An extensible platform for enabling artificial intelligence guided design of catalysts and materials

Advances in artificial intelligence (AI), machine learning (ML), and automated experimentation are poised to vastly accelerate and reshape polymer science research. The difficulties of experimental data and polymer structure representation form a significant barrier in developing foundational AI models in polymer chemistry. These issues are further compounded by a lack of accessible and extensible software tools for experimental researchers to effectively leverage their historical data for AI development. To address these difficulties, we developed an extensible domain specific language, termed Chemical Markdown Language (CMDL), to enable flexible and consistent representation of disparate experiment types. CMDL can be extended to cover new data types and provides built-in support for the graph representation of polymers and automated experimentation platforms such as continuous flow reactors facilitating their ingestion into AI/ML pipelines. CMDL enabled seamless use of historical experimental data to train new AI models for catalyst and materials design. Here, the utility of this approach was first demonstrated through the successful synthesis and experimental validation of novel AI-generated ring-opening polymerization catalysts. Next, we derived a first-of-kind generative model for block and statistical co-polymer structures from their CMDL graph representations. This model preserved critical functional groups within the polymer structure, allowing them to be readily validated experimentally. These results reveal how the versatility of CMDL facilitates translation of historical experimental data into meaningful predictive and generative models which produce experimentally actionable output.