Researchers at University of Southern California, Airforce Research Lab/RX, Kumamoto University and TetraMem have demonstrated graphene-based memristors that remain reliable at temperatures where conventional electronics fail. Their Gra/HfOx/tungsten (W) devices operate up to 700 °C with an ON/OFF current ratio greater than 103, data retention beyond 50 hours, and endurance exceeding 109 switching cycles, while switching in tens of nanoseconds at about 1.5 V. At 700 °C the devices showed no intrinsic limit; the maximum temperature was set by the test equipment.
Most commercial electronics begin to fail when pushed past roughly 200 °C, which has long limited systems for applications such as Venus landers, deep‑earth geothermal drilling, and nuclear or fusion energy environments. In this context, a non‑volatile memory technology that can operate stably at 700 °C marks a major advance for extreme‑environment electronics. The new device is a memristor, a nanoscale two‑terminal component that stores information and can perform computing operations. It is built as a tiny sandwich: a tungsten top electrode, a hafnium oxide (HfOx) switching layer, and a one‑atom‑thick graphene bottom electrode. All three materials can withstand very high temperatures without structural breakdown, enabling robust operation in regimes far beyond conventional silicon‑based chips.
The key to the high‑temperature stability is interfacial engineering. In conventional Pt/HfOx/W memristors, high heat drives tungsten atoms from the top electrode through the HfOx layer into the bottom platinum electrode, eventually forming a permanent conductive path that leaves the device stuck in the ON state. Transmission electron microscopy on these structures after high‑temperature annealing reveals pronounced W diffusion into the inert Pt electrode, directly associated with this failure mechanism. In Gra/HfOx/W devices, similar analysis does not show W penetration at the graphene interface, and the devices continue to switch normally. First‑principles calculations attribute this to weaker W adsorption and higher surface diffusion barriers on graphene than on metals like Pt: tungsten atoms that reach the graphene surface cannot anchor and instead migrate away, preventing the formation of a permanent short. This demonstrates how 2D materials like graphene can suppress diffusion‑driven degradation and highlights the critical role of interface design in high‑temperature NVM.
The discovery itself arose while the team was attempting to build a different graphene‑based device that did not behave as expected. “To be honest, it was by accident, as most discoveries are,” said Joshua Yang, who led the study. “If you can predict it, it’s usually not surprising, and probably not significant enough.” The researchers then combined advanced electron microscopy, spectroscopy, and quantum‑level simulations to link the macroscopic performance-700 °C operation, >103 ON/OFF ratio, >50 hour retention, and >109 cycles—to the microscopic physics at the graphene–tungsten interface. This mechanistic understanding turns a single serendipitous result into a design principle that can guide the search for other 2D/metal combinations with similarly suppressed interdiffusion.
Beyond harsh‑environment memory, the device is naturally suited for artificial intelligence workloads. In modern AI systems, including large language models, most of the computation is matrix multiplication. A memristor crossbar can perform these operations physically via Ohm’s law: voltage multiplied by conductance yields current, so matrix–vector products are obtained directly from the current flowing through an array of programmable conductances. “Over 92 percent of the computing in AI systems like ChatGPT is nothing but matrix multiplication,” Yang said. “This type of device can perform that in the most efficient way, orders of magnitude faster and at lower energy.” With co‑authors Qiangfei Xia, Miao Hu, and Ning Ge, Yang has co‑founded TetraMem to commercialize room‑temperature memristor chips for AI computing, already used in the lab for machine‑learning tasks that conventional hardware cannot match. The high‑temperature Gra/HfOx/W version could extend these capabilities into environments where standard chips cannot function, enabling spacecraft, probes, or industrial sensors to process data directly on site.
Although these devices were hand‑fabricated at sub‑microscale and memory alone does not constitute a full computing system, two of the three materials (tungsten and hafnium oxide) are already standard in semiconductor manufacturing, and graphene has been demonstrated at wafer scale and appears in foundry roadmaps. For Yang, the publication of this work signals a broader shift: “Space exploration has never been so real, so close, and at such a large scale,” he said. “This paper represents a critical leap into a much larger, more exciting frontier.”
