Machine Learning-Based Design of Ionic Liquids at the Atomic Scale for Highly Efficient CO2 Capture

Ionic liquids (ILs) are considered excellent substitutesfor aqueousalkanolamine solutions in CO2 capture systems. However,the smart design of ILs, facing the small sparse data set and complexionic structures, poses a huge challenge. To address this issue, anovel machine learning method based on a syntax-directed variationalautoencoder (SDVAE), deep factorization machine (DeepFM), and gradient-basedparticle swarm optimization (GBPSO) was proposed in this work. TheSDVAE converts the molecular structure and chemical space of the ILs,and then DeepFM predicts the solubility of each coordinate in thechemical space representing an IL. Finally, GBPSO identifies the coordinatesthat represent ILs with ideal properties. Our main optimization objectiveis a high solubility difference for CO2 between its absorptionand desorption conditions in commercial plant capture systems, whichrepresents the CO2 capture ability. The best IL generatedhas a predicted solubility difference that is 35.3% higher than thatof the best one in the data set. A synthetic novel IL [EMIM][TOS]from the generated results was experimentally evaluated; it has asufficiently high solubility difference to be a capture solvent withlow energy consumption. Our model has proved to be a high-efficiencymolecular design model that can be used for sparse small data sets. Ionic liquids are promising alternativesolvents of alkanolaminesolutions for CO2 capture. This work establishes a reliabledesign method at the atomic scale to improve CO2 captureability.