We will present novel approaches to representing, visualizing, and analyzing chemicals and materials as well as insights from model interpretation to enable accelerated discovery of novel functional chemicals and materials. Recently, we and others have developed novel structural and electronic materials descriptors that can be used to visualize the materials space (materials cartogram) and identify regions of materials with specific compositions and properties [1], as well as develop Quantitative Materials Structure-Property Relationships (QMSPR) models. Using the data from the AFLOW repository of high-throughput ab-initio calculations for inorganic molecules, we have generated QMSPR models to predict critical material properties such as metal/insulator classification, bulk modulus, Fermi energy, and band gap energy [2]. As an experimental proof-of-concept, we have employed the QMSPR approach to identify a novel photocathode material for dye-sensitized solar cells (DSSCs) [3]. In parallel, we and others have developed a novel computational strategy based on deep and reinforcement learning techniques for de-novo design of molecules with desired properties. Our strategy integrates two deep neural networks – generative and predictive – that are trained separately but employed jointly, with an added component of reinforcement learning, to generate novel chemical structures with the desired properties. In the proof-of-concept study [4], we have employed this strategy (termed Reinforcement Learning for Structure Evolution, or ReLeaSE) to design chemical libraries biased toward compounds with specific ranges of physical properties, such as melting point and hydrophobicity, as well as to develop novel compounds selective against specific targets such as kinase inhibitors. I will discuss recent advances in the ReLeaSE technology including visualization of model outcomes and talk about opportunities for combining ReLeaSE with property filters and robotic chemistry to accelerate practical design and discovery of novel chemical entities with the desired properties.

Emerging AI and machine learning approaches for designing novel chemicals and materials with the desired properties

2019 Fall National Meeting

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.