A few months ago, we brought news of a bendable CPU, termed Plastic ARM, that was built of amorphous silicon on a flexible substrate. The use cases for something like this are extremely low-powered devices that can be embedded in clothing or slapped on the surface of irregular objects, allowing them to have a small amount of autonomous computing. But to meet the low power requirements, a minimalist processor is not enough—all the components have to sip power as well. And that makes for a poor fit for traditional RAM technology, which needs power to maintain the state of the memory.

But a group from Stanford now has that covered. The researchers have built a form of flexible phase-change memory, which is closer in speed to normal RAM than flash memory but requires no power to maintain its state. And, while their work was initially focused on getting something that’s flexible to work, the principles they uncovered during their work should apply to phase-change memory in general.

A change of phase

People have made flexible forms of memory before, including flash and ferroelectric RAM, and resistive RAM can be made from materials that are also bendable. But phase-change memory has myriad advantages. It works by connecting two electrodes via a material that can form crystalline and amorphous states, depending on how quickly it’s cooled down after heating. These two states differ in how well they conduct electricity, allowing them to be distinguished.

The electrodes provide convenient ways of reading and writing. By putting in high levels of current, you can heat the material; shutting off the current suddenly will let it cool in an amorphous state, while ramping the current down slowly allows a crystalline state to form. Once that’s done, the state can be read by passing a much smaller current through and reading the resistance; it’s even possible to store more than one bit per device by adjusting the heating to create several discrete resistance levels. Critically, no current is required to maintain the bit(s) stored in one of these devices, since the crystalline/amorphous difference is stable.

The problem is that resetting the device requires enough current to partially melt the material. So, while the average power use is low, it’s quite high at critical points. This poses a challenge for devices that may be powered by a trickle of charge harvested from environmental sources. So, making phase-change memory out of flexible materials is not enough. You also have to match its performance to the typical use cases for flexible devices.

Conveniently, part of the process of making it flexible also provided the solution for improving its performance.

Making it flex

A lot of flexible electronics are built on polymer substrates rather than rigid materials like silicon. In addition to being flexible, most polymers are insulators—they don’t conduct electricity or heat very well. And that turned out to be critical to boosting the efficiency of the phase-change memory.

The gist of the new device is that the phase-change hardware is surrounded by materials that don’t conduct heat well. This helps trap the heat required to partially melt the device where it’s needed, which means that you don’t have to generate as much heat in the first place. And that in turn means that you need to put less power in to reset the device.

The device was made by drilling a hole in aluminum oxide. The hole was then filled with alternating layers of tin telluride and tin/gallium telluride, which acted as the phase-change material. Electrodes ran across the aluminum oxide to wire up the two ends of the device, and it was built on top of a flexible polymer material.

Modeling showed that the combination of aluminum oxide and the polymer trapped heat in the hole where the phase-change material was located. This was confirmed by showing that the power requirements for resetting the device dropped as the the researchers increased the amount of aluminum oxide surrounding the device. At its best, the device’s power requirements were over 100 times lower than current devices fabricated on a silicon substrate.

All of that would be useless if the device didn’t function well. But the researchers showed that it could be wrapped around a metal rod that’s eight millimeters across and still operate normally. Performance was the same after 200 bending and straightening cycles, and the stability of its storage was confirmed to be good out to over 1,000 reads. Finally, multibit storage was demonstrated using different resistance levels. So, all in all, it seems to be doing about what you’d want from phase-change memory.

The researchers note, however, that the basic principle here—lowering energy use by thermally insulating the material that stores data—could also be used in more conventional rigid phase-change memory. And that could have some useful applications beyond memory, since other teams have shown that it’s possible to train a neural network in phase-change memory, rather than relying on repeated rounds of calculations. That process is already more energy efficient than using traditional computers for the same task, so boosting the energy efficiency of phase-change materials could make it an even better option.

Science, 2021. DOI: 10.1126/science.abj1261  (About DOIs).

https://arstechnica.com/?p=1793369