Low Temperature Pyrolysis of Plastics Based on Small Sample Machine Learning: Product Prediction, Optimization Conditions, and Model Interpretability Analysis

This study proposes a novel data-driven strategy that integrates machine learning (ML) with low-temperature catalytic pyrolysis to optimize the selection of plastic feedstocks and catalyst designs. Using a custom-built experimental data set based on polyethylene (PE) model compounds and zeolite ZSM-5 catalysts, we trained and evaluated three ML models─TabPFN, CatBoost, and XGBoost─on a small-sample data set of 105 orthogonally designed experiments. TabPFN outperformed conventional gradient boosting models in both classification and regression tasks, achieving an R2 of 0.982 and RMSE of 4.37 on test data, with strong generalization capacity. SHAP analysis and Pearson correlation jointly revealed that temperature and the Si/Al ratio were the dominant factors influencing C2–C6 olefins selectivity, with the latter exhibiting an inverted U-shaped influence on olefin selectivity. The trained TabPFN model successfully identified optimal reaction conditions, which were experimentally validated with a high C2–C6 olefin yield of 82.5% and minimal coke formation. This work demonstrates the potential of ML-based small-sample frameworks for rapidly optimizing catalytic systems, and the analysis of SHAP interpretation provides possible insights into the reaction pathway.