Colloquium: Dr. Murat Keceli, Argonne National Laboratory

The dawn of the exascale era and the machine learning boom provide tremendous opportunities for computational scientists. The first exascale machine is expected in less than two years, but are we ready to utilize the unprecedented computational power with existing algorithms and codes? The driving force behind the success of machine learning is big data, but do we have accurate and diverse data sets to employ deep learning algorithms? 

In this talk, I will try to provide answers to these questions based on my research on atomic simulations and electronic structure calculations. I will describe the opportunities provided by the leadership computing facilities and challenges to be addressed as we approach the exascale era. In the first part, I will discuss how we can improve the computational performance of density functional theory based methods using locality of chemical interactions and sparsity in linear algebra algorithms. I will then describe a novel valence bond based method that can achieve very high strong scaling efficiency. In the second part, I will talk about developing efficient and resilient workflows for high throughput computing to generate data sets for machine learning and how we can reduce the training time on supercomputers for deep learning algorithms with large datasets. Finally, I will discuss the benefits of using modern software engineering practices (version control, auto-documentation, continuous integration) and productivity tools (Jupyter notebooks, integrated development environments, containers) in scientific software development.