As artificial intelligence finds more and more applications within scientific contexts, the question on how to utilize it without sacrificing scientific integrity comes up naturally. In this context, FrevaGPT is a novel system that leverages LLMs such as GPT-4o and GPT-4o-mini to enable users to perform advanced analyses. It allows the loading and analysis of climate datasets by the LLM and moves the basis of truth to generated code, which can be checked by the user. Its backend was developed and deployed using modern software components (e.g. Rust, Python, Podman), focussing on correctness and reliability. The backend of FrevaGPT and its API is presented and the way it integrates into the larger Freva ecosystem as well as the role it plays in the improvements of ad-hoc analyses for climate data is discussed. Additionally, a suite of scientific prompts is explored to evaluate the capabilities of GPT-4o and GPT-4o-mini and how they compare in climate data analysis tasks. The prompts differ both in difficulty and complexity as well as in the requested output type: from a single number, to a graph, to a plot. This evaluation revealed that while both models demonstrated potential, GPT-4o outperformed GPT-4o-mini in handling more complex tasks involving diverse knowledge domains and programming requirements. GPT-4o-mini exhibited a higher tendency for errors and struggled with issues such as mismatched data dimensions, yet it remained a competitive, cost-effective alternative for simpler tasks. The findings highlight FrevaGPT as a significant step towards integrating advanced AI technologies into Earth sciences, bridging the gap between computational complexity and accessibility. 

How to cite: Willmann, S., Ludwig, T., and Kadow, C.: Evaluation of GPT-4o and GPT4o-mini for Climate Data Analysis with a novel tool-call software connecting different LLMs with an HPC, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-18345, https://doi.org/10.5194/egusphere-egu25-18345, 2025.

The chabot writing poems can do climate analysis? Large Language Models (LLMs) promise a paradigm shift as chat-based geoscientific research transformers (chatGRT) by removing technical barriers and empowering scientists to focus on deeper, more innovative inquiries. We introduce FrevaGPT, an LLM-driven “scientific assistant” integrated into Freva, the Free Evaluation System for climate data analysis on high performance computers. FrevaGPT automatically translates natural language questions into traceable, editable, and reusable scripts; retrieves relevant data and publications; executes the analyses; and visualizes the results - the scientist can focus on what matters most: science. By tapping into a wide repository of climate datasets, FrevaGPT ensures transparent, reproducible workflows and lowers the threshold for advanced data handling. Its co-pilot functionality not only delivers answers, tables, and plots, but also proactively suggests next steps, points to relevant climate modes and events, and presents associated scientific findings. Through integrated approaches to model evaluation and observational data comparisons, FrevaGPT accelerates scientific discovery and fosters interdisciplinary collaboration. Real-world use cases highlight FrevaGPT’s capacity to guide researchers beyond routine analysis, freeing them to explore innovative questions and deepen their understanding of complex climatic phenomena. As a pioneering application of LLMs in climate science, FrevaGPT illustrates how such tools can fundamentally reshape research processes, unleashing new possibilities for efficiency and creative exploration in the geosciences.

How to cite: Kadow, C., Saynisch-Wagner, J., Willmann, S., Lentz, S., Baehr, J., Sieck, K., Oertel, F., Wentzel, B., Ludwig, T., and Bergemann, M.: FrevaGPT: A Large Language Model-Driven Scientific Assistant for Climate Research and Data Analysis, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-15507, https://doi.org/10.5194/egusphere-egu25-15507, 2025.

We present LLM-Enhanced CMIP6 Search, a Python-based tool built with LangChain and LangGraph frameworks that simplifies the discovery of and access to Coupled Model Intercomparison Project Phase 6 (CMIP6) climate data through natural language processing. By combining Large Language Models (LLMs) with retrieval-augmented generation (RAG), our system translates user queries into precise CMIP6 search parameters, bridging the gap between researchers' information needs and CMIP6's structured metadata system. The tool employs a single LLM agent coordinating three specialized tools: a search tool that maps natural language to CMIP6 parameters (such as model, experiment, and variable identifiers), an access tool that both verifies data availability and generates ready-to-use Python code for retrieval, and an adviser tool that helps refine search criteria. To improve search accuracy, we developed a refined database of CMIP6 metadata descriptions, optimizing vector-based similarity matching between user queries and technical CMIP6 terminology, providing a foundation for more intuitive climate data discovery.

How to cite: Shapkin, B., Pantiukhin, D., Kuznetsov, I., Jost, A. A., and Koldunov, N.: LLM-Enhanced CMIP6 Search, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-7059, https://doi.org/10.5194/egusphere-egu25-7059, 2025.

PANGAEA GPT is a Large Language Model (LLM) multi-agent framework that aims to streamline the work of geoscientists with the diverse Earth system datasets held in the PANGAEA archive (pangaea.de), a widely used data repository in Earth and Environmental Sciences. Built on top of the LangChain library and the LangGraph framework, it uses a multi-agent collaboration approach with a centralized supervisor agent that interprets incoming user queries and then coordinates specialized agents according to task requirements. These specialized agents include the Search Agent, which performs data lookups via API requests to PANGAEA and locates related publications via Crossref (to further answer questions about what has been published based on a particular dataset). They also include an orchestra of Data Agents configured in different modes - such as "oceanographer," "ecologist," or "geologist" - to perform dataset-specific analyses. Each Data Agent operates within a dedicated Python environment that allows for code manipulation, data analysis, visualization, and iterative refinement of results. The Supervisor Agent then aggregates the output from these Data Agents and delivers a consolidated response back to the user (including generated analysis scripts). The current framework has been shown to excel at providing a list of relevant datasets, locating related publications, and performing statistical analysis upon user request, greatly simplifying data discovery and use for geoscientists. In addition to the rapid discovery, analysis, and visualization of heterogeneous datasets, a particularly valuable end goal of PANGAEA GPT is to generate concise documentation for historical or underutilized datasets that currently lack related publications, ensuring that their valuable information endures and drives further scientific discoveries.

How to cite: Pantiukhin, D., Shapkin, B., Kuznetsov, I., Jost, A. A., Jung, T., and Koldunov, N.: PANGAEA GPT: A Coordinated Multi-Agent Architecture for Earth System Data Discovery and Analysis, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-13656, https://doi.org/10.5194/egusphere-egu25-13656, 2025.

ClimSight is an innovative open-source climate information system that integrates large language models (LLMs) with geographical and climate data to provide climate information to everyone, everywhere. This description builds upon the original paper [1] by presenting the system’s recent developments and updated methodologies. By leveraging high-resolution data, including local conditions and climate projections, combined with retrieval-augmented generation systems (based on climate reports, scientific literature, and other sources), and an agent-based architecture, ClimSight addresses the limitations of general-purpose LLMs in climate data analysis, ensuring accurate, reliable, and reproducible outputs. This presentation details the enhanced methodologies employed in ClimSight to deliver climate assessments for specific locations and activities. The system utilizes the LangGraph and LangChain packages to manage agents and LLM calls, providing flexibility in selecting different LLM models, with current implementations relying on OpenAI’s models. The effectiveness of ClimSight is demonstrated through selected examples and evaluations, highlighting its potential to democratize access to localized climate information.

[1] Koldunov, N., Jung, T. Local climate services for all, courtesy of large language models. Commun Earth Environ 5, 13 (2024). https://doi.org/10.1038/s43247-023-01199-1

How to cite: Kuznetsov, I., Jost, A. A., Pantiukhin, D., Shapkin, B., Rajput, M. M., Jung, T., and Koldunov, N.: ClimSight: Leveraging LLMs for Revolutionizing Climate Services, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-8220, https://doi.org/10.5194/egusphere-egu25-8220, 2025.

The increasing availability of social media data offers valuable opportunities for real-time crisis monitoring and disaster management. However, extracting actionable insights from these unstructured, multilingual, and often ambiguous data sources remains a significant challenge, particularly in non-English contexts. In this context, natural language processing (NLP) and machine learning techniques are key tools to automated data extraction and enhance situational awareness for crisis managers, particularly during flash floods and earthquakes.

In crisis management, the rapidly processing and transformation of unstructured social media data into actionable information is essential for effective decision-making. While the literature  highlights the value of social media for improving the situational awareness of decision-makers, extracting relevant information remains resource-intensive, especially for most French crisis management units, which lack the necessary tools and resources. Although, several systems exist for extracting automatically information in social media, only few of them deal with French language. One of the main challenges with social media data lies in its inherent ambiguity including semantic variability (context-dependent meanings of words and idioms), informal language (abbreviations, typos, emojis, and neologisms), entity ambiguity (e.g., locations or organizations with identical names), and a high proportion of noisy or irrelevant content.  

The French ReSoCIO project addresses these challenges by bringing together experts in earth sciences, AI, social sciences and specialists and software developers in risk management and forecasting  to develop a novel approach to social data disambiguation for geospatial visualization of crisis situations. This study introduces an innovative pipeline that combines filtering, entity linking, and geolocation integration to enhance data disambiguation and tailored for real-time predictions. The pipeline first employs a supervised classifier to filter out unrelated tweets. Relevant messages are then processed through an entity linking module, where detected entities are disambiguated by matching them with Wikidata entries. This process leverages embeddings from Wikipedia and compares them with tweet embeddings using CamemBERT, enriching extracted data with contextual and geospatial information. The final step employs large language models LLMs to summarize and linked the extracted information, ensuring that stakeholders receive concise and accurate overviews validated against structured event reports. By characterizing and predicting the impacts and damages of crisis events, this pipeline provides a robust framework for transforming fragmented online data into structured, actionable knowledge.

The system's performance aligns with state-of-the-art models, effectively identifying entities that correspond with the spatiotemporal patterns of actual natural disasters. While this suggests the system's potential utility in enhancing situational awareness for crisis managers by providing timely and accurate geolocated information extracted from social media posts, experimental observation conducted during the ReSoCIO project confirms the contribution of this disambiguation pipeline to French crisis managers.

How to cite: Montarnal, A., Gracianne, C., Caillaut, G., Sabouni, A., Adrot, A., Chave, S., Rigart, L., Faï, F., and Auclair, S.: ReSoCIO: Towards geospatial visualization of Social Media Data by AI-driven Disambiguation. Application  to Crisis Management in the French Context., EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-18732, https://doi.org/10.5194/egusphere-egu25-18732, 2025.

A natural hazard event that highly impacted a society might trigger a wave of post-disaster research analysis, which looks into the cause of the disaster, the types of impact, or any lessons learned to prevent similar events in the future. In short, post-disaster research contains valuable knowledge that should be utilized in disaster risk management. However, in the past 70 years, the scientific community published around 600,000 articles on hydro-hazards, such as floods, droughts, and landslides. Finding articles that describe specific disaster events and synthesizing their knowledge is not humanly possible anymore due to near exponentially increasing numbers of publications. However, recent advancements in large language models allow the analysis and extraction of described disaster events in the scientific literature.

Here we make use of the Wealth over Woe scientific abstract dataset (Stein et al. 2024), with abstracts that were automatically annotated for hydro-hazards and geolocation.  It allows us to track publication trends and to identify disaster events that triggered a wave of new research. We additionally make use of the large language model Llama 70B to extract specific hazard events mentioned in each abstract (e.g. 2003 summer drought in Europe, Pakistan flood in 2010, 2002 Elbe flood, etc.) as well as other described details surrounding the event.

While we know that hydro-hazard research is biased against low-income countries, exceptional disaster events can shift research priorities for several years. The additional funding can support valuable local post-disaster research. The named event recognition can therefore help us answer questions such as: What kind of hydro-hazards are studied in detail and where? What are the key research foci for post-disaster analysis? And are there regional differences to these answers?

How to cite: Stein, L., Pfitzmann, B. M., Mukkavilli, S. K., Ozturk, U., Staar, P. W. J., Berrospi, C., Brunschwiler, T., and Wagener, T.: Automated disaster event extraction to understand lessons learned: A large-scale text analysis on the scientific literature of floods, droughts, and landslides. , EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-6762, https://doi.org/10.5194/egusphere-egu25-6762, 2025.

An increasing volume of abstracts across geoscience is presented annually at the EGU General Assembly (GA). To manage thousands of abstracts, the conference is structured into divisions, thematic sessions, and individual sessions. However, creating rigid organizational boundaries that separate research contradicts commonly demanded interdisciplinary research: researchers may be only exposed to ideas within their peer group, reinforcing existing perspectives. Such phenomena of filter bubbles and selective exposure to information have been observed in various contexts to limit creativity and innovation. Yet, it persists and remains underexplored in the context of large scientific conferences like the EGU GA. 
In this contribution, we demonstrate how natural language processing allows for breaking the scientific silos to encourage interdisciplinary interaction at EGU GA. We use sentence embeddings (SBERT) to evaluate the semantic similarity between scientific abstracts and identify closely related ones. We analyzed 5,000 randomly selected abstracts per EGU GA, identifying the 10 most similar abstracts. The results show that participants who focus exclusively on abstracts within their thematic session potentially overlook 44% of the ten most relevant contributions to their research, underscoring the risk of missed interdisciplinary connections. Beyond those findings, we will outline existing projects and plans for improving the conference experience and making geoscience research more interdisciplinary.

How to cite: Sodoge, J., Carvalho, T. M. N., and de Brito, M. M.: Encouraging interdisciplinary connections at EGU through text mining , EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-3179, https://doi.org/10.5194/egusphere-egu25-3179, 2025.

Of the generative Artificial Intelligence (AI) systems, Retrieval Augmented Generation (RAG) has attracted a lot of attention for its ability to support natural language queries into large text corpora with the help of Large Language Models (LLM). In a pilot project, we explored RAG and LLM finetuning as tools for exploring the abstracts of the EGU General Assembly as a text corpus.

To ingest the text corpus, we built a processing pipeline to convert the abstract corpus from XML to JSON in a structure that would make it easy to import the data into a vector storage system. For additional context, we added the association of an abstract with the scientific divisions of the EGU, including co-organisation between two or more divisions. This information was not available at the time of this project and had to be scraped from archived versions of the conference online programme.

The RAG system is designed to read various model formats, such as GGUF, GPTQ, and Transformers models. It also integrates with a vector storage solution to read and use conference abstracts to provide enriched responses. Its implementation uses Apptainer for containerised execution.

The first responses from the RAG system to natural language queries produced promising results. The inclusion of links to the source materials allowed us to compare the query response with the information in the source materials. However, evaluating generative AI models is not trivial since one query can produce multiple results. Using a well-understood text corpus and being able to trace the probable origin texts of the results allows us to evaluate the quality of the results and better understand the origin of deficient RAG responses.

How to cite: Klump, J., Hille, J., Guglielmo, M., and Gardner, B.: Building a RAG system for querying a large corpus of conference abstracts, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-7690, https://doi.org/10.5194/egusphere-egu25-7690, 2025.

Efforts worldwide aim to collect detailed information on the spatial and temporal distribution of natural hazards to improve our understanding of their occurrence and ultimately prevent their impacts. However, data on the location, timing, and impact of hazards remain scarce in many regions, even in the most exposed ones. Data collection methods are usually framed around earth observation approaches, sometimes combined with citizen science. Such approaches can be time-consuming, resource-intensive, and may fall short regarding data needs, especially at large scales. Combining these methods with complementary approaches could better address these challenges. We introduce a multilingual tool that uses natural language processing techniques to extract information on geo-hydrological hazards from online news articles. The tool is developed based on a worldwide application where we processed ~ 5.8 million articles published between 2017 and 2023 across 58 languages. The articles were extracted from GDELT (Global Database of Events, Language, and Tone), a global database monitoring events through online news articles. Using large language models, the tool analyzes articles at the paragraph level through three major steps: (1) filtering paragraphs for relevancy, (2) extracting information on the location (down to street level), timing, and impact, and (3) clustering information into events. This multilingual approach enabled the tool to extract and analyze 12.438 flood events, 1.312 landslide events, and 1.086 flash flood events globally for 2023 alone, providing ~ 20 times more data than current disaster databases and improving the coverage worldwide. In regions such as South and Central America, Europe, and Asia, where English is not the primary reporting language, non-English texts were the most important source of information. Especially in South and Central America, where non-English (primarily Spanish and Portuguese) paragraphs outnumbered English paragraphs by a factor of five. The proposed tool provides a new way to extract an unprecedented level of data on geo-hydrological hazards, forming a complementary source of information to existing methodologies. Beyond geo-hydrological hazards, the tool can be used to document other hazards, including earthquakes, wildfires, or volcanic activity. In addition, with this specific application, we provide a new extensive global dataset on impactful geo-hydrological hazards, which offers new opportunities for improving our understanding of these processes and their impact on continental to global scale.

How to cite: Valkenborg, B., Dewitte, O., and Smets, B.: A multilingual tool for the documentation of impactful geo-hydrological hazards using online news articles: a worldwide application, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-6548, https://doi.org/10.5194/egusphere-egu25-6548, 2025.

Spain and the Mediterranean coast are largely affected by flash floods, which are generated by intense, localized storms within smaller basins (Gaume et al., 2016). In Spain, floods are the country's primary recurring natural disaster, accounting for nearly 70% of the compensation amount issued by the Consorcio de Compensación de Seguros (CCS, 2021).  Improving early warning systems is crucial to reducing risks associated with floods. Comprehensive and up-to-date databases of past flood events serve as essential tools for developing such systems.

This study presents the implementation of an AI-based text-mining tool designed to automate the creation and updating of flood event databases using information extracted from newspapers. This tool is tailored to enhance and expand INUNGAMA, an impact database of flood events in the Catalonia region (Barnolas and Llasat, 2007), by extracting data from ‘La Vanguardia’, a major Catalan newspaper. The text-mining tool involves several steps, starting with the retrieval of potentially relevant news through keyword-based queries on the newspaper’s online archive. To eliminate irrelevant news, a natural language processing (NLP) model filters the initial dataset. Impact data of flood events are extracted by analyzing the newspaper text with an advanced NLP model; the extracted information is saved in a machine-readable and consistent format. Finally, the tool integrates the extracted data with the pre-existing INUNGAMA database, either by merging new information with existing events or by creating entries for previously undocumented events.

The tool was calibrated and tested using the INUNGAMA database. Its ability to download and filter relevant articles was assessed over six non-consecutive months, demonstrating excellent performance in identifying and distinguishing flood events. Furthermore, the AI model exhibited high accuracy in extracting impact data from the text when tested over one year of newspaper data.

The proposed AI-based tool offers a powerful solution for automating the creation and updating of flood impact databases, providing a solid foundation for developing early warning systems aimed at risk reduction. The text-mining tool is designed to complete the INUNGAMA database and to update it up to the present. Moreover, it can be adapted for creating new databases in other regions using different newspaper sources.

This research has been done in the framework of the Flood2Now project, Grant PLEC2022-009403 funded by MCIN/AEI/10.13039/501100011033 and by the European UnionNextGenerationEU/PRTR and the I-CHANGE Project from the European Union’s Horizon 2020 Research and Innovation programme under grant agreement 101037193.

How to cite: Guzzon, C., Marcos Matamoros, R., Marinelli, D., Llasat-Botija, M., and Llasat-Botija, M. C.: An AI-Based Text-Mining Tool for flood impact data extraction from newspaper information, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-15719, https://doi.org/10.5194/egusphere-egu25-15719, 2025.

Water governance systems in many river basins require improvement to adapt to changes in environmental and socioeconomic landscapes. However, water governance reformation is a complex and challenging process. In particular, policymakers and water managers need a comprehensive understanding of the fundamental components that form the current water governance systems. Only then can new rules be introduced to alter the governance characteristics of these systems. This process is especially challenging in the case of interstate rivers that flow across multiple states, where governance systems are characterized by complex interstate water agreements and/or laws that cover various cross-state water management affairs and regulate stakeholders from different states. We use the institutional grammar (IG) to parse water agreements and laws, generating text-based data for assessing the institutional characteristics of interstate water governance systems. The IG decomposes written statements in the documents into different syntactic components. Based on these components, the functions of the statements can be identified and categorized into one of seven types of institutional rules, as defined by the rule concepts of the institutional analysis and development (IAD) framework. By analyzing these findings with indicators of governance characteristics, we are able to assess the allocation of water governance responsibilities and the degree of coordination within a water governance system to identify its institutional characteristics. We applied this method to analyze the water-related laws that form the governance systems of the Yellow River Basin (YRB) in China. The findings reveal that the YRB’s water governance system has undergone five major stages of structural evolution since 1987. During this process, the basin’s focus in water governance has shifted from flow regulation to water consumption governance, as well as expanding its governance scope to include interstate water administration and drought management. Currently, the YRB’s water governance systems are dominated by centralized governance structures characterized by the centralization of water governance responsibilities and a high degree of stakeholder coordination. The method demonstrates that text-based data generated through parsing water agreements and laws can systematically analyze the complex institutional characteristics of water governance systems. This research contributes to the advancement of text-based method for water governance analysis.

How to cite: Lai, C. H. and Zhao, J.: Institutional grammar as a text-based method for water governance analysis , EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-7072, https://doi.org/10.5194/egusphere-egu25-7072, 2025.