Software Development

IKST team and its collaborators are involved in the development of simulation environments and source codes. The package CINEMAS has been developed by IKST which can be used for setting up manipulating visualizing and post-processing the density functional theory based calculations. To study the transport in semiconductor we have developed the AMMCR code which is interfaced with interfaces with Vienna ab-initio simulation package (VASP). It is written in C++ and is extremely fast. The work was done in collaboration with IIT-Bombay. To study the transition states and activation barriers for the chemical reactions and diffusion processes, we have developed in collaboration with Indian Institute of Science (IISc), a python based code, namely PASTA which is interfaced with multiple DFT codes. Further collaboration with IISc team and IKST resulted a post processing code for the widely used wannier90 package which can be employed for studying chemical bonding in covalently bonded materials..

A wealth of information is available in modern age in relatively unstructured ways such as scientific publications, web content, online documents, etc. Text mining and natural language processing (NLP) approaches are extremely useful in extracting valuable information from textual resources, in a sophisticated manner managing extracted information, and in building new knowledge through them. In recent years such methods have enabled automated text and table data mining across a wide range of scientific documents.Through collaboration with IITKGP, we are in the process of developing a suit which can create auto-generated database of computational methods and the corresponding results from the online and offline scientific literature. The relevant scientific literature will be in the form of portable document format (PDF) or text.

Predicting different material properties using machine learning and deep learning methods has recently attracted considerable research interest. Recently proposed crystal graph convolutional neural network (CGCNN) is an efficient way of representing the crystalline materials through multigraphs or crystal graphs. By adapting Graph Convolution Neural Network (GCN) to these graphs one can generate the representation of each node and finally pooling layer is used to generate a single latent representation for the entire structure.

The CGCNN's main drawback in its current form is that large data set is required to train the properties. To overcome this issue by we have planned to develop a method which will use transfer learning approach within CGCNN framework. In addition, the integration of explicit features from certain online tools such as Matminer will increase the predictability. For such purpose we plan to develop a package in collaboration with IITKGP which is about to be scheduled.