Current position: Home >> Scientific Research >> Research Field

Machine Learning for Quantum Chemistry

Machine learning (ML) is revolutionizing quantum chemistry by enabling faster, more accurate predictions and deeper understanding of molecules. Here, we explore key areas of development:

1. Molecular Representations:We design methods to represent molecules as vectors based on their 2D or 3D structure, including atomic charges (Z) and spatial coordinates (R). The representation plays an essential role in machine learning models.

2. Building Molecular Big Data:High-quality data is crucial for ML success. We focus on constructing large datasets of molecules with high-precision properties.

3. Exploring Chemical Space:The vast chemical space offers a seemingly infinite number of molecules. We develop techniques to efficiently navigate this space by reducing its dimensionality and identifying promising regions.

4. Machine Learning Models:We build multi-level models that leverage combinatorial mathematics to achieve high-precision predictions. These models can be trained with a combination of a small amount of high-precision data and a larger amount of lower-precision data.

Applications of Machine Learning Models:

a) High-fidelity Force Fields:We develop accurate machine learning force fields (models describing interatomic interactions) that enable simulations of complex systems, such as organic reactions in solution under realistic conditions.

b) Direct Property Prediction:We build models that directly predict quantum chemical properties based solely on a molecule's structure, bypassing complex quantum calculations.

c) Accelerating Ab Initio Methods:We aim to accelerate core steps in solving the Schrödinger equation (the governing equation of quantum mechanics) using machine learning, leading to faster ab initio calculations (calculations from first principles).

d) Inverse Molecular Design:We leverage deep learning models to design molecules with desired properties. By providing the target properties, the model can suggest optimal molecular structures.

Models and Methods in Theoretical Chemistry

We also focus on developing simpler models to understand fundamental chemical principles and predict properties:

i) Electronic Structure Calculations:We develop and implement theoretical methods to calculate the electronic structure of complex systems like solid surfaces. This helps establish structure-property relationships for such systems where traditional methods are limited.

ii) Conceptual Density Functional Theory:We employ concepts within the framework of density functional theory to understand the initial stages of chemical reactions. This approach is exemplified by the activity theory of metal surfaces.

iii) Nature of Chemical Bonds:We explore the nature of chemical bonds using simplified models like valence bond theory or semi-empirical molecular orbital theory to quantitatively describe bond strength.

Surface Catalysis and Electrocatalysis

We investigate the mechanisms of important reactions in surface catalysis, including thermal catalysis (gas-solid interface) and electrocatalysis (solid-liquid interface). This includes reactions with significant value like oxygen reduction and CO2 reduction, which play a role in clean energy technologies.

Profile

Professor Bing Huang (group website https://bhuang.whu.edu.cn) received his B.S. degree in June 2009 and Ph.D. degree in June 2015 from Wuhan University, Wuhan, China, where he studied and developed reactivity theories about solid surfaces under the supervision of Prof. Zhuang Lin. Subsequently, he carried out postdoctoral research at the Department of Chemistry, University of Basel, in collaboration with Prof. Anatole von Lilienfeld, while shifting his research interests towards the development of machine learning models and methods in quantum chemistry to explore compound space. Since 2021, he moved with Anatole to the Department of Physics at the University of Vienna, Austria, to continue his postdoctoral research, and in March 2024, he joined Wuhan University to establish an independent group (under the large team of the Laboratory of Hydrogen Electrochemistry, led by Prof. Lin Zhuang), with research interests in the areas of chemical space exploration and molecular inverse design, theoretical and computational chemistry methods combined with machine learning, electronic structure theory and calculations, conceptual density functional theory (e.g. reactivity theory), and theoretical studies of surface catalysis and electrocatalysis.

Currently, he has published more than 30 scientific papers, among which many papers have been published as first author and/or corresponding author in several high-impact scientific journals, including Science & Nat. Chem. His scientific papers have been cited more than 3100 times, and his personal H-factor is 22.

Selected publications:

(1) Bing Huang*, Guido Falk von Rudorff*, and O. Anatole von Lilienfeld*. “The central role of density functional theory in the AI age”, Science 381, no. 6654 (2023): 170-175. (see URL)

(2) Bing Huang, and O. Anatole Von Lilienfeld. “Ab Initio Machine Learning in Chemical Compound Space”, Chemical Reviews 121, no. 16 (2021): 10001-10036. (see URL)

(3) Bing Huang, and O. Anatole von Lilienfeld. “Quantum machine learning using atom-in-molecule-based fragments selected on the fly”, Nature Chemistry 12, no. 10 (2020): 945-951. (see URL) (highlighted by chemistryworld, see URL)

(4) B. Huang, L. Xiao, J. Lu, L. Zhuang, “Spatially Resolved Quantification of the Surface Reactivity of Solid Catalysts”, Angewandte Chemie International Edition, 2016, 55, 6239-6243. (see URL) (hot paper, see URL)

(5) B. Huang, L. Zhuang, L. Xiao, J. Lu, “Bond-Energy Decoupling: Principle and Application to Heterogeneous”, Chemical Science, 2013, 4, 606-611. (hot paper, see URL)

(6) Bing Huang, and O. Anatole von Lilienfeld. “Communication: Understanding molecular representations in machine learning: The role of uniqueness and target similarity”, Journal of Chemical Physics, 2016, 145, 161102 (see URL)