A search that starts with 32,000,000
An artificial intelligence system is software trained to find patterns in large amounts of data. Microsoft and the Pacific Northwest National Laboratory, a United States government laboratory, built one and set it loose on a search for a new battery electrolyte, the material inside a battery that carries charged particles between its 2 terminals. The system screened 32,000,000 candidate materials in a few days and narrowed them, through rounds of computer checks, down to 18 candidates promising enough for battery development. Laboratory scientists then picked 6 of those to actually try to make. 1 worked. The whole process, from screening to a synthesized material, took 80 hours. The material is a solid state electrolyte, a solid rather than a liquid, that combines lithium and sodium and cuts the lithium a battery needs by up to 70%. It was made into prototype battery cells, working versions built to test an idea rather than finished products, and it was still awaiting further testing when Microsoft published the result in January 2024.
Irrespective of whether it's a viable battery in the long run, the speed at which we found a workable battery chemistry is pretty compelling.
Brian Abrahamson, chief digital officer at the Pacific Northwest National Laboratory. Source 1.
Attempts against successes
The gap between candidates screened and materials made is not unique to the Pacific Northwest National Laboratory project. Comparing 3 different artificial intelligence projects on the same terms, how many materials each one attempted to make in a lab against how many it actually made, shows the same pattern every time.
The Pacific Northwest laboratory figure is the 6 candidates its scientists vetted for synthesis, source 1. The Berkeley A Lab figure is the 57 targets it attempted over 17 days, source 4. The MatterGen figure is the 4 candidates it attempted after narrowing 8,192 generated structures, source 5.
Show the numbers
| Pacific Northwest laboratory | 6 |
| Berkeley A Lab | 57 |
| MatterGen | 4 |
Google DeepMind built a separate system called GNoME, which predicted 381,000 new stable materials in November 2023, meaning materials whose atomic structure holds together instead of falling apart. That list includes 528 possible lithium ion conductors, materials that let the charged lithium particles inside a battery move through them. A robotic laboratory at the University of California Berkeley, called the A Lab, runs chemistry experiments without a person present. It used the GNoME list to attempt 57 of the predicted materials and made 36 of them, in 17 days of continuous operation.
A newer Microsoft Research system called MatterGen, not built for batteries specifically, generated 8,192 candidate structures aimed at 1 target property, narrowed them to 75, attempted 4 in the lab, and made 1, a compound of tantalum, chromium and oxygen. Energy storage is named in the MatterGen paper as one of the uses the method could eventually serve.
The Pacific Northwest laboratory made 1 material out of 6 attempted, source 1. The Berkeley A Lab made 36 out of 57 attempted, source 4. MatterGen made 1 out of 4 attempted, source 5.
Show the numbers
| Pacific Northwest laboratory | 1 |
| Berkeley A Lab | 36 |
| MatterGen | 1 |
One of the main challenges is that people publish their success stories, not their failure stories.
Vijay Murugesan, materials sciences group lead at the Pacific Northwest National Laboratory. Source 1.
What has not shipped
None of the 5 sources behind this pattern describe a machine learning discovered material reaching a shipping electric vehicle battery. The furthest any of them has gone is a working prototype cell, still awaiting more testing rather than a finished product. The technology speeds up finding a candidate. It does not shorten the multi year process of qualifying 1 for a real product.