Multi-Score Reinforcement Learning for High-Tg Polyimide Design.

Area of Science:

• Materials Science
• Computational Chemistry
• Polymer Science

Background:

• Designing polymers with high glass transition temperatures (Tg) is crucial for advanced applications.
• Traditional methods often struggle to efficiently explore the vast chemical space for optimal polymer structures.

Purpose of the Study:

• To develop and evaluate a computational framework for generating polyimides with high Tg (>750 K) using reinforcement learning (RL).
• To investigate the impact of combining multiple scoring functions on the quality and diversity of generated polymers.

Main Methods:

• A systematic computational framework integrating multiple scoring functions within an RL approach for molecular design.
• Analysis of individual and combined scores, including predictive models, naı̈ve high-Tg scores, Tanimoto similarity, and various aggregation methods (arithmetic, geometric, harmonic means, and ExpAgg).

Main Results:

• Score combination significantly influences the quality and diversity of generated polymers.
• The novel ExpAgg aggregation function demonstrated superior performance across multiple RL configurations.
• The multiscore framework successfully generated chemically reasonable high-Tg polyimide candidates, even when the predictive model underestimated out-of-distribution values.

Conclusions:

• Score aggregation strategies are critical for successful molecular reinforcement learning outcomes in targeted polymer design.
• The ExpAgg function offers a promising approach for fusing complementary scores.
• Guidelines for selecting aggregation functions are provided, emphasizing the need to move beyond simple averages for enhanced molecular design.