Research

Research Overview

Our research bridges quantum chemistry (QC), machine learning (ML) and artificial intelligence (AI), and chemical understanding to solve the world’s most pressing problems: Energy supply and climate change. Our approaches build on the growing computing power to achieve Virtual High Throughput Screening (VHTS), which is complemented to the ample literature precedents to arrive at Property-constrained Deep Generative Models (pDGMs) for molecular design or is directly trained for Machine Learning (ML) Property Prediction. We further combine QC and ML to derive chemical understanding to foster comprehensions of chemical reactions and molecular design.

De Novo Molecular Design

Virtual High Throughput
Screening (VHTS)

The vast chemical space has hindered the exhaustive exploration of it for any particular property or reaction of interest. With the emergence of quantum chemistry (QC) and the growing computing power, thousands of QC simulations can be readily performed per day to test candidate molecules for the desired properties and reactions in silico. Virtual high throughput screening (VHTS) hence has become a routine to narrow down regions of the vast chemical space prior to synthetic explorations. Challenges our group aim to solve are pertained to global energy supplies with particular focuses on 1) selective oxidation of CH₄ to CH₃OH, 2) H₂O oxidation, and 3) oxidation of NH₃ to N₂.

Property Prediction
with Machine Learning

Machine learning (ML) models, such as kernel methods, Gaussian Process (GP), Support-vector machine (SVM), random forests (RF), and artificial neural network (ANN), have emerged as a new form of prediction of molecular properties compared to the traditional physics-based Ab Initio methods. The relative black-box nature of ML methods allows faster property predictions (around seconds to minutes vs. hours to days) of the properties of interest by providing a direct mapping between the encoding of molecules (e.g. molecular fingerprints) and the learnt properties (excitation energies, redox potentials, etc). Our group’s interest focuses on applying ML property prediction to find better organic light emitting diodes (OLED) for display applications and catalysts for selective oxidation of CH₄ to CH₃OH, H₂O to O₂, and NH₃ to N₂.

Generative AI models

The growing number of available experimental data and the advancement of computing power have enabled the application of deep learning (DL) to molecular discovery. Among DL methods, deep generative neural network (DGNN) has emerged as a leading technique to find promising candidate molecules. By studying the distribution of the input data (experimental, computational, or both), DGNN samples from the distribution generated by the NN model to create novel molecules. By applying property constraints during model training, we guide the DGNN model to generate molecules with the desired properties. Particularly, we apply semi-supervised DGNN to find better organic light emitting diodes (OLED) for display applications and catalysts for selective oxidation of CH₄ to CH₃OH, H₂O to O₂, and NH₃ to N₂.

Understanding Chemical Reactions

Quantum Chemical Modeling

Quantum chemical (QC) simulations can provide great insights on experimental observations to further our understanding on the driving force behind the observed reactivity, both electronically and sterically. Our group has the specialty in modeling reaction mechanisms of homogenous catalysis and systems with complex electronic structure using both Density Functional Theory (DFT) and high-level Ab Initio methods such as Complete Active Space SCF (CAS-SCF).

Machine Learning Facilitated
Chemical Insights

Machine learning (ML) methods realize direct mappings between features of molecules and properties of interest, thus allowing fast prediction of properties of new molecules. On the other hand, we can determine which of the features of the molecules the ML model was trained on are important for the prediction of the desired properties. This helps us decide, in combination with physics-based knowledge, which features, such as effective nuclear charge, spin state, and types of ligands, are to be optimized, yielding chemical insights of driving forces behind the properties of interest.