Ms. Elnaz Azmi successfully completed her dissertation on 8.2.2024. Since 2017, she has been working on the challenge of reducing the resource requirements of environmental simulations by using machine learning methods. One day later, Oskar Taubert was also able to celebrate the successful defense of his dissertation. Since 2018, Mr Taubert has been working on the machine learning of sparse data problems in biology and biologically inspired optimization algorithms adapted to high-performance computing (HPC).
Ms. Azmi has worked extensively with simulations in the environmental sciences, which are essential for the understanding of complex natural phenomena. Due to their high spatial and temporal resolution, these often impose high demands on available computing resources. In her research, Ms Azmi worked on increasing the efficiency of computationally intensive simulations by applying approximation and optimization approaches to examples from hydrology and climate science. The aim was to use machine learning to recognize similarities in time and space within the simulations and thus develop a method to reduce redundancies. By integrating an unsupervised learning module directly into the simulation code and replacing part of the simulation with a neural network, Ms. Azmi showed that by identifying model redundancies and reducing computational complexity, the efficiency of the simulations can be increased, thereby reducing resource requirements.
Mr. Taubert applied machine learning to the prediction of the structure of biomolecules. In order to make the best possible use of the limited training data available, he combined various methods: self-supervised neural networks, fine-tuning and gradient boosted decision trees. The parameterization of these complex model processes is made possible by algorithms adapted to the computing environment, which use the given computing resources efficiently to propose new models and train them. The contributions researched by Mr. Taubert on models generated with sparse data and on model architecture search should also be used in scientific machine learning in the future. Mr. Taubert's work was funded by the Helmholtz Analytics Framework (HAF) project, a "Google Faculty Research Award" received in the Multiscal Biomolecular Simulation research group and the Helmholtz Artificial Intelligence Cooperation Unit (Helmholtz AI) platform.
SCC congratulates Ms. Azmi and Mr. Taubert on successfully completing their doctorates and wishes them all the best for their future careers.
Dissertation of Elnaz Azmi: Approximation and Optimization of Compute-Intensive Environmental Simulations through Machine Learning Methods
Dissertation of Oskar Taubert: Machine Learning from Evolution