August 17, 2022

Hydrogen-soaked crystal lets neural networks expand to match a problem

Hydrogen-soaked crystal lets neural networks expand to match a problem
Hydrogen-soaked crystal lets neural networks expand to match a problemHydrogen-soaked crystal lets neural networks expand to match a problem

Training AIs remains very processor-intensive, in part because traditional processing architectures are poor matches for the sorts of neural networks that are widely used. This has led to the development of what has been termed neuromorphic computing hardware, which attempts to model the behavior of biological neurons in hardware.

But most neuromorphic hardware is implemented in silicon, which limits it to behaviors that are set at the hardware level. A group of US researchers is now reporting a type of non-silicon hardware that’s substantially more flexible. It works by controlling how much hydrogen is present in an alloy of nickel, with the precise amount of hydrogen switching a single device among four different behaviors, each of which is useful for performing neural-network operations.

Give it the gas

The material being used here is one of a class of compounds called perovskite nickelates. Perovskite is a general term for a specific arrangement of atoms in a crystalline structure; a wide variety of chemicals can form perovskites. In this case, the crystal is formed from a material that’s a mix of neodymium, nickel, and oxygen.

The crystal structure has enough open space that it can readily absorb and hold onto hydrogen. Once the hydrogen is incorporated, its electron will often end up being transferred to one of the nickel atoms. This changes the electrical properties of the atom and, in doing so, changes the conductivity of the material in general. The degree to which they change depends on how much hydrogen is present.

Since the hydrogen ends up with a positive charge after giving up its electron, it can be controlled by externally applied electric fields. So, by controlling the electrical environment, it’s possible to redistribute the hydrogen within the perovskite structure. That will then change the conductive properties of the material.

The researchers show that these states are meta-stable: they’ll change if an external force is applied but will remain stable for up to six months without the need to refresh the hydrogen. It’s not clear whether it needs to be refreshed at that point or whether that’s simply the latest they checked.

In any case, the researchers create a device simply by hooking up the perovskite to electrodes in a hydrogen atmosphere. (Getting the hydrogen into the material requires one electrode to be made from platinum or palladium.) From there, they demonstrated that it can be reliably switched among four states.

One state allows it to act as a resistor, meaning the device can act as a memristor. Similarly, it’ll behave as a memcapacitor, holding charge if set in that state. When in spiking neuron mode, it will accumulate multiple signals, at which point its resistance changes dramatically. This mimics how a neuron requires incoming spikes to exceed a threshold before it switches into an active state. Finally they had a configuration that acted like a synapse (at least in neural-network terms), transforming an input based on its strength.

Obviously, it’s possible to do similar things with dedicated devices for each of the four functions if you’re willing to activate and shut off different parts of a chip when needed. But many of these behaviors are analog, something that silicon requires even more hardware to emulate. Here, all this is done with a single bit of material between two electrodes.