Autonomous multi-objective design optimization using Bayesian approaches and active learning
View/
Keywords
Bayesian optimization ; Active learning ; Multi-objective optimization ; Design space reduction ; Nanoporous materials ; Nanoporous materials Pareto optimization materials discovery ; Pareto optimization materials discovery ; Artificial IntelligenceAbstract
The discovery and design of advanced functional materials often involve extremely large
combinatorial design spaces, making exhaustive exploration through traditional
computational or experimental approaches impractical. In recent years, machine learning
and data-driven optimization techniques have emerged as promising tools for accelerating
materials discovery by enabling efficient exploration of complex design spaces. Among
these approaches, Bayesian Optimization (BO) has gained significant attention due to its
ability to optimize expensive black-box functions using a limited number of evaluations.
However, the performance of BO can deteriorate when applied to very large search spaces,
as the algorithm must allocate evaluations across many potentially unpromising regions.
To address this challenge, this thesis investigates the integration of Active Learning–based
space reduction with Bayesian optimization for efficient materials design.
Specifically, the proposed framework first applies Bayesian Optimization to explore the
initial design space and gather informative samples. Subsequently, a classification-based
Active Learning method (DAGS) is used to identify and filter low-performing regions of
the design space, effectively reducing the search domain. Bayesian Optimization is then
re-applied within the reduced design space in order to improve optimization efficiency.
The proposed pipeline is evaluated in both single-objective and multi-objective
optimization settings.
Experimental results demonstrate that the proposed space-reduction strategy can
significantly improve optimization efficiency by focusing the search on high-potential
regions of the design space while maintaining competitive solution quality. The findings
highlight the potential of combining Active Learning and Bayesian Optimization as a
unified framework for data-efficient exploration of large materials design spaces,
particularly in the context of nanoporous materials such as Covalent Organic Frameworks
(COFs).