VSR Seminar with two talks

1st talk: Understanding emergent phenomena using massively parallel computers

Speaker: Prof. Dr. Eva Pavarini, PGI-2/IAS-3, Forschungszentrum Jülich, Germany

Strongly-correlated materials are characterized by emergent phenomena, the origin of which typically remains debated several decades after the experimental discovery. Classical examples are the Mott metal-insulator transition, orbital physics and super-exchange effects as well as entangled states of matter. Understanding these phenomena is a grand challenge in condensed-matter physics. In this talk I will illustrate how in the last decade considerable progress could be achieved thanks to the synergy between modern algorithms and massively parallel supercomputers. As representative application I will discuss the case of orbital super-exchange effects in transition-metal oxides.

2nd talk: Leveraging HPC and Quantum Systems for Efficient Hyperparameter Optimization

Speaker: Marcel Aach, Dr. Andreas Lintermann, JSC, Forschungszentrum Jülich, Germany

For all kinds of machine learning methods, the model performance is highly sensitive to the choice of hyperparameters that have to be set manually before starting the training process. These decisions involve not only the general architecture of the network and the parameters of the optimizer but also the selection of parameters for data pre-processing and regularization methods. The goal of the raise-ctp2 project is to identify and improve the most efficient hyperparameter optimization methods and apply them across different models and datasets on the JURECA-DC-GPU partition. First results show variations of random search with early stopping and Bayesian optimization techniques to be the most promising. The models to be optimized include Convolutional Neural Networks, Graph Neural Networks, and Transformers, applied to particle datasets from High-Energy Physics in cooperation with CERN and remote sensing datasets. Additionally, the novel Quantum Annealer JUPSI is used in a hybrid quantum-classical scheme to perform performance prediction tasks and save resources on JURECA-DC-GPU.