Modern computational-chemistry software suites are proven to have powerful predictive capacity. However, many methods remain time consuming and labor intensive for the analysis of new systems. Machine learning (ML) provides an alternative computing technique that has the potential to cut computational cost while maintaining comparable accuracy. Machine learning models have already been used to predict electronic properties of small organic molecules. We have previously demonstrated a computational modeling method for determining the Flory-Huggins ? parameter between polymers that uses a suite of molecular modeling software. This software suite uses ab initio methods, DFT calculations, and MD simulations to generate molecular geometry and miscibility data. The output is the ? parameter between two small molecules. We propose that it is possible to use a neural network to replace the blends modeling software in this method. A primary challenge in machine learning for chemistry is the meaningful representation of molecules in a dataset. Therefore, our focus will be to generate a molecular dataset that can successfully train an ML model to predict ? parameters, following the example of Montavon et al. (New Journal of Physics, Vol 15, 2013). The ML model would intake Cartesian coordinates and partial atomic charges of each atom in a pair of molecules, and output the ? parameter for that pair. We will generate a preliminary neural network to test our dataset that will be expanded upon in future work. This dataset and model will provide a foundation for efficiently determining the interaction between new systems of polymer molecules.<br/>

Machine learning dataset and model applied to calculating the Flory-Huggins

255th National Meeting (2018)

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