The team is developing a method for computing with neural networks in water

Microprocessors in smartphones, computers and data centers process information by manipulating electrons through solid semiconductors, but our brain has a different system. They rely on manipulating ions in liquids to process information.

Inspired by the brain, researchers have long attempted to engineer “ions” in an aqueous solution. While ions in water move more slowly than electrons in semiconductors, scientists believe the diversity of ionic species with different physical and chemical properties could be harnessed for richer and more diverse information processing.

However, ionic computing is still in its infancy. So far, labs have only developed individual ionic devices such as ionic diodes and transistors, but until now nobody has assembled many such devices into a more complex circuit for computers.

A team of researchers from the Harvard John A. Paulson School of Engineering and Applied Sciences (SEAS), in collaboration with DNA Script, a biotech startup, has designed an ionic circuit comprising hundreds of ionic transistors and performed a core process of neural network computing .

The research is published in Advanced Materials.

The researchers began building a new type of ion transistor using a technique they recently pioneered. The transistor consists of an aqueous solution of quinone molecules connected like a porthole with two concentric ring electrodes with a central disc electrode. The two ring electrodes electrochemically lower and tune the local pH around the central disc by generating and trapping hydrogen ions. A voltage applied to the center disk causes an electrochemical reaction to produce a flow of ions from the disk into the water. The reaction rate can be accelerated or decreased – by increasing or decreasing the ion current – by adjusting the local pH. In other words, the pH controls (gates) the ionic current of the disc in the aqueous solution, creating an ionic counterpart of the electronic transistor.

They then engineered the pH-controlled ion transistor so that the plate current is an arithmetic multiplication of the plate voltage and a “weight” parameter representing the local pH that controls the transistor. They organized these transistors into a 16×16 array to extend the analog arithmetic multiplication of individual transistors to analog matrix multiplication, with the array of local pH values ​​serving as the weight matrix found in neural networks.

“Matrix multiplication is the most widely used computation in artificial intelligence neural networks,” said Woo-Bin Jung, a postdoctoral researcher at SEAS and first author of the paper. “Our ionic circuit performs matrix multiplication in water in an analogous manner, based entirely on electrochemical machines.”

“Microprocessors digitally manipulate electrons to perform matrix multiplication,” said Donhee Ham, Gordon McKay Professor of Electrical Engineering and Applied Physics at SEAS and senior author of the paper. “While our ionic circuitry may not be as fast or as accurate as digital microprocessors, electrochemical matrix multiplication in water is inherently appealing and has the potential to be energy efficient.”

Now the team is trying to increase the chemical complexity of the system.

“Until now, we have only used 3 to 4 ion species, such as hydrogen and quinone ions, to enable gating and ion transport in the aqueous ion transistor,” said Jung. “It will be very interesting to use more diverse ionic species and see how we can use them to make the content of the information to be processed richer.”

The study was co-authored by Han Sae Jung, Jun Wang, Henry Hinton, Maxime Fournier, Adrian Horgan, Xavier Godron and Robert Nicol.