Accelerating Screening of Phase Separation Agents for Carbon Dioxide Capture: A Machine-Learning Framework with Interpretable Quantum Chemical Insights

Against the backdrop of addressing climate change and
reducing
carbon emissions, carbon dioxide capture technology has gained increasing
significance as a key approach toward achieving global carbon neutrality.
However, in the development of phase change absorbents, current screening
methods for phase separation agents (PSAs) predominantly rely on static
empirical parameters, neglecting electronic effects and phase equilibrium
mechanisms, which somewhat restrict the improvement of screening efficiency
and accuracy. This study innovatively proposes a novel paradigm integrating
quantum chemistry and machine learning. A total of 434 experimental
systems with amine, PSA, and water at a 1:2:2 ratio were constructed,
from which 358 valid data sets were selected. A three-dimensional
quantum chemical descriptor system encompassing 48 dynamic electronic
parameters was established, accompanied by an innovative anion-centered
molecular characterization method for the reaction states. A two-stage
machine learning framework was employed: the random forest algorithm
evaluated three types of descriptors and identified 34 key features,
with particular emphasis on anion-related descriptors such as Orbital
Discrete Index and Electrostatic Potential. After the six machine-learning
models were compared, CatBoost was identified as the optimal classifier,
achieving a test accuracy of 89.7% and a phase change recall of 96.3%.
Interpretive analysis based on SHAP revealed that PSAs with lower
mean Orbital Discrete Index values facilitate phase separation by
weakening the hydrogen bond network, while the geometric feature (longest
interatomic distance) and electronic property (minimum electrostatic
potential) of organic amine anions synergistically drive the reconstruction
of the phase interface. Validation in a 3-(dimethylamino)propylamine
system demonstrated a prediction accuracy of 83.3%, significantly
reducing experimental costs. This study provides an accurate and interpretable
solution for the rational design of PSAs, strongly promoting the development
of high-efficiency carbon dioxide capture materials.