(356g) Fully Convolutional Neural Network Models for Materials Science Applications

In silico design can be an efficient and cost effective method for discovering new materials. Traditionally, quantum mechanics, classical approaches, and quantitative structure property relationships (QSAR) are used to design materials using the structure of the materials as an input. Although these methods are helpful in designing materials, they all suffer from a common drawback: they estimate properties from an input structure which usually necessitates iteration between synthesis and modeling to achieve the desired performance metric. Moreover, that process is also dependent on the ingenuity of design team to identify potential materials for computational evaluation which may be very different structurally from the benchmark material. Here we describe a generative deep learning approach to address some of these drawbacks: a fully convolutional neural network (FCNN) molecular generative model. The efficiency and reconstruction accuracy of this method was found to be comparable with Junction Tree variational autoencoder (JT-VAE) and superior to a SMILES-based variational autoencoder (CVAE). Furthermore, this approach (p-FCNN) is tuned to generate molecules which incorporate a larger fraction of heteroatoms (heavy atoms other than C’s) compared to those from CVAE. Detailed analysis will be provided to understand the relative distribution of the newly generated molecules vs. the training set. In addition, the latent space generated from FC-NN will be organized according to ALogP and the response surface will be smoothed using a Gaussian process model. Gradient search will be performed on the organized latent space to extract molecules near the optimal value of ALogP. In conclusion, FCNN and p-FCNN can be considered as an efficient generative deep learning approach to discover materials with an optimized performance metric.