Machine learning takes biochar dye cleanup to the next level
image: Enhanced machine learning prediction of biochar adsorption for dyes: Parameter optimization and experimental validation
Credit: Chong Liu, Paramasivan Balasubramanian, Xuan Cuong Nguyen, Jingxian An, Sai Praneeth, Pengyan Zhang & Haiming Huang
Industrial dyes used in textiles, plastics, paper, and cosmetics make wastewater vividly colored and potentially toxic. Many of these dyes resist normal treatment, threatening aquatic life and human health. Now researchers have harnessed machine learning to predict how well biochar, a carbon-rich, low-cost material made from plant waste, can remove these stubborn pollutants from water.
The study, published in Carbon Research, compared nine advanced machine learning models to forecast how different types of biochar capture various dyes under changing environmental conditions. The research team found that a model known as CatBoost gave the most accurate results, achieving a coefficient of determination (R²) of 0.988 and a root-mean-square error of 0.0839. These scores indicate a near-perfect match between the model’s predictions and laboratory data.
“Biochar is already a promising material for cleaning wastewater, but its performance varies widely depending on how it is made and what it is used to remove,” said corresponding author Chong Liu of Dongguan University of Technology. “By integrating machine learning, we can understand these complex relationships faster and design more efficient biochars for real-world applications.”
To build the model, the researchers assembled 685 data sets covering 43 biochars and 15 dyes from published studies. They standardized all measurements, corrected missing values, and removed outliers to ensure a reliable database. The program evaluated chemical and physical properties of each biochar, the pH and temperature of dye solutions, and dye molecular characteristics. It then used Bayesian optimization to fine-tune model parameters and five-fold cross-validation to confirm stability.
The analysis revealed that experimental conditions such as dye concentration and temperature exert the greatest influence on adsorption efficiency, accounting for about half of the predictive power. Biochar characteristics contributed roughly one-third, while dye type made up the remainder. Among all factors, the initial dye concentration relative to biochar dosage emerged as the most critical.
To verify the predictions, the team tested biochar made from cotton straw on three common dyes: methylene blue, congo red, and malachite green. The experimental data closely matched the CatBoost model, with a validation R² of 0.904. This strong agreement demonstrates that machine learning can reliably guide the design of next-generation biochars.
The researchers also created a user-friendly computer interface using PySimpleGUI that allows scientists and engineers to input their own data and instantly estimate dye removal performance. The open-source code and dataset are available on GitHub to encourage transparency and collaboration.
According to co-author Haiming Huang, the work shows how artificial intelligence can accelerate sustainable solutions for water pollution. “Our goal is to make predictive modeling accessible so that more laboratories and industries can adopt data-driven tools to treat wastewater efficiently and at lower cost,” Huang said.
The authors plan to expand the database with additional real-world cases and test the model under variable field conditions. Such advances could bring machine-learning-guided biochar design from the laboratory to full-scale wastewater treatment, helping transform waste biomass into an effective weapon against water contamination.
===
Journal reference: Liu, C., Balasubramanian, P., Nguyen, X.C. et al. Enhanced machine learning prediction of biochar adsorption for dyes: Parameter optimization and experimental validation. Carbon Res. 4, 46 (2025). https://link.springer.com/article/10.1007/s44246-025-00213-9
===
About Carbon Research
The journal Carbon Research is an international multidisciplinary platform for communicating advances in fundamental and applied research on natural and engineered carbonaceous materials that are associated with ecological and environmental functions, energy generation, and global change. It is a fully Open Access (OA) journal and the Article Publishing Charges (APC) are waived until Dec 31, 2025. It is dedicated to serving as an innovative, efficient and professional platform for researchers in the field of carbon functions around the world to deliver findings from this rapidly expanding field of science. The journal is currently indexed by Scopus and Ei Compendex, and as of June 2025, the dynamic CiteScore value is 15.4.