Recent advancements in complementary metal-oxide-semiconductor (CMOS) based microelectrode arrays (MEAs) have enabled simultaneous, high-resolution neural signal recording and stimulation across thousands of channels. However, current systems rely on external computers for signal processing, introducing latency. As the number of recording sites increases, so does the volume of data to be processed, adding further power and bandwidth requirements. To address these challenges, research efforts focus on integrating analog, on-chip neural signal processing using memristive devices. A hybrid platform combining a CMOS MEA with a memristive device array would offer a high potential for realizing on-chip, low-latency, low-power neural signal processing. This work presents the development of CMOS-compatible memristive devices for seamless integration into the back-end-of-line (BEOL) of a CMOS MEA chip. Two major constraints must be addressed for such integration: • Thermal budget for BEOL integration: Memristive devices must be fabricated below a thermal budget of 400°C. • Operating voltage constraints: The CMOS MEA platform is fabricated using a 180 nm technology node, which limits the allowed device operating voltage to a maximum of 3.3 V. Two types of memristive devices were developed: (1) filamentary electrochemical metallization memory (ECM) devices based on W/SiO2/Cu stacks, and (2) non-filamentary, interfacial/bulk switching-type devices based on amorphous GaOx thin films. The ECM devices were fabricated using low-temperature, BEOL-compatible processes. W/SiO2/Cu cells with 10 nm SiO2 layers exhibit stable, forming-free switching at low voltages (VSET = 0.47 V, VRESET = -0.06 V) and high ON/OFF (~106) ratios. Transient current measurements show sub-milisecond switching times at voltages larger than 3.3 V. Scaling down the SiO2 thickness (8, 7, and 6 nm) reduces the switching voltages but introduces larger device-to-device and cycle-to-cycle variability and a forming step requirement. The ECM devices exhibit binary switching behavior with limited multi-level conductance control and stochasticity, which together limit their application for analog signal encoding. To address these limitations, non-filamentary Ti/GaOx/W devices were developed using sub-stoichiometric GaOx films grown by plasma-enhanced atomic layer deposition (PE-ALD) at 250 °C. These devices exhibit area-dependent, self-rectifying switching characteristics. The electrical properties of the GaOx film can be tuned by controlling the oxygen plasma exposure time during deposition, which modulates the oxygen vacancy concentration. Transmission electron microscopy (TEM) analysis revealed the formation of an interfacial TiOx layer critical to the switching. Devices with TiN and Pt bottom electrodes confirm the role of the TiOx interfacial oxide in switching. The Ti/GaOx/W devices exhibit multi-level analog conductance modulation with more than a thousand stable switching cycles and excellent cycle-to-cycle reproducibility. However, their operation voltage (8.0 V) exceeds the 3.3 V CMOS constraint. To reduce the operation voltage, bilayer stacks consisting of GaOx/HfO2 or GaOx/Al2O3 were introduced. GaOx/HfO2 stack achieves an operation voltage of 3.2 V but suffers from poor endurance ( 104), further reduced operation voltage (2.0 V), stable analog conductance modulation under identical pulse conditions, fast switching (50 µs), and low programming current (~0.02 A/cm²), all within the CMOS constraints. To evaluate neural activity detection, the optimized devices were tested using pre-recorded in-vitro neural signals provided by Naturwissenschaftliches und Medizinisches Institut (NMI) in Tübingen. Finally, a hybrid memristive device – CMOS MEA emulator was designed and implemented on a custom PCB at the Technical University of Berlin by the group of Prof. Dr. -Ing. Roland Thewes. The chips, consisting of GaOx/Al2O3 memristive device arrays, were fabricated and wire bonded to the CMOS MEA. Preliminary measurements using in-vitro neural cells were conducted together with the project collaborators Alex Ronja Wiemhoefer and Tom Stumpp at NMI. This proof-of-concept platform is intended to validate the functionality and efficiency of on-chip neural signal processing using optimized GaOx/Al2O3 bilayer memristive devices, establishing a foundation for future closed-loop neuromodulation systems.
Onur Toprak (Thu,) studied this question.