In a recent study published in the journal Energy Conversion and Management: X, researchers Kusum Yadav, Lulwah M. Alkwai, Shahad Almansour, and Mehrdad Mottaghi investigated new methods to optimize the production of biochar. Biomass stands out as a critical renewable resource that can help industries transition away from fossil fuels and minimize global warming. Turning this plant matter into biochar through pyrolysisPyrolysis is a thermochemical process that converts waste biomass into bio-char, bio-oil, and pyro-gas. It offers significant advantages in waste valorization, turning low-value materials into economically valuable resources. Its versatility allows for tailored products based on operational conditions, presenting itself as a cost-effective and efficient More, a thermal degradation process, yields a solid carbon residue with immense utility for agricultural soil enhancement, bioenergy generation, and sustainable waste management. However, because the final output varies dramatically depending on the specific plant properties and the heating conditions, manufacturing biochar efficiently has historically been a challenge. To address this problem, the authors developed a highly sophisticated machine learning system capable of forecasting exact production amounts.

The team gathered a large database containing over two hundred experimentally validated observations from peer-reviewed scientific studies to train their digital models. This dataset mapped out fourteen distinct characteristics of different raw materials, such as their elemental composition, surface area, and specific processing settings like maximum heating temperature and residence timeResidence time refers to the duration that the biomass is heated during the pyrolysis process. The residence time can influence the properties of the biochar produced.   More. They applied a core machine learning technique known as gradient boosting decision trees, which works by connecting multiple weak computer algorithms in a sequence to continuously correct errors and tackle complex, non-linear patterns. To make this core model as precise as possible, the researchers tested four different nature-inspired optimization strategies to find the perfect operational settings. These strategies included systems mimicking the social dynamics of bird flocks, the bubble-net hunting methods of humpback whales, a probabilistic metal-cooling approach, and the path-finding behavior of foraging ants.

The final results revealed clear differences in accuracy and reliability among the optimized computer models. The configuration that incorporated the ant colony optimization strategy achieved the highest predictive accuracy by a wide margin, delivering a test predictability score of 0.709 and the lowest overall average forecasting errors when exposed to completely unseen data. This particular method uses virtual pheromone trails to guide the computer swarm toward the best possible mathematical settings, effectively preventing the program from making premature conclusions. In contrast, the model based on whale hunting strategies suffered from severe overfitting. While the whale model memorized the training data almost perfectly, it failed to generalize effectively when faced with new scenarios, proving that aggressive pattern-matching does not guarantee robust real-world performance. Additionally, while the bird flocking and metal-cooling methods ran significantly faster, taking only fifty seconds compared to over eleven minutes for the ant algorithm, they yielded much lower prediction accuracy on independent tests.

To ensure the computer models were not just black boxes delivering numbers without context, the authors integrated an explainable artificial intelligence tool to uncover the underlying chemical logic. This advanced sensitivity analysis verified that the model had correctly internalized fundamental thermochemical principles. The analysis proved that a material’s initial ash fraction, the total duration of heating, and the peak processing temperature are the three dominant factors governing biochar formation. Specifically, higher peak temperatures trigger an aggressive volatilization of organic matter, which systematically reduces the final solid biochar yield while boosting its structural strength. Conversely, longer heating durations and higher intrinsic ash components strongly correlate with an increase in solid carbon retention. By successfully combining rigorous data screening, multi-algorithmic tests, and transparent AI explanation tools, this research establishes a reproducible digital foundation that helps engineers refine biochar manufacturing, supporting global climate action and sustainable agriculture.

Source: Yadav, K., Alkwai, L. M., Almansour, S., & Mottaghi, M. (2026). On the construction of hybrid algorithms to predict biochar yield as a function of pyrolysis parameters. Energy Conversion and Management: X, 102047.