Advancing Bayesian Optimization: Efficient Strategies for Accelerated Materials Discovery.

Supervisors: George Giannakopoulos, Panagiotis Krokidas, Christoforos Rekatsinas

Description:

Bayesian Optimization (BO) is an AI-driven technique designed to efficiently explore and sample large design spaces, particularly in domains where data are sparse, experiments are costly, and simulations are computationally intensive. This approach is especially valuable in the development of novel materials, enabling the discovery of optimized materials with enhanced performance while significantly reducing the number of physical or computational experiments required. However, traditional BO methodologies are constrained by high memory demands, largely due to the reliance on Gaussian Processes (GPs) as surrogate models. This limitation leads to a computational complexity that scales as O(n³), making it impractical for large-scale or high-dimensional problems. To address these challenges, next-generation BO algorithms must incorporate more efficient strategies. This thesis will focus on investigating advanced approaches to improve the efficiency of BO, such as batch sampling techniques, dynamically optimized batch sizes, adaptive exploration-exploitation trade-offs, efficient surrogate models, and multi-fidelity methods inspired by state-of-the-art advancements. The goal is to develop and evaluate innovative methods that maintain the predictive power of BO while significantly reducing memory and computational overhead, ultimately enabling its broader applicability in materials science and other high-impact fields.

Qualifications required: Python programming, Machine Learning.
Qualifications desired: Machine learning and Deep learning tool-kits (e.g. PyTorch); genetic algorithms, Bayesian optimization, active learning.