As many emerging techniques appear on market, more emphasis has been imposed on the development of various functional materials. However, materials discovery is still a harsh process that needs long time and high cost. In this respect, much attention has been paid on machine learning to accelerate the process. Its role has been extended from the prediction of materials properties in high-throughput screening to the direct design of target materials. In this study, an evolutionary molecular design methodology based on deep learning which automatically searches the routes for molecular modification was developed to alter the properties of initial molecules to target directions. Genetic algorithm was employed to make perturbations on the fingerprint vectors of starting molecules. By predicting the materials properties using deep neural network, only favorable mutations for target directions were screened, and subsequent crossovers using the filtered mutations further strengthened the changes. In every stage, the modified fingerprints were transformed into simplified molecular-input line-entry system strings using recurrent neural network not only to confirm molecular structures and properties with DFT simulation but also to update the neural networks. Successive iterations gradually enhanced the properties and generated automatically optimized molecules. The effectiveness of the methodology was validated by designing light-absorbing organic molecules using PubChem library with the objectives of increasing and decreasing maximum absorbance wavelength. The starting molecules had been evolved successfully as intended and the model also provided us the rational design clues for target directions. The closed-loop evolutionary workflow guided by deep learning was effective to find appropriate design paths from large amount of data and could accelerate the overall design process compared with previous reports processed only by simulation. We expect that the proposed methodology can be a promising tool in exploring chemical space and discovering novel materials by exploiting inherent knowledge in data.

Evolutionary design of organic molecules based on deep learning and genetic algorithm

256th National Meeting (2018)

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