This title appears in the Scientific Report :
2023
SiSc Lab 2022, Project 6. A machine learning playground in quantum mechanical simulation.
SiSc Lab 2022, Project 6. A machine learning playground in quantum mechanical simulation.
Density functional theory (DFT) is one of the most widely used simulation techniques. About a third of world supercomputing time is spent each year on such calculations. DFT approximates the solution to the Schrödinger equation, to elucidate the electronic structure of materials and molecules. While...
Saved in:
Personal Name(s): | Wasmer, Johannes (Corresponding author) |
---|---|
Rüssmann, Philipp / Chen, Po-Yen (Contributor) / Burdulea, Ilinca (Contributor) / A, Lixia (Contributor) | |
Contributing Institute: |
Quanten-Theorie der Materialien; PGI-1 Quanten-Theorie der Materialien; IAS-1 |
Imprint: |
2022
|
Conference: | RWTH Aachen University, Aachen (Germany), 2022-11-01 - 2023-03-01 |
Document Type: |
Lecture |
Research Program: |
Helmholtz School for Data Science in Life, Earth and Energy (HDS LEE) Topological Matter |
Publikationsportal JuSER |
Density functional theory (DFT) is one of the most widely used simulation techniques. About a third of world supercomputing time is spent each year on such calculations. DFT approximates the solution to the Schrödinger equation, to elucidate the electronic structure of materials and molecules. While it makes quantum many-body problems in a lot of systems of interest tractable, it still is computationally demanding. It typically scales in O(N^3) with the number of electrons in the system, limiting its application to systems with a few thousand atoms at most. Over the last 15 years, the development of surrogate models based on machine learning (ML) has steadily gained momentum in the field of atomistic simulation. In ab initio molecular dynamics for instance, machine-learned interatomic potentials at a fraction of the cost and comparable accuracy of mechanistic methods have already become mainstream. Now these surrogate models also start to increasingly be developed to predict the underlying electronic structure properties of atomic systems directly.In this project, the students will be given the chance to play around with a wide array of state-of-the-art models in this field, from traditional kernel methods to deep graph convolution networks. They will be provided with a computational infrastructure and training datasets from DFT calculations. The challenges ladder they will climb has the rungs a) understanding the electronic structure data, b) understanding the reasoning behind the various surrogate ML modeling approaches, c) discovering common features of the atomic systems to come up with clever optimizations of the model architectures, and d) achieving reasonable prediction accuracy for the targeted electronic structure properties. The project goals will be adjusted, in a reasonable range given the limited time frame, according to the speed of progress and the particular interests of the students.Expected prerequisites: Applied Quantum Mechanics, basic Python skills. Desired, but optional: Physics track, some hands-on ML experience.Advisors : Johannes Wasmer, Philipp Rüßmann , Stefan Blügel |