UBITECH’s Energy Digitalization Group (https://ubitech.eu/edg/) announced today the publication of a peer-reviewed research paper introducing mAIEnergy, an open-access, multimodal dataset purpose-built to support Large Language Model (LLM) and Retrieval-Augmented Generation (RAG) applications in the energy domain. The paper, titled “A Multimodal Dataset for Large Language Model Applications in the Energy Domain,” was authored by Costas Mylonas and Dr Magda Foti and has been accepted and published in Scientific Data, a journal of Nature Portfolio. The work addresses a long-standing gap in energy-sector data infrastructure: while numerical, geospatial, textual, and imagery data relevant to energy systems have long existed in scattered, single-modality repositories, no openly available corpus had previously integrated all four modalities into a single, harmonized, AI-ready resource. mAIEnergy closes that gap, providing a foundational knowledge base that energy researchers, utilities, regulators, and technology developers can use directly or extend with their own proprietary data.
The mAIEnergy dataset spans four distinct data modalities. Its textual corpus draws on Wikipedia, news coverage retrieved through the GNews API, scientific literature from arXiv, and official governmental and regulatory documents from EU institutions, including the Directorate-General for Energy and the Agency for the Cooperation of Energy Regulators (ACER). Its imagery collection combines Copernicus satellite tiles, aerial imagery from the INRIA Aerial Image Labeling dataset, energy-efficiency labels from the European Product Registry for Energy Labelling (EPREL), building façade imagery from the Irregular Façades (IRF) dataset, and contextual imagery from Wikimedia Commons. Its numerical archive brings together electricity load, generation, and market-price data from the ENTSO-E Transparency Platform, energy balances and macroeconomic indicators from Eurostat, building-stock performance data from the EU Building Stock Observatory, and historical weather observations from Open-Meteo. Its geospatial and relational layer maps power plants, substations, transmission infrastructure, and EV charging stations sourced from OpenStreetMap and the GridKit European Transmission Grid, alongside the Global Power Plant Database, the ENTSO-E TSO network, and the CORDIS database of EU-funded energy research projects.
Rather than simply co-locating these sources, the research team harmonized every record into a consistent schema with normalized units and full provenance metadata, then linked the four modalities through shared entity keys, including country codes, ENTSO-E bidding zones, and geographic coordinates. A property graph built in Neo4j ties infrastructure assets, statistics, imagery, and documents that describe the same geographic context, while the entire corpus is embedded and indexed across OpenSearch, Milvus, and Neo4j to enable semantic, lexical, image-based, and graph-based retrieval through a single hybrid query interface.
The publication includes an extensive technical validation of the dataset, covering file integrity, schema compliance, deduplication, language consistency, plausibility bounds, and topological consistency across all four modalities. Beyond these per-file checks, the authors evaluated the corpus as an integrated retrieval system, issuing a 50-question benchmark of energy-domain queries to the hybrid retriever. The system returned a relevant record within the top five results for 88 percent of questions overall, with a mean reciprocal rank at ten of 0.73, demonstrating that the dataset functions not merely as a static archive but as a genuinely retrieval-ready resource suited to real-world RAG deployments.
The paper illustrates this capability with a worked example: for the question “What is the role of renewable energy in Austria’s electricity system?” the hybrid retriever simultaneously surfaced a relevant Wikipedia article, a Eurostat renewable-energy-share series, a Copernicus land-cover tile, and a power-plant node from the knowledge graph, allowing a language model to compose a grounded, evidence-based answer drawing on all four modalities at once.
In line with FAIR (Findable, Accessible, Interoperable, Reusable) data principles, the mAIEnergy dataset is openly archived on Zenodo under a persistent digital object identifier, with per-source licensing recorded transparently in the accompanying metadata. The complete data-retrieval and preparation pipelines, together with the OpenSearch, Milvus, and Neo4j database back-ends used to index the corpus, have been released as open-source code on GitLab, allowing the wider research community to validate, reproduce, and extend the workflow to additional regions, time periods, or proprietary data sources.
While the current release is predominantly focused on the European Union, reflecting its reliance on European institutional open-data sources such as ENTSO-E, Eurostat, and Copernicus, the authors note that the underlying schema and retrieval pipelines are region-agnostic and can be re-run against equivalent sources for other geographies.
The mAIEnergy dataset is designed to underpin a range of applications across the energy value chain, including continual pre-training and domain-adaptive fine-tuning of LLMs on energy-specific language, retrieval-augmented generation for conversational and decision-support assistants, text-to-image retrieval of satellite tiles and appliance energy labels, knowledge-graph analytics of infrastructure and cross-border interconnections, and conventional numerical tasks such as load and price forecasting.
“Energy is one of the most consequential domains for AI to get right, yet it has historically been one of the most fragmented in terms of open data. With mAIEnergy, we set out to give the research and developer community a single, harmonized foundation that spans text, imagery, numerical time series, and geospatial infrastructure data, so that language models can reason about the energy system the way experts do, by drawing on multiple kinds of evidence at once. Publishing this work in Scientific Data, and releasing the dataset and full retrieval pipeline openly, reflects our commitment to advancing transparent, reproducible AI research that can genuinely support the energy transition,” said Dr Magda Foti, Head of the Energy Digitalization (EDG) Research Group of UBITECH, and co-author of the study.
About the publication
Mylonas, C. & Foti, M. “A Multimodal Dataset for Large Language Model Applications in the Energy Domain.” Scientific Data (2026). The article is available online at:
https://www.nature.com/articles/s41597-026-08075-5
The mAIEnergy dataset is openly archived on Zenodo at:
https://zenodo.org/records/16401633
The data-retrieval pipelines and vector database back-ends are available on GitLab at:
https://gitlab.com/maienergy-data-retrieval
https://gitlab.com/maienergy-vector-databases
This work was supported by the European Union-funded project DIGITISE under grant agreement No 101160671, and by the FFplus project, funded by the European High-Performance Computing Joint Undertaking (JU) under grant agreement No 101163317.
