AI is expanding its horizons in materials science. A research group from Google’s DeepMind has recently created GNoME (Graph Networks for Materials Exploration), a deep learning AI that successfully discovered 2. An estimated 2 million new chemical compositions of inorganic crystals are possible if synthesized, while only 380,000 are expected to be thermodynamically stable.
These discoveries, which scientists say are still only “treading water” and are equivalent to 800 years of traditional research, could be used to create new technologies ranging from improved batteries to superconductors. It shows that researchers are progressing much faster than ever with the help of Artificial Intelligence, which can revolutionize the industries that rely on these materials.
How Google’s AI is speeding up a process that used to take years of research
Conventional material search has always been a process that is time-consuming and often very exhaustive. Other researchers would begin by taking known crystalline structures and experimenting with them, using much trial and error that could take as long as a few months or even years.
Organic materials depend on rigid atomic structures, which are the building blocks of today’s advanced technologies, such as solar panels, computer chips, and batteries.
However, synthesizing these new materials in the lab is a tedious process that usually takes hours, perhaps days. Yet, there is no assurance that the correct formula will be discovered from all these experiments.
The enumerated limitations become more intense when searching for materials with higher stability and performance characteristics. Based on the data from the Materials Project, a collaborative database hosted at the Lawrence Berkeley National Laboratory, it was stated that human-driven experiments led to the discovery of nearly 20,000 inorganic crystals.
The technology behind GNoME and how it is changing material discovery forever
This is where the theory of GNoME (Graph Networks for Materials Exploration) comes in. DeepMind’s system is designed to predict the structure of previously unknown inorganic crystals at an unprecedented speed and with high levels of accuracy. GNoME uses two pipelines: a structural pipeline and a compositional pipeline. The structural pipeline is based on actual crystal structures and optimizes them to predict new materials; the compositional pipeline is based on random distributions of atoms within compositions.
Subsequently, GNoME assesses the stability of these created potentials with density functional theory, a standard computational tool employed in chemistry and physics domains. In any case, whether the outcome is positive or negative, the AI uses its prediction results and incorporates them into future versions as a form of learning. This learning loop repeats hypothesis formation and experimentation but at a much faster speed than a human could ever gain.
Fully automated laboratories are now a reality with AI-driven technology and robotics
Unfortunately, GNoME’s predictions can be helpful, but most materials have yet to be synthesized and experimentally investigated. To address this, researchers at Lawrence Berkeley National Laboratory have adopted autonomous robotics. Their A-Lab, a fully automated laboratory system, can run each experiment independently of a human operator. It can autonomously place elements, produce samples, and evaluate outcomes, driven by the AI algorithm for the subsequent procedure of developing new materials. It synthesizes and invents far more rapidly than researchers.
After 17 days, A-Lab achieved 71% success in synthesizing 41 of the 58 materials that GNoME projected. This contribution epitomizes how synergistic AI predictions and robotics are in materials science. The dynamics shown in the success of A-Lab also suggest the possibility of fully automated labs that could simultaneously work with intelligent models such as GNoME to produce new materials at a higher speed, mentally verify their predictions, and even adjust mechanisms according to outcomes.
