What do cats and batteries have in common? Not much, you might think. After all, cats are cuddly and purr. Batteries? They power your flashlights and cellphones, but no one wants a battery sitting on their lap while they watch TV.
Cats were the subject of a recent, surprising news item. A group of computer scientists at Google and Stanford University fed YouTube videos to a computer that was running a “machine learning” program. This program “trains” on the input to find clusters of similar images and once it’s trained, the computer can classify new images as belonging to one of the clusters. After training on images from ten million YouTube videos, the computer learned to reliably identify images of cats. Like a newborn baby, the computer started with no knowledge but learned to identify objects – in this case cats – based on what it had already seen. This exercise illustrates the ability of machine learning to enable recognition tasks such as speech recognition, as well as classification tasks such as identifying cat faces as a distinct category of images.
Batteries deserve attention on this website because of their essential role in any strategy for sustainable energy. Batteries are a primary means for storing, transporting and accessing electrical energy. For example, they provide storage of excess energy from wind and solar sources and enable electrical power for cars and satellites. Today’s hybrid and electric vehicles depend on lithium-ion batteries, but the performance of these vehicles is limited by the energy density and lifetime of these batteries. To match the performance of internal combustion vehicles, researchers estimate that the energy density of current batteries would need to increase by a factor of 2 to 5.
Strategies for achieving these gains depend on identifying new materials with higher energy densities. The traditional method for finding new materials is to propose a material based on previous experience, fabricating the new material and measuring its properties, all of which can be expensive and time consuming. More recently, computational methods, such as density functional theory, have been used to accurately predict the properties of hypothetical materials. This removes the fabrication step but can involve large-scale computing. Although both of these methods have produced many successful new materials, the time and expense of the methods limit their applicability.
Cats – more precisely, the machine learning program that recognized cats – could come to the rescue. Instead of watching YouTube videos, a machine learning method could train on existing databases (from both experiment and computation) of properties for known materials and learn to predict the properties for new materials. Once the machine learning method is trained (which can be a lengthy process), its prediction of material properties should be very fast. This would enable a thorough search through chemical space for candidate materials. Machine learning methods have not yet been used for finding materials for batteries, but they have been used for prediction of structural properties, atomization energies, and chemical reaction pathways. Their use in materials science is growing rapidly, and we expect that they will soon be applied to materials for batteries and other energy applications.
Russ Caflisch, Director
Institute for Pure & Applied Mathematics (IPAM)