Carbon dioxide is the greenhouse gas that contributes the most to global warming overall. According to the Intergovernmental Panel on Climate Change, the average global temperature would rise by around 1.5 degrees Celsius by 2100 if nothing was done. Researchers and businesses attempting to tackle global warming have struggled to find efficient ways to absorb and store carbon dioxide; Amir Barati Farimani has been trying to change that.
According to Barati Farimani, an assistant mechanical engineering professor at Carnegie Mellon University, “machine-learning models bear the promise of discovering new chemical compounds or materials to combat global warming.” “Machine learning models can accomplish precise and effective virtual screening of CO2 storage candidates and may even generate preferable compounds that never existed before.”
Barati Farimani has made progress in the identification of ionic liquid molecules using machine learning. Ionic liquids (ILs) are classes of molten salt that, due to their excellent chemical stability and high CO2 solubility, remain in a liquid state at normal temperature, making them the perfect choice for CO2 storage. The characteristics of ILs are strongly influenced by the ion combination. Nevertheless, it is very difficult to exhaust the design space of ILs for effective CO2 storage through conventional studies due to the combinatorial possibilities of cations and anions.
In drug development, graph neural networks (GNNs), which interpret molecules as graphs and utilise a matrix to identify molecular bonds and related features, are frequently used in conjunction with machine learning to create what are known as molecular fingerprints. Barati Farimani has created GNNs and fingerprint-based ML models that can both forecast the CO2 absorption in ionic liquids for the first time.
“Our GNN method achieves superior accuracy in predicting the CO2 solubility in ion liquids,” claimed Barati Farimani. GNN directly learns the features from molecular graphs, in contrast to earlier ML techniques that depend on manually created features.
Just as crucial as identifying the chemical characteristics of molecules is understanding how machine-learning models make decisions. This explanation gives researchers more information about how, from a data-driven standpoint, the molecule’s structure impacts the properties of ionic liquids. For instance, Barati Farmimani’s team discovered that molecule fragments with a chemical rather than a physical interaction with CO2 are more significant. Furthermore, those with a weaker nitrogen-hydrogen bond may be better suited to formalising a stable chemical relationship with CO2.
