| Abstract Scope |
The need for novel and advanced materials including polymers has never been greater, whether used in biodegradable plastics or in medical implants. Copolymers offer a good option for many such applications. While tools such as SA_Score help with determining synthesizability of homopolymers, there haven’t been any computational tools to evaluate copolymer synthesizability. This research was conducted to fill this gap.
To evaluate copolymer synthesizability, a synthesizability model was created by fine-tuning a ChemBERTa transformer with copolymer training data including PSMILES of 2,150 known, synthesizable linear copolymers from sources such as PolyInfo and an additional 40,830 PSMILES of empirically synthesizable linear copolymers created using augmentation techniques. The fine-tuning of the synthesizability transformer was performed using masked language modeling with a custom loss function, achieving a synthesizability accuracy of 92%. The research confirmed synthesizability evaluation for linear copolymers using AI, which helps reduce time and costs of discovering new polymers. |