Biochar may look like little more than charcoal, but inside its carbon-rich structure it can harbor chemically active species that persist long after production. A study from Kunming University of Science & Technology has now combined laboratory experiments with interpretable machine learning to reveal how the elemental composition of lignocellulose-derived biochar influences persistent free radicals (PFRs)—long-lived molecular fragments capable of triggering powerful chemical reactions. The findings suggest that measurements as simple as hydrogen, carbon, and oxygen content could help predict both how many persistent radicals a biochar contains and what kinds of radicals are present, offering a potential new way to assess its performance and environmental safety before it is deployed.
Biochar is produced by heating plant-based biomass under oxygen-limited conditions, a process known as pyrolysis. Depending on the feedstock and temperature, pyrolysis rearranges the molecular architecture of cellulose, hemicellulose, and lignin, creating a porous solid dominated by carbon. This material is increasingly investigated for improving soil, capturing pollutants, storing carbon, and supporting wastewater treatment. Yet biochar is not chemically inert. Defects in its carbon framework, oxygen-containing groups, and aromatic structures can stabilize unpaired electrons, forming PFRs that may survive for days, weeks, or even months. Those radicals can be useful because they may promote the formation of reactive oxygen species that degrade contaminants. The same chemistry, however, could generate oxidative stress in organisms if poorly characterized biochar enters soil or water.
The researchers set out to determine whether readily measured elemental properties could serve as reliable indicators of PFR behavior. They compiled 263 paired records reporting both PFR concentration and electron paramagnetic resonance g-Factor values from previously published studies. The g-Factor is a spectroscopic parameter that helps describe the electronic environment of an unpaired electron. Because its value shifts according to the atom or molecular structure surrounding the electron, it can provide clues about whether a radical is primarily carbon-centered or associated more strongly with oxygen-containing groups. By examining both concentration and g-Factor, the researchers could investigate not only how much radical activity biochar possessed, but also whether its chemical character changed as its composition changed.
The team compared six machine-learning approaches: XGBoost, gradient boosting, support vector regression, a shallow neural network, random forest, and ensemble learning. These algorithms were trained to identify nonlinear relationships between elemental descriptors and the two radical-related outcomes. Unlike a conventional linear model, machine learning can detect interactions in which one variable changes the effect of another—for example, a situation in which hydrogen content matters differently at different oxygen-to-carbon ratios. The researchers then applied interpretable machine-learning techniques to move beyond prediction and examine the chemical logic behind the models. This distinction is important: a highly accurate algorithm may forecast an outcome without explaining it, whereas an interpretable model can help researchers understand which measurable features are driving the forecast.
The strongest predictors of PFR concentration were the hydrogen-to-carbon ratio, known as H/C, and the overall oxygen content of the biochar. In general, lower H/C values and lower oxygen content were associated with higher concentrations of persistent radicals. The H/C ratio is influenced by the degree of carbonization and the transformation of hydrogen-rich biomass components during pyrolysis. A lower value often indicates a more condensed, aromatic carbon structure, which can contain defects and electronically active sites capable of stabilizing unpaired electrons. Oxygen content, meanwhile, reflects the presence of oxygenated functional groups and other oxygen-bearing structures that can alter the distribution and stability of electrons across the carbon matrix. Their combined influence indicates that PFR formation is controlled by the evolving chemical architecture of biochar rather than by carbon content alone.
The g-Factor followed a different pattern. Oxygen content and the oxygen-to-carbon ratio, or O/C, were the most influential predictors of this measurement. Increasing oxygen content and O/C generally corresponded to higher g-Factor values, suggesting a shift toward radicals with greater oxygen-centered character. In electron paramagnetic resonance analysis, this shift can reflect stronger interaction between the unpaired electron and oxygen atoms or oxygen-containing functional groups. The result gives researchers a potential compositional signal for distinguishing changes in radical type, not simply changes in radical abundance. The analysis also uncovered a more subtle interaction: when O/C values were similar, biochars with higher hydrogen content tended to show higher g-Factor values. Such relationships could be easily missed by simple correlation tests, but they emerged through the models’ ability to examine multiple variables simultaneously.
The predictive performance was substantial despite the complexity of the underlying chemistry. The best models produced test R² values of 0.7797 for PFR concentration and 0.7647 for g-Factor, indicating that the models explained a large share of the variation in previously unseen data. Their ability to distinguish samples with relatively high and low values was even stronger, with receiver operating characteristic area-under-the-curve scores above 0.96. An ROC-AUC near 1 indicates that a model can reliably separate two categories, in this case samples exhibiting comparatively high or low radical-related measurements. These results do not mean that elemental composition fully determines PFR behavior, since pyrolysis conditions, mineral content, surface structure, moisture, and measurement protocols can also matter. They do suggest, however, that elemental analysis contains a powerful and practical chemical signal.
To test whether the computational findings were reflected in real materials, the researchers prepared biochars from five different lignocellulosic sources: cellulose, lignin, peanut hull, rice straw, and pine sawdust. The feedstocks were subjected to different pyrolysis temperatures, creating materials with distinct chemical compositions and carbon structures. Independent laboratory measurements supported the main model-derived trends. Spearman correlation analysis, which evaluates whether two variables move consistently together without requiring a strictly linear relationship, confirmed that H/C and oxygen content were most closely associated with PFR concentration. Oxygen content and O/C showed the strongest relationships with g-Factor. This experimental validation strengthens the case that the machine-learning results were not merely artifacts of a heterogeneous published dataset.
The study offers a practical framework for designing biochars with deliberately tuned reactivity. A producer seeking high radical activity for pollutant degradation might adjust feedstock selection and pyrolysis conditions to encourage the structural features associated with greater PFR concentration. Conversely, applications involving direct contact with soil, plants, animals, or water may require materials whose radical abundance and oxygen-centered character are more carefully controlled. Because elemental composition can be measured relatively quickly compared with detailed spectroscopic and mechanistic characterization, H/C, oxygen content, and O/C could become useful screening indicators. The approach may also help researchers compare biochars made under different conditions and identify materials that require additional toxicological testing before environmental use.
The researchers emphasize that their results are best viewed as a bridge between data science and experimental chemistry, not as a replacement for direct measurement. Persistent free radicals can evolve during storage and react with minerals, water, oxygen, and pollutants, meaning that a biochar’s behavior in the field may differ from its behavior immediately after pyrolysis. Even so, linking composition to radical concentration and g-Factor provides a faster route to meaningful predictions and reveals chemical interactions that would otherwise remain hidden. By showing that a few elemental ratios can help explain the amount and identity of long-lived radicals in biochar, the study brings scientists closer to producing carbon materials that are not only effective, but also predictable and environmentally responsible.
Subject of Research: Persistent free radicals in lignocellulose-derived biochar and their relationship to elemental composition.
Article Title: Elemental composition-based prediction of persistent free radicals concentration and g-Factor in lignocellulose-derived biochar combining interpretable machine learning and experimental analysis
News Publication Date: 17-Aug-2026
Web References: https://doi.org/10.48130/bchax-0026-0020
References: Gao L, Xu C, Li M, Xu Z, Tao W. 2026. “Elemental composition-based prediction of persistent free radicals concentration and g-Factor in lignocellulose-derived biochar combining interpretable machine learning and experimental analysis.” Biochar X 2: e022. DOI: 10.48130/bchax-0026-0020.
Image Credits: Linjian Gao, Chengcheng Xu, Mengzi Li, Zhongda Xu and Wenmei Tao.
Keywords
Biochar, persistent free radicals, g-Factor, lignocellulose, pyrolysis, machine learning, interpretable artificial intelligence, elemental composition, environmental remediation, oxidative stress.
