Material property prediction and structural inverse design with modern deep learning techniques

[ACCESS RESTRICTED TO THE UNIVERSITY OF MISSOURI AT REQUEST OF AUTHOR.] To manipulate the mechanical and physical properties of bulk materials, like metals, ceramics, and semiconductors, the introduction of structural defects on the atomic or nano-level scales to the material has been widely adopted [1]. Properties that can be altered using this strategy include mechanical [2-4], magnetic [5-7] and electronic [8-11] properties. Recent development in the area of machine learning (ML) and deep learning (DL) generated new insights in the area of material research. ML models have been applied to the prediction of material properties of stoichiometric inorganic crystalline materials [12]. With the motivation to resolve the challenge of applying ML and DL in problems to correlate the predicted properties to its corresponding material structures, our lab previously proposed a brand new Concatenate Convolutional Network (CCN) [1] for predicting electronic properties, i.e. bandgaps, for doped graphene, a 2D material with widely tunable properties by doping different atoms. The proposed DL network provided a very promising performance in the prediction of electronic properties of graphene and boron-nitride (BN) hybrids, a well-known 2D bulk material. To take one step further in the performance of the prediction, as well as providing more insights into the structure-property relationships, we recently have applied a modified version of Google Inception V2 [13] network to the previously proposed problem and achieved much more improvements on predicting the electronic properties. The success of a highly accurate prediction of electronic properties led to the possibility of inverse design of the material using an even newer DL structure, Generative Adversarial Network (GAN). Since the possibility of hybridized graphene is enormous due to the dopant atom species, concentrations, and configurations, searching through such a vast material dimensional space in a high throughput manner would largely prompt the usage of doped graphene in the field of electronics, photo-electronics and robotics. More prominently, direct inverse design based on a desiring target functionality is highly anticipated. We recently have proposed a brand-new GAN structure solving the problem of inverse material design provided with the desired properties [14]. The new GAN structure we proposed can generate material data conditioned on a given electronic property, which is a continuous quantitative label. To the best of our knowledge, current existing GAN structures cannot generate data with the functionalities of regressional and conditional, some of the previous trials have ended up in either poor performance or non-fully autonomous generation. With the modified Google Inception Network to predict the electronic property of graphene-BN hybrids (h-BN) and the new regressional and conditional GAN (RCGAN) to design h-BN upon a desired electronic property, we wondered if the strategies and neural networks can be extended further into other material property related problems. Similar strategies were applied for mechanical property related problems of h-BN, however, instead of training a neural network from scratch, transfer learning has been adopted. The new network borrowed the prediction power from the network used for the electronic property prediction through sharing the same set of weights on the convolutional layers. The new network also achieved higher accuracy in predicting mechanical properties for graphene-BN hybrids, while required less resource in training the network and converged to a stable performance with a higher efficiency. After predicting mechanical properties of h-BN graphene successfully, inverse design based on desired mechanical properties has been achieved using RCGAN. To further explore applications of ML and DL in the world of material science, several ML and DL models have been applied to resolve the problem of predicting methane uptake based on material dimensions and environmental conditions. Two major questions raised are, what is the key factor affecting the methane uptake and how are they affecting it. They have been solved using the feature importance vectors output from the ML model with the highest predicting accuracy, along with the visualization of methane uptake through a contour plot created using the DL model. In all, throughout my latest year in research topics combining ML and DL with material problems, we have proposed several feasible strategies and mechanisms for gaining more insights into material and its correlated properties. We will focus more on the area of autonomous chemical structural design and discovery upon a desiring property using modified RCGAN and generate SMILES [15] representing chemical structures corresponding to the desired property in the future research.