Generative Multiobjective Bayesian Optimization with Scalable Batch Evaluations for Sample-Efficient De Novo...

Area of Science:

• Computational chemistry and materials science.
• Machine learning applications in molecular discovery.

Background:

• Designing molecules with multiple objectives is challenging due to vast chemical space and simulation costs.
• Existing machine learning methods often face scalability issues with continuous latent spaces.

Purpose of the Study:

• To develop a modular "generate-then-optimize" framework for de novo multiobjective molecular design.
• To improve the efficiency and scalability of molecular discovery using limited data.

Main Methods:

• Utilizes a generative model to create diverse molecular candidates.
• Introduces a novel acquisition function, qPMHI (multipoint Probability of Maximum Hypervolume Improvement), for optimal batch selection.
• Employs Monte Carlo sampling for scalable batch selection based on probability ranking.

Main Results:

• Demonstrates significant improvements over state-of-the-art methods on synthetic benchmarks and application-driven tasks.
• Successfully identifies novel, diverse, and high-performing organic cathode materials for aqueous redox flow batteries.
• The qPMHI acquisition function enables exact and scalable batch selection.

Conclusions:

• The proposed framework offers a powerful and efficient approach to multiobjective molecular discovery.
• This method accelerates the identification of promising molecular candidates for specific applications like sustainable energy storage.
• The modular design and scalable selection process address key limitations of previous methods.