It’s official—artificial intelligence from Google and Lawrence Berkeley National Laboratory has achieved the equivalent of 800 years of human research in materials science

The science of materials seems to have welcomed artificial intelligence with open arms. Recently, GNoME (Grapo Networks for Materials Exploration) was created by researchers at DeepMind, Google, together with the Lawrence Berkeley National Laboratory. This is a deep learning AI that has been able to discover nearly 2 million chemical compositions of inorganic crystals resulting from the synthesis process, of which 380,000 could be thermodynamically stable. The significance of this event is better understood when it is noted that the global scientific database only contained 48,000 stable crystals, meaning that the AI has done work equivalent to 800 years of traditional research.

The data obtained from the robots at the Berkeley Laboratory can be found publicly, and demonstrate the discovery of new 2D materials similar to graphene, 25 times more conductive to lithium, essential for battery production, and new lithium-manganese leisure compounds, an alternative to lithium-cobalt that has generated much controversy in its use in batteries. These discoveries shed light on various aspects such as the creation of cleaner energy, improvements in computing with new semiconductors and potential superconductors, and enhancements in medical technology materials. This union between quantum physics, robotics, and AI facilitates obtaining more precise and intelligent predictions.

GNoME (Graph Networks for Materials Explotration)

The union of robotics, quantum physics, and artificial intelligence is opening very promising doors in the world of research. This is demonstrated by the creation of GNoME (Graph Networks for Materials Exploration), developed by researchers from DeepMind at Google and the Lawrence Berkeley National Laboratory. It is a deep learning AI that has been able to discover, in 17 days, 2.2 million new inorganic crystalline structures, of which 380,000 are expected to be thermodynamically stable.

At first glance, these numbers may not seem significant, which is why it is important to mention that the global scientific database contained 48,000 stable crystals before this discovery, and the 17 days would translate into 800 years of traditional research. “This represents almost a tenfold increase over previously known stable inorganic crystals,” said Michael Nuñez in VentureBeat.

New discoveries

The studies carried out with robots in the Berkeley laboratory showed how they were able to test 58 predictions, managing to synthesize 41, with a 71% success rate. The results achieved in this project are included in the materials project database, which anyone can access and find that it includes:

• 52,000 potential new 2D materials similar to graphene.
• 25 times more solid lithium conductors than previous studies–essential for the next generation of batteries.
• New lithium-manganese oxide compounds, a promising alternative to the controversial lithium-cobalt used in batteries today.

GNoME’s Engine

Some voices have spoken out questioning the functioning of these robots, casting doubt on whether it is truly intelligence or alchemy. The reality is that GNoME works thanks to the use of graph neural networks (GNNs), models that represent atoms and their bonds as graphs. This is what allows it to make predictions in a matter of seconds, and to determine whether a structure will be stable or not. The process is as follows:

1. Two parallel pipelines are taken: one explores variations of known crystals; the other ventures into new chemical combinations.
2. A ultra-fast filtering is carried out: each proposal is analyzed with Density Functional Theory (DFT) simulations, a standard method in physics and chemistry.
3. Active learning: the results are fed back to the model, which is refined with each cycle, increasing accuracy and speed.

What repercussions does this advancement have?

It’s not just about being able to make faster and more accurate predictions (which is significant in itself). This finding has repercussions in several other areas, such as:

• Generating cleaner energy with longer-lasting batteries.
• Advancements in computing: new semiconductors and even potential superconductors, like Mo5GeB2, for the IT sector.
• Developing materials that make medical equipment more durable and precise.