AutoClimDS: Climate Data Science Agentic AI -- A Knowledge Graph is All You Need

AutoClimDS is an Agentic AI system that leverages a curated climate knowledge graph as its essential structural foundation to automate end-to-end climate data analysis and reproduce scientific figures from natural language, overcoming the limitations of general-purpose LLMs in identifying authoritative datasets and constructing valid scientific workflows.

The Gist

Imagine you are trying to bake a complex, famous cake (like a climate study) that was published in a cookbook. You have a brilliant, creative chef (an AI) who knows how to bake anything, but you've handed them a kitchen with no recipe book, no labeled ingredients, and no idea where the flour is stored.

If you ask this chef, "Make me the NYC sea-level cake," they might guess the ingredients, try to bake with sand instead of flour, or simply give up because they don't know where to find the right tools. This is the current problem with climate data science: we have amazing AI, but the data is scattered, messy, and hard to find.

AutoClimDS is the solution. It's a new system that gives the AI chef a super-organized, magical library (called a Knowledge Graph) before they start cooking.

Here is how it works, broken down into simple analogies:

1. The Problem: The "Lost in the Library" Scenario

Right now, climate data is like a massive library where books are thrown on the floor, written in different languages, and have no titles.

• The Old Way: If you ask a smart AI (like a general chatbot) to find a specific dataset, it might guess the title, hallucinate a fake book, or get stuck because it doesn't know the specific rules for checking out that book (like needing a special key or password).
• The Result: The AI fails to do the science because it lacks the "map" to the data.

2. The Solution: The "Master Blueprint" (The Knowledge Graph)

The authors built a Knowledge Graph (KG). Think of this not just as a list of links, but as a living, breathing blueprint of the entire climate data universe.

• It knows the "What": It knows exactly what "Precipitation" means across different datasets.
• It knows the "Where": It knows exactly which server holds the data for New York City.
• It knows the "How": This is the magic part. It doesn't just say "Data is here." It says, "To get this data, you must click this button, use this password, and then convert the file from this format to that format." It encodes the procedures (the steps) just like a recipe.

3. The Team: The "Conductor" and the "Specialists"

AutoClimDS isn't just one AI; it's a team of agents working together, directed by a conductor:

• The Conductor (Orchestrator): Takes your simple request ("Show me sea levels in NYC") and breaks it down into steps.
• The Librarian (Discovery Agent): Uses the Blueprint to find the exact right books (datasets) in the massive library. It ignores the fake ones and finds the real, high-quality data.
• The Runner (Acquisition Agent): Goes to the server, uses the correct keys (passwords/APIs) found in the Blueprint, and grabs the data. If the first door is locked, it knows to try the back door.
• The Chef (Modeling Agent): Takes the raw ingredients, washes them (preprocessing), cooks them (analysis), and plates the dish (creates the chart).

4. The Magic Trick: "A Knowledge Graph is All You Need"

The paper's title sounds bold: "A Knowledge Graph is All You Need."
This doesn't mean we don't need the AI chef. It means that without the Blueprint, the chef is useless.

• The Experiment: The team asked a top-tier AI (GPT-5.1) to recreate a famous scientific chart about sea levels. Without the Blueprint, the AI failed. It couldn't find the right data or follow the steps.
• The Win: When they gave the same AI the Blueprint (AutoClimDS), it successfully recreated the exact same chart, with perfect numbers, just by listening to a simple sentence.

5. Why This Matters

• Democratization: You don't need to be a computer expert to do climate science anymore. You just need to speak naturally.
• Reliability: Because the system follows a strict, pre-verified map, it doesn't "hallucinate" (make things up). It produces results that scientists can trust and reproduce.
• Speed: What used to take a human researcher weeks of hunting for data and writing code, this system does in minutes.

The Bottom Line

Think of AutoClimDS as giving a super-intelligent robot a GPS and a detailed instruction manual for the climate world. Before, the robot was smart but blind. Now, with the Knowledge Graph as its eyes and memory, it can navigate the complex world of climate data, find the right information, and do the science for you.

It proves that to build the future of AI science, we don't just need smarter brains (AI models); we need better libraries (Knowledge Graphs) to teach them how to think.

Technical Summary

1. Problem Statement

Climate data science faces significant barriers to entry and reproducibility due to:

• Data Fragmentation: Observational and simulation datasets are scattered across heterogeneous sources (e.g., NASA, NOAA, CMIP6) with inconsistent formats and metadata.
• Technical Complexity: Researchers require deep expertise to navigate authentication protocols, API structures, and preprocessing workflows.
• Limitations of Current AI: General-purpose Large Language Models (LLMs) lack the "structured scientific memory" required for autonomous workflows. Without specific guidance, they hallucinate dataset names, fail to identify authoritative sources, or cannot construct valid retrieval pipelines.
• Static Retrieval Systems: Existing knowledge graphs (KGs) often provide only conceptual vocabularies or keyword search, lacking the procedural knowledge (executable access paths, variable mappings, authentication steps) needed for agentic reasoning.

2. Methodology: AutoClimDS Architecture

The authors propose AutoClimDS, a Minimum Viable Product (MVP) Agentic AI system that integrates a curated Climate Knowledge Graph (KG) with generative AI agents. The system operates on the premise that while LLMs provide reasoning capabilities, the KG provides the essential structural substrate for reliable execution.

A. Knowledge Graph Construction

The KG serves as the "structured scientific memory," encoding not just data locations but procedural workflow knowledge.

• Data Sources: Integrated metadata from NASA CMR, NOAA OneStop, ERA5, and CMIP6.
• Scale: ~1.48 million nodes (45 types) and ~5.8 million edges (39 relationship types), covering ~208,000 datasets.
• Key Innovations in the Schema:
  • Procedural Links: Encodes executable access links, authentication protocols (e.g., NASA Earthdata tokens), and API types (OpenDAP, ERDDAP, REST).
  • Semantic Variable Mapping: Uses a fine-tuned ClimateBERT model (99.17% accuracy) to map natural language queries to standardized climate variables (e.g., mapping "rainfall" to specific CESM variable IDs).
  • Geospatial Hierarchy: Classifies spatial scopes (global, continental, country) and enables traversal via hasLocation edges.
  • Link Scoring: Assigns weights (1–5) to data links based on accessibility (e.g., direct download = 5, visualization service = 1) and verifies endpoint reliability.

B. Agentic AI System

The system utilizes LangChain with ReAct (Reasoning + Acting) logic, powered by Bedrock Claude Sonnet 4, and orchestrates three specialized agents:

1. Data Discovery Agent:
   • Converts user queries into 384-dimensional vectors for semantic search within the KG (using Amazon Neptune Analytics).
   • Performs multi-criteria filtering (temporal, spatial, resolution) by combining vector search with structured OpenCypher queries.
   • Reranks retrieved datasets based on link weights to prioritize direct, authenticated access.
2. Data Acquisition Agent:
   • Executes retrieval protocols using pre-configured tokens or dynamic discovery (web scraping documentation to find new API endpoints if standard links fail).
   • Handles heterogeneous formats (NetCDF, HDF, CSV) via automated transformation functions ($T: D \to \hat{D}$) to standardize data.
   • Operates in a sandboxed Python environment with strict network whitelisting (NASA, NOAA, AWS S3) for security.
3. Climate Modeling/Analytics Agent:
   • Specialized for simulation data (ERA5/CMIP6), handling multi-model ensembles, spatial subsetting, and unit conversions.
   • Generates visualizations and statistical summaries.

C. Error Recovery & State Management

• Fallback Mechanisms: If data acquisition fails, the system iterates through reranked links, attempts dynamic web discovery, or traverses the KG to find alternative datasets with the same variables.
• Guardrails: Implements cycle detection (terminating if thought embeddings are too similar), maximum iteration limits, and circuit breakers to prevent infinite loops.
• State Persistence: Uses SQLite and LangChain memory buffers to maintain session continuity across complex, multi-step workflows.

3. Key Contributions

1. Procedural Knowledge Graph: Unlike prior KGs that offer static vocabularies, this system encodes executable workflows, including authentication, variable mapping, and access protocols, making the graph "machine-interpretable" for agents.
2. Agentic Workflow Reproduction: Demonstrates that an AI agent, guided by this KG, can autonomously reproduce published scientific figures from natural language instructions alone, a task where state-of-the-art LLMs (e.g., GPT-5.1 without KG) fail.
3. Hybrid Search Architecture: Combines vector-based semantic search with structured graph traversal (OpenCypher) to handle complex, multi-constraint scientific queries.
4. Democratization Framework: Provides a pathway for non-experts to access complex climate data by lowering technical barriers through human-AI collaboration.

4. Results & Case Studies

The system was validated through two primary case studies, successfully replicating published analyses without human intervention in data selection or preprocessing:

• Case Study 1: Sea Level Trends (Observational Data)

  • Task: Reproduce figures from the NYC Climate Risk Information 2022 (NPCC4) report using only natural language prompts.
  • Outcome: AutoClimDS successfully calculated sea level trends (0.112 in/yr vs. reported 0.11 in/yr) and Vertical Land Motion contributions with high statistical fidelity.
  • Comparison: Baseline models (LinkClimate, GPT-5.1) failed to locate authoritative datasets or construct valid workflows, often hallucinating sources.

• Case Study 2: Temperature Projections (Simulation Data)

  • Task: Analyze future temperature projections for NYC using CMIP6 and ERA5 datasets.
  • Outcome: The system autonomously queried the KG, authenticated with ESGF/CDS APIs, performed spatial subsetting, computed ensemble means, and generated uncertainty visualizations.
  • Significance: Demonstrated the system's ability to handle complex multi-model ensemble data, a task where GPT-5.1 failed without KG guidance.

5. Significance and Conclusion

• The "Knowledge Graph is All You Need" Thesis: The paper argues that while LLMs are powerful, they are insufficient for autonomous scientific reasoning without a structured memory layer. The KG acts as the irreplaceable structural foundation that prevents hallucination and enables reliable execution.
• Cloud-Native & Extensible: The system is designed to be cloud-agnostic (though implemented on AWS) and modular, allowing the community to extend the KG with new datasets, tools, and ontologies.
• Impact: By encoding procedural knowledge, AutoClimDS transforms general-purpose AI into domain-competent scientific agents. This approach accelerates discovery, ensures reproducibility, and democratizes climate research by enabling policy-makers, educators, and citizen scientists to interact with complex data through natural language.

In summary, AutoClimDS proves that integrating a rich, procedurally encoded Knowledge Graph with Agentic AI is the critical missing link for achieving autonomous, reproducible, and accessible climate data science.