Predicting Soil Carbon Gains from Biochar Application Using Interpretable Machine Learning

Soil organic carbon (SOC) contributes to soil fertility and agricultural productivity and is an important part of the global carbon cycle. Increasing the amount stored in agricultural soils can improve soil quality and help mitigate climate change. Biochar, a carbon-rich material produced by heating biomass under oxygen-limited conditions, is one option for increasing soil carbon. It can also improve nutrient retention and influence microbial activity, but its effects vary considerably between agricultural sites.

Estimating how much carbon a soil will gain from biochar is not straightforward. Application rate and production temperature both matter, as do soil texture, initial carbon content, pH, climate, and cropping practices. Even identical treatments can produce substantially different results from one location to another. Choosing an appropriate application rate requires a better understanding of how these factors interact.

Traditional soil carbon models describe the biological and chemical processes involved in carbon cycling, often requiring detailed field measurements and extensive calibration. Machine learning takes a different approach, using experimental data to identify relationships without explicitly describing every underlying process. Accurate predictions across different agricultural conditions remain challenging, especially when the influence of individual variables is not fully understood.

In a recent research paper published in GCB Bioenergy, Dr. Avedananda Ray, Dr. Yujuan Chen, and Professor Dafeng Hui from Tennessee State University, together with Dr. Xin Li and Dr. Wenju Zhang from the Chinese Academy of Agricultural Sciences and Dr. Xinyao Yang from Xi’an Jiaotong-Liverpool University, developed a machine learning model combining ExtraTrees, LightGBM, and CatBoost to predict SOC changes following biochar application. They also used SHAP analysis to interpret the predictions and created a graphical interface for estimating soil carbon responses under different field conditions.

Experiments and Findings

The researchers analyzed 800 observations from 104 published field studies, comparing SOC measurements in biochar-treated soils with those from untreated controls. The dataset also contained information about the biochar, soil properties, climate, and agricultural management. Carbon changes were expressed as logarithmic response ratios so that findings from different experiments could be compared.

Several machine learning algorithms were evaluated before the team selected three for the final ensemble. Each uses a different method to identify relationships in the data, and their predictions were averaged to produce a single estimate. Most observations were used to train the models, with the remaining data reserved for testing.

The ensemble achieved a coefficient of determination (R²) of 0.86 on the test dataset, indicating close agreement between predicted and measured SOC responses. Its accuracy was slightly higher than that of the individual models, including CatBoost, which also performed well. In a separate evaluation using eight independent observations, prediction errors remained below 25% in every case. The team performed SHAP analysis which identified biochar application rate as the strongest predictor of carbon gains, followed by crop type and soil characteristics.

The authors also found the relationship between application rate and SOC was not linear and carbon gains increased as more biochar was added, but the additional gains became smaller. The model identified an important change in the response around 40 tonnes per hectare, with gains approaching saturation at approximately 120 tonnes per hectare.

Biochar produced larger relative increases in soils with low initial carbon content than in soils already rich in carbon. Soil pH influenced the response as well, with higher predicted gains in slightly acidic to alkaline soils than in more acidic soils.

Differences also appeared among cropping systems and soil types. Legume and fruit crops were associated with greater SOC increases than cereal and vegetable crops. Fine-textured soils responded more strongly than coarse-textured soils, possibly reflecting differences in organic matter inputs, microbial activity, and the protection of carbon by soil minerals.

Predicted carbon gains also varied with biochar production temperature and initial soil nitrogen content, with the response tending to level off at higher nitrogen levels. Longer field experiments recorded greater carbon accumulation.

Engineering Significance

Average results from field experiments cannot always predict how biochar will perform at a particular agricultural site. The model developed by Professor Dafeng Hui and colleagues accounts for differences in soil properties, climate, cropping practices, and biochar characteristics to provide more specific estimates of carbon gains.

For soil management, identifying where the response to increasing biochar application begins to level off may support more efficient use of the amendment. Researchers can use the model to compare potential treatments and select conditions for further field experiments.

SHAP analysis makes the predictions easier to interpret by identifying the variables most strongly associated with carbon accumulation and these relationships describe patterns in the experimental data rather than establishing the biological or chemical processes responsible for them.

The independent validation produced encouraging results, and further testing across different soils and agricultural systems will provide a clearer assessment of the model’s performance under a wider range of conditions. To make the model accessible, the team developed a graphical interface where users can enter biochar properties, soil characteristics, and application conditions to obtain predicted SOC responses. Researchers and agricultural managers can use this tool to explore treatment options and plan experiments without needing experience in machine learning.