Highlight Talk #1: Anoop Krishnan, IIT Delhi

Rational Design of Glasses using Deep Learning

Glasses are used ubiquitously for a wide range applications from aerospace to nuclear industry. Despite the extensive usage, designing glasses with tailored properties remain an open challenge due to their extremely non-linear composition–property behaviour. Here, using deep learning, we develop composition–property models for glasses with large number oxide components (more than 40). Specifically, the model can be used to predict relevant properties of glasses, namely, density, Young’s modulus, shear modulus, bulk modulus, refractive index, abbe number, glass transition temperature, hardness, liquids temperature, and thermal expansion coefficient. Further, using these models, we develop multi-property charts, named as glass selection charts. These charts enable the rational design of oxide glasses. All the models developed are shared as a GUI-based software packaged Python for Glass Genomics (PyGGi). Finally, we also implement the design of tailored glasses using various surrogate-model based single and multi-objective optimization schemes. We believe that the approach presented here can significantly accelerate the design of glasses reducing the design to deploy period by more than a few years, thereby accelerating the design of novel glass compositions.

Highlight Talk #2 Livia Bartok-Partay, University of Warwick

High-throughput computational thermodynamics

In recent years we have been working on adapting a novel computational, sampling technique, called nested sampling, to study the potential energy surface of atomistic systems from a new perspective. Nested sampling automatically generates all the relevant atomic configurations, unhindered by high barriers, and without advance knowledge of the potentially stable structures. One of the method's most appealing advantages is that the global partition function can be calculated at arbitrary temperatures very easily, as a simple post-processing step, thus thermodynamic properties become accessible. Nested sampling may be fully automated, allowing high-throughput calculations of phase transformations and phase diagrams of different materials: clusters, metals or alloys.

Highlight Talk #3 – Thomas Hammerschmidt, ICAMS, Ruhr University Bochum

Machine-learning material properties with domain knowledge of the
interatomic bond

Machine learning the properties of materials relies on descriptors of local atomic environments. In a recent work we introduced descriptors that are based on the moments of the local electronic density of states.
These moments are explicitly linked to the crystal structure and the chemistry through the moments theorem. This makes it possible to construct electronic-structure based descriptors of the local atomic environment that have an immediate relation to the binding energy.
This domain-knowledge of the interatomic bond rules the descriptors highly competitive for machine-learning applications. Here, we demonstrate their performance using simple tight-binding Hamiltonians as the basis for the moments calculation.