Over the past few years, there has been a rising trend in the use of data-driven approaches for the discovery and design of new chemistry. These strategies encompass virtual high-throughput exploration of chemical space and analysis of the generated data using data mining techniques, and our group has developed a cyberinfrastructure for accomplishing these tasks. However, in order to accelerate this process, it is critical to identify and optimize several steps in our research paradigm and we therefore employ a Genetic Algorithm (GA) for generating high quality solutions to search and optimization problems.<br/> <br/>Rational model selection and hyperparameter optimization are important concerns for the efficient and successful use of machine learning but have so far largely remained unexplored by this community. To address these issues, we use GA to automatically perform these tasks and present results for training an artificial neural network to predict refractive indices of one hundred thousand small organic molecules. Further, we benchmark its performance against other algorithms popular in the data science community and the results show that our implementation outperforms these methods both in terms of time and accuracy. The effectiveness of our implementation is further demonstrated via a scenario involving multi-objective optimization for model selection.<br/> <br/>Next, we use the GA-optimized machine learning model for small organic molecules to ‘reverse engineer’ new molecules that will have a high refractive index. Since chemical space is infinite and its exhaustive exploration is impractical for any real application, it is desirable that this exploration proceeds in a direction biased towards a set of target properties. In principle, a fitness function for GA can be designed such that it penalizes sub-structures that suppress the desired property in a molecule. However, the biggest challenge in using GA for high-throughput screening is the encoding of molecules into a GA chromosome. We modify our implementation of GA to not only handle this by using a node-network of python objects for every molecule, but also keep track of the connections between molecules which can later be used for creating structural descriptors in a machine learning model. Using GA, we propose new molecules with high refractive index and validate the predicted values using Density Functional Theory.

Adapting evolutionary algorithms for autonomous machine learning in chemistry

Macromolecular Chemistry: The Second Century

[Author Benefits](https://scimeetings.acs.org/?utm_source=pubs_content_marketing&utm_medium=website&utm_campaign=0320_MCF_NPI_Launch_Spring_Homepage&ref=pubs_content_marketing)

[How to use SciMeetings](https://storage.googleapis.com/dl.morressier.com/acs/SciMeetings-ACSSpring2020-HowToGuide_M3020-Final.pdf)

[Frequently Asked Questions](https://storage.googleapis.com/dl.morressier.com/acs/SciMtng_ACSNMSp2020_FAQ_M3120_Final.pdf)

ACS Spring 2020 National Meeting & Expo

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