Physics-informed machine learning prediction of char mass evolution in the catalytic pyrolysis of polyetherimide/graphite nanocomposites

Abstract

Carbonaceous structures can be produced via pyrolysis of polymeric precursors for applications in gas separation membranes, energy storage, flexible electronics, electromagnetic interference shielding foams, etc. Maximizing char yield is a primary objective, determined by precursor chemistry, composition, pyrolysis conditions, and kinetics. The complex, non-linear relationships among these factors favor machine learning (ML) for process design and optimization. A physics-informed, transformer-based ML model was developed to predict char mass evolution (thermal decomposition) of transition-metal-catalyzed polyetherimide (PEI)/graphite (Gr) nanocomposites from thermogravimetric analysis (TGA) data. The dataset included 38 formulations with varying Gr and catalyst (Fe, Ni, Co) contents, heating rates, and pyrolysis temperatures. Additional features captured Gr and catalyst structural and electronic properties (crystal system, d-orbital free electrons, lattice parameters, cohesive energy, carbide formation energy, electrical conductivity at 20 °C) and kinetic parameters from 2D/3D Avrami–Erofeev models (pre-exponential factor, activation energy). Data were split into “seen” catalysts (Fe, Ni) for training/validation and an “unseen” catalyst (Co) for testing. Hyperparameters and feature selection were optimized via the random forest method. The model achieved $R²$ > 0.98 on unseen data, accurately predicting TGA curves and kinetic trends. Experimental and ML-predicted curves showed close agreement, with successful extrapolation to Co-containing nanocomposites. This study integrates kinetics modeling with advanced ML to enhance prediction of pyrolysis behavior in polymer nanocomposites, providing a practical framework for developing carbonaceous materials with tailored properties.