Using Experimental Data in Computationally-Guided Rational Design with Machine Learning

I will discuss our efforts to use machine learning (ML) to accelerate the computational tailoring and design of complex materials by leveraging experimental datasets. The breadth of transition metal chemical space covered by databases such as the Cambridge Structural Database is not conducive to application-specific modeling and the development of structure–property relationships. I will discuss how we have employed both supervised and unsupervised natural language processing (NLP) techniques to link experimentally synthesized compounds in the CSD to their respective applications - for catalysis, photophysical activity, biological relevance, and magnetism. I will describe how we have analyzed the chemical substructures within each dataset to reveal common chemical motifs in each of the designated applications. Next, I will describe how we have used large subsets of materials from the CSD to train machine learning models, leading to the design mechanically active components polymers from the ground up that lead to novel network scale toughness. I will describe how we have leveraged the large amount of experimental data available for metal-organic framework (MOF) materials. I will describe how we have trained machine learning models to predict their stability in terms of heat, activation, water, and in acidic/basic mixtures. I will show how we use these models in combination with computed properties for the multi-objective optimization of MOFs with unprecedented properties. Finally, I will conclude with a perspective on some of the challenges of working with and extracting experimental data for machine learning accelerated materials discovery.