Illustration of ESA’s Sentinel constellation, the six satellite families that feed continuous Earth-observation data to Europe’s Copernicus program. (Image credit: ESA)
Over the past two years, 218-year-old publisher Wiley has repositioned itself in the AI landscape, focusing on specialized “vertical” models over general chatbots. Explaining the company’s proactive strategy, SVP Josh Jarrett stated Wiley’s view is “we learn more by doing and potentially help shape AI’s development, rather than just be shaped by it.” This approach paved the way for partnerships like one with Rome-based Pi School, which tapped Wiley’s Earth-science catalog for the European Space Agency’s EVE project.
So what’s EVE, and who are the key players? EVE, short for Earth Virtual Expert, is a fine-tuned language model the European Space Agency (specifically its Φ-lab innovation hub) aims to develop into a powerful query engine. It will tap peer-reviewed research with citations built-in to answer questions about Earth science, potentially integrating vast datasets like satellite imagery and climate records over time. The technical heavy lifting falls to Pi School, a Rome-based firm specializing in applied AI projects. ESA selected Pi School to train and deploy the model, which will be piloted with 200 experts starting soon. “Ultimately, the intention is to release EVE as an open-weight model for the broader benefit of society,” said Wiley’s SVP of AI Growth, Josh Jarrett. The project combines Pi School’s technical chops with curated content from sources like Wiley’s Earth-science journals.
Jarrett says EVE’s first job is to earn scientists’ trust. Rather than scraping the open internet, the model trains on a hand-picked trove of 71 peer-reviewed journals, a dozen Earth-science books and Wiley courseware Q&As, including material from the American Geophysical Union. To validate this approach, ESA’s 200-person beta test will go beyond simple fact-checking, assessing whether researchers trust EVE’s outputs enough to confidently cite them in their own work. This reliance on structured, varied content aims for a level of scholarly integration often difficult with general-purpose models, and certain formats within that data, Jarrett suggests, are especially well suited to the fine-tuning process itself.
Josh Jarrett
Jarrett said the model’s training set matters more than its size for this phase. He told R&D World that structured formats like “Q&A formats, worked examples, [and] step-by-step instructions” seem “particularly effective for deepening expertise” during fine-tuning. In other words, good pedagogical design for humans can also be relevant for AI training.
The sheer volume and diversity of Earth science data — including decades of satellite imagery, weather patterns, climate records, and published research across multiple disciplines — present a significant integration challenge for researchers, policymakers and the public alike. Information often resides in separate silos. EVE aims to break down these barriers by acting as a synthesis engine. “You have temperature readings, weather patterns, the last 50 years of satellite imagery. You can imagine stacking all that and looking for patterns at massive scale,” Jarrett explained. The potential applications are numerous for both the public and dedicated researchers. “We hope this unlocks new scientific findings and provides practical utility for the general public, answering questions from ‘What’s the wildfire risk where I live?’ to ‘What crops are suitable for this region over time?'”
What’s next? The immediate focus is the pilot phase where 200 experts will test EVE. If this trial demonstrates the model can reliably answer complex Earth-science queries with trustworthy citations, the path opens further. Jarrett confirmed the ultimate “intention is to release EVE as an open-weight model for the broader benefit of society” once its performance is validated. Success could also solidify Wiley’s vertical AI strategy, paving the way for similar models tackling data challenges in other research-intensive fields.
