Physics-Informed Data-Analysis Frameworks for Thermochemical Conversion Processes

Doctoral thesis, 2026

The escalating global production and consumption of plastics pose a significant environmental threat, demanding innovative waste management solutions. Thermochemical conversion via steam cracking offers a promising alternative to mechanical recycling, enabling the processing of highly heterogeneous plastic waste streams and recovery of carbon into monomeric species and valuable chemicals suitable for producing virgin-quality materials. The resulting product slate from steam cracking includes syngas, aliphatics, aromatics and soot, and is intrinsically linked to reactor conditions and feedstock’s chemical characteristics. From a data analysis perspective, this distribution holds valuable information of the process that can be leveraged for instrument validation and estimation of unmeasured species, and relevant process variables. However, the high-dimensional, partially observed, and structurally diverse nature of the data demands robust, physically consistent analytical frameworks.

This thesis contributes to the theoretical and practical understanding of thermochemical conversion processes from a data analysis perspective. It introduces a constraint-aware methodology for developing data-driven models that can serve as tools for process optimization and to gain insights into the relationship between feedstock characteristics and process outcomes. The central idea is to merge first-principles constraints with compact statistical representations to build mathematical frameworks known as Constraint Networks (CN), enabling high-dimensional experimental systems to be transformed into low dimensional, tractable, and physically consistent models.

The research encompasses the development and validation of two complementary data-driven models. The Parametric System Model (PSM) transforms product species distributions into low-dimensional, physics-informed representations using discrete probability functions. By embedding conservation laws and topological constraints into a CN framework, the PSM enables physically meaningful estimation of unmeasured species, assessment of data quality, and validation of experimental setup. The Carbon Bond Group (CBG) model reduces both feedstock and product spaces into chemically meaningful vectors based on bond group environments. This dimensionality reduction allows steam cracking to be represented as a column-stochastic transformation between feed and products, facilitating cross-feed comparisons and enabling structure-based predictions through machine learning implementations. The models’ development and validation were done using a pool of experimental data generated from steam cracking of various polymeric feedstocks under different operating conditions in a semi-industrial scale dual fluidized bed (DFB) reactor.

Overall, this work summarizes efforts to create generalizable data analysis frameworks aligned with physical principles for high-temperature thermochemical conversion systems. It contributes a scalable, interpretable, and constraint-centric modeling approach that supports the development of digital tools for process control and design, helping to pave the way for the future technology integration into circular economy strategies.