First-Principles-Based Machine Learning Models for Phase Behavior and Transport Properties of CO2

Inthis work, we construct distinct first-principles-based machine-learningmodels of CO2, reproducing the potential energy surfaceof the PBE-D3, BLYP-D3, SCAN, and SCAN-rvv10 approximations of densityfunctional theory. We employ the Deep Potential methodology to developthe models and consequently achieve a significant computational efficiencyover ab initio molecular dynamics (AIMD) that allowsfor larger system sizes and time scales to be explored. Although ourmodels are trained only with liquid-phase configurations, they areable to simulate a stable interfacial system and predict vapor-liquidequilibrium properties, in good agreement with results from the literature.Because of the computational efficiency of the models, we are alsoable to obtain transport properties, such as viscosity and diffusioncoefficients. We find that the SCAN-based model presents a temperatureshift in the position of the critical point, while the SCAN-rvv10-basedmodel shows improvement but still exhibits a temperature shift thatremains approximately constant for all properties investigated inthis work. We find that the BLYP-D3-based model generally performsbetter for the liquid phase and vapor-liquid equilibrium properties,but the PBE-D3-based model is better suited for predicting transportproperties.

Automated summary

In this work, distinct first-principles-based machine-learning models of CO2 were constructed, allowing for stable interfacial system simulation, prediction of vapor-liquid equilibrium properties, and improved computational efficiency. The SCAN and SCAN-rvv10 models exhibit temperature shifts, while the BLYP-D3 model performs better for liquid phase and vapor-liquid equilibrium properties, and the PBE-D3 model is better suited for predicting transport properties.