In the modern world, the discovery of new materials has not only revolutionized technology, but also greatly impacted various aspects of our daily lives, from clean energy to advanced computing. Traditionally, scientists find new useful materials by synthesizing various potential structures in the lab and then rigorously testing each out. This has been a costly and time-consuming trial-and-error process. Recently, however, a study by Google Deepmind and Google Research has drastically improved the efficiency and scope of material discovery by leveraging the power of deep learning. Deep learning, a subset of artificial intelligence, involves training computer systems to perform tasks by learning from large amounts of data, mimicking the way humans gain knowledge. The team’s breakthrough in predicting over 2.2 million new materials, including 380,000 stable ones, signifies a pivotal advancement in the realm of materials science.

At the core of this scientific breakthrough is a two-step approach: generation and filtration. The process begins with the generation of diverse candidate crystal structures that are similar to known stable materials. Then a state-of-the-art graph neural network model trained on extensive datasets (dubbed GNoME for Graph Networks for Materials Exploration) filters these structures by predicting their stability. The structures identified by GNoME are then iteratively refined. This investigation led to the discovery of over 2.2 million new crystal structures, a significant increase compared to the 28,000 known previously. Importantly, many of these materials have been experimentally created and validated, demonstrating the practical effectiveness of this method.

The implications of this research extend far beyond the laboratory. The vast expansion of stable materials opens the door for technological advancements – imagine faster creation of better solar panels, more efficient batteries, and cleaner energy sources. Moreover, the use of deep learning models heralds a new age in materials science, where the discovery of novel materials is faster, cheaper, and more efficient than ever before. This showcases the untapped potential of deep learning and artificial intelligence in scientific discovery, accelerating the pace of innovation across a multitude of fields.

This study was led by researchers at Google Deepmind and Google Research, Mountain View, CA, USA, including Amil Merchant, Simon Batzner, Samuel S. Schoenholz, Muratahan Aykol, Gowoon Cheon, and Ekin Dogus Cubuk.

Managing Correspondent: Rosella (Qian-Ze) Zhu

Press Article: Millions of new materials discovered with deep learning (Deepmind)

Original Journal Article: Scaling deep learning for materials discovery (Nature)

Image Credit: Pexels/Jonny Lew