Enhancing soil science research with multi-agent artificial intelligence systems

Abstract

Soil science is entering a new era characterized by the integration of artificial intelligence (AI) multi-agent systems, extending the field beyond traditional machine learning (ML) applications such as digital soil mapping and spectroscopy. While current ML tools are effective for specific tasks, they often lack the reasoning, contextual integration, and adaptability required to address complex, dynamic soil systems. We propose multi-agent AI systems—autonomous, interactive software agents capable of perceptual processing, planning, and scientific reasoning—as a novel framework to support and accelerate soil science research. These agents can fulfill diverse roles, including synthesizing data from field sensors and remote sensing to create dynamic digital soil twins, generating hypotheses, designing experiments, and simulating climate-driven changes in soil function. To illustrate this approach, we tasked a multi-agent system with creating research hypotheses on the topic of mineral-associated organic carbon saturation in soils. The agents generated five hypotheses on effective versus theoretical saturation thresholds, biological and chemical controls, climate influence, interdisciplinary feedback, and actionable management strategies. Each hypothesis was evaluated for empirical grounding, conceptual breadth, and scientific rigor by experts and a simulated peer review. Our findings highlight the potential of multi-agent AI systems, guided by human experts, to accelerate early-stage discovery, support interdisciplinary exploration, and emulate the scientific review process. Nonetheless, challenges remain, particularly around data quality, model transparency, epistemic overtrust, computational cost, ethical implications, and the retention of foundational scientific knowledge. We emphasize AI as an augmentative partner, not a replacement, for human-led discovery.

Key points

Introduction

In recent years, soil scientists have rapidly transitioned from machine learning (ML) to broader artificial intelligence (AI) applications to address various disciplinary challenges (1). ML, a subset of AI, focuses on pattern recognition and predictive modeling from data, but AI encompasses more advanced capabilities such as reasoning, autonomous decision-making, and the integration of diverse data sources through generative models such as large language models (LLMs) and multi-agent systems. AI and ML have become increasingly important in tackling global soil security challenges (1–3).

ML continues to play a central role in spatializing and monitoring soil conditions using proximal and remote sensing data, generating digital soil maps at scales ranging from field plots to global landscapes (4, 5). These tools support precision agriculture and inform decisions related to nutrient management and soil carbon stock assessment.

The transition to AI has introduced new capabilities that go beyond mapping. AI systems, especially those incorporating generative capability such as LLMs, can now autonomously synthesize scientific literature. However, most current ML and AI applications in soil science are limited to pattern recognition or prediction based on existing soil data (1, 6). As powerful as these tools are, they typically focus on specific tasks and lack the flexibility and reasoning capacity required to fully address the complexity of soil systems. The next leap forward lies in intelligent automation: the development of multi-agent AI systems that not only process and predict but also have the technical abilities to emulate perception, reasoning, planning, and collaboration. These systems have the potential to work alongside human experts, navigating complex scientific questions, integrating diverse data sources, generating hypotheses, and designing adaptive experiments. Embracing this paradigm would allow soil science to harness AI’s full potential—not to replace researchers but to empower them in addressing the global challenges of soil security, climate resilience, and sustainable land use.

Currently, AI systems generally lack the integration, adaptability, and reasoning capabilities necessary to address the multifaceted and interconnected nature of soil systems. Thus, it remains an unexplored area, partly due to the lack of enabling tools. However, inspired by rapid advancements in other scientific domains, such as biomedicine and materials science (7–10), we propose a perspective shift toward a new paradigm: the use of multi-agent AI in soil science (5).

An AI agent is a software system that uses AI to interact with its information environment, gather information, and perform tasks to achieve specific goals. A multi-agent system consists of multiple AI agents working independently, collaboratively, or autonomously to address a complex problem (11). In such systems, AI is able to make independent decisions and generate novel ideas (12).

We envision multi-agent AI systems not merely as predictive models or data processors but as integrated, intelligent systems capable of perception, reasoning, planning, learning, and collaboration. These agents could coordinate diverse tools, from proximal sensors and satellite imagery to laboratory instruments and complex simulation models, to break down complex soil science problems into manageable and actionable steps. This approach strengthens research design and interpretation before engaging with real-world soil conditions, facilitating the scaling up of ideas or findings.

This paper explores these ideas by examining the potential of multi-agent AI systems in soil science. Specifically, we address the following aims:

• to explain and test a multi-agent AI framework that emulates collaborative scientific reasoning by integrating specialized agents for literature analysis, hypothesis generation, peer review, and experimental design;

• to explore the potential application of multi-agent AI systems in soil science;

• to demonstrate the capability of AI agents to accelerate early-stage soil scientific discovery through structured problem decomposition, interdisciplinary reasoning, and iterative evaluation cycles;

• to evaluate the scientific plausibility, novelty, feasibility, and scalability of AI-generated hypotheses for a specific soil science problem using multi-agent peer review and human-in-the-loop feedback.

AI in soil science: advances and limitations

Recent soil science advances have produced numerous data sources, analytical tools, and models that can be enhanced by ML and AI algorithms (Figure 1), including:

Figure 1

• networks of in-situ sensors measuring soil moisture, temperature, and physical and chemical properties (13–15);

• sophisticated laboratory equipment for analyzing particle-size distribution, mineralogy, chemical and biological properties, and various spectroscopic methodologies from radio to gamma waves (potentially automatable) (16–18);

• process-based simulation models [e.g., for water, solute, heat and gas transport, organic matter (OM) dynamics, nutrient transport, and crop growth] (19, 20);

• existing ML models for specific predictive tasks, such as pedotransfer functions, soil spectra functions, and spatial prediction models (21–23);

• remote sensing platforms providing data on land cover, topography, relief parameters, and spectral properties (24–27);

• extensive databases of soil properties and experimental results (28, 29).

Despite these successes, current AI applications in soil science are largely task-specific and are developed and applied in isolation (30, 31). These tools typically operate with limited autonomy, requiring substantial human input for data preparation, model training, and interpretation. The reasoning capabilities of many ML and AI models are often embedded implicitly in their architecture rather than explicitly articulated. As a result, many applications are typically geared toward execution, such as in soil spectroscopy, where the focus is on achieving specific predictive objectives. However, they often bypass the traditional scientific process of hypothesis formulation and experimental design (32). For example, a model may accurately link a spectral response to a soil property, but it will rarely address why this relationship exists. As a result, there is a growing concern that much of the ML-based soil science literature lacks robust investigation of the underlying cause-and-effect mechanisms (33). We argue that this gap highlights the need for research that integrates mechanistic understanding with data-driven methods, ensuring that AI not only predicts but also contributes to scientific explanation.

LLMs are currently used to enhance scientific writing and knowledge retrieval. While they demonstrate promising capabilities for conversational interaction, ensuring their reliability and depth of understanding in specialized soil science domains still requires significant human validation and expertise. For example, Khanifar (34) assessed the performance of LLMs in answering soil science-related questions and found that, on average, they could correctly answer only up to 65% of questions from advanced soil science examinations.

This existing foundation, with its strengths and weaknesses, sets the stage for the development of next-generation AI agents capable of more integrated, autonomous, and collaborative scientific exploration and management in soil science. The vision of AI agents transforming scientific discovery is gaining momentum in fields such as chemistry, materials science, and biomedicine, where complex systems, vast datasets, and multidisciplinary knowledge must be integrated (5, 35, 36).

The concept of an “AI scientist” or “co-scientist”—an AI system capable of skeptical learning, reasoning, and collaboration—has been proposed as a framework for generating novel scientific hypotheses aligned with researcher objectives (8, 37, 38). Rather than fully automating discovery, these agents are envisioned as partners that combine human creativity and domain expertise with AI’s capacity for multi-scale data analysis, hypothesis exploration, and task execution.

Central to the AI agent concept is the ability to break down complex problems (39). Soil systems are notoriously complex, involving coupled hydrological, biogeochemical, and physical processes operating across multiple spatial and temporal scales (40). An AI agent could break down a grand challenge, such as predicting the impact of climate change on global soil carbon storage, into manageable subtasks (Figure 2): i) analyzing regional climate projections, ii) modeling soil moisture dynamics, iii) simulating microbial decomposition rates under different temperature regimes or soil mineralogical domains, iv) integrating remote sensing data on vegetation cover, and v) assessing the effects of various land management practices on subsoil rooting and carbon storage at depth. Each subtask could potentially be handled by specialized modules or even distinct agents within a larger multi-agent system.

Figure 2

Multi-agent AI systems

Effective problem break down relies on the agent’s ability to integrate and coordinate a diverse array of tools. A multi-agent system can act as an orchestrator, selecting, configuring, and executing these tools within a coordinated workflow, moving beyond the current fragmented use of individual technologies. The agents possess perceived reasoning and planning capabilities that extend well beyond simple data processing (7).

In recent years, the use of multi-agent AI systems has emerged as a transformative approach in scientific research, particularly in fields grappling with vast, complex datasets and interdisciplinary challenges, such as in biomedicine (9, 10). Unlike single-agent models, which are often constrained by limited capabilities, multi-agent frameworks distribute tasks across specialized agents, each equipped with domain-specific knowledge, reasoning strategies, and tools. These systems emulate collaborative scientific workflows involving brainstorming, critique, consensus-building, and iterative refinement. As a result, multi-agent AI systems can accelerate discovery and enhance research quality (5).

Several multi-agent models have been proposed for use in scientific discovery, and Table 1 summarizes key design features, agent roles, and contributions to automated and augmented scientific inquiry across various disciplines.

These models share several core principles:

• task decomposition. The system segments a complex research problem into manageable subtasks for specialized agents.

• collaboration among agents. Decomposed tasks are delegated to agents with different areas of expertise, promoting collaborative problem-solving akin to human teams.

• iterative reasoning and refinement. The system supports cyclical processes of hypothesis generation and evaluation, utilizing feedback from internal and external sources.

• autonomy with human optionality. The model can operate autonomously but also allows for human-in-the-loop interactions (e.g., co-pilot mode or human review stages).

• tool integration. Agents interface with external tools, databases, and libraries to perform domain-specific tasks.

Here, we propose a multi-agent system for hypothesis generation and experiment design that integrates key components from leading multi-agent AI models (Figure 3). The system begins with a Planner Agent, which breaks down the research problem and delegates data segments to specialized supporting agents. These include a Literature Review Agent for retrieving relevant prior work, a Data Analyst Agent for exploratory data analysis, and a Tool/Database Agent for accessing external information and models. The outputs from these agents are passed to a team of specialized Scientist Agents, each tasked with generating structured hypotheses within defined research areas. To foster creativity and interdisciplinary insight, an Analogy and Synthesis Agent searches for related concepts across disciplines, enriching the hypothesis pool.

Figure 3

The generated hypotheses are evaluated through multiple lenses: a Reflection Agent simulates peer review, a Critic Agent assesses empirical validity and methodological rigor, a Ranking Agent performs a tournament-style evaluation, and a Proximity and Grouping Agent identifies and eliminates redundancies and promotes diversity. Evaluation Agents in multi-agent AI systems can engage in expert consultation (e.g., one agent can query another specialized agent and/or a human can critique the results), structured debates (weighing competing hypotheses), and deliberate discussions (reaching consensus on management strategies) (7, 10). Importantly, real scientists should be part of this process by interacting with the system, providing inputs, and critiquing the generated hypotheses.

A Meta-Review Agent aggregates these evaluation results and ranks the hypotheses. A Dialogue Coordinator Agent manages agent communications during deliberation rounds, while a Memory Agent stores prior outputs and insights to support continual learning. This architecture supports an end-to-end loop of hypothesis generation, evaluation, and refinement, simulating the dynamics of a collaborative and adaptive scientific research team.

Potential roles and applications of multi-agent AI systems in soil science

Building on the current foundation of AI tools and inspired by agent-based models from other disciplines, we see huge potential for multi-agent AI systems to assist in soil science research. These agents, functioning as integrated systems, could help address complex challenges that often exceed the capacity of scientists or are too difficult to solve using isolated tools. With the ability to coordinate data collection, conduct advanced analyses, formulate hypotheses, and interact with digital, human, and potentially even physical systems, AI agents offer new avenues for investigating and understanding soil processes. Below, we highlight several promising applications.

Integrated soil digital twin agents

A multi-agent AI system could serve as a central hub for soil monitoring, continuously orchestrating data acquisition from diverse sources, including networks of in-situ and proximal sensors, as well as satellite data streams capturing multi- or hyperspectral imagery.

Continual learning and adaptation are essential for agents operating in dynamic soil environments. As new data are streamed from sensors or experiments, agents must be able to incorporate this information to refine their internal models, update predictions, and adjust ongoing strategies. For instance, an agent managing irrigation could learn from moisture sensor feedback and weather forecasts to optimize water delivery, adapting its approach throughout the growing season. Similarly, an agent monitoring temporal soil conditions could adapt its assessment based on long-term trends observed in sensor data and remote sensing imagery, potentially identifying emerging degradation issues earlier than traditional methods. It could also simulate future scenarios based on different management strategies.

A multi-agent AI system could be tasked with fusing this multimodal data to create a digital twin for monitoring and generating in silico experiments (27, 41). Algorithms could detect subtle spatial and temporal patterns based on real-time data, identifying anomalies indicative of issues such as nutrient depletion, water stress, compaction, or erosion hotspots (42). Beyond simple detection, the agent could perform integrated soil assessments from surface to subsoil conditions using multiple indicators. When combined with process-based models, an AI agent could run in silico experiments to test new land management options or model soil system responses to climate extremes. The system could then generate dynamic, high-resolution maps and actionable reports for land managers and researchers, far surpassing current manual or semi-automated approaches.

Soil microbiome analysis agents

In the biological frontier, multi-agent AI systems could tackle the complexity of soil microbial communities and their functions. Integrating high-throughput sequencing data (metagenomics or metatranscriptomics) with detailed environmental parameters (i.e., soil chemistry, mineralogy, moisture, temperature, or vegetation types), the system could employ ML models to identify key microbial taxa or functional genes that serve as sensitive indicators of soil health, nutrient cycling efficiency, or disease suppressiveness (43). It could help unravel complex microbe–microbe, plant–microbe, or soil–microbe interactions; predict the functional consequences of community shifts caused by land management (such as the use of herbicides or amendments) or climate change; and guide the development of targeted microbial inoculants (44).

AI could enhance soil microbiome research by uncovering the structural and functional potential of vast microbial protein datasets. Soil microbiomes contain vast, poorly characterized microorganisms, many of which produce proteins with unknown functions. AI models could be used to predict protein structures from metagenomic data and infer protein function by identifying structural similarities to known proteins. This would contribute to a better understanding of microbial ecology and ecosystem services and to the development of precision soil management tools grounded in protein-level insights.

Multi-agent AI could assist in formulating testable hypotheses based on existing knowledge and incoming data, such as whether a specific microbial consortium enhances nutrient availability under drought conditions, and design strategies for testing them. Multi-agent systems could therefore help researchers to better understand continent-wide microbial distributions and how these relate to soil-plant-climate interactions, thereby assuming a “whole-ecosystem approach”. They could then plan efficient strategies to test these hypotheses, perhaps by designing optimal field sampling campaigns, specifying parameters for targeted laboratory experiments, or setting up in silico experiments using simulation models.

Climate change impact and adaptation agents

To address global challenges, a multi-agent AI system could integrate climate model projections with soil process models to forecast the impacts of rising temperatures and altered precipitation patterns on key soil properties, such as organic carbon stocks, moisture regimes, and nutrient availability across different regions. It could also evaluate the effectiveness of various adaptation strategies, such as drought-tolerant cropping systems or water conservation techniques, under future climate scenarios. Importantly, feedback on soil responses to climate change [such as loss of soil organic carbon (SOC) and increased heat transfer in the soil] is needed. Furthermore, such an agent could provide essential support for monitoring, reporting, and verification of soil carbon sequestration (45). Current broad nutrient applications could be transformed through AI-designed soil sampling strategies, producing more detailed information that would enable fertilization rates to be optimized in light of both local conditions that reduce leaching and greenhouse gas emissions.

The common thread across these applications is a shift toward integrated, reasoning systems capable of handling complexity and automating workflows beyond the capacity of current tools. However, a key concern is how AI agents will perform under novel or unprecedented scenarios, such as extreme flooding or other climate-induced events, that fall outside the domain of their training data. In the absence of physical model constraints, predictive reliability may degrade under such conditions. This emphasizes the continuing need for expert oversight, domain knowledge, and hybrid approaches that combine data-driven insights with mechanistic understanding.

Hypothesis generation and experimental design agents

Multi-agent AI could facilitate knowledge discovery through automated literature analysis and synthesis (46–48). Leveraging LLMs and knowledge graphs, AI could continuously mine the vast body of soil science literature, experimental databases, and real-time monitoring data to identify knowledge gaps and propose novel, testable hypotheses about soil processes.

These systems could then move from hypothesis generation to experimental design for validation. Critically, multi-agent systems must function collaboratively with human soil scientists. The goal is not to replace human expertise but to augment it (5). Agents should be able to communicate their reasoning, present findings clearly, and solicit and incorporate expert feedback. The following section demonstrates the use of such a system.

Scientific hypotheses generation case study: soil carbon saturation research

To demonstrate the feasibility of multi-agent AI systems for hypotheses generation in soil science, we provide an example in which an AI agent system was tasked with a research question. SOC sequestration is increasingly recognized as a key strategy for mitigating climate change and enhancing soil health (49). A central concept of organic carbon persistence is soil carbon saturation, particularly regarding mineral-associated organic carbon (MAOC) (50). This concept suggests that mineral surfaces, especially within clay and silt fractions, have a finite capacity to form organo-mineral complexes due to limited reactive surface area. However, its operational relevance is under debate. Empirical studies by Begill et al. (51), Heinemann et al. (52), and Poeplau et al. (53) challenge the idea of a clear saturation plateau, showing near-linear increases in MAOC even under high carbon inputs, implying that biological and environmental constraints may often limit SOC accrual before theoretical saturation is reached. Conversely, Cotrufo et al. (54) maintain the physical validity of the saturation concept, arguing that observed inconsistencies stem from methodological artefacts, such as the misclassification of particulate organic matter (POC) as MAOC, and emphasize the need to identify real-world factors that prevent soils from reaching their saturation potential.

To explore this issue, we presented this problem to Manus AI, a general-purpose multitask AI agent (55). Manus AI is a multi-agent system built with a transformer-based large language model, organized into three coordinated agents, namely Planner, Execution, and Verification, that collaboratively break down, perform, and validate tasks. It is designed to execute complex, multi-step tasks across various domains independently. It can use external tools, including web browsers, to gather knowledge and information.

For our purposes, we used Manus AI 1.4 to generate a set of scientific hypotheses aimed at improving our understanding of soil carbon saturation and whether it is a useful concept, particularly in the context of agricultural soils under climate change. The prompt and interactions with Manus AI are provided in the Supplementary Material. The AI system was prompted to draw on recent scientific literature (both within and, more importantly, beyond soil science), generate a set of contrasting scientific hypotheses, simulate a peer review process (Reflection, Critic, Proximity, and Grouping) (Figure 3), and rank the hypotheses based on novelty, feasibility, plausibility, and scalability.

Manus AI provided the following steps in its process:

• literature review. Identification and analysis of recent scientific publications on soil carbon saturation, mineral-organic matter interactions, adsorption phenomena, and related topics.

• hypothesis generation. Development of five pairs of contrasting hypotheses addressing key uncertainties and debates around soil carbon saturation.

• simulated review:

  • Reflection Agent. Assessed hypotheses for clarity, logical consistency, and alignment with the research question.

  • Critic Agent. Evaluated hypotheses for empirical validity, testability, and scientific rigor based on current evidence.

  • Proximity and Grouping Agent. Analyzed hypotheses for conceptual overlap, redundancy, and overall diversity.

• hypothesis ranking. Synthesized the reviews and evaluated the hypotheses against criteria of novelty, feasibility, plausibility, and scalability.

Hypotheses generation on soil organic carbon saturation

Based on its analysis of recent literature from soil science and related fields (51, 54, 56, 57), Manus AI generated five pairs of hypotheses (Table 2). The hypotheses explore contrasting views on soil carbon saturation, focusing on the following themes:

• the relevance of effective versus theoretical MAOC capacity (H1);

• biological and chemical controls on MAOC formation (H2);

• climate change impacts on carbon persistence and vulnerability (H3);

• interdisciplinary mechanisms involving adsorption and co-precipitation (H4);

• management strategies based on input quality versus saturation status (H5).

The themes span both fundamental concepts reframed for practical relevance and more novel, cross-disciplinary ideas. For example, H1 offers a pragmatic reinterpretation of the carbon saturation concept, and H3 proposes responses to climate change based on a soil’s saturation status. H2 challenges the traditional mineral-centric view of MAOC formation by emphasizing microbial processing and OM chemistry as primary constraints. H4 draws on mechanisms from surface chemistry and geochemistry, which are less explored in soil carbon models, and H5, while the most operationally relevant, builds on established management principles.

Among the hypotheses outlined within the themes, H1b, H3b, and H5a stand out as novel (Table 2), as they highlight underexplored dimensions of soil carbon research. H1b emphasizes the long-term significance of theoretical MAOC capacity under high inputs or land-use scenarios, a concept often overlooked in favor of short-term, field-based studies. H3b introduces a saturation-sensitive risk perspective, suggesting that near-saturated soils may be more vulnerable to carbon loss under climate change. Some of the hypotheses are interrelated as, in practice, the findings from H3a and H3b would influence the feasibility of H1b, for example. H5a emphasizes the practical importance of maximizing high-quality carbon inputs in unsaturated soils, a widely accepted practice yet rarely formalized or directly compared to saturation-based management strategies in research.

Simulated agent reviews

The generated hypotheses then underwent a simulated multi-agent review process to evaluate them from various perspectives, as summarized in Table 3. The Reflection Agent reviewed the hypotheses for clarity, logical consistency, internal coherence, and alignment with the overall research question. The Critic Agent evaluated the hypotheses for empirical validity, testability, and scientific rigor, drawing on the analyzed literature and identifying potential challenges or contradictions. The Proximity and Grouping Agent reviewed the hypotheses for conceptual overlap, redundancy, and overall diversity and suggested potential groupings or refinements.

In summary, the simulated multi-agent reviews found the hypotheses to be clear, coherent, and diverse, empirically varied but scientifically valuable, with several requiring advanced methods for testing, and well-differentiated with no major redundancy, supporting a broad and balanced research agenda.

Hypotheses ranking

Based on simulated reviews and evaluations of novelty, feasibility, plausibility, and scalability, the hypotheses were ranked according to:

• Novelty: degree to which the hypothesis challenges existing paradigms.

• Feasibility: practicality of testing with current methods and resources.

• Plausibility: alignment with existing empirical and theoretical knowledge.

• Scalability: potential to generalize across soil types, climates, and management regimes.

The final ranking (Table 4) highlighted H1 (effective vs theoretical saturation) and H3 (climate change impacts) as particularly promising avenues for research due to their high relevance, plausibility, and scalability, despite moderate feasibility challenges. H5 (management implications) also ranked highly due to its high feasibility and practical relevance. H2 (biological/chemical controls) and H4 (interdisciplinary mechanisms), while highly novel and plausible for advancing fundamental understanding, face greater hurdles in terms of empirical testing and scaling.

A comparison with expert opinions

After our AI task was completed, a review by Georgiou et al. (58) was published, summarizing the current conceptual understanding of soil carbon saturation. The paper discusses factors controlling carbon inputs and outputs to soil MAOC. The review reinforces the view that MAOC formation depends on the quantity and chemical composition of inputs from plant and microbial sources, as well as the physical and chemical properties of the mineral matrix. MAOC formation and destabilization are further influenced by soil conditions and disturbances. Table 5 summarizes the conceptual understanding of soil MAOC saturation based on Georgiou et al. (58), comparing it with the AI-generated hypotheses.

Most of the concepts reviewed by Georgiou et al. (58), except for the steady-state concept, are included in the AI-generated hypotheses (Table 2). H1a and H1b align with core soil carbon saturation concepts but emphasize different aspects. H1a prioritizes effective capacity as the most relevant limit of carbon sequestration under real-world climate and management constraints, deeming it more practical than theoretical models. In contrast, H1b stresses the long-term significance of the theoretical mineral capacity as a boundary for maximum sequestration, especially under high carbon inputs or land-use change.

Hypotheses H2a, H2b, and H5a collectively emphasize that carbon input–response saturation is governed by more than just mineral surface availability. H2a highlights a microbial bottleneck, H2b focuses on the importance of OM chemistry in governing soil carbon saturation, and H5a supports increased, diverse carbon inputs as a practical strategy for carbon-depleted agricultural soils. However, it also indicates that, without addressing microbial and chemical constraints, saturation thresholds may still limit sequestration efficiency. The role of carbon use efficiency and microbial processing for MAOC formation (H2a) aligns with the Microbial Efficiency-Matrix Stabilization framework by Cotrufo et al. (59).

Hypotheses H3a and H3b engage with the concepts of soil saturation and vulnerability (58) in the context of climate change. H3a suggests that climate change (e.g., warming and altered precipitation) reduces a soil’s effective MAOC capacity by accelerating decomposition and altering biotic and abiotic conditions. H3b highlights that soils near effective saturation are especially vulnerable to climate pressure, as destabilization of less-protected POC and some MAOC becomes more likely. Regarding hypothesis H3a, Rocci et al. (60), in a meta-analysis, could not find a significant effect of elevated CO₂ and temperature on MAOC, while the effect of changes in precipitation regime were inconclusive due to a lack of sufficient data. Their findings underscore the need to pursue this line of research and support AI ranking. In relation to H3b, a study indicates that SOC vulnerability increases near saturation, especially when stabilization mechanisms rely on weaker or outer-layer interactions (61). Martin et al. (62) recommended prioritizing residue returns in unsaturated cropland soils and to protect SOC in forests and unimproved grasslands, given their high stocks and relatively high saturation levels. Georgiou et al. (58) therefore conclude that the effect of MAOC saturation on MAOC vulnerability remains uncertain and is highly context-dependent, requiring further research.

Hypothesis H4a (surface chemistry analogy) posits that classical surface chemistry principles, such as adsorption isotherms (e.g., Langmuir and Freundlich models), predict MAOC formation and saturation more accurately than simple linear regressions based solely on fine fraction content (63). This is further confirmed by Georgiou et al. (58), who advocate for mechanistic, process-based modeling frameworks that account for heterogeneous surface reactivity, organo–organic layering, and non-linear carbon loading behavior.

Summary

This process demonstrates how a multitask AI system can effectively support scientific hypothesis generation and evaluation through a structured, multi-agent framework. One of the most immediate benefits is the acceleration of early-stage research. AI systems can help researchers gather information and rapidly generate hypotheses that are grounded in empirical evidence and theoretical frameworks. This could reduce the time and effort typically required to transition from exploratory reading to focused scientific inquiry.

Equally important is the enhanced rigor introduced through simulated peer review. This needs to be further improved through a step-by-step sequence where researchers evaluate the basis of the hypotheses. This step requires human input and interaction. For example, soil scientists stated that they would like to see a more fundamental understanding of interactions between clay minerals and OM, as most studies use only clay content to determine or evaluate carbon stock. However, it remains uncertain whether the hypotheses generated by AI are entirely novel or simply well-synthesized from existing knowledge. The “black box” nature of the AI system also does not allow for probing of where the knowledge was derived from.

Challenges and future directions

While the potential for AI agents to revolutionize soil science is vast, realizing this vision requires overcoming significant technical, practical, and ethical challenges. As with any tool, multi-agent AI systems require appropriate training and guidelines to ensure responsible and effective implementation. Addressing these limitations will be crucial for developing robust and widely accessible agent-based systems. Minasny et al. (5) outlined a framework on how AI could be used to assist soil science research.

Data remains a fundamental challenge. AI agents, particularly those employing deep learning, thrive on large, high-quality datasets (64). However, comprehensive soil data are often scarce, fragmented, or inconsistent across different regions and measurement protocols. Issues of data heterogeneity (combining point samples, sensor streams, remote sensing imagery, and laboratory analyses), lack of standardization, and limited data-sharing practices hinder the development of generalizable models (28). Consequently, the success of agent-based systems is often limited by the availability of well-documented, ground-truth reference data rather than by data volume from proximal and remote sensors. Future efforts must focus on establishing standardized data formats and protocols, promoting open data initiatives, and developing AI techniques that are adept at learning from sparse, noisy, or multimodal data. Privacy-preserving approaches such as federated learning, secure multiparty computation, and differential privacy can enable collaborative model development without direct disclosure of data (4). Blockchain can serve a limited governance role by supporting provenance tracking, access control, and auditability of model updates (65). These approaches could address private policy issues and consolidate private data from companies.

Model interpretability and trustworthiness are paramount, especially for agents designed to support critical decision-making in agriculture and environmental management. The “black box” nature of many complex AI models can be a barrier to adoption. Developing and implementing explainable AI (XAI) methods tailored to soil science applications is essential (66). Current XAI for LLMs and generative models mainly provide transparency and provenance rather than explanations of internal reasoning. While this can improve trust and usability, it does not offer causal insight into decision processes, reinforcing the need for human oversight and uncertainty-aware deployment (67).

Predictive uncertainty in multi-agent workflows remains insufficiently explored. Sensitivity to prompts, stochastic planning, variable evidence quality, and measurement errors can propagate through agent chains, increasing the risk of spurious correlations in a data-poor and methodologically heterogeneous domain such as soil science. Uncertainty extends well beyond data quality to include structural assumptions and the reliability of the scientific literature itself. These challenges underscore the necessity of uncertainty-aware workflows and clearly defined human-in-the-loop processes, where expert judgement guides problem formulation, evaluates evidence quality, and contextualizes AI outputs. A framework is suggested in Minasny et al. (5).

Integrating AI agents with physical systems presents another barrier. While agents can excel at analyzing data and running simulations, extending their capabilities to interact with the real world (for instance, by controlling automated soil sampling robots, managing sensor networks, or directly interfacing with laboratory equipment) requires significant advances in robotics, sensor technology, and cyber-physical systems integration (68). Bridging the gap between digital intelligence and physical action is key to realizing concepts like the “self-driving lab” in soil science.

Computational cost and accessibility are practical concerns. Training large AI models and running complex agent systems can be computationally intensive, requiring significant resources that may not be available to all researchers or end-users. Developing more efficient algorithms, leveraging cloud computing platforms, and creating user-friendly interfaces are necessary to ensure equitable access to the benefits of AI agent technology. Nevertheless, AI computations require substantial energy and water for cooling systems, resulting in a significant carbon footprint (69).

Ethical considerations must be proactively addressed throughout the development and deployment lifecycle. Concerns include potential biases in algorithms trained on unrepresentative data, issues of data privacy and ownership, the equitable distribution of benefits derived from AI, and establishing clear lines of responsibility when agent-driven decisions lead to unintended consequences. Scientists also need to assess the AI systems’ outputs and the real-life implications for local communities within a given cultural and socio-economic context. Full use of AI in the scientific method, from hypothesis generation to publication, raises questions about whether researchers truly internalize knowledge produced by AI. There is a risk that scientists may not fully understand AI-derived insights, especially as systems grow more complex. The transfer of knowledge from AI to human researchers is crucial. Without this, there will be ambiguity in attributing authorship and intellectual ownership. An additional risk lies in excessive reliance on AI-generated hypotheses. As Messeri and Crockett (70) argue, assuming that AI has a deeper understanding, broader hypothesis coverage, or more objectivity than it does may lead to a “monoculture of knowers” and bias the scientific process. Scientists should remain active learners and critical thinkers, using AI as an assistive—not substitutive—tool. A framework for responsible AI development in soil science is needed.

Furthermore, effectively grounding AI agents in fundamental scientific knowledge remains a challenge. While LLMs can process vast amounts of text, ensuring they correctly apply core soil science principles [i.e., climate, organisms, relief, parent material, and time, collectively known as the CLORPT factors of soil formation (71)], it still requires sophisticated knowledge integration techniques, potentially involving knowledge graphs or hybrid AI approaches. Agents must be more than pattern recognizers; they need a degree of scientific understanding.

Creativity as an essential component of the (human) scientific process should also be nurtured, especially since this remains a limitation of multi-agent AI systems. AI models are fundamentally dependent on the data they are trained on. While these agents can autonomously search for and retrieve literature, their effectiveness is shaped by how they access, filter, and prioritize sources. This data-dependency constrains their ability to think “outside the box”, a quality often crucial for scientific breakthroughs. Although AI excels at recombining and synthesizing existing ideas, its capacity for true creativity and the generation of genuinely novel concepts remains uncertain. One potential strength, however, lies in AI’s ability to explore knowledge across disciplines that are rarely considered by soil scientists, thereby expanding the conceptual search space. The central question remains: can AI genuinely create or does it merely reconstruct and reconfigure what is already known?

A pressing issue emerging from the rise of ML and AI over the past two decades is how scientists can maintain a strong foundation in traditional soil science while leveraging AI technologies. A noticeable trend has been the decline in fundamental disciplinary education, as students and professionals are increasingly required to navigate the sophisticated world of ML and AI. This shift has affected the depth of background knowledge among emerging professionals, sometimes making it difficult to interpret or explain research outcomes meaningfully (72).

In this context, AI has the potential to ease the burden by automating time-intensive preparatory tasks, such as literature review and scenario development (as illustrated in the previous section). While AI may contribute to this “fast science” trend, it could also support the “slow science” movement by improving the work-life balance by saving time. Researchers could focus more on deepening their understanding of core soil science principles.

The scientific community must establish mechanisms to ensure foundational knowledge is not lost in the process. Overreliance on AI tools risks cognitive offloading (73). Balancing AI-driven efficiency with rigorous disciplinary training is crucial to maintaining scientific integrity and achieving long-term progress. Imagination and creativity need to be fostered. AI can supercharge soil science but only if we pair its speed with core field knowledge.

Finally, strengthening collaboration is key. Progress requires breaking down disciplinary silos and encouraging close collaboration between soil scientists, AI researchers, computer scientists, engineers, and social scientists. Platforms that facilitate this interaction, such as the AI Co-Scientist, could help develop solutions that are not only technically sound but also relevant, usable, and adopted by the end-user community. Building shared infrastructure, benchmark datasets, and open-source tools will further accelerate progress. Adopting open-access, open-source, and open-data practices, guided by the FAIR principles (findable, accessible, interoperable, and reusable), provides an essential foundation for transparency, reproducibility, and inclusive participation, helping to broaden access to the benefits of AI-enabled soil science (74).

Conclusion: cultivating soil discovery with intelligent partners

Soil science is on the brink of a transformative era, fueled by the expanding capabilities of multi-agent AI systems. Unlike traditional ML tools that focus on isolated tasks, these advanced, integrated systems offer a powerful new way to address the complexity and ever-changing nature of soil ecosystems. AI agents can seamlessly integrate multiple sources of data to build dynamic models, generate and test hypotheses, and design adaptive strategies for monitoring and sustainable management. Crucially, they work in partnership with human scientists—augmenting but not replacing their expertise.

These intelligent collaborators can dramatically accelerate the pace of discovery, transforming vast datasets into meaningful insights. By exploring multiple solution pathways and building on existing knowledge, AI can expand the boundaries of scientific exploration while still relying on human judgment to interpret, refine, and apply the findings.

While AI can emulate aspects of expert reasoning, it cannot replace the contextual judgment, creativity, and critical interpretation that human scientists bring to the research process. As such, AI should be viewed as an augmentative tool, enhancing but not substituting human-led scientific inquiry.

To fully realize this promising future, we need to foster interdisciplinary collaboration, ensure equitable access to AI tools, and thoughtfully address ethical challenges. By bridging digital innovation with real-world application, we can unlock new levels of understanding, stewardship, and security for our soils, one of the planet’s most vital and existential resources.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fsci.2026.1721295/full#supplementary-material

References

• 1

  WadouxAMC. Artificial intelligence in soil science. Eur J Soil Sci (2025) 76(2):e70080. doi: 10.1111/ejss.70080

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 2

  GrunwaldS. Artificial intelligence and soil carbon modeling demystified: power, potentials, and perils. Carbon Footprints (2022) 1(1):5. doi: 10.20517/cf.2022.03

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 3

  WeindorfDCChakrabortyS. Balancing machine learning and artificial intelligence in soil science with human perspective and experience. Pedosphere (2024) 34(1):9–12. doi: 10.1016/j.pedsph.2023.09.010

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 4

  GalliosGTsakiridisNTziolasN. Federated learning applications in soil spectroscopy. Geoderma (2025) 456:117259. doi: 10.1016/j.geoderma.2025.117259

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 5

  MinasnyBMcBratneyAB. Machine learning and artificial intelligence applications in soil science. Eur J Soil Sci (2025) 76(2):e70093. doi: 10.1111/ejss.70093

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 6

  MaYMinasnyBMaloneBPMcbratneyAB. Pedology and digital soil mapping (DSM). Eur J Soil Sci (2019) 70(2):216–35. doi: 10.1111/ejss.12790

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 7

  GaoSFangAHuangYGiunchigliaVNooriASchwarzJRet al. Empowering biomedical discovery with AI agents. Cell (2024) 187(22):6125–51. doi: 10.1016/j.cell.2024.09.022

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 8

  GottweisJWengWHDaryinATuTPalepuASirkovicPet al. Towards an AI co-scientist. arXiv [preprint] (2025). doi: 10.48550/arXiv.2502.18864

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 9

  SaeediDBucknerDAponteJCAghazadehA. Astroagents: a multi-agent AI for hypothesis generation from mass spectrometry data. arXiv [preprint] (2025). doi: 10.48550/arXiv.2503.23170

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 10

  WangHHeYCoelhoPPBucciMNazirAChenBet al. Spatialagent: an autonomous AI agent for spatial biology. bioRxiv [preprint] (2025). doi: 10.1101/2025.04.03.646459

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 11

  DresnerKStoneP. A multiagent approach to autonomous intersection management. J Artif Intell Res (2008) 31:591–656. doi: 10.1613/jair.2502

  • CrossRef
  • Google Scholar

• 12

  HarariYN. NEXUS: a brief history of information networks from the Stone Age to AI. New York, NY: Penguin Books. (2024).

  • Google Scholar

• 13

  LloretJSendraSGarciaLJimenezJM. A wireless sensor network deployment for soil moisture monitoring in precision agriculture. Sensors (2021) 21(21):7243. doi: 10.3390/s21217243

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 14

  WidyastutiMTPadarianJMinasnyBWebbMTaufikMKiddD. Mapping near-real-time soil moisture dynamics over Tasmania with transfer learning. Soil (2025) 11(1):287–307. doi: 10.5194/soil-11-287-2025

  • CrossRef
  • Google Scholar

• 15

  YinHCaoYMarelliBZengXMasonAJCaoC. Soil sensors and plant wearables for smart and precision agriculture. Adv Mater (2021) 33(20):2007764. doi: 10.1002/adma.202007764

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 16

  DaiLWangZZhuoZMaYShiZChenS. Prediction of soil organic carbon fractions in tropical cropland using a regional visible and near-infrared spectral library and machine learning. Soil Tillage Res (2025) 245:106297. doi: 10.1016/j.still.2024.106297

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 17

  NgWMinasnyBJeonSHMcBratneyA. Mid-infrared spectroscopy for accurate measurement of an extensive set of soil properties for assessing soil functions. Soil Secur (2022) 6:100043. doi: 10.1016/j.soisec.2022.100043

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 18

  PengYBen-DorEBiswasAChabrillatSDemattêJAGeYet al. Spectroscopic solutions for generating new global soil information. Innovat (2025) 6(5):100839. doi: 10.1016/j.xinn.2025.100839

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 19

  MaggiFTangFHTubielloFN. Agricultural pesticide land budget and river discharge to oceans. Nature (2023) 620(7976):1013–17. doi: 10.1038/s41586-023-06296-x

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 20

  XieXYanHLuYZengL. Simulating field soil temperature variations with physics-informed neural networks. Soil Tillage Res (2024) 244:106236. doi: 10.1016/j.still.2024.106236

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 21

  JiangZYangSSmithPPangQ. Ensemble machine learning for modeling greenhouse gas emissions at different time scales from irrigated paddy fields. (2023) 292:108821. doi: 10.1016/j.fcr.2022.108821

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 22

  WeberTKDWeihermüllerLNemesABechtoldMDegréADiamantopoulosEet al. Hydro-pedotransfer functions: a roadmap for future development. Hydrol Earth Syst Sci (2024) 28(14):3391–433. doi: 10.5194/hess-28-3391-2024

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 23

  PoggioLDe SousaLMBatjesNHHeuvelinkGBKempenBRibeiroEet al. SoilGrids 2.0: producing soil information for the globe with quantified spatial uncertainty. Soil (2021) 7(1):217–40. doi: 10.5194/soil-7-217-2021

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 24

  DemattêJASafanelliJLPoppielRRRizzoRSilveroNEQMendesWDSet al. Bare earth’s surface spectra as a proxy for soil resource monitoring. Sci Rep (2020) 10(1):4461. doi: 10.1038/s41598-020-61408-1

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 25

  FenzSNeubauerTFriedelJKWohlmuthML. AI-and data-driven crop rotation planning. Comput Electron Agric (2023) 212:108160. doi: 10.1016/j.compag.2023.108160

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 26

  HuBXieMZhouYChenSZhouYNiHet al. A high-resolution map of soil organic carbon in cropland of Southern China. Catena (2024) 237:107813. doi: 10.1016/j.catena.2024.107813

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 27

  TsakiridisNLSamarinasNKalopesaEZalidisGC. Cognitive soil digital twin for monitoring the soil ecosystem: a conceptual framework. Soil Syst (2023) 7(4):88. doi: 10.3390/soilsystems7040088

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 28

  BatjesNHCalistoLde SousaLM. Providing quality-assessed and standardised soil data to support global mapping and modelling (WoSIS snapshot 2023). Earth Syst Sci Data (2024) 16(10):4735–65. doi: 10.5194/essd-16-4735-2024

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 29

  SmithPFalloonPDKörschensMShevtsovaLKFrankoURomanenkovVet al. EuroSOMNET–a European database of long-term experiments on soil organic matter: the WWW metadatabase. J Agric Sci (2002) 138(2):123–34. doi: 10.1017/S0021859601001800

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 30

  CastroJPGaschCKFloresP. Artificial intelligence in soil science: where do we go now? Agric Environ Lett (2024) 9(2):e20134. doi: 10.1002/ael2.20134

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 31

  XuYLiuXCaoXHuangCLiuEQianSet al. Artificial intelligence: a powerful paradigm for scientific research. Innovation (2021) 2(4):100179. doi: 10.1016/j.xinn.2021.100179

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 32

  MaYMinasnyBDemattêJAMcBratneyAB. Incorporating soil knowledge into machine-learning prediction of soil properties from soil spectra. Eur J Soil Sci (2023) 74(6):e13438. doi: 10.1111/ejss.13438

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 33

  NardiPDi MatteoGHarteminkAE. Hypotheses presence and acceptance in seven soil science journals. Geoderma (2015) 243:10–7. doi: 10.1016/j.geoderma.2014.12.005

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 34

  KhanifarJ. Evaluating AI-generated responses from different chatbots to soil science-related questions. Soil Adv (2025) 3:100034. doi: 10.1016/j.soilad.2025.100034

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 35

  GhafarollahiABuehlerMJ. Automating alloy design and discovery with physics-aware multimodal multiagent AI. Proc Natl Acad Sci (2025) 122(4):e2414074122. doi: 10.1073/pnas.2414074122

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 36

  SongTLuoMZhangXChenLHuangYCaoJet al. A multiagent-driven robotic AI chemist enabling autonomous chemical research on demand. J Am Chem Soc (2025) 147(15):12534–45. doi: 10.1021/jacs.4c17738

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 37

  LuCLuCLangeRTFoersterJCluneJHaD. Towards end-to-end automation of AI research. Nature (2026) 651:914–9. doi: 10.1038/s41586-026-10265-5

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 38

  YamadaYLangeRTLuCHuSLuCFoersterJet al. The AI Scientist-v2: workshop-level automated scientific discovery via agentic tree search. arXiv [preprint] (2025). doi: 10.48550/arXiv.2504.08066

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 39

  EppeMGumbschCKerzelMNguyenPDButzMVWermterS. Intelligent problem-solving as integrated hierarchical reinforcement learning. Nat Mach Intell (2022) 4(1):11–20. doi: 10.1038/s42256-021-00433-9

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 40

  YoungIMCrawfordJW. Interactions and self-organization in the soil-microbe complex. Science (2004) 304(5677):1634–7. doi: 10.1126/science.1097394

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 41

  SilvaLRodríguez-SedanoFBaptistaPCoelhoJP. The digital twin paradigm applied to soil quality assessment: a systematic literature review. Sensors (2023) 23(2):1007. doi: 10.3390/s23021007

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 42

  ParewaiIKöppenM. A digital twin approach for soil moisture measurement with physically based rendering simulations and machine learning. Electronics (2025) 14(2):395. doi: 10.3390/electronics14020395

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 43

  XuePMinasnyBWadouxAMCDobarcoMRMcBratneyABissettAet al. Drivers and human impacts on topsoil bacterial and fungal community biogeography across Australia. Global Change Biol (2024) 30(3):e17216. doi: 10.1111/gcb.17216

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 44

  XuePMinasnyBRomán DobarcoMWadouxAMCPadarian CampusanoJBissettAet al. The biogeography of soil bacteria in Australia exhibits greater resistance to climate change than fungi. Global Change Biol (2025) 31(6):e70268. doi: 10.1111/gcb.70268

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 45

  SmithPSoussanaJFAngersDSchipperLChenuCRasseDPet al. How to measure, report and verify soil carbon change to realize the potential of soil carbon sequestration for atmospheric greenhouse gas removal. Global Change Biol (2020) 26(1):219–41. doi: 10.1111/gcb.14815

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 46

  BlanchyGAlbrechtLKoestelJGarréS. Potential of natural language processing for metadata extraction from environmental scientific publications. Soil (2023) 9(1):155–68. doi: 10.5194/soil-9-155-2023

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 47

  FureyJDavisASeiter-MoserJ. Natural language indexing for pedoinformatics. Geoderma (2019) 334:49–54. doi: 10.1016/j.geoderma.2018.07.050

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 48

  PadarianJFuentesI. Word embeddings for application in geosciences: development, evaluation, and examples of soil-related concepts. Soil (2019) 5(2):177–87. doi: 10.5194/soil-5-177-2019

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 49

  SmithP. Soils and climate change. Curr Opin Environ Sustainabil (2012) 4:539–44. doi: 10.1016/j.cosust.2012.06.005

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 50

  CotrufoMFRanalliMGHaddixMLSixJLugatoE. Soil carbon storage informed by particulate and mineral-associated organic matter. Nat Geosci (2019) 12(12):989–94. doi: 10.1038/s41561-019-0484-6

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 51

  BegillNDonAPoeplauC. No detectable upper limit of mineral-associated organic carbon in temperate agricultural soils. Global Change Biol (2023) 29(16):4662–9. doi: 10.1111/gcb.16804

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 52

  HeinemannHDonAPoeplauCMerbachIReinschTWelpGet al. No saturation of soil carbon under long-term extreme manure additions. Plant Soil (2025) 512:1367–84. doi: 10.1007/s11104-024-07146-z

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 53

  PoeplauCDechowRBegillNDonA. Towards an ecosystem capacity to stabilise organic carbon in soils. Global Change Biol (2024) 30(8):e17453. doi: 10.1111/gcb.17453

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 54

  CotrufoMFLavalleeJMSixJLugatoE. The robust concept of mineral-associated organic matter saturation: a letter to Begill et al. Global Change Biol (2023) 29(21):5986–7. doi: 10.1111/gcb.16921

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 55

  ShenMYangQ. From mind to machine: the rise of Manus AI as a fully autonomous digital agent. arXiv [preprint] (2025). doi: 10.48550/arXiv.2505.02024

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 56

  BreureTVan WesemaelBVanwindekensFCarnolMFollainSGoidtsS. Revisiting the soil carbon saturation concept to inform a risk index in European agricultural soils. Nat Commun (2025) 16:2538. doi: 10.1038/s41467-025-57355-y

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 57

  LiXSchindlerMZhouJSamaradiwakaraSWuL. Interaction between metal(loid)s and soil mineral-organic matter associations. Crit Rev Environ Sci Technol (2024) 55(9):624–47. doi: 10.1080/10643389.2024.2434328

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 58

  GeorgiouKAngersDChampinyRECotrufoMFCraigMEDoetterlSet al. Soil carbon saturation: what do we really know? Global Change Biol (2025) 31(5):e70197. doi: 10.1111/gcb.70197

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 59

  CotrufoMFWallensteinMDBootCMDenefKPaulE. The Microbial Efficiency-Matrix Stabilization (MEMS) framework integrates plant litter decomposition with soil organic matter stabilization: do labile plant inputs form stable soil organic matter? Global Change Biol (2013) 19:988–95. doi: 10.1111/gcb.12113

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 60

  RocciKSLavalleeJMStewartCECotrufoMF. Soil organic carbon response to global environmental change depends on its distribution between mineral-associated and particulate organic matter: a meta-analysis. Sci Total Environ (2021) 793:148569. doi: 10.1016/j.scitotenv.2021.148569

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 61

  MeyerNBornemannLWelpGSchiedungHHerbstMAmelungW. Carbon saturation drives spatial patterns of soil organic matter losses under long-term bare fallow. Geoderma (2017) 306:89–98. doi: 10.1016/j.geoderma.2017.07.004

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 62

  MartinMPDimassiBRomán DobarcoMGuenetBArrouaysDAngersDAet al. Feasibility of the 4 per 1000 aspirational target for soil carbon: a case study for France. Glob Change Biol (2021) 27:2458–77. doi: 10.1111/gcb.15547

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 63

  ShiJLvJPengYYaoYWeiXWangX. Mechanisms controlling the stability and sequestration of mineral associated organic carbon upon erosion and deposition. Catena (2024) 242:108119. doi: 10.1016/j.catena.2024.108119

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 64

  NgWMinasnyBMendesWDSDemattêJAM. The influence of training sample size on the accuracy of deep learning models for the prediction of soil properties with near-infrared spectroscopy data. Soil (2020) 6(2):565–78. doi: 10.5194/soil-6-565-2020

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 65

  PadarianJMcBratneyAB. A new model for intra-and inter-institutional soil data sharing. Soil (2020) 6(1):89–94. doi: 10.5194/soil-6-89-2020

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 66

  WadouxAMCMolnarC. Beyond prediction: methods for interpreting complex models of soil variation. Geoderma (2022) 422:115953. doi: 10.1016/j.geoderma.2022.115953

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 67

  ParkCLeeHLeeSJeongO. Synergistic joint model of knowledge graph and llm for enhancing xai-based clinical decision support systems. Mathematics (2025) 13(6):949. doi: 10.3390/math13060949

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 68

  NorbyJWangSWangHDengSJonesNMishraAet al. Path to autonomous soil sampling and analysis by ground-based robots. J Environ Manage (2024) 360:121130. doi: 10.1016/j.jenvman.2024.121130

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 69

  TamburriniG. The AI carbon footprint and responsibilities of AI scientists. Philosophies (2022) 7(1):4. doi: 10.3390/philosophies7010004

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 70

  MesseriLCrockettMJ. Artificial intelligence and illusions of understanding in scientific research. Nature (2024) 627:49–58. doi: 10.1038/s41586-024-07146-0

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 71

  JennyH. Factors of soil formation: a system of quantitative pedology. New York, NY: McGraw-Hill (1941).

  • Google Scholar

• 72

  DemattêJAMinasnyBHarteminkAE. Vital for sustainable agriculture: pedological knowledge and mapping. Eur J Soil Sci (2025) 76(1):e70040. doi: 10.1111/ejss.70040

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 73

  GerlichM. AI Tools in Society: Impacts on cognitive offloading and the future of critical thinking. Societies (2025) 15(1):6. doi: 10.3390/soc15010006

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 74

  HuertaEABlaiszikBBrinsonLCBouchardKEDiazDDoglioniCet al. FAIR for AI: an interdisciplinary and international community building perspective. Sci Data (2023) 10(1):487. doi: 10.1038/s41597-023-02298-6

  • Pubmed Abstract
  • CrossRef
  • Google Scholar