March 18-20, Zhangjiagang, Jiangsu, scan the QR code to register

More than 300 million tons of plastic waste are generated worldwide each year, of which only 21% is recycled or incinerated, and the rest enters the environment in large quantities, causing chain problems such as microplastic pollution and ecological damage. Traditional physical recycling technology is difficult to handle mixed plastics, while thermochemical conversion technology - especially catalytic cracking - is considered the key to solving plastic pollution because of its strong adaptability and controllable products. However, there are three major problems in the catalytic cracking process:

1. Complex products : Cracking products include solid wax, liquid oil and gas. Although liquid oil is of high value, its composition is complex;
2. Difficulty in catalyst screening : Different plastics require matching specific catalysts (such as molecular sieves), and traditional experimental methods are time-consuming and labor-intensive;
3. Reaction conditions are difficult to optimize : parameters such as temperature and catalyst ratio require repeated trial and error, which is inefficient.

the Institute of Urban Environment of the Chinese Academy of Sciences took a different approach and combined machine learning (ML) with molecular sieve catalysis technology to develop an intelligent reverse design framework, which successfully achieved efficient cracking of waste plastics into high-value fuel. The relevant results were published in the top international journal "Applied Energy" , attracting widespread attention from academia and industry.

Editor's note : Research by the Chinese Academy of Sciences team has proven that machine learning can not only accelerate catalyst screening and process optimization, but also break through the boundaries of traditional cognition and discover "counterintuitive" optimal solutions (such as low temperature and high yield) . This technology is expected to be extended to green chemical fields such as biomass conversion and CO2 catalysis, providing underlying technical support for the "dual carbon" goals. Plastic waste is no longer a burden, but the starting point of the "urban oil field".

— 1 —

Technological breakthrough: How machine learning can tame plastic waste

1. Data-driven: Building the world’s largest waste plastic cracking database

The team extracted 280 sets of liquid oil yield data and 142 sets of C5-C12 hydrocarbon yield data from the literature since 2000, covering key parameters such as plastic types (PP, PE, PS, etc.), molecular sieve properties (specific surface area, silicon-aluminum ratio, etc.), reaction conditions (temperature, catalyst ratio), etc. Through data cleaning, missing value filling (such as k-nearest neighbor algorithm) and normalization processing, the foundation of a high-precision prediction model was built.

Figure 1. System diagram , including detailed steps of prediction, interpretation, and optimization for machine learning model development.

2. Model selection: XGBoost stands out

The team tested three tree models: Random Forest (RF), Gradient Boosting Regression (GBR), and Extreme Gradient Boosting (XGBoost). The results show that XGBoost performs best in predicting the yield of liquid oil and C5-C12 hydrocarbons, with test set R² reaching 0.85 and 0.87 respectively, and errors (RMSE) of only 7.72% and 7.73%, far exceeding the "single factor optimization method" of traditional experiments.

Figure 2. Statistical analysis of raw data on the yields of liquid oils and C5-C12 hydrocarbons in zeolite-catalyzed cracking of waste plastics according to (a) plastic type, (b) zeolite framework type code, and (c) identified zeolite class.

3. Feature analysis: Four key factors emerge

Through SHAP value analysis and feature importance ranking, the team found that:

• Polyethylene (PE) ratio : PE content exceeding 10% will significantly reduce the yield of liquid oil and light hydrocarbons;
• Reaction temperature : 380–550°C is the high-yield range for liquid oil, but C5-C12 hydrocarbons only reach their peak at 380–430°C;
• Molecular sieve specific surface area : Too high or too low is not good, the best value is about 350 m²/g;
• Silicon-to-aluminum ratio (Si/Al) : A high silicon-to-aluminum ratio (such as 200) is beneficial to liquid oil, while a low silicon-to-aluminum ratio (such as 25) promotes the formation of light hydrocarbons.

Figure 3. PCC of active metal descriptors (a) and their importance in predicting liquid yield (b) and yield of C5-C12 hydrocarbons (c).

4. Reverse design: AI guides experiments, with higher yields than expected

Combined with the particle swarm optimization (PSO) algorithm, the team selected the optimal combination: polystyrene (PS) as raw material, ZSM-5 molecular sieve (Si/Al=200, specific surface area 352 m²/g) as catalyst, and reaction temperature 413°C . The model predicted an oil yield of 80.85%, while the experimental verification result was as high as 87.82% , with an error of only -7.93%. Even more amazingly, the temperature was nearly 200°C lower than that reported in the literature, greatly reducing energy consumption.

— 2 —

Experimental verification: the leap from data to practice

The team built a two-stage catalytic cracking device (Figure S1), using PS plastic provided by Fujian Petrochemical and molecular sieve catalyst from Nankai University . The experiment showed:

Figure S1. Schematic diagram (a) and experimental setup (b) of zeolite-catalyzed pyrolysis of waste plastics. (1. Nitrogen bottle; 2. Mass flow controller; 3. Temperature controller; 4. Laboratory two-stage catalytic pyrolysis tube furnace; 5. Plastic; 6. Catalyst; 7. Condenser; 8. Liquid oil collection bottle; 9. Oil washer (acetone); 10. Air bag

• C5-C12 hydrocarbons account for 48% of liquid oil and can be used directly as a gasoline component;
• The gaseous products (such as methane and ethylene) can be recycled to provide energy, making the process self-sustaining;
• Compared with traditional high-temperature cracking (600°C), energy consumption is reduced by more than 30%.

This result not only verifies the reliability of the model, but also reveals the huge potential of machine learning in complex chemical systems.

— 3 —

Future Outlook: Challenges from Laboratory to Industrialization

Despite the remarkable results, the team pointed out three major challenges to be solved:

1. Data size limitation : There are only a few hundred data sets available, and it is necessary to jointly build a shared database with global laboratories;
2. Insufficient mechanism explanation : In the future, we plan to integrate density functional theory (DFT) calculations to reveal the microscopic mechanism of molecular sieve catalysis;
3. Engineering scale-up : There are differences between laboratory tests and industrial-scale reactors, and a dynamic adaptive ML model needs to be developed.

《Zeolite-catalytic pyrolysis of waste plastics: Machine learning prediction, interpretation, and optimization 》

Scan the QR code to download the full text

(Source: Applied Energy , original text available if needed)

Disclaimer: This public account is intended to disseminate industry news and scientific research results, and to promote exchanges in the field of plastic recycling. It is not intended for any commercial purpose. If there are any copyright issues, please contact us and we will modify or delete them in a timely manner. 

— 4 —
ChinaReplas2025 The 31st China Plastic Recycling and Regeneration Conference March 18-20, 2025 · Zhangjiagang, Suzhou