Running AI models means repeating huge matrix operations. In today’s computers, memory and processor are separate, so data keeps moving between them, and this movement costs more time and energy than the calculation itself. In-memory computing avoids the problem by doing the calculation inside the memory.
For this, we need devices that can hold an analog value and change it on demand, much like a synapse in the brain. Placed in a crossbar array, each device stores one weight, and the whole array computes at once. We grow these devices as epitaxial thin films by pulsed laser deposition, and by controlling growth conditions and composition, we optimize their memory characteristics.
A memristor stores information as resistance. In most memristors, switching depends on conductive paths that form at random inside the film, which makes the device behave differently from cycle to cycle.
We focus on making switching predictable. In complex oxides, ions can move along paths set by the crystal structure. In ferroelectrics, polarization can change the whole electrode interface at once. By controlling defects during growth, we tune how and where switching happens. The goal is a device that updates its weight evenly, lasts a long time, and works in large arrays without extra selector devices.
A memcapacitor stores information as capacitance and is read by the charge it holds. Almost no current flows, so power use is very low and neighboring cells in an array do not interfere with each other.
The main limit has been the small difference between capacitance states at 0 V. We use ferroelectric oxides, where polarization and defects together can widen this difference, so that several states can be stored in one device. We also work on circuits to read these states in a real system.