Categories
Research

Machine Learning

High-Throughput Simulation and Machine Learning Aided Molecular Discovery

To enable computer-aided chemical discovery, one needs to explore the vast chemical space efficiently. Data-driven models, including the emerging machine learning techniques, provide faster-than-fast tools for traversing the chemical space. However, the accuracy of these models relies on the quality of first-principle calculation datasets used for model training. Therefore, we need automated workflows to enable high-throughput first-principle simulation of numerous systems at an appropriate level of theories. These workflows will be applied to the discovery of molecules with catalytic or photophysical functionalities.

Automated workflow for reaction kinetics study in reaction networks

Transition state (TS) search is crucial for understanding reaction kinetics, but computationally challenging because the result highly depends on the conformers of the reactants and products. I have developed an automated toolkit, AutoNEB, for studies of reaction pathways in complicated reaction networks involving multiple conformers. This workflow avoids the typical biases in TS search, and automatically post-processes simulation results to discover the most favorable path. This toolkit has been applied to the reaction network study of gas-phase Pomeranz-Fritsch synthesis of isoquinoline.

Related Publications

S. Banerjee,§ F. Liu,§ D.M. Sanchez, T. J. Martínez, and R. N. Zare, Pomeranz-Fritsch Synthesis of Isoquinoline: Gas-Phase Collisional Activation Opens Additional Reaction Pathways, J. Am. Chem. Soc. 139, 14352 (2017) [ §These two authors contribute equally]

L.-P. Wang, A. Titov, R. McGibbon, F. Liu, V. S. Pande, and T. J. Martínez. “Discovering chemistry with an ab initio nanoreactor.” Nature chemistry 6 (2014)

Machine learning aided workflow for method selection of transition metal complexes

A crucial step to accurately model transition metal chemistry is to choose between single- and multi-reference based methods. As a MolSSI software fellow, I developed a Python API to calculate widely used multi-reference (MR) character diagnostics for any given molecule. With this tool, I curated a dataset for the MR diagnostics of thousands of octahedral complexes. I have built machine-learning models based on this dataset to predict MR character without first-principle calculation. This will enable automated method selection for modeling transition-metal complexes.

Related Software

https://github.com/hjkgrp/MultirefPredict (Website)

https://multirefpredict.readthedocs.io/en/latest/ (Documentation)

Categories
Research

Electronic Structure

Fast and Accurate Electronic Structure Methods for Correlated Electronic Systems

Open shell transition metal complexes form a gigantic pool for the discovery of functional materials or catalysts. Excited electronic states are essential for photochemistry processes related to new energy and human health, including photosynthesis, biosensors, and human vision. Unfortunately, first-principle simulations for these systems are notoriously difficult because of the challenges in describing the correlated electrons both accurately and efficiently. I aim to build fast and accurate electronic structure methods for correlated electronic systems.

In the excited state, TDDFT is incapable of correctly describing the topography of potential energy surfaces at conical intersections, an essential component of studying non-adiabatic dynamics in photochemistry. The state-interaction state-averaged restricted ensemble-referenced Kohn-Sham (SI-SA-REKS) method is a computationally efficient alternative. It retains the attractive scaling features of its KS-DFT sibling, while explicitly accounting for non-dynamic correlation through the ensemble DFT formalism. I derived and implemented the analytical energy derivatives for SI-SA-REKS on GPUs to enable efficient non-adiabatic simulation of photochemistry. My implementation affords mean-field computational cost while allowing an accurate description of conical intersections. Collaborating with my colleagues, we have successfully applied this method to the non-adiabatic QM/MM dynamics study of the ultrafast photoisomerization of Channelrhodopsinand Bacteriorhodopsin. To tackle more strongly correlated electrons manifesting as a Nagaoka transition, I have recently made a vital contribution to the ab initio modeling of the quantum analog simulator as a precursor for quantum computing.

Related Publications

F. Liu, M. Filatov, and T. J. Martínez, Analytical Derivatives of the Individual State Energies in Ensemble Density Functional Theory Method II: Implementation on Graphical Processing Units (GPUs), Preprint chemrxiv.79856 (2019)

M. Filatov, F. Liu, K. S. Kim, and T. J. Martínez, Self-Consistent Implementation of Ensemble Density Functional Theory Method for Multiple Strongly Correlated Electron PairsJ. Chem. Phys. 145, 244104 (2016)

M. Filatov, F. Liu, K. S. Kim, T. J. Martínez, Analytical Derivatives of the Individual State Energies in Ensemble Density Functional Theory Method. I. General formalism, J. Chem. Phys. 147, 034113 (2017)

Yu, Jimmy, Ruibinng Liang, Fang Liu, and Todd J. Martinez. “Characterization of the Elusive I Fluorescent State and the Ultrafast Photoisomerization of Retinal Protonated Schiff Base in Bacteriorhodopsin by Nonadiabatic Dynamics Simulation.” J. Am. Chem. Soc. 141 (2019): 18193.

Liang, Ruibin, Fang Liu, and Todd J. Martínez. “Nonadiabatic Photodynamics of Retinal Protonated Schiff Base in Channelrhodopsin 2.” J. Phys. Chem. Lett. 10 (2019): 2862

Electronic structure of transition metal containing molecules

In the ground state, transition metal complexes are notoriously difficult to study accurately with approximate DFT. Few systematic studies have been done for the quantification and elimination of density delocalization error (DDE), although it determines many molecular properties. I focused on understanding and controlling DDE in DFT simulations of transition metal chemistryTo quantify the DDE and its impact on chemical properties, I curate accurate reference densities with correlated wave function theory (WFT) for a broad range of transition metal compounds. With the references, I investigated the DDE reducing effects of various techniques, including DFT+U and global/range-separated hybrid tuning. Based on these method advances, I further investigated the spin-state energetics of transition metal single-atom catalysts (SACs). To determine the uncertainty of first-principle calculations in predicting the energetics of SACs, I developed accurate benchmarks from correlated WFT. These benchmarks enable the parameter tuning of DFT functional for predicting spin ordering in larger SACs. This work provides broad recommendations for modeling of open-shell transition metal SACs.

Related Publications

F. Liu, and H. J. Kulik “Impact of Approximate DFT Density Delocalization Error on Potential Energy Surfaces in Transition Metal Chemistry.” J. Chem. Theory Comput. (2019).

F. Liu, T. Yang, J. Yang, E. Xu, A. Bajaj, and H. J. Kulik. “Bridging the homogeneous-heterogeneous divide: modeling spin and reactivity in single atom catalysis.” Frontiers in Chemistry 7 (2019): 219.

A. Bajaj,  F. Liu, and H. J. Kulik. “Non-empirical, low-cost recovery of exact conditions with model-Hamiltonian inspired expressions in jmDFT.” The Journal of chemical physics 150 (2019): 154115.

Categories
Research

Solvent Models

Deciphering the Role of Solvent in Molecular Properties and Reactions

The solvent environment plays an essential role in chemical processes, including catalysis, photochemical reactions, and the stability of materials. Recent years, new experimental techniques have been introduced to characterize liquid-phase reactions. This places an urgent need for efficient computational models to further predict the effect of solvents on the rate and product distribution. I aim at developing methods and protocols for efficient modeling of reactions in realistic experimental conditions.

I developed the graphical processing units (GPUs) algorithms to accelerate the widely used implicit solvent model, the conductor-like polarizable continuum model (C-PCM) With 10X-100X speed-ups for ground-state Hartree Fock (HF) or density functional theory (DFT) calculation, my algorithm keeps the world’s fastest record. More importantly, these award-winning method developments enable more realistic simulation of large biomolecular systems in solution phase: the bulk solvent effects can now be simulated with little computational overhead.

Related Publications

F. Liu, N. Luehr, H. J. Kulik, and T. J. Martínez. “Quantum chemistry for solvated molecules on graphical processing units using polarizable continuum models.” Journal of chemical theory and computation 11 (2015): 3131-3144.

B. D. Mar., H. W. Qi, F. Liu, and H. J. Kulik. “Ab Initio screening approach for the discovery of lignin polymer breaking pathways.” The Journal of Physical Chemistry A 119 (2015): 6551-6562.

Solvent effects for excited state properteis and photophysical processes

The significances of solvent effects also lie in the simulation of spectroscopy and photochemistry, which involves evaluation of excited-state properties. To this end, time-dependent density functional theory (TDDFT) has been increasingly employed in recent years due to its superior computational scalability. To study the excited-state properties of solvated molecules, I implemented the TDDFT algorithm in the C-PCM solvent model with GPU acceleration, with 40-150X speed-ups. The implementation includes both the equilibrium and non-equilibrium solvation schemes that can describe fluorescence and vertical excitations, respectively. This method enables PCM-TDDFT for systems with ca. 1000 atoms, making the excited-state simulation in combined implicit/explicit solvent model practical.

Related Publications

F Liu, D. M. Sanchez, H J. Kulik, and T. J. Martínez. “Exploiting graphical processing units to enable quantum chemistry calculation of large solvated molecules with conductor-like polarizable continuum models.” International Journal of Quantum Chemistry 119 (2019): e25760