History of Materials Informatics
From the origin of the term “Materials Informatics” to the build-out of data infrastructure and the leap brought by GNNs — tracing the journey of MI.
The Birth of “Materials Informatics” (2005)
Krishna Rajan, “Materials Informatics,” Materials Today 8, 38 (2005).
In 2005, Dr. K. Rajan published a paper titled “Materials Informatics” in Materials Today. The idea of combining databases, machine learning, and theory to derive new knowledge — an idea that still defines MI today — was already articulated more than two decades ago. Its foresight is striking.
Today's Materials Informatics has grown along the very framework this paper outlined: the integration of databases × machine learning × theory (computation).
Materials Genome Initiative (2011)
In 2011, the U.S. launched the Materials Genome Initiative (MGI), propelling MI research forward dramatically. Harnessing data science requires data: materials data from first-principles calculations were gathered, and databases such as The Materials Project and AFLOW were built.
Python-based data-analysis platforms such as matminer and pymatgen were also developed. All are freely and publicly available — it is no exaggeration to say they laid part of the foundation for today's MI research.
The Rise of Graph Neural Networks (2017–present)
The platforms established by the MGI made it possible to build property-prediction models based on a material's composition. Yet such models cannot distinguish, for example, diamond, carbon nanotubes, and graphite (pencil lead) — all made of the very same carbon atoms.
In 2017, several machine-learning models based on graph neural networks (GNNs) were proposed, overcoming this limitation. A GNN represents a material as a graph in which atoms are nodes and interatomic interactions are edges. By incorporating structural information, GNNs achieved a dramatic leap in prediction accuracy and now play a central role in MI research.