Fertilizer or pollutant: analyzing the effects of biochar on soil organisms using machine learning

In the context of carbon neutrality targets, biochar is widely promoted as a soil amendment to sequester organic carbon in soils. Although a wealth of research has illustrated the benefits of biochar to plants, its potential toxicity to soil fauna and microbes requires serious consideration. The aim of this study was to perform a meta-analysis of experimental data on biochar effects (i.e. percentage change in endpoints after biochar application compared to the control group) on plants, animals, and microorganisms. The experimental data were extracted from 61 papers and consists of 1329 data points. In a next step, machine learning was used to develop a classifier to predict, whether biochar has positive or negative consequences on soil organisms based on biochar and soil properties. The meta-analysis shows that the effect of biochar is negatively correlated with the biochar application rate, biochar pH, pyrolysis temperature, and soil pH. A random forest classifier was then developed to classify whether biochar was “beneficial” or “hazardous” based on four types of descriptors: biochar properties, soil properties, test organism, and endpoint type. The accuracy of the best model achieved an R2 of 0.79. In the next step, a quantitative model was developed to predict the effect with an R2 of 0.48. The model is of great significance for understanding the role of biochar in soil and improving the quality control strategy for biochar production. Positive and negative effects of biochar on soil biota were extracted from 61 studies A model was developed to predict whether biochar has positive or negative consequences 4 types of descriptors: biochar properties, soil properties, test organism, endpoint type The accuracy of the best model achieved an R2 of 0.79 Positive and negative effects of biochar on soil biota were extracted from 61 studies A model was developed to predict whether biochar has positive or negative consequences 4 types of descriptors: biochar properties, soil properties, test organism, endpoint type The accuracy of the best model achieved an R2 of 0.79