Machine learning-based identification of key biotic and abiotic drivers of mineral weathering rate in a complex enhanced weathering experiment

Iris Janssens, Thomas Servotte, Tullia Calogiuri, Steven Mortier, Harun Niron, Thomas Corbett, Reinaldy Poetra, Lukas Rieder, Michiel Van Tendeloo, Abhijeet Singh, Steven Latré, Stefan Vlaminck, Jens Hartmann, Jan Willem van Groenigen, Anna Neubeck, Alix Vidal, Ivan A. Janssens, Mathilde Hagens, Sara Vicca, Tim VerdonckOpen 2025 · DOI

TL;DR

This study uses machine learning to untangle the complex web of factors influencing mineral weathering, a key process for carbon dioxide removal. By analyzing data from a large-scale experiment with 30 potential drivers, researchers identified that while abiotic factors like rock type and surface area are dominant, biological components like earthworms and microbes also play a significant role, highlighting the need for a holistic approach to enhanced weathering.

What Problem Does This Paper Address?

Enhanced mineral weathering (EW) is a promising technology for removing carbon dioxide (CO2) from the atmosphere, a critical step in combating climate change. However, optimizing EW requires a deep understanding of what drives the rate of mineral weathering. This process is influenced by a multitude of factors, including the type and surface area of minerals (abiotic), as well as the presence and activity of organisms like earthworms, fungi, and bacteria (biotic). These factors can interact in complex ways, making it difficult to predict and control weathering rates. Previous studies have often focused on only a few drivers, leaving a significant knowledge gap regarding the combined effects of numerous biotic and abiotic influences in a realistic, complex system. This paper aims to address this gap by simultaneously investigating 30 potential drivers of mineral weathering.

Methods

The researchers employed a sophisticated approach combining a large-scale experimental setup with advanced machine learning (ML) techniques and explainability methods.

  • Experimental Design: An extensive 8-week column experiment was conducted with 200 columns per round, running for 10 rounds. Each column was a unique combination of 30 potential drivers, including various rock types and surface areas, irrigation settings, biochar, organic amendments, and different types and quantities of biota (earthworms, fungi, bacteria, enzymes).
  • Machine Learning Models: Three ML regression models were used to predict carbon dioxide removal (CDR) indicators:
  • LASSO (Least Absolute Shrinkage and Selection Operator): A linear model useful for feature selection in high-dimensional datasets.
  • Random Forest (RF): An ensemble method that builds multiple decision trees to improve prediction accuracy and robustness.
  • eXtreme Gradient Boosting (XGB): An advanced boosting technique that iteratively corrects errors, often leading to superior performance.
  • Explainability: To understand why the ML models made their predictions, SHapley Additive exPlanations (SHAP) values were used. SHAP quantifies the contribution of each feature to the prediction for individual data points and can reveal interactions between features.

Datasets

The study utilized data generated from their extensive 8-week column experiment. This dataset is characterized by its high dimensionality, with 30 potential drivers and numerous experimental combinations. The paper also references previous work (Calogiuri et al., 2023) for details on the experimental setup and materials.

Key Insights and Ideas

  • Abiotic Dominance, Biotic Importance: While abiotic factors (like rock type and surface area) were consistently the most influential drivers across most indicators, biotic factors were not mere bystanders. They made a significant and sometimes crucial contribution to the predictions, underscoring the need to consider the biological component in EW.
  • Indicator-Specific Drivers: Different drivers influenced each of the four measured indicators (dissolved inorganic carbon, solid inorganic carbon, total inorganic carbon, and total alkalinity) in distinct ways. This highlights that optimizing EW requires understanding which drivers affect which specific outcome.
  • Steel Slag's Complex Role: Steel slag emerged as a key driver, particularly for solid inorganic carbon (SIC) and total inorganic carbon (ATIC). However, its presence also significantly altered the impact of other factors, such as irrigation, suggesting complex interactions that can either enhance or hinder carbon sequestration depending on the specific conditions.
  • Earthworms' Dual Impact: Earthworms, both alive and dead, influenced carbon removal. Live earthworms positively impacted dissolved inorganic carbon (ADIC) and total alkalinity (ATA), likely through organic matter decomposition and nutrient cycling. Dead earthworms also showed a positive effect on ADIC and ATA, suggesting that decomposition itself releases valuable nutrients.
  • Machine Learning as a Navigator: The study demonstrates the power of ML and SHAP values in navigating highly complex, multifactorial datasets that would overwhelm traditional statistical methods, providing a robust framework for understanding intricate biogeochemical systems.

Key Results

  • Model Performance: XGBoost generally outperformed Lasso and Random Forest in predicting the four CDR indicators (ADIC, ASIC, ATIC, and ATA), achieving the best average performance metrics.
  • Dominant Drivers: For ATIC (the most representative indicator of CDR), steel slag addition and mixed rock grain sizes were the primary drivers. Earthworms and microbes also had a substantial impact.
  • Rock Type Matters: The choice of rock type was a major factor. Peridotite and basalt positively influenced ADIC, while steel slag, despite reducing ADIC, significantly increased ASIC and ATIC.
  • Irrigation Nuances: Irrigation flow rate and frequency were important. In columns without steel slag, higher irrigation increased ADIC, ATIC, and ATA. However, in columns with steel slag, higher flow rates led to lower ADIC, ASIC, and ATIC, likely due to increased precipitation.
  • Organic Amendments Boost CDR: Biochar, straw, and alfalfa mixture generally had a positive impact on CDR indicators, particularly ASIC, ATIC, and ATA.

Limitations and Future Work

  • Correlation vs. Causation: The ML and SHAP analysis reveals correlations, not direct causation. Future work could focus on experimental designs that isolate causal relationships.
  • Temporal Scope: The 8-week experimental window provides a snapshot of early-stage weathering. The long-term durability of carbon sequestration and the evolving role of drivers over time remain areas for further investigation.
  • Soil-less Environment: The experiment was conducted without soil to control conditions. Future studies should explore these processes in more realistic soil environments to understand how soil context might alter driver importance.
  • Direct CDR Measurement: The study inferred CDR from indirect indicators. Future work should aim for more direct measurements of CDR to refine understanding and optimize strategies.
  • Model Deployment: While the nested cross-validation provides robust performance estimates, deploying a final model would benefit from validation on a separate hold-out test set.
  • Conceptual Models: The authors acknowledge the difficulty in extracting a single, comprehensive conceptual model due to the complexity and interactions observed. However, they suggest that future studies focusing on subsets of the data could contribute to developing such models.

Summary updated

Abstract

<ns3:p>Background The optimization of enhanced mineral weathering as a carbon dioxide removal technology requires a comprehensive understanding of what drives mineral weathering. These drivers can be abiotic and biotic and can interact with each other. Therefore, in this study, an extensive 8-week column experiment was set up to investigate 30 potential drivers of mineral weathering simultaneously. Methods The setup included various combinations of rock types and surface areas, irrigation settings, biochar and organic amendments, along with various biota and biotic products such as earthworms, fungi, bacteria and enzymes; each varying in type or species and quantity. The resulting changes in dissolved, solid, and total inorganic carbon (∆TIC), and total alkalinity were calculated as indicators of carbon dioxide removal through mineral weathering. Three machine learning models, Least Absolute Shrinkage and Selection Operator (LASSO), Random Forest and eXtreme Gradient Boosting (XGB) regression, were used to predict these indicators. Dominant drivers of the best performing model were investigated using SHapley Additive exPlanations (SHAP). Results SHAP analysis revealed that each CDR indicator was influenced by different factors. However, key drivers were consistently abiotic, though biota also made a significant contribution to the predictions. The most representative CDR indicator, ∆TIC, was predominantly driven by steel slag addition and mixed rock grain sizes but was also substantially impacted by earthworms and microbes. Conclusions These findings provide valuable insights into the complex interplay of numerous abiotic and biotic factors that affect mineral weathering, highlighting the potential of machine learning to unravel complex relationships in biogeochemical systems.</ns3:p>
BibTeX
@inproceedings{janssens2025machine,
  author = {Iris Janssens and Thomas Servotte and Tullia Calogiuri and Steven Mortier and Harun Niron and Thomas Corbett and Reinaldy Poetra and Lukas Rieder and Michiel Van Tendeloo and Abhijeet Singh and Steven Latré and Stefan Vlaminck and Jens Hartmann and Jan Willem van Groenigen and Anna Neubeck and Alix Vidal and Ivan A. Janssens and Mathilde Hagens and Sara Vicca and Tim Verdonck},
  title = {{Machine learning-based identification of key biotic and abiotic drivers of mineral weathering rate in a complex enhanced weathering experiment}},
  booktitle = {Open},
  year = {2025},
  doi = {10.12688/openreseurope.19252.3},
}
Figure