Training AI to Understand Nature

I. Training AI to Understand Nature

Artificial intelligence is shaped by the datasets it learns from. As AI becomes increasingly embedded into culture, creativity, and everyday life, we began asking a fundamental question:

What kinds of datasets should future creative systems be built upon?

Most contemporary AI models learn from vast collections of human-generated text and images. At Refik Anadol Studio, we became interested in a different possibility. What would happen if an AI model learned not primarily from human culture, but from nature itself? What kinds of relationships, patterns, and forms of intelligence might emerge?

These questions led to the creation of the Large Nature Model (LNM), a nature-based multimodal AI model trained on one of the world's largest ecological archives. Developed for DATALAND, the LNM combines more than 1.2 billion ecological data points, environmental recordings, scientific observations, and field collections gathered through long-term collaborations with museums, research institutions, scientists, conservationists, and environmental organizations around the world. The goal of the LNM is to build a richer understanding of the living systems that surround us. Unlike conventional image datasets that treat photographs as isolated visual information, the LNM is designed to learn relationships: between species and habitats, climate and behavior, sound and environment, structure and growth. The model is not only trained on how nature looks, but also on how it behaves.

II. From Images to Ecological Understanding

Building the Large Nature Model required a fundamentally different approach to data collection and processing. We wanted to move beyond a library of images toward a living archive capable of representing ecological complexity.

To do this, we developed a workflow that combines large-scale environmental datasets with advanced AI systems. Working with partners including Google Cloud, we built processes that allow the model to identify species, recognize habitats, generate scientifically informed descriptions, and understand relationships across billions of images and environmental records. What emerges is a structured memory of the natural world.

This distinction matters because the quality of an AI model is shaped by what it is able to perceive. A model trained on isolated images learns visual patterns. A model trained on relationships can begin to understand context.

That context ultimately becomes part of the artwork itself.

III. Why Field Expeditions Matter

While many of the datasets used to train the Large Nature Model come from long-term partnerships with museums, research institutions, and environmental organizations, we also believed it was essential to engage directly with the environments we hoped to represent.

Beginning in 2023, our team undertook a series of field expeditions across some of the world's most biodiverse ecosystems. These journeys were intended to document dimensions of nature that are often absent from conventional datasets: spatial relationships, environmental rhythms, ecological interactions, and the hidden processes that allow living systems to function.

In the Gondwana Rainforests of Australia, one of the oldest surviving rainforest systems on Earth, we documented ancient forest structures, environmental conditions, wildlife soundscapes, and atmospheric behaviors across multiple layers of the ecosystem. In Indonesia's tropical forests, we worked alongside local researchers to capture the complex relationships between climate, vegetation, and biodiversity unique to the region. Across California's diverse biomes, from redwood forests and coastal environments to wetlands, waterfalls, and mountain ecosystems, we conducted long-duration recordings that captured how landscapes transform throughout entire day-night cycles.

Each expedition contributed a different dimension of ecological intelligence to the model. Using LiDAR scanning and photogrammetry, we captured forests as dense three-dimensional point clouds composed of millions of spatial measurements. Rather than interpreting a tree as a flat image, the model learns the architectural relationships between canopy layers, understory vegetation, root systems, and surrounding environments. This information contributes to the depth, scale, and structural complexity that emerge within DATALAND's visual worlds.

Sound recordings reveal another layer of ecological knowledge. Dawn choruses, insects, rainfall, streams, birds, and changing weather patterns document the temporal rhythms of ecosystems and how species interact with one another across time. Collected using ambisonic microphones, binaural recording systems, and long-duration environmental rigs, these recordings help shape the evolving sonic environments of the artwork, allowing visitors to experience the acoustic identity of places that may be thousands of miles away.

Environmental sensors extend perception even further. Working with researchers in locations including Chile, Australia, and California, we collected data describing rainfall, soil moisture, atmospheric conditions, sap flow, and plant-water relationships. These systems document ecological processes that often remain invisible to human observation and reveal how forests regulate themselves, respond to environmental change, and sustain complex networks of life.

IV. Ethical Data as Creative Infrastructure

The future of AI datasets is both a question of scale and responsibility.

For us, ethical data collection means more than obtaining permission. It means building long-term relationships with the institutions, researchers, conservationists, and local experts who preserve these environments. It means minimizing ecological disturbance during data collection. It means recognizing that environmental information carries scientific, cultural, and ecological significance beyond its use in artistic systems.

These choices directly shape the outputs of the Large Nature Model. The quality of an artwork is inseparable from the quality of the relationships, knowledge, and care embedded within the data itself.

V. Inside DATALAND

The Large Nature Model serves as the foundation of Machine Dreams: Rainforest, the inaugural exhibition at DATALAND.

When visitors encounter the evolving visual worlds, soundscapes, scents, and immersive environments generated by the LNM, they are engaging with a model trained on an unprecedented archive of ecological relationships. The artwork is an attempt to create new ways of perceiving its complexity.

Our goal is not to replace direct experiences of the natural world with technology. Rather, we see AI as a tool for expanding perception, revealing hidden patterns, and creating new forms of connection between humans and the living systems that sustain us.

As the Large Nature Model continues to grow through new partnerships, datasets, and field research, we remain committed to exploring a simple but increasingly urgent question: What might artificial intelligence learn if nature became one of its primary teachers?

Special Thanks

Special thanks to the conservationists, scientists, universities, research labs, botanists, ecologists, field recordists, sensor specialists, LiDAR and photogrammetry teams, local guides, rangers, environmental organizations, and production crews who contributed to these ongoing expeditions and collaborations.

Credits

Prof. Kim Calders, scientist at Ghent University in Belgium, contributed a range of lidar datasets from forests around the world. His team at Q-ForestLab focuses on measuring forests in 3D using laser scanners, both from the ground and from within the canopy. Their research uses these 3D data to build virtual forests that help improve our understanding of satellite observations, biodiversity, and forest functioning.

2018 Robson Creek terrestrial laser scanning data provided by prof. Kim Calders and Dr. Louise Terryn (UGent). This dataset was acquired with support BELSPO (Belgian Science Policy Office) in the frame of the STEREO III programme - project 3D-FOREST (SR/02/355) and TERN Australia.

Terrestrial laser scanning (TLS) data of TERN's Cumberland Plain SuperSite, Australia, collected in 2022 using a RIEGL VZ-400i scanner. Dataset provided by prof. Kim Calders (UGent) in collaboration with TERN Australia.

Terrestrial laser scanning (TLS) data of Berchtesgaden Endstal, Germany, collected in 2023 using a Riegl VZ-400i scanner. Dataset provided by prof. Kim Calders & Karun Dayal (UGent) in collaboration with prof. Cornelius Senf and Andreas Hanzl (TU Munich).

Terrestrial laser scanning (TLS) data of Litchfield TERN site, Australia, collected in 2018 using a Riegl VZ-400 scanner. Dataset provided by prof. Kim Calders (UGent) in collaboration with TERN and the data acquisition was funded by BELSPO (Belgian Science Policy Office) in the frame of the STEREO III programme - project 3D-FOREST (SR/02/355).

2013 Lopé National Park, Gabon (plot LPG-01) terrestrial laser scanning data provided by Prof. Mathias Disney (UCL) and Prof. Kim Calders (UGent), collected with the collaboration of Agence Nationale des Parcs Nationtax(ANPN), Gabon.

Terrestrial laser scanning (TLS) data of mangrove forest plots in Suriname, collected in 2022 using a Riegl VZ-400i scanner as part of the GCCA+ Phase 2 project. It includes high-resolution 3D point clouds of riverine mangrove stands dominated by Rhizophora mangle (Red Mangrove).

Acknowledgements when data are used: Dataset provided by Jasper Feyen and prof. Kim Calders (UGent) in collaboration with the Foundation of Forest Monitoring and Production Control Suriname (SBB).

2023 Centenary Fig Tree of the Atlantic Rainforest at ARAÇÁ Project, RJ, Brazil, terrestrial laser scanning data provided by prof. Kim Calders (UGent) in collaboration with Barbara D’hont (UGent), Wout Cherlet (UGent), Alexandre Antonelli (ARAÇÁ Project) and Thomas Berg (ARAÇÁ Project).

2022 EucFACE terrestrial laser scanning data provided by prof. Kim Calders (UGent) in collaboration with Western Sydney University. This dataset was acquired with funding from the European Union (ERC-2021-STG Grant agreement No. 101039795 SPACETWIN).

2023 Gandalf’s staff terrestrial laser scanning data provided by prof. Kim Calders (UGent) in collaboration with Barbara D’hont (UGent) and Arko Lucieer and Leonard Hambrecht (University of Tasmania) and Steven Pearce (The Tree Projects). This dataset was acquired with funding from the European Union (ERC-2021-STG Grant agreement No. 101039795 SPACETWIN).

2023 Lathamus Keep terrestrial laser scanning data provided by prof. Kim Calders (UGent) in collaboration with Barbara D’hont (UGent) and Arko Lucieer and Leonard Hambrecht (University of Tasmania) and Steven Pearce (The Tree Projects). This dataset was acquired with funding from the European Union (ERC-2021-STG Grant agreement No. 101039795 SPACETWIN).

2015 Wytham Woods terrestrial laser scanning data provided by prof. Kim Calders (UGent) in collaboration with Niall Origo and Joanne Nightingale (National Physical Laboratory, UK) and Mat Disney (University College London). This dataset was acquired with funding support from NERC National Centre for Earth Observation (NCEO), UCL Geography and Metrology for Earth Observation and Climate project (MetEOC-2), grant number ENV55 within the European Metrology Research Programme (EMRP).

–

ICT International Pty Ltd is an Australian company specialising in environmental monitoring solutions for plant, soil, water, and climate research for over 45 years. Their team of scientists, engineers, and technicians supports environmental, forestry, agricultural, and conservation projects worldwide through scientific expertise, technical support, and innovative measurement technologies. Project contributors included: Peter O. Cull – Director, Susan Cull – CEO, Ed Lefley – Operations/R&D, Australia, Toby Partridge – Engineer, Australia, Francesca Rigato – Applications Scientist, Australia, Sam Fisher - Environmental Scientist, Australia.

–

Special thanks to Mutiara K. Pitaloka, Frisca Damayanti, and Muhammad Efendi (Research Center for Applied Botany, BRIN, Indonesia) for their contributions to environmental data collection in Indonesia

–

Andrés Plaza-Aguilar, Ph.D. (Principal Investigator) and his Lab Teams at Climate Adaptation and Forest Management Laboratory (ACGBLab), Faculty of Forest Sciences and Conservation, Universidad de Chile; and the Estación Experimental Frutillar, Faculty of Forest Sciences and Conservation, Universidad de Chile. The Climate Adaptation and Forest Management Laboratory (ACGBLab), based at the Faculty of Forest Sciences and Conservation of the Universidad de Chile, investigates how forests and forest plantations respond to climate variability and change, with a focus on improving forest health, resilience, and sustainable management. The team integrates data from multiple sources and scales, including in-situ sensors, ecophysiological measurements at the individual tree and species level, remote sensing, and field sampling, to analyse and model forest ecosystem dynamics. The environmental and physiological data collected for the Dataland project, including temperature, humidity, rainfall, soil conditions, tree sapflow, and stem diameter growth, were gathered at the Estación Experimental Frutillar, a 33-hectare research station located in Frutillar Bajo in the Los Lagos Region and operated by the same Faculty since 1959. The station harbours one of the few remaining forests on the central plain of the Los Lagos Region, including centenarian native trees and more than 150 plant species, and serves as a long-term platform for forest research, conservation, and environmental education.