“Generative AI” refers to the ability of some kinds of machine learning system, most commonly large language models (LLMs, very large neural network models trained on enormous text datasets), to create multimodal content (e.g. text, images, video, audio) that is fast becoming indistinguishable from human-created content. Such models are the basis for popular AI applications like chatbots, image-generators, virtual assistants and analytical tools, which are beginning to appear in the climate services domain (e.g. refs. 6,7,8,9). Subject to critical evaluation, such tools might be integrated with climate services workflows to help synthesise information and support human decision-making.
The AI tools so far reported in the climate services literature (e.g. refs. 6,7,8,9) have diverse purposes, datasets, and approaches to LLM integration. As an illustration from a UK perspective, here we consider two such applications from our own work. Figure 1 shows schematics for a chatbot designed to communicate information from the UK Climate Projections (UKCP18) (Fig. 1a) and for a tool for generating climate service “recipes” based on a large evidence base of climate models and adaptation solutions (Fig. 1b). Both tools adopt an “agentic” approach (i.e. modular systems which decompose the larger task into smaller sub-tasks that are each handled by semi-autonomous LLM-based agents). Both also use “human-in-the-loop” mechanisms for resolution of ambiguities. Both were co-created with end users at the UK Met Office, with users involved in shaping and evaluating the systems.
Fig. 1: Schematic diagrams showing workflows for illustrative examples of agentic AI tools for climate services.
a Multimodal chatbot and data analysis tool for UKCP. 1. User interacts with chat interface and data workspace. 2. Chat interface provides multimodal outputs and data visualisation. 3. Data workspace holds extracted datasets, analysis/visualisations, and chat history. 4. RAG agent supports multimodal question-answering based on UKCP archive. 5. Data agent enables data extraction and generates code to perform analysis/visualisation. b Agentic workflow for generative climate service recipes. 1. User provides input: templated task/query and project description. 2. Reasoning LLM is prompted to create a “naïve” draft recipe. 3. Knowledge graph collates climate service case studies and climate models/datasets. 4. Multiple LLM agents (supervisor & specialists) augment/improve naïve recipe with specific climate models/datasets/procedures, constrained by knowledge graph. 5. Output is an actionable “final” recipe provided back to the human user.
UKCP18 was produced by the UK Met Office to provide robust climate projections and aid planning and decision-making for the UK until 2100. It consists of scientific reports (containing text, images, tables and graphs), large numerical datasets of climate model outputs (in gridded/tabular form), and various ancillary content (e.g. visual summaries, documentation). Despite many applications, many users highlight difficulties in applying UKCP18 outputs without technical support10. Our experimental chatbot (Fig. 1a) seeks to overcome this usage barrier by using natural language interactions to provide user-friendly and fast access to UKCP18 data/information, with question-answering, data retrieval, and data analysis functionalities. It uses retrieval-augmented generation (RAG)8 to ground LLM responses in the UKCP18 knowledge base and reduce the risks of inaccuracy or hallucination (generating a false or unfounded answer that is not based on real information11). The RAG tool is multimodal and can answer with relevant images/graphs or datasets as well as text. Human-in-the-loop validation is used when the system detects an ambiguous query, especially during data selection tasks where precision on spatial location, time period and climate variables is needed. The chatbot also supports interactive analysis of raw climate model data, generating on-the-fly code that enables users to (in natural language) select different climate variables, locations and time periods for visualisation, comparison, and statistical analyses. Many users previously required input from a (human) climate consultant to access these functions.
Our second example tool (Fig. 1b) is an agentic system designed to automatically generate climate service recipes (i.e. detailed workflows for application of climate science in a given scenario or context). This tool aims to help human climate service providers to easily identify a method for using climate data in a particular client context, reducing the time and cost of delivery. The tool combines LLM reasoning with contextual knowledge from a curated knowledge graph12 that maps relationships between climate variables, emission scenarios, indices, hazards, sectors, and key datasets (e.g. CORDEX, CMIP5, UKCP18). The agentic workflow begins with a description of a climate scenario; human-in-the-loop confirmation is used to augment the initial description and complete a scenario template for further processing. The architecture involves a memory-enabled supervisor agent that orchestrates multiple specialised agents to plan workflows, create draft recipes, query the knowledge graph, source external (non-climate) domain knowledge, and summarise the final output. This multi-agent approach effectively generates “recipes” that specify appropriate methodologies and datasets, with clear traceability back to the knowledge graph to encourage user trust. The approach shows good potential to increase efficiency of climate service design across multiple hazards and domains, enabling human consultants to deliver climate services more rapidly.