Research
Our primary research objectives involve the development of theoretical methods to address existing limitations in quantum chemistry simulations and to offer theoretical insights into macroscopic systems through the application of machine learning theories at ab initio level.
Universal ab initio machine learning (ML) potential
Adv. Energy Mater. 12, 2202279 (2022), Adv. Energy Mater. 12, 2201497 (2022)., Phys. Rev. Lett. 128, 045301 (2022), Phys. Rev. B 103, 214102 (2021).
We develop sparse Gaussian process ML potential and Bayesian sampling models. Based on the ML potential, we aim to overcome current quantum computational limitations and understand the quantum properties of open macroscopic material systems and provide scientific insights. By leveraging machine learning potential, we strive to perform quantum simulations at nano/micrometer scales, enabling simulations by several orders of magnitude faster and larger than before. Our quantum mechanical machine learning models are highly universal and can be applied to various materials, opening up research opportunities in physics, chemistry, materials, energy, and biology.
Diagrammatic Approach for Quantum Chemistry
Addressing the exact solution of the Schrödinger equation has been a pivotal challenge in quantum chemistry for nearly a century. Nevertheless, this challenge persists, primarily due to the curse of dimensionality in the Hilbert space, which hinders direct acquisition of an exact wavefunction solution. To circumvent this problem, one approach is to investigate the electron density solution of the Schrödinger equation based on the Hohenberg-Kohn density functional , referred to as density functional theory (DFT). Owing to its O(N3) scalability, DFT has established itself as the go-to method for simulating large-scale, realistic systems across a diverse array of materials. However, it remains plagued by functional dependency and frequently produces substantial errors due to the lack of electron correlations. Dynamical Mean Field Theory (DMFT) offers an alternative insight that circumvents these problems and accommodates both metallic and insulating systems (unlike post Hartree-Fock methods that can only treat insulating system) by relying on the Luttinger-Ward functional of electron Green’s function and calculate electron correlations with all possible Feynman diagrams. By directly calculating exchange and correlations using this diagrammatic approach, I am implementing a real-space cluster-DMFT framework for chemical systems. Furthermore, I have been developing DMFT solvers, such as non-equilibrium non-crossing and one-crossing approximations, to address non-equilibrium transport problems on a mesoscopic scale
AI-driven materials design & robotics synthesis
Energy Environ. Sci. 14, 3455 (2021); J. Phys. Chem. C 124, 8905-8918 (2020).
We are developing AI models to integrate with experimental synthesis procedures, using Bayesian AI models to predict synthesizability and stability in both theory and experiment. These AI models are built on high-throughput computational and experimental data to predict and synthesize novel energy materials such as catalysts, solar cells, and metal-organic frameworks. Our approach aims to revolutionize the way we discover and create innovative energy solutions.
Energy Materials Theory
Nature 609, 942–947 (2022), Nat. Energy 3, 773–782 (2018), Nat. Commun. 10, 5195 (2019), Adv. Mater. 33, 2005400 (2021), ACS Energy Lett. 3, 1294-1300 (2018), Adv. Energy Mater. 8, 1702898 (2018).
Our research focuses on materials theory for fundamental insights and deep understanding, including first-principles studies on solar cells, light-emitting diodes, catalysts, and batteries. We investigate crystallization, phase diagrams, and kinetics of materials to gain a comprehensive understanding of their properties. Additionally, we explore structure and reaction pathway searches for complex interfaces, enabling more efficient and effective material development.
