AI Creates 41 New Materials in Just 17 Days
-
On November 30, Nature published two groundbreaking studies: The latest AI-driven platform GNoME (Graph Networks for Materials Exploration) can independently discover and synthesize new inorganic compounds, including identifying over 2.2 million stable structures and creating 41 new materials in just 17 days, with both speed and accuracy far surpassing human capabilities.
Technological advancements have improved computer programs' ability to identify new materials, but the main obstacle in this process is how learning algorithms adapt to results that contradict their training, as new discoveries inherently require novel and creative ways to interpret data.
The DeepMind team proposed a computational model that enhances the efficiency of material discovery through large-scale active learning. This program is trained using existing literature to generate diverse potential compound candidate structures, which are then continuously refined through iterative learning. GNoME identified over 2.2 million stable structures, improving the accuracy of stability predictions to over 80%. In predicting compositions, the accuracy increased to 33% per 100 trials, compared to just 1% in previous work.
In the second study, a team from the University of California, Berkeley, developed an automated laboratory (A-Lab) system. This A-Lab is trained on existing scientific literature and, combined with active learning, can create up to five initial synthesis recipes for proposed compounds. It then uses robotic arms to execute experiments, synthesizing compounds in powder form. If a recipe yields less than 50%, the A-Lab adjusts the recipe and continues experimenting, stopping only after achieving the target or exhausting all possible recipes. After 17 days of continuous experimentation, the A-Lab conducted 355 experiments, producing 41 out of 58 proposed compounds (71%). In contrast, human researchers would take months of trial and error.
The two studies demonstrate that AI training, combined with rapid advancements in computational power and existing literature, holds immense promise for using learning algorithms to assist in the discovery and synthesis of inorganic compounds. Future autonomous laboratories will be able to uncover new materials with minimal human intervention and at unprecedented speeds.