A2MC-elm: An AI-Driven Autonomous Calibration Workflow for Earth System Models

A2MC-elm is an intelligent adaptive multi-objective calibration workflow developed by Lawrence Berkeley National Laboratory, targeting E3SM land models (e.g., ELM-FATES, ELM-ReSOM). It uses AI agents to achieve autonomous parameter calibration, integrating interpretable diagnostics, RAG/GraphRAG knowledge retrieval, hypothesis-driven optimization, and adaptive memory mechanisms. Its goal is to address the problems of traditional manual calibration, which is time-consuming and struggles to find global optimal solutions.

Earth System Models (ESMs) are key tools for understanding and predicting climate change. Among them, the land component ELM of E3SM is responsible for simulating processes such as vegetation growth and carbon cycles. However, parameter calibration for land models faces huge challenges: models contain hundreds to thousands of parameters, traditional manual calibration is time-consuming and hard to find global optimal solutions; with the emergence of complex models like ELM-FATES and ELM-ReSOM, the problem becomes more severe.

The A2MC method introduces LLM agents into the field of calibration. Its core components include: 1. AI agent-driven autonomous calibration (intelligent scheduling, understanding physical meanings); 2. Multi-objective optimization framework (finding optimal trade-offs on the Pareto front); 3. Interpretable diagnostic system (transparently presenting decision logic); 4. RAG/GraphRAG knowledge retrieval (obtaining support from literature and documents);5. Hypothesis-driven optimization cycle (scientific method paradigm);6. Adaptive memory mechanism (recording strategies for continuous evolution).

A2MC-elm is a modular framework that adapts to multiple ELM configurations: ELM-FATES (simulating vegetation dynamics and community succession), ELM-ReSOM (detailed simulation of soil carbon cycles), and standard ELM (basic land processes). Its flexibility can serve different research needs from site to global scales.

The significance of A2MC-elm includes: accelerating scientific discovery (reducing calibration time from months to days/hours); improving model reliability (reducing human bias and identifying structural defects); knowledge accumulation and inheritance (capturing expert knowledge to help novices); interdisciplinary integration (combining AI technology with Earth system science to demonstrate the potential of AI for Science).

Current limitations: high computational cost (both LLM calls and ESM simulations are intensive); limited generalization ability (mainly targeting the ELM family); strict verification of AI calibration schemes is needed. Future directions: multi-model integrated calibration, real-time data assimilation, uncertainty quantification, and expansion to other ESM components such as atmosphere/ocean.

A2MC-elm represents an important progress in the intersection of AI and Earth system science, demonstrating the role of LLM agents in scientific computing with deep domain knowledge. It suggests that future scientific computing may be a new paradigm where humans and AI agents collaborate to explore complex problems, which is worthy of attention from researchers in climate simulation, ecological modeling, and scientific AI applications.