Chemistry and local environment adaptive representation graphs as material descriptors

Machine learning (ML) offers promising methods for identifying hidden correlations among variables in large parameter spaces in materials science, enhancing the design and discovery of materials. However, accelerating materials design and property prediction via supervised learning involves manually generating feature vectors or performing complicated atom coordinates transformation, which restricts the model to a limited set of crystal structures or makes it challenging to provide chemical insights. Finding an appropriate descriptor that can represent the complete behavior of a material structure is a significant challenge.

In a new study, IISc researchers led by Abhishek Kumar Singh at the Materials Research Centre have developed a unique low-dimensional featurisation technique known as “Chemistry and Local Environment Adaptive Representation” (CLEAR) graphs to automatically learn the properties of materials through the chemistry of the atom connections. CLEAR is an adaptive featurisation method that maps the chemistry of atoms with their local atomic environment using Voronoi nearest neighbors (NN). Different atom-environment pairs are identified using Voronoi NN which dominates the structural information in the formalism of CLEAR. The chemistry weighted atom-environment interactions eventually build up to a high-dimensional feature vector, which is flattened out via a pooling process to reduce the complexity, generating low-dimensional CLEAR descriptors.

The CLEAR descriptors are built for a high entropy alloy (HEA) database developed using first-principles density functional theory (DFT) calculations. Prof. Singh’s group have applied this approach to study the stability of compositional and configurational diverse high entropy alloys (HEAs). Furthermore, explainable machine learning is utilised to unravels the potential of CLEAR descriptors to obtain numerous scientific insights in HEAs. CLEAR has envisioned an automated and data-centered exploration of materials space for the accelerated design of novel materials. The proposed framework also provides a universal low-dimensional representation with chemistry-informed local environment descriptors, which outperforms the prediction of phases and formation energies in HEAs.

Reference:
Swetlana S, Singh AK, Chemistry and Local Environment Adaptive Representation Graphs as Material Descriptors, Acta Materialia (2024)
https://doi.org/10.1016/j.actamat.2024.120122

Website:
http://mrc.iisc.ac.in/abhishek/