Neural Network Model Chemistries

Doctoral Dissertation


For the past few decades, theoretical simulation was proved to be an important tool for understanding the underlying mechanisms, predicting the molecular proper- ties/reactivities and designing new materials. Despite its huge success and development, atomistic force-field based methods and ab-initio methods remain to be the two workhorses in theoretical simulation. Each suers from its own limitation: the low accuracy of force-field based methods and the high cost of ab-initio methods.

With the ability to perform highly nonlinear transformation and model com- plicated system, neural network is becoming a promising tool to develop model chemistries for performing simulations. Neural network o↵ers a new way to cal- culate the molecular properties with near ab-initio accuracy at low cost.

Several dierent approaches of applying neural networks have been tried in this thesis, such as constructing neural network based kinetic energy functional, combining neural networks with the many-body expansion, defining bond energies and developing a neural network based molecular simulation package TensorMol. Each approach has addressed certain aspects of the problem of building neural network model chemistries. More investigations are needed to develop a widely accepted neural network model chemistry for general purpose simulations.


Attribute NameValues
Author Kun Yao
Contributor Gregory Hartland, Committee Member
Contributor J. Daniel Gezelter, Committee Member
Contributor John Parkhill, Research Director
Degree Level Doctoral Dissertation
Degree Discipline Chemistry and Biochemistry
Degree Name PhD
Defense Date
  • 2018-03-23

Submission Date 2018-03-29
  • Machine Learning

  • Neural Network

  • Computational Chemistry

  • Theoretical Chemistry

  • English

Record Visibility Public
Content License
  • All rights reserved

Departments and Units


Please Note: You may encounter a delay before a download begins. Large or infrequently accessed files can take several minutes to retrieve from our archival storage system.