Access Official Technical PDF

Share Summary:LinkedInEmail

Monolithic 3D CMOS integration offers a way to stack active devices rather than placing every transistor on a single plane. In a demonstrated monolithic 3D CMOS inverter, researchers formed an NMOS transistor from molybdenum disulfide (MoS2) and a PMOS transistor from tungsten diselenide (WSe2) on separate, sequentially fabricated levels. The devices shared a common gate, and the inverter reached a peak voltage gain of about 45 V/V at a 1.5 V supply voltage and a 1 µm gate length.

For automation and semiconductor decision-makers, the work is a device-level demonstration—not evidence of a production-ready logic platform. Its significance lies in combining complementary transistors made from layered semiconductors in a vertical architecture, while exposing practical issues such as threshold-voltage alignment and usable switching behavior.

Why build CMOS circuits in three dimensions?

Conventional planar CMOS places NMOS and PMOS transistors side by side. As device dimensions shrink, variability, parasitic resistance, and capacitance in devices and interconnects can limit circuit performance. Monolithic 3D integration instead fabricates active device layers sequentially. Stacking can reduce average interconnect length and capacitance, with the potential to support faster circuits and lower power dissipation.

A vertical approach can also bring different active materials together. That is relevant to transition metal dichalcogenides (TMDs), a family of layered semiconductors that includes MoS2 and WSe2. Their thickness can be controlled down to a monolayer, potentially helping address channel-thickness variation in ultra-thin-body devices. The source also notes that the relatively low dielectric constant of some TMDs may reduce drain-to-channel coupling, a consideration for short-channel behavior.

How the monolithic 3D CMOS inverter was built

The inverter paired a MoS2 NMOS on the lower level with a WSe2 PMOS above it. Both transistors used a shared gate, while their source and drain connections were electrically isolated. This distinction matters beyond the inverter: in logic gates such as NAND, NOR, XOR, and XNOR, transistor source and drain terminals cannot always be shared, so electrical isolation is needed in a general device architecture.

Fabrication began with MoS2 and WSe2 flakes transferred to a silicon dioxide substrate. The MoS2 channel and its source/drain contacts were formed first, followed by a zirconium dioxide (ZrO2) gate dielectric and the common gate. Another ZrO2 layer was then deposited, and the WSe2 flake was placed on top to form the second transistor level. Platinum and gold contacts provided the WSe2 source and drain connections.

A forming-gas anneal was performed after WSe2 transfer. The reported purpose was to clean the WSe2 surface and reduce the PMOS output conductance. This illustrates how processing steps between stacked layers can affect the electrical behavior of the completed circuit, not just the properties of an individual transistor.

Device performance and inverter gain

The measured transistor characteristics included on-to-off current ratios above 106 for the MoS2 NMOS and 107 for the WSe2 PMOS. Reported field-effect mobilities were 37 cm2/Vs for electrons in MoS2 and 236 cm2/Vs for holes in WSe2. The devices also showed output saturation, with output conductance below 1 nS/µm at the specified operating point.

In an inverter, voltage gain around the switching region depends on the transconductance of both devices and their output resistance. In simplified terms, higher transconductance and higher output resistance support greater gain. The measured peak gain was approximately 45 V/V at a 1.5 V supply and 1 µm gate length. The researchers attributed improved PMOS output conductance, relative to an earlier implementation, in part to the added forming-gas anneal.

These results demonstrate a high-gain switching response in a vertically integrated test structure. They do not, by themselves, establish manufacturing yield, long-term reliability, or performance in a complete control or computing system.

Threshold voltage remains a circuit-level challenge

High gain is only one requirement for a useful logic inverter. Ideally, the output should switch between the supply rails, with the switching point near half the supply voltage. In the reported device, the NMOS threshold voltage was negative, so the inverter did not switch across the full 0-to-1.5 V range. As a result, noise margins could not be calculated for that inverter configuration.

The study examined substrate back-biasing as one way to shift the MoS2 NMOS threshold voltage. Applying a more negative back bias made the threshold voltage less negative and moved the inverter switching point toward more positive input voltages. Across the tested conditions, reported gain rose from 26 V/V to 45 V/V as back bias changed from −50 V to −70 V. The authors also identified gate work-function engineering, channel doping, and local back-biasing as possible approaches to threshold-voltage control.

What the demonstration means for integration teams

For engineers evaluating emerging semiconductor architectures, the work highlights both an opportunity and a design constraint. Sequentially integrating complementary TMD transistors can produce a functional common-gate inverter with substantial voltage gain, and layered materials offer properties of interest for scaled devices. At the same time, circuit usefulness depends on more than peak gain: threshold alignment, rail-to-rail switching, noise margins, and the integration of isolated source/drain connections all matter.

The practical takeaway is to treat monolithic 3D TMD logic as a promising research direction rather than a drop-in replacement for established CMOS. The demonstrated MoS2-WSe2 inverter shows that vertically stacked complementary devices can operate together, while its threshold-voltage limitation underscores the need to optimize the full circuit architecture. For future integration work, device performance and process compatibility must be assessed alongside the requirements of the intended logic function.

Access Official Technical PDF

About The Author