AI Co-scientists Help Uncover the Secrets of Soil

Soil is More Than Just Dirt!

People often use the word “dirt” in a negative way—someone might try to “dig up dirt” on a person to make them look bad or call an unpleasant job “dirty work”. You might have also heard the phrase “dirt cheap” to describe something extremely inexpensive. But the soil beneath our feet is much more than “dirt”, and it is certainly not bad or worthless!

Soil is the thin, living skin of the Earth, holding up forests, grasslands, farms, and even cities. Almost everything we eat depends on the soil: crops grow in it, and animals graze on its plants. Soil filters rainwater, helping to keep streams and drinking water clean, and it provides a home for countless organisms, from earthworms, insects, and fungi to the tiniest bacteria. On a larger scale, soil influences the health of the planet itself. When soil is healthy—rich in organic matter, well-structured, and full of life—ecosystems thrive. However, when it is damaged by erosion, pollution, or poor management, crops fail, water becomes polluted, and carbon stored in the soil escapes back into the air, contributing to climate change. In short, soil is one of the most valuable resources on Earth (Check out some songs we wrote for World Soil Day here and here)!

Digging Into Soil Science

For all its importance, soil remains one of the least understood parts of our planet. Every handful is a mixture of minerals, air, water, and living things interacting in complex ways. Although it is not always easy to see from the surface, no two soils are exactly the same (Figure 1). Soil varies from place to place and changes constantly, shaped by weather, plants, animals, and how people use the land. A single field might be rich and fertile one season and dry and compacted the next. Understanding how soil’s living and non-living components work together—and how they might respond to Earth’s changing climate—is one of science’s biggest ongoing challenges.

Figure 1 - No two soils are exactly the same—soil varies from place to place based on the sizes and kinds of particles it contains. Major types of soil include: (A) clayey soil; (B) loamy soil; (C) sandy soil; (D) rocky soil; and (E) peaty soil (All images sourced from Creative Commons).

To understand soil, scientists gather data from many sources: laboratory tests reveal its texture, minerals, and nutrient content; sensors buried in the ground record changes in temperature and moisture; and satellites capture how vegetation and land use change over time. The result is a mountain of data—thousands of measurements for every field or forest, collected over years. Each type of information tells part of the story, but bringing all those clues together to understand how soils work, and how they might respond as the planet warms, is a major challenge. Making sense of so much data is slow, demanding work, even for large research teams.

AI Co-scientists—A New Kind of Teammate

Science takes a long time. Before trying a new experiment, scientists read many scientific papers to understand what others have discovered. They also compare data from various sources and discuss possible explanations, called hypotheses, for what they observe. Then they design experiments to test their hypotheses, and when the results come in, other scientists review the work to make sure it holds up.

Artificial intelligence (AI) already helps with some of these time- consuming tasks. Today’s AI tools are very good at learning from data and spotting patterns. In soil science, for example, AI can analyze satellite images to map soil types or use sensor data to estimate how wet or nutrient rich a field is [1]. AI tools can process far more data than a person can, and they do it much faster. However, most AI tools do just one job at a time—they might predict soil properties, classify images, or summarize information. They usually do not “understand” what they are looking at or explain why something happens or how one part of a system affects another. But soil science often depends on exactly this kind of reasoning.

To tackle these harder problems, scientists are developing AI multiagent systems. Instead of one AI tool working alone, a multiagent system is a team of AI co-scientists, called agents, that each take on a different part of the scientific process—the same way members of a human research team might (Figure 2) [1–3]:

• A planner agent decides what tasks are needed to answer the scientific question.

• A literature agent searches thousands of scientific papers for useful information.

• A data agent analyzes soil data from sensors, laboratory tests, or satellite images.

• A scientist agent generates new hypotheses to test.

• A reviewer agent evaluates those hypotheses, compares them, and identifies the most promising ones.

• A memory agent keeps track of what has been learned so far.

Figure 2 - A multiagent system is a team of AI co-scientists, called agents, that each take on a different part of the scientific process—the same way members of a human research team do. It is important to remember that a human scientist leads and guides the work. The goal is for multiagent systems to provide more time for human scientists to use their creativity and think about the big picture, while AI handles the repetitive, time-consuming steps.

Because the AI co-scientists can “talk” to each other—sharing results and questioning one another’s reasoning—the whole system can act more like a real research group than a single AI tool can.

Importantly, a human scientist still leads the work: choosing the question, judging whether the AI’s ideas make sense, and deciding what to test in the real world. The goal is not to replace scientists, but to let multiagent systems handle repetitive, data-heavy work so human scientists can focus on creativity, big-picture thinking, and understanding the world beneath our feet.

Teaming Up With AI to Study Soil Carbon

As we briefly mentioned earlier, soil stores enormous amounts of carbon from decomposing plants and animals (organic matter)—more carbon than is found in all living plants and the atmosphere combined. This natural, long-term storage, called carbon sequestration, helps to prevent the carbon from returning to the air, balancing the gases in the atmosphere and keeping Earth’s temperature stable (Figure 3A) [4].

Figure 3 - (A) The amount of carbon soil can hold over the long term, called carbon sequestration, can be affected by changes in Earth’s climate. Heavy rainfall, rising temperatures, and severe droughts can reduce the ability of soil to sequester carbon. Importantly, the more carbon that is released into the air, the worse global warming becomes. (B) We asked our team of AI co-scientists whether there is a point where soil becomes “full” of carbon. The AI agents gathered various kinds of information and proposed hypotheses that we could test in real-world experiments in the future.

Scientists are still not sure whether there is a limit to how much carbon soil can store before it starts releasing it again. Understanding what controls that limit, and what makes soils gain or lose carbon, could be key to protecting soil health and slowing climate change. So, our research group used a team of AI co-scientists [5] to help us understand how much carbon soil can really hold, and for how long. We gave the multiagent system a clear question: Is there a point where soil becomes “full” of carbon?

Our AI co-scientists got to work (Figure 3B). First, the literature agent gathered the latest papers and data on carbon storage. Scientist agents proposed ideas—for example, that some soils may never reach a limit, while the limit of other soils might depend on the types of minerals or microbes they contain. The reviewer agent checked whether each idea made sense based on current evidence. In the end, the multiagent system produced a set of competing hypotheses, each with strengths and weaknesses. Some seemed easy to test; others raised new possibilities scientists had not considered. Human experts then stepped in to judge which ideas were most realistic and how they could be tested in real soils.

Our work showed that AI multiagents can speed up early scientific brainstorming, helping soil scientists generate new hypotheses and decide which are worth pursuing. But it is important to remember that these AI co-scientists cannot tell us which ideas are true—only careful experiments in the real world can do that.

“Digging” Even Deeper With AI

The same human–AI teamwork that helped explore how soil stores carbon could also help scientists study how soils respond to climate change more broadly—how rainfall, temperature, and land use affect their ability to store nutrients, support crops, and stay healthy over time.

One way to do this is by creating digital soil twins—virtual versions of real soil systems, like those in fields, forests, or other environments, built from soil data that has already been collected [6, 7]. These virtual versions update as new information arrives, allowing multiagent systems to track how soil conditions evolve in real time. By comparing sensor readings and satellite images, for example, the system could spot early signs of erosion, drought, or nutrient loss. It could even run virtual experiments, testing how different farming or conservation methods might affect soil health decades into the future.

Another fascinating area is the study of the soil microbiome—the billions of microorganisms that live underground and drive nearly every soil process. These microbes release nutrients that plants need, store carbon, and help soil recover from damage. AI agents could combine different types of data to learn how the soil microbiome responds to environmental stresses such as drought, pollution, or excessive fertilizer use. These insights could help farmers improve crop yields while protecting soil health and biodiversity.

Two Kinds of Intelligence, One Goal

Soil may be ancient, but the tools for studying it are changing fast. With help from AI multiagent systems, soil scientists can explore questions that once seemed impossible. These AI co-scientists can search, analyze, and compare information far more quickly than people can, freeing human scientists to focus on understanding the bigger picture.

But technology alone will not solve the challenges facing our soils. AI depends on the data it receives, and the goals set by the people who design and use it. If that data is incomplete or biased, the results will be too. Human scientists must still decide which questions matter most and test those ideas in the real world. In the end, the most powerful discoveries may come not from machines or humans working alone, but from their collaboration. AI multiagents bring speed, pattern recognition, and memory; humans bring curiosity, creativity, and judgment. Together, they can form a new kind of research partnership—one that could help protect the thin layer of ground that sustains life on Earth.

Glossary

Organic Matter: ↑ Pieces of once-living things in the soil, such as dead leaves, roots, and tiny organisms. Organic matter helps soil store nutrients, water, and carbon.

Hypotheses: ↑ Possible explanations for what scientists observe. A hypothesis is an idea that can be tested with experiments or by collecting more evidence.

Artificial Intelligence: ↑ Computer programs that can learn from information, recognize patterns, and make predictions or suggestions without being directly told what to do.

AI Multiagent Systems: ↑ Groups of AI programs that work together like a research team. Each “agent” takes on a different scientific task and shares results with the others.

Carbon Sequestration: ↑ Long-term storage of carbon in soil or other natural places, which helps keep carbon out of the air and slows climate change.

Digital Soil Twins: ↑ Virtual versions of real soils, built from measurements and data. They can update over time and allow scientists to test ideas without disturbing real land.

Soil Microbiome: ↑ The community of tiny living organisms in soil—bacteria, fungi, and others—that recycle nutrients, store carbon, and help soil stay healthy.

Conflict of Interest

The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Acknowledgments

We wish to thank Dr. Susan Debad for providing us with a first draft and for her continued collaborative input as co-author. We would also like to thank the coauthors of the original manuscript: José A.M. Demattê, Mercedes Roman Dobarco, and Pete Smith. AM acknowledges the support of the Australian Research Council through its Discovery and Laureate programs. BM acknowledges the support of the Australia-India Scientific Research Fund AIRXV000066 – Development of AI based monitoring platform for soil carbon sequestration.

AI tool statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

↑Minasny, B., McBratney, A., Demattê, J. A. M., Román Dobarco, M., and Smith, P. 2026. Enhancing soil science research with multi-agent artificial intelligence systems. Front. Sci. 4:1721295. doi: 10.3389/fsci.2026.1721295

[1] ↑ Minasny, B., and McBratney, A.B. 2025. Machine learning and artificial intelligence applications in soil science. Eur. J. Soil Sci. 76:e70093. doi: 10.1111/ejss.70093

[2] ↑ Gottweis, J., Weng, W. H., Daryin, A., Tu, T., Palepu, A., Sirkovic, P., et al. 2025. Towards an AI co-scientist. arXiv. [preprint]. arXiv:2502.18864. doi: 10.48550/arXiv.2502.18864

[3] ↑ Yamada, Y., Lange, R. T., Lu, C., Hu, S., Lu, C., Foerster, J., et al. 2025. The AI Scientist-v2: workshop-level automated scientific discovery via agentic tree search. arXiv. [preprint]. arXiv:2504.08066. doi: 10.48550/arXiv.2504.08066

[4] ↑ Smith, P. 2012. Soils and climate change. Curr Opin Environ Sustain. 4:539–44. doi: 10.1016/j.cosust.2012.06.005.

[5] ↑ Shen, M., and Yang, Q. 2025. From mind to machine: the rise of Manus AI as a fully autonomous digital agent. arXiv. [preprint]. arXiv:2505.02024. doi: 10.48550/arXiv.2505.02024

[6] ↑ Silva, L., Rodríguez-Sedano, F., Baptista, P., and Coelho, J. P., 2023. The digital twin paradigm applied to soil quality assessment: a systematic literature review. Sensors. 23:1007. doi: 10.3390/s23021007

[7] ↑ Tsakiridis, N. L., Samarinas, N., Kalopesa, E., and Zalidis, G. C. 2023. Cognitive soil digital twin for monitoring the soil ecosystem: a conceptual framework. Soil Syst. 7:88. doi: 10.3390/soilsystems7040088