Hafnium oxide is the high-k gate dielectric that already sits inside every advanced CMOS transistor, and in doped or hafnia-zirconia form it is also ferroelectric, making it the only CMOS- and BEOL-compatible route to non-volatile memory and analogue in-memory compute at sub-10 nm dimensions.
Summary
Hafnium oxide (HfO2, hafnia) is described in the literature as the cornerstone high-k dielectric of modern silicon technology: it replaced silicon dioxide as the transistor gate insulator because its higher dielectric constant allows a physically thicker layer at the same electrical thickness, suppressing tunnelling leakage. Because the material itself cannot realistically be swapped out inside existing silicon process flows, the engineering lever is dopant incorporation, tuning band gap and dielectric constant with additions such as Al, Si and Y. That work finds formation energy, band gap and dielectric constant are decoupled parameters, so dopant choice has to be prioritised per application rather than jointly optimised.
The second, newer function is ferroelectricity. Doped HfO2 and HfxZr1-xO2 can be stabilised in a polar orthorhombic phase, which gives a switchable remanent polarisation in a film only a few nanometres thick. The significance is compatibility: doped-hafnia ferroelectricity makes FeRAM CMOS- and back-end-of-line-compatible and sub-10 nm scalable, unlike legacy PZT, and it supports logic-in-memory and compute-in-memory device styles. Recent device work integrates HfO2/ZrO2 ferroelectric nanolayers in the BEOL of CMOS, reaching nanosecond multilevel switching below 5 V, endurance above 10^11 cycles under partial switching, and ten-year retention of memcapacitance states.
The parameters that decide the technology are all defect and phase parameters. Performance is governed by the interplay of oxygen vacancies, dopants and structural defects, which drive polar phase stabilisation and the reliability failure modes known as wake-up, fatigue and imprint; these remain an active area of investigation, to the point that atom probe tomography is being proposed to map dopants and vacancy clusters in three dimensions. Doping trades polarisation against endurance directly: in epitaxial Ce-doped HfO2, remanent polarisation falls from about 15 to 3.8 microcoulombs per square centimetre as Ce content rises from 5% to 20%, while cycling stability improves up to 10^8 cycles, with the orthorhombic-to-tetragonal transition temperature dropping from around 800 degrees C at 5% to around 300 degrees C at 15%.
Adjacent functions are emerging from the same material family: multiferroicity in V-doped and Cr-doped hafnia tied to oxygen deficiency, and pyroelectric and electrocaloric response in oxygen-deficient core-shell Hf0.5Zr0.5O2 nanoparticles of about 7 nm average size. The dielectric role is not free of physics penalties either: modelling of a 1.6 nm double-gated silicon nanosheet with SiO2/HfO2 gate stacks finds room-temperature mobility much lower than bulk-phonon assumptions predict, once phonons are treated as clamped at the SiO2/HfO2 interfaces and remote-phonon scattering is included.
Evidence base
- 2026-06-03: BEOL-integrated HfO2/ZrO2 ferroelectric capacitors achieved nanosecond multilevel switching below 5 V, endurance above 10^11 cycles under partial switching and ten-year retention, but the memcapacitance window collapses above 1 MHz read frequency, requiring a 20 ps leakage-based readout.
- 2026-07-31: in epitaxial Ce-doped HfO2, remanent polarisation fell from about 15 to 3.8 microcoulombs per square centimetre as Ce went from 5% to 20%, while endurance improved to 10^8 cycles, and the orthorhombic-tetragonal transition temperature dropped from around 800 to around 300 degrees C.
- 2026-08-13: HfO2 is characterised as the cornerstone high-k dielectric in modern silicon, with dopant incorporation the principal available lever because the material itself cannot be replaced within existing flows; formation energy, band gap and dielectric constant are decoupled.
- 2024-10-15: Yole projects FeFET and HfO2-FeRAM growing from below 0.5% of embedded NV memory units in 2024 to at least 5% by 2030, with GF 22FDX FeFET reaching volume customer adoption in 2026 to 2027, and no displacement of HBM through 2030 2024 Yole Emerging Memories 2024.
- 2026-06-22: doped-HfO2 ferroelectricity is credited with making FeRAM CMOS- and BEOL-compatible and sub-10 nm scalable unlike legacy PZT, while embedded flash is confirmed to end at 28/22 nm on mask-count economics (six to eight extra masks versus about three for MRAM).
- 2026-06-16: performance of HfO2-based ferroelectrics is still governed by an incompletely understood interplay of oxygen vacancies, dopants and structural defects driving wake-up, fatigue and imprint, with atom probe tomography proposed as the 3D atomic-scale tool and demonstrated only as proof of concept.
Open questions
- Can the memcapacitance window collapse above 1 MHz be fixed at the material or stack level, or does every high-speed HfO2 FeCap application depend on indirect leakage-based readout schemes?
- Is the polarisation-versus-endurance trade-off seen in Ce-doped HfO2 general to all dopants, or can a dopant or strain route give both high remanent polarisation and beyond-10^8 cycling?
- Will atom probe tomography actually resolve the vacancy and dopant distributions well enough to give a predictive model of wake-up, fatigue and imprint, or does it remain a proof of concept?
- Does the remote-phonon mobility penalty at SiO2/HfO2 interfaces become a hard limit for sub-2 nm silicon channels, and if so what replaces the stack?
Assessment drafted 2026-08-31 from up to 11 KB sources using the technology-scorecard framework; scores are a draft read pending review.