Thermodynamics Guided Machine Learning Models for CO₂ desublimation temperature prediction

Abstract

Accurate prediction of the carbon dioxide desublimation temperature (CDDT)—the threshold below which CO₂ transitions directly from vapor to solid phase—is critical for cryogenic carbon-capture systems, natural-gas processing, and pipeline safety. This paper presents a unified, three-tier study covering:  a classical thermodynamic model based on the Peng–Robinson equation of state (PR EoS) solved via Lagrange’s analytically stable cubic solver;  machine learning (ML) approaches including Decision Tree (DT), Gaussian Process Method (GPM), Adaptive Neuro-Fuzzy Inference System (ANFIS), and Genetic Programming (GP) applied to a consolidated dataset of 430 experimental measurements; and a physics-informed hybrid machine learning (PI-HML) framework that embeds fugacity equilibrium, Clausius–Clapeyron consistency, and monotonicity constraints directly into the learning objective. Together, these frameworks provide a complete, validated toolkit for managing CO₂ solidification in industrial cryogenic operations.