Artificial intelligence driven design of catalysts and materials for ring opening polymerization using a...

Area of Science:

• Polymer Science
• Materials Informatics
• Computational Chemistry

Background:

• Machine learning (ML) and automated experimentation can accelerate polymer science research.
• Current data models lack flexibility for diverse polymer experiment and data types, hindering ML integration.
• Leveraging historical data for ML development in polymer science is challenging due to data representation limitations.

Purpose of the Study:

• To introduce a flexible and extensible domain-specific language, Chemical Markdown Language (CMDL), for representing diverse polymer science data.
• To demonstrate how CMDL facilitates the use of historical experimental data for training ML models.
• To showcase the application of CMDL in generative molecular design and experimental validation.

Main Methods:

• Development of the Chemical Markdown Language (CMDL) for data representation.
• Fine-tuning regression transformer (RT) models using historical data encoded in CMDL.
• Experimental validation of ML-generated catalysts and polymers from ring-opening polymerization.

Main Results:

• CMDL provides a flexible, extensible, and consistent data representation for various experiment types and polymer structures.
• CMDL enables seamless integration of historical data to fine-tune ML models for generative tasks.
• The CMDL-tuned model successfully generated and experimentally validated catalysts and polymers, preserving key functional groups.

Conclusions:

• CMDL offers a versatile solution for representing complex polymer science data, overcoming limitations of rigid models.
• This approach accelerates the translation of historical data into predictive and generative ML models.
• CMDL facilitates the production of experimentally actionable outputs, advancing polymer research and development.