Chemistry Seminar: Fang Liu

Event interval: Single day event
Campus location: Chemistry Building (CHB)
Campus room: CHB 102
Accessibility Contact: chem59x@uw.edu
Event Types: Academics,Lectures/Seminars
Link: https://flgroup.emorychem.science/

"Synergizing GPU-Accelerated Quantum Chemistry and Machine Learning for Molecular Discoveries in the Condensed Phase"
Assistant Professor Fang Liu – Department of Chemistry, Emory University
Host: Xiaosong Li

Machine learning (ML) and big data play increasingly critical roles in chemical discovery. However, datasets (both computational and experimental) and ML models for condensed-phase molecular systems, such as solvated molecules and molecule assemblies, remain scarce. My research group leverages GPU-accelerated quantum chemistry and machine learning to address these gaps. 

Many crucial solvent-solute interactions, like hydrogen bonds, cannot be captured by the implicit solvent models routinely used in quantum chemistry calculation, and require explicit solvent treatment. To streamline the simulation workflow for arbitrary organic and organometallic solute molecules in explicit solvent molecules, we developed AutoSolvate, an open-source toolkit. To further enhance accessibility, we launched AutoSolvateWeb, a chatbot-assisted, cloud-based platform that automates simulation setup and execution using cloud resources. These tools have enabled the efficient generation of diverse computational datasets for solvated molecules. Leveraging these datasets, we trained Δ-ML models to enhance the accuracy of low-cost computational methods against experimental measurements.  

For molecular assemblies, we addressed computational challenges in predicting excited-state properties. We developed a size-transferable machine-learned exciton model that significantly reduces computational costs by tens of thousands of folds without sacrificing accuracy. Additionally, we aim to bridge the gap between simulated and experimental datasets by leveraging large volumes of computational data to train ML models for real-time analysis in autonomous experiments. As a proof of concept, we successfully trained an ML model to detect material phase transitions in situ using angle-resolved photoemission spectroscopy (ARPES).