论(即观察、假设与证伪)的启发,MAGE 将偏微分方程的发现过程构建为一个由置信度驱动的假设验证循环,显著提升了方程发现的准确性与鲁棒性。
看见、假设与验证:用于发现控制偏微分方程的多模态智能体框架
arXiv ID: arXiv:2608.27869 [cs.AI]
DOI: 10.48550/arXiv.2608.27869
提交时间: 2026年8月28日
主要学科: 人工智能 (cs.AI)
作者: Sarang Manoj Pekhale, Amartya Roy, Rajat Sarkar, Souvik Chakraborty
📌 总结 (Summary)
Discovering governing partial differential equations (PDEs) from observational data is a fundamental challenge across scientific domains. Traditional approaches—such as sparse-regression, symbolic-regression, and early Large Language Model (LLM) techniques—frequently suffer from limitations such as rigid predefined libraries, noise sensitivity, hallucinations, and a lack of iterative refinement.
To overcome these issues, the authors introduce MAGE (Multimodal Agentic G}overning Equation Discovery). Inspired by the scientific method (observation, hypothesis, and falsification), MAGE frames PDE discovery as a confidence-governed hypothesis validation loop**.
从观测数据中发现底层的控制偏微分方程(PDE)是科学领域的一项基本挑战。传统方法(如稀疏回归、符号回归和早期的大语言模型(LLM)技术)经常受到诸如刚性预定义函数库、对噪声敏感、产生幻觉以及缺乏迭代改进等限制。
为了克服这些问题,作者引入了 MAGE(Multimodal Agentic Governing Equation Discovery,多模态智能体控制方程发现框架)。受科学方法(观察、假设和证伪)的启发,MAGE 将 PDE 的发现过程构建为一个由置信度控制的假设验证循环。
🔬 MAGE 框架 (The MAGE Framework)
MAGE orchestrates four role-specialized collaborative agents to discover governing laws library-free:
- Differential Observer: Computes derivatives and generates essential diagnostic visualizations.
- Phenomenology Extractor (VLM-powered): Distills qualitative insights and cues directly from multimodal diagnostics.
- Governing Law Synthesizer (LLM-driven): Proposes candidate equations without relying on a predefined library.
- Equation Arbiter: Fits coefficients and assigns confidence scores to hypotheses.
Discovery proceeds iteratively via an explicit accept-reject protocol until the top candidate clears a user-specified confidence threshold.
MAGE 协调四个角色各异的协作智能体,在无需预定义函数库的情况下发现控制规律:
- 微分观察器(Differential Observer): 计算导数并生成必要的诊断可视化图表。
- 现象学提取器(Phenomenology Extractor,由 VLM 驱动): 直接从多模态诊断结果中提炼定性见解和线索。
- 控制律综合器(Governing Law Synthesizer,由 LLM 驱动): 在不依赖预定义函数库的前提下提出候选方程。
- 方程仲裁器(Equation Arbiter): 拟合系数并为假设分配置信度分数。
发现过程通过显式的接受-拒绝协议进行迭代,直至最佳候选方程达到用户指定的置信度阈值。
📊 核心结果 (Key Results)
- Exact Structural Recovery: Achieved 8/8 exact structural recovery on the evaluated canonical PDE suite.
- Coefficient Accuracy: Delivered the lowest coefficient error among compared methods on 7/8 systems, achieving improvements of up to 4 orders of magnitude (with a geometric-mean improvement of ~3 orders of magnitude).
- Complex Geometries & Real-World Data: Successfully recovered expected operators across two complex geometries and selected a cubic restoring-force model with a held-out \(R^2 = 0.98538\) on a real laboratory sensor record.
- 精确结构恢复: 在评估的标准 PDE 测试集上实现了 8/8 的精确结构恢复。
- 系数准确性: 在 7/8 的系统中实现了对比方法中最低的系数误差,实现了高达 4个数量级 的提升(几何平均提升约为 3 个数量级)。
- 复杂几何与真实世界数据: 成功在两种复杂几何结构中恢复出预期的算子,并在真实的实验室传感器记录上挑选出了保留测试集 \(R^2 = 0.98538\) 的三次恢复力模型。
🔗 全文与访问链接 (Full-Text & Access Links)
🌐 参考文献与外部工具 (References & External Tools)
- Bibliographic Tools: NASA ADS, Google Scholar, Semantic Scholar
- Code & Catalogs: CatalyzeX Code Finder, Hugging Face, DagsHub
- Community Platforms: alphaXiv, ScienceCast
- 文献计量工具: NASA ADS, Google Scholar, Semantic Scholar
- 代码与目录: CatalyzeX Code Finder, Hugging Face, DagsHub
- 社区平台: alphaXiv, ScienceCast