Designing Chemical Processes with Multicomponent Solvents through Self-Evolving Solubility Databases and Neural Networks

点击翻译按钮获取中文摘要

With support from the Chemical Theory, Models and Computational Methods (CTMC) program in the Division of Chemistry, Seonah Kim of the Department of Chemistry at Colorado State University is developing novel machine-learning (ML) predictive models for solubility in complex multicomponent systems. Solubility is essential to designing various chemicals such as pharmaceutical drugs, renewable fuels, electrolytes, and polymers. Further, solubility also plays a key role in the synthesis and separation processes employed to obtain these chemicals. In this project, the Kim group will build an extensive self-evolving solubility database from computational and experimental data and develop an ML solubility prediction model using graph neural networks (GNNs) for more rapid and accurate solubility predictions. A novel innovation in this approach will be expansion to multicomponent solvent systems and polymer solubility. The project will offer graduate and undergraduate students an opportunity to acquire a diverse range of relevant skills for the modern scientific enterprise. Additionally, the Kim research team will actively participate in outreach events, connecting with local schools and reaching a nationwide audience through in person and virtual activities. The proposed research will first develop innovative self-evolving solubility databases and GNNs. The approaches will adopt semi-supervised self-training, which combines experimental and computational databases to rectify the discrepancies between these two data sources. A unique aspect of this approach will be the correction of solubilities calculated from quantum chemistry (QC) methods prior to integration into the database. This data augmentation process will be performed for a broad scope of solutes and solvents. Multiple QC methods will be employed, such as the Conductor-like Screening Model (COSMO) and density functional theory (DFT) with an implicit Solvation Model based on Density (SMD), for a reliable augmentation. The self-training scheme will then be expanded to multicomponent systems. Various GNN variants will be examined by incorporating the descriptors for complex interactions among multiple components and the model will be leveraged to generate new solvent parameters for SMD. Prediction of polymer solubility will be followed by experimental validation as another practical application. Dr. Seonah Kim’s group will develop extensive, self-evolving, and reliable solubility databases and accurate solubility prediction GNNs for a wide range of solutes, including polymers dissolved in single and multicomponent solvents.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.