Physics-informed framework for predictive modeling of product distributions in steam cracking of heterogeneous polymeric waste

Artikel i vetenskaplig tidskrift, 2027

Thermochemical recycling of heterogeneous plastic waste via steam cracking offers a scalable pathway for carbon circularity, yet predictive control of product distributions remains limited by feedstock variability and complex radical chemistry. This work introduces a physics-informed data-driven model, referred to as the Carbon Bond Group (CBG) model, that describes feedstock–product relationships at a reduced but chemically meaningful level of abstraction. The approach reformulates thermochemical conversion as a constrained transformation between compact structure–based carbon spaces, where feedstocks and products are represented by key carbon bonding environments and linked through a column–stochastic operator matrix enforcing conservation laws. Operating conditions are incorporated through a temperature–dependent parameterization, enabling the mapping to evolve systematically while preserving physical admissibility. The model is trained on a multi–year experimental dataset from a semi–industrial Dual Fluidized Bed steam cracker spanning diverse waste streams and temperatures (720–835 °C). It accurately reproduces product group distributions (R² up to 0.98 for COx and aliphatics, and 0.90 for aromatics) and demonstrates robust performance under cross–validation. Analysis of the temperature-dependent conversion matrix reveals chemically consistent carbon redistribution trends, such as increasing routing of oxygen-bound carbon toward COx and enhanced gasification of aromatic carbon at high temperatures, demonstrating the model’s ability to provide insights into the system’s carbon conversion behavior. By embedding structural descriptors, physical constraints, and operating variability, the CBG model provides an interpretable, scalable surrogate for thermochemical conversion behavior of complex feedstocks. The approach supports predictive feedstock assessment, scenario exploration, and development of reduced–order digital twins for heterogeneous waste steam-cracking systems.