Abstract:The widespread use and continual upgrading of electronic devices have generated a large amount of electronic solid waste. As devices are upgraded and replaced, the quantity of discarded electronic products has increased year by year. Thermal conversion processes are an important route for valorizing electronic solid waste; however, brominated flame retardants present in such waste can readily transform into pollutants, including dioxins. Accurately predicting the formation of brominated products during thermal conversion is therefore crucial for controlling pollutant emissions. This study focused on the pyrolysis of waste printed circuit boards, a representative type of electronic solid waste. We compiled 653 pyrolysis data points from 76 research articles published since 2000 and applied five common machine learning models—extreme gradient boosting regression (XGBoost), random forest (RF), multilayer perceptron (MLP), Gaussian process regression (GPR), and support vector regression (SVR)—to predict the relative content of brominated phenols, brominated bisphenol A (Br-BPA), brominated polycyclic aromatic hydrocarbons (Br-PAHs), hydrogen bromide (HBr), and aliphatic compounds in the pyrolysis products. Key input variables included pyrolysis parameters (e.g., temperature, heating rate, and residence time), feedstock parameters (e.g., elemental composition), and characteristics of inorganic components (e.g., average atomic properties). The hyperparameters were optimized by grid search combined with five-fold cross-validation. The prediction results showed that ensemble models, particularly XGBoost, achieved superior performance in this multi-feature, data-limited setting: for all targets on the test set, the coefficient of determination (R2) exceeded 0.8 and the root mean square error (RMSE) was below 3.0%. The XGBoost model provided the best predictive accuracy for the relative content of aliphatic compounds, whereas its performance was least satisfactory for Br-BPA; for the latter, the test-set R2 and RMSE were 0.820 and 2.247%, respectively. Feature-importance analysis was then used to explore the mechanisms of bromine migration and removal. Pyrolysis parameters such as temperature and residence time strongly influenced the bromine content in liquid pyrolysis products: high temperatures and long residence times promoted the removal of organic bromine and its conversion to inorganic bromine. Feedstock properties, notably carbon content, affected the types of brominated compounds formed: high carbon content favored the production of brominated phenols and Br-BPA, while high bromine content facilitated the formation of Br-PAHs. Inorganic components had a significant influence on HBr formation. Non-brominated aliphatic products were only weakly affected by debromination conditions and exhibited trends similar to pyrolysis products from conventional organic polymers. These findings demonstrate the potential of machine learning methods for electronic solid waste treatment and provide predictive guidance for process optimization and for directional control of pollutant formation during thermal conversion of electronic solid waste. Close-