Professor Jason Goodpaster of University of Minnesota is supported by an award from the Chemical Theory, Models and Computational Methods program in the Division of Chemistry to develop computational methods which enable the accurate calculation of energies of molecules and solids utilizing machine learning. Highly accurate calculations of large molecules use a lot of expensive computer time and thus, our ability to understand the properties of these compounds is limited. Machine learning offers the ability to leverage many small calculations and combine them to accurately calculate the properties of large molecules, such as proteins or industrial catalysts. Professor Goodpaster and his research group are developing machine-learning methodologies that utilize regression and neural network models to accurately predict the properties of larger chemical systems. This methodology may open the door to improved understanding of these complex molecules. Professor Goodpaster is applying his computational techniques to the study of molecular magnets and metal oxides; however, the machine learning methodology is generalizable and can be made be applicable to a wide variety of molecular and solid materials. This project bolsters industries of the future as it develops new ways of designing chemicals and materials with specific properties. The Goodpaster group is also developing an outreach program targeting K-12 students from historically underrepresented and underprivileged groups. His educational program seeks to encourage these students to pursue higher education, specifically in STEM fields. Part of this program involves training high school students to perform chemistry research.
In this project, Professor Goodpaster and his research team will produce new state-of-the-art electronic structure methods with improved accuracy and computational efficiency. The project will develop and assess the accuracy of different machine learning (ML) methods, specifically different machine-learning features and regression algorithms, in predicting the Full Configuration Interaction energy. The researchers will also develop deep learning networks to predict the Full Configuration Interaction energy, and produce computationally efficient algorithms and code for these approaches. With the ML methodologies developed, they will be directly applied and combined with quantum embedding theory to study a variety of complex electronic structure problems, such as molecular magnets and metal oxides to benchmark the accuracy of these new methods and gain fundamental insights in these complicated systems.
This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.
