Daniel Saedi Nia Post Baccalaureate Intern, Oak Ridge National Laboratory, USA

Title: An Agentic Approach to CMIP Data Discovery using Large Language Models (LLMs)

Abstract: Earth system models and observational datasets hosted through the Earth System Grid Federation (ESGF), including the Coupled Model Intercomparison Project (CMIP), are critical resources for Earth science research. However, complex metadata conventions and structures often hinder data discovery, particularly for researchers unfamiliar with ESGF/CMIP workflows. To lower these barriers, the ESGF2-US team is exploring large language model (LLM)-driven approaches for natural-language interaction with ESGF services and CMIP data. Initial efforts focused on fine-tuning an open-source foundation model (Meta’s Llama 3.1 8B) for code generation using instruction-output pairs derived from ESGF and CMIP workflows (Saedi Nia et al., 2025). Building on this, we introduce an agentic system that combines LLM reasoning with specialized tools for dataset discovery, metadata interpretation, and CMIP variable identification across CMIP6 and emerging CMIP7 infrastructures. Furthermore, the system can translate natural-language queries into executable Python workflows using the intake-esgf library while dynamically interacting with metadata tools to resolve search facets. We demonstrate this capability through an end-to-end workflow in which a user moves from an initial question to discovering available CMIP variables, refining constraints, and ultimately generating runnable code to access the selected datasets. This work highlights the potential of agentic AI systems to streamline scientific workflows and enable next-generation data access within evolving ESGF ecosystems.

Tom Bearpark Post-Doctoral Fellow, University of Exeter, UK

Title: The economic geography of climate risk

Abstract: Projected temperature changes are variable in both their magnitude and geography. This paper studies how nonlinear damages and general equilibrium forces filter this climate risk across time and space. To do so, we build a tractable dynamic spatial model linking countries through trade and migration, and derive analytical first- and second-order welfare approximations that decompose the mean and variance of welfare changes into damage function, trade, and migration components. Using an ensemble of temperature projections from CMIP-6 and an empirically estimated damage function, we show that climate change-induced welfare risk is large. The standard deviation of country-level projected welfare loss across temperature projections is on average over 8% — compared to an average welfare loss of 13\% — and is spatially unequal. Climate risk inequality is half as large as global income inequality. We show that spatial linkages reshape not only the level of climate damages, but also the spatial distribution of climate risk: accounting for trade and migration can reduce the standard deviation of welfare changes by up to 40% in low-income, internationally integrated nations that can diversify their exposure to local shocks. 

Anastasia Romanou Physical Research Scientist, NASA, Columbia University, USA

Title: Bridging ESMs and ML: High-Fidelity Climate Emulation with ModelEMU

Abstract: Robust and accurate multi-scale predictions of changes in surface properties, beyond just surface air temperature and precipitation, need to be communicated quickly and efficiently to decision makers but also scientists. Such predictions, typically obtained from Earth system models (ESMs), use a small set of standardised emission scenarios that are available once in a decade rather than more frequently released bespoke scenarios, which however can be computationally prohibitive. Machine learning (ML) emulators of climate model predictions trained on sparse scenarios offer an efficient alternative, but their predictive accuracy and uncertainty quantification is inconsistent. In this paper we assess the skill of the CNN-LSTM emulator within the established benchmark framework ClimateBenchv1.0 focusing on its ability to emulate climate responses of a broad set of climate diagnostics and extreme event metrics. We aim to assess whether such climate emulators can be used to produce operational scenario predictions. The ML architecture (modelEMU) is trained on low- and high-end CMIP forcing scenario simulations and optimised to a single parameter, here surface air temperature, in one mid-range future scenario simulated with the NASA-GISS modelE climate model. ModelEMU shows good predictive skill for the ensemble mean of several ESM variables in the seen and unseen scenarios, but the skill is reduced in overshoot scenarios and for the more non-linear fields such as deep ocean temperatures and air-sea fluxes of CO2. Optimising the emulator to a single parameter, such as the surface air temperature, provides an effective diagnostic tool for interpreting regional biases of the emulated fields.