ViscNet: Neural network for predicting the fragility index and the temperature-dependency of viscosity

Artificial intelligence and machine learning in glass science and technology: 21 challenges for the 21st century

“…While this approach may yield satisfactory results for a select glass composition, the models trained on one glass composition will not be transferable to another set of compositions with different components. A more generic approach would be to develop novel physics‐based features—some examples of this include topology—, 20 interatomic potential parameter—, 57 and periodic table‐based descriptors 8,21 . The advantage of such descriptors is that the models developed have the potential to be universal and transferable.…”

“…To infuse “common‐sense” into the model, which allows reasonable extrapolation using the basic physical laws available. This alternate approach, known as the gray‐box neural network, learns the parameters associated with a functional relationship between the input‐output, 8,21 instead of directly learning the input and output. For example, a recent work used the neural network to learn the parameters associated with the MYEGA equation, which was then used to predict the viscosity 8 .…”

“…A more generic approach would be to develop novel physics-based features-some examples of this include topology-, 20 interatomic potential parameter-, 57 and periodic tablebased descriptors. 8,21 The advantage of such descriptors is that the models developed have the potential to be universal and transferable. Furthermore, these models might be able to provide deeper insights into the factors controlling the respective properties.…”

“…Due to these reasons, the glass community has also started adopting the AI/ML approaches to tackle a variety of problems such as property‐prediction, tailored design, understanding the physics, accelerating the modeling, to name a few. Among these, some of the specific applications include the prediction of Young's modulus, 12‐14 liquidus temperature, 15 solubility, 16 glass‐transition temperature, 17,18 dissolution kinetics, 19,20 viscosity 8,21 . Other recent works have developed composition–property models for several important thermal, optical, and mechanical properties of glasses based on the available experimental dataset 22‐25 .…”

“…Among these, some of the specific applications include the prediction of Young's modulus, [12][13][14] liquidus temperature, 15 solubility, 16 glass-transition temperature, 17,18 dissolution kinetics, 19,20 viscosity. 8,21 Other recent works have developed composition-property models for several important thermal, optical, and mechanical properties of glasses based on the available experimental dataset. [22][23][24][25] A recent work has bundled these models, 22,23 along with a database and an optimization module, in a first-of-its-kind software package for accelerating glass discovery, namely Python for Glass Genomics (PyGGi, see http://pyggi.iitd.ac.in).…”