A review published in ENGINEERING Energy systematically evaluates how artificial intelligence is transforming waste plastic pyrolysis. By integrating kinetic parameter identification, machine learning product prediction, model interpretability, and AI-driven reactor design, researchers outline an advanced roadmap toward the autonomous, intelligent chemical recycling of global plastic waste.
With global plastic consumption projected to nearly triple by 2060 and less than 10% currently recycled, thermal conversion via plastic pyrolysis offers a promising industrial solution for transforming post-consumer waste polymers into valuable liquid fuels and raw chemical feedstocks. However, scaling up plastic pyrolysis has long been bottlenecked by variable feedstock streams, complex non-linear degradation mechanisms, and severe heat and mass transfer limitations inside reactors.
In a comprehensive review published in ENGINEERING Energy , researchers from Shanghai Jiao Tong University, Guangdong Technion-Israel Institute of Technology, and the Technion-Israel Institute of Technology systematically evaluate how artificial intelligence (AI) is solving these foundational challenges—from fundamental reaction kinetics to scale-up reactor design and fully automated, agent-driven workflows.
Traditional laboratory and computational methods, such as density functional theory (DFT) or pure Computational Fluid Dynamics (CFD), often struggle with high computational costs and exponential complexity when applied to real-world, mixed plastic streams. AI models bypass these traditional computation bottlenecks by directly processing high-dimensional, non-linear experimental and physical data.
To overcome the challenges of process optimization and fundamental exploration, the review highlights several key scientific breakthroughs and methodologies across five critical areas:
- Accelerated Kinetic Parameter Identification: Advanced metaheuristic algorithms, such as Genetic Algorithms (GA), Particle Swarm Optimization (PSO), and Whale Optimization Algorithms (WOA), successfully extract kinetic parameters from thermogravimetric analysis (TGA) curves, achieving high numerical fit qualities (exceeding 95%–99%) across polyolefins like polyethylene (PE), polypropylene (PP), and polystyrene (PS).
- Physics-Embedded Neural Architectures: The integration of chemical reaction neural networks (CRNNs) embeds physical principles—such as the Law of Mass Action and the Arrhenius law—directly into network weights, enabling the direct discovery of underlying reaction pathways and rate constants from raw experimental data without predefining complete reaction networks.
- Data-Driven Yield & Composition Prediction: Supervised machine learning models, particularly tree-based gradient boosting algorithms (such as XGBoost and CatBoost), reliably forecast three-phase product yields (gas, liquid oil, and solid char) as well as specific chemical fraction distributions (C1–C4 gases, gasoline-range C5–C12 hydrocarbons, and aromatics like BTEX).
- Model Interpretability and Feature Attribution: Post-hoc interpretability tools, specifically Shapley Additive Explanations (SHAP), open the "black box" of machine learning predictions. SHAP quantifies how feed characteristics (such as Si/Al ratios, specific surface areas, and elemental composition) and operating variables dynamically govern product selectivity, bridging data-driven models with fundamental reaction chemistry.
- Surrogate-Driven Reactor Optimization: Coupling CFD simulation data with AI surrogate models and multi-objective optimization algorithms (such as NSGA-II) enables efficient reactor design. Applying these AI-CFD frameworks to fluidized-bed reactors processing high-density polyethylene (HDPE) achieved an 8.10% increase in liquid oil yield alongside a 6.38% reduction in pressure drop.
- LLM-Driven Pyrolysis Agents: The authors propose a forward-looking conceptual architecture for Large Language Model (LLM)-driven pyrolysis agents. By combining natural language processing (NLP), automated literature extraction, knowledge graphs, and real-time interaction with laboratory robotics, LLM agents create an autonomous closed-loop paradigm for process design, kinetic retrieval, and real-time reactor control.
Despite these significant technical advancements, the review highlights critical gaps that must be addressed before wide industrial implementation. Current models suffer from small-sample data limitations, lack standardization across literature datasets, and frequently fail to account for real-world reactor dynamics such as catalyst deactivation, vapor residence time, dynamic coke formation, and heat-and-mass transfer limitations inherent to scale-up.
To address these hurdles, the research team emphasizes the urgent need to build standardized, high-quality open-access databases, integrate thermodynamic and physical constraints into hybrid machine learning models, and advance AI-CFD digital twin technologies for physical reactor optimization.
About the Research
Journal: ENGINEERING Energy
