Viscosity, as a measure of internal resistance to flow, is the most fundamental characteristic property of all liquids. Viscosity plays a critical role in the transport processes in a wide range of engineering applications, including household products, pharmaceutical, and chemical industry. Despite this broad and profound relevance, the viscosity values of pure ionic liquids (pILs) are currently largely determined by experimental measurements that limit the rate of data acquisition. The total number of 3,066 pILs with available viscosity measurement is less than 1% of 452,751 pILs (combinatorially enumerated 1344 cations and 357 anions reported in literatures). Researchers have long sought to build an extensive library of pILs viscosity data to meet rapidly increasing demands from researchers in both academics and industries.
Fig. Insights into the Properties and Potential Applications of Renewable Carbohydrate-Based Ionic Liquids: A Review
In recent years, the extensive data-driven machine learning efforts are being made in the field of molecular and materials science. In contrast to experimentation or computations/simulations based on fundamental equations at different time and length scales, a large set of machine learning models have been developed to enable rapid predictions based purely on available data. The versatile machine learning methodology provides a desirable platform to speed up the evaluation and reveal the correlation. Leveraging recent research progress on viscosity, we will develop a generic graph convolution neural network (G2-CNN), and new descriptors based on correlation matrix of intermolecular properties, such as electronegativity, polarizability, hydrophobicity, hydrogen bond, and van der Waals volume of groups in each ion, for better description of many body effects in pILs to model the viscosity. An extensive library containing temperature-dependent viscosity data of 259,350 pILs (combinatorially enumerated 6-cation and 10-anion families with 910 cations and 285 anions) will be built based on the G2-CNN model. A reinforcement learning enhanced variational autoencoder will be integrated to the G2-CNN responding to the quest for designing more diverse morpholinium-based pILs. Validation will be conducted using coupled molecular dynamics simulations with highly efficient dressed diffusion methodology.
The proposed deep learning models will generate more interests of the chemistry community to communicate with experts of artificial intelligence to stimulate more exciting research and accelerate many traditionally cost-intensive chemistry research and development work. The scientific merit of this proposal not only comes from the large amount of pILs viscosity data, but also in the artificial intelligence powered platforms developed in this project, which can be further developed towards inverse design of molecules and materials in a wide range of technologies.
Komentáře