Innovation@IITD: Machine learning for glass compositions

Researchers at IIT Delhi have developed a first of its kind machine learning software -- Python for Glass Genomics (PyGGi) for predicting and optimizing glass compositions.

Priced at Rs 75,000, the first version of PyGGi was launched on August 2. Besides cutting on the cost, this machine learning algorithms will also help in reducing the manufacture time because of giving away of the usual trial and errors.

"In glasses, understanding and predicting the composition–structure–property relationship is the key to developing novel glassy systems. Data-driven approaches such as machine learning can prove key to predict structure and composition of glasses for tailored applications. PyGGi is a software package developed using python, for predicting and optimizing the properties of inorganic glasses," says Prof N. M. Anoop Krishnan, one of the Project Investigators (PI)

What is PyGGi?

PyGGi is trained with numerous experimental data sets. It has an ability to detect and predict glass properties like Poisson’s ratio, bulk modulus, shear modulus, elastic modulus and density. Predicting properties at the mere tap of a button, PyGGi uses machine learning algorithms like Gaussian Process and Neural Network. It has a built-in range function which analyses the changing compositions and thus the variation of the glass properties over a wide range of its oxide proportions. An interface is developed that can easily allow the users to choose glass compositions of interest, predict the properties, plot them, and finally, download them.


Why is it important?

Glasses are archetypical disordered materials that can be formed by the fast quenching of a liquid. Due to the disordered structure, glasses can accommodate a wide-range of compositions with almost any element in the periodic table. This makes it extremely challenging to predict the mechanical properties of glasses, as they exhibit highly nonlinear behavior with respect to the glass composition. As an alternative route, data driven approaches such as machine learning can prove key to predict structure and composition of materials for tailored applications.


Application fields?

PyGGi can be highly useful, both for industry and academia, to develop novel compositions for targeted applications and to significantly reduce the time and resources that go into developing such compositions. The approach is scalable and can be extended to a vast variety of materials. Publicly available datasets can be exploited to deploy machine learning techniques and not only design, but also discover or characterize materials.


What next?

The plan is to introduce new properties and an optimization module to predict glass compositions for targeted properties.The researchers are also trying to extend this scalable approach to other materials as well with aim of accelerating materials discovery for healthcare, energy, and automotive applications


28th November 2019