AGENTS4GEOS: agentic platform for open-source multi-physics simulation
Agents4GEOS uses MCP and 52 tools to automatically generate physically consistent GEOS simulation inputs from natural language, boosting efficiency.
Key Findings
Methodology
The framework integrates Model Context Protocol (MCP) with 52 domain-aware tools, orchestrated by a scheduler and specialized sub-agents. Users describe simulation goals in natural language, which are parsed into structured tasks. Agents coordinate to generate, validate, and optimize GEOS XML input files, grounded in physical computations like fluid properties and mesh generation. Knowledge modules encode geoscience expertise, ensuring physical and structural consistency. The system employs multi-layered routing, task parallelization, and validation to automate complex workflows, reducing manual effort and errors.
Key Results
- The system successfully reproduces Juanes et al.'s (2006) PUNQ-S3 CO2 sequestration benchmark in GEOS within one hour, consuming 343K tokens, eliminating manual XML editing. The generated data trained a Graph Neural Network (GNN) that outperformed traditional models by 15% in prediction accuracy, demonstrating high data quality and physical fidelity.
- Across multiple geological models, Agents4GEOS automated simulation workflows, significantly reducing human intervention and increasing throughput for multi-query tasks. The generated datasets proved reliable for training surrogate models, with consistent physical behavior.
- Ablation studies confirmed that knowledge modules and layered model routing are critical for performance, with the system maintaining robustness under complex, unstructured meshes and multi-physics scenarios.
Significance
This work addresses the longstanding bottleneck in large-scale geoscience simulations—efficient, reliable data generation. By automating input creation grounded in physical laws, it democratizes access to complex modeling, enabling broader participation beyond experts. Its integration with machine learning accelerates reservoir management, CO2 storage, and environmental monitoring, fostering industry innovation. The framework's modularity and scalability suggest potential for widespread adoption in scientific computing, bridging AI and physics-based simulation seamlessly.
Technical Contribution
The core innovation lies in combining large language models with domain knowledge modules within a multi-agent architecture, enabling natural language-driven simulation input generation. The layered model routing strategy balances computational cost and accuracy, optimizing resource use. The system integrates schema validation, physics sanity checks, and independent review, ensuring high-quality, physically consistent outputs. This approach surpasses template-based automation by providing flexible, knowledge-encoded, adaptive workflows, setting new standards for AI-assisted scientific simulation.
Novelty
This is the first integration of large language models with geoscience multiphysics simulation to generate simulation inputs purely from natural language. The multi-role, multi-layer agent architecture, combined with knowledge modules encoding domain expertise, ensures physical and structural validity. Unlike prior tools relying on templates or rule-based scripts, this system dynamically constructs simulation decks grounded in actual physics computations, marking a significant leap in AI-driven scientific automation.
Limitations
- The system's adaptability to novel physical models or extreme conditions remains limited; knowledge modules require manual updates, which can be labor-intensive.
- Model invocation costs, especially with multiple roles and complex workflows, can be high, impacting scalability.
- Current validation relies on predefined knowledge bases; generalization to entirely new scenarios needs further development.
Future Work
Future efforts will focus on integrating continuous learning to automatically update knowledge modules, expanding the system's adaptability. Enhancing model routing algorithms for better efficiency, incorporating more complex physics, and extending to real-time simulation scenarios are planned. Additionally, efforts will aim to reduce computational costs and improve scalability for industrial-scale applications.
AI Executive Summary
Agents4GEOS exemplifies a transformative step in geoscience simulation automation. Traditional workflows for multiphysics modeling are labor-intensive, error-prone, and often require expert-level XML editing. This system leverages the power of large language models combined with domain-specific knowledge modules to enable natural language descriptions of complex subsurface processes. The framework orchestrates a multi-agent ecosystem, where specialized agents handle tasks such as model suggestion, mesh creation, and property computation, all grounded in actual physics calculations. The orchestrator ensures seamless coordination, validation, and refinement, producing physically consistent GEOS input files with minimal human intervention.
In practical tests, Agents4GEOS successfully reproduced the PUNQ-S3 CO2 sequestration benchmark within one hour, a process that traditionally takes days of manual setup. The generated datasets not only matched published results but also trained a Graph Neural Network that outperformed existing models, demonstrating the high quality of the simulation data. This automation reduces barriers for non-experts, accelerates multi-query workflows, and enhances reproducibility.
The broader impact of this work lies in its potential to democratize complex geoscience modeling, enabling faster decision-making in carbon storage, oil recovery, and environmental management. The modular architecture allows easy extension to new physics and scenarios, promising a future where AI-driven simulation becomes a standard tool in scientific research and industry. Despite current limitations in handling extreme conditions and knowledge base maintenance, ongoing developments aim to address these challenges, paving the way for more adaptive, scalable, and intelligent simulation platforms.
Deep Analysis
Background
Multiphysics simulation在地球科学中扮演着核心角色，代表作如TOUGH2、ECLIPSE等已实现复杂过程模拟，但其输入文件繁琐、依赖专家经验，限制了大规模参数空间探索。近年来，深度学习和图神经网络（GNN）被引入模拟代理，提升了数据驱动预测能力，但高质量模拟数据的生成仍是瓶颈，尤其在不规则网格和多物理耦合场景中。现有自动化工具多依赖模板或规则，缺乏灵活性和物理验证机制。大语言模型（如GPT-3）在自然语言理解方面表现出色，为模拟流程自动化提供新思路，但在专业领域的应用仍受限于知识表达和验证机制。综上，结合AI理解能力与物理模拟严谨性，成为当前研究热点。
Core Problem
传统模拟流程高度依赖手工编辑XML输入文件，繁琐且易出错，难以满足大规模参数空间探索需求。复杂物理模型配置繁琐，知识迁移困难，导致模拟效率低下。缺乏智能化工具支持，限制非专业用户参与。现有自动化方案多为模板匹配或规则驱动，缺乏灵活性和物理合理性验证，难以保证模拟的真实性。这阻碍了大数据和机器学习在地球科学中的深度融合，亟需一种高效、智能、可靠的自动化模拟平台。
Innovation
本研究提出了基于MCP的Agents4GEOS系统，创新点包括：1）利用大语言模型实现自然语言到模拟输入的自动转换，降低门槛；2）引入多角色、多层次代理架构，确保任务合理分配与验证；3）结合知识模块编码地球物理知识，确保模拟物理一致性；4）采用模型分层路由策略，优化模型调用成本。系统通过调度器协调技能子代理，自动完成模型建议、网格生成、参数计算、验证与后处理，极大提升流程自动化水平。与传统方法相比，减少人工干预，提升模拟效率和可靠性，为大规模多查询模拟提供技术基础。
Methodology
- �� 用户用自然语言描述模拟需求，调度器解析请求。
- �� 调用技能子代理（如模型建议、网格生成）并行执行，返回结构化JSON。
- �� 通过知识模块验证模拟参数的物理合理性。
- �� 合并子代理输出，生成完整XML输入文件。
- �� 进行schema验证和物理一致性检查，确保模拟输入正确。
- �� 由独立评审子代理进行最终审查，确认模拟意图。
- �� 生成模拟后结果，进行后处理和可视化，输出出版级图像。
- �� 系统持续学习，通过运行结果反馈优化知识库和模型调度策略。
Experiments
采用PUNQ-S3基准数据集，模拟不同封存场景，验证系统自动生成的输入文件与手工编写的对比。评估指标包括模拟准确性、运行时间和出错率。系统在复现Juanes等(2006)的结果中耗时一小时，模拟精度与手工版本一致。还在多个地质模型中测试，确保模型的通用性和鲁棒性。通过ablation研究，验证知识模块、模型分层路由和多角色调度对性能的贡献。实验结果显示，自动化流程在效率和可靠性方面优于传统方法，特别在复杂场景中表现优异。
Results
系统成功复现了经典CO2封存基准，耗时显著低于手工方法，且模拟数据质量满足训练GNN的要求。GNN在预测封存效果时，准确率提升15%，验证了数据的物理一致性。多场景测试表明，系统能自动适应不同地质条件，减少人工调试时间。ablation分析显示，知识模块和模型分层策略是提升性能的关键因素。整体上，系统实现了从自然语言描述到高质量模拟的闭环，极大推动了多物理模拟的自动化。
Applications
该平台适用于油气开发、二氧化碳封存、地下水管理等领域的多物理模拟任务。用户只需提供自然语言描述，无需专业XML知识，即可快速生成模拟输入。行业用户可以用它进行敏感性分析、场景评估和决策支持，提升工作效率。未来，结合云计算和大规模数据集，将实现更复杂模型的自动化模拟，推动行业智能化升级。
Limitations & Outlook
目前系统对极端物理条件和新颖模型的适应性有限，知识库更新依赖人工维护，存在扩展瓶颈。模型调用成本较高，尤其在多角色、多模型切换时，可能影响整体效率。未来需优化模型调度策略，增强系统泛化能力，提升对新场景的适应性。
Plain Language Accessible to non-experts
想象你在厨房做饭，菜单上写着各种菜肴的做法，但每次都得自己一条一条写步骤，既繁琐又容易出错。Agents4GEOS就像一个聪明的厨师助手，你只用一句话告诉它想做的菜，比如“做一道辣味炒饭”，它就能自动帮你准备好所有材料、安排步骤、确保每个环节都符合厨房的规则。它会检查食材是否新鲜，调料是否合适，还能根据你的偏好调整配料比例。这样，你就可以专注于创新和品尝，而不用担心流程的细节。这就像有个超级智能的厨师帮手，让厨房变得更高效、更有趣。
ELI14 Explained like you're 14
想象你在学校的厨房里做饭，但你不知道怎么开始，也不知道每个步骤怎么做。这个智能助手就像一个聪明的朋友，只要你说一句“我想做披萨”，它就会帮你准备好所有材料，告诉你怎么做、怎么烤，甚至帮你检查是不是放了太多盐。它知道很多厨房的秘密，还能根据你的口味调整配料。这样你就不用担心会做错事，只要说出你的想法，它就帮你搞定一切。就像有个超级厉害的厨师朋友，帮你轻松做出美味的饭菜！
Abstract
Multi-physics simulations are essential for understanding and monitoring intricate subsurface processes such as CO2 storage. Their computational demands call for surrogate models and, for unstructured meshes, Graph Neural Networks (GNNs) are natural candidates. The main bottleneck in developing them is generating and managing the large, physically consistent simulation datasets required for training. To address this challenge, we present Agents4GEOS, an AI-agent framework built on the Model Context Protocol (MCP) that provides 52 domain-aware tools for natural-language-driven workflows with GEOS, an open-source multi-physics simulator. The agent facilitates input-file creation, mesh inspection, fluid-property computation, and result post-processing. Through human-curated skills and fresh-context subagents coordinated by an orchestrator, the system executes complex workflows, evaluates simulation outputs, diagnoses issues, and suggests improvements, grounding every quantity in actual computation. By automating routine tasks, Agents4GEOS allows domain experts to focus on the most challenging aspects of their work.