Ferroelectrics are insulators whose electric polarisation can be flipped by a voltage and stays flipped without power, and the current research front is dominated by hafnium-zirconium oxide for CMOS-compatible embedded memory plus a fast-growing set of layered van der Waals and quantum ferroelectrics for memory, photonics and magnetoelectric devices.
Summary
A ferroelectric is an insulator with a spontaneous electric polarisation that can be switched between two or more energetically equivalent states by an applied field, and that retains its state when the field is removed. That makes it the natural material basis for non-volatile memory, for tunable capacitors and piezoelectric actuators, and for devices where a static polarisation is used to gate a semiconductor channel, a junction barrier or a magnetic layer. Physically the polarisation comes from a polar distortion of the crystal (a condensed polar phonon mode, or a set of ionic displacements); the height of the barrier between the equivalent states sets the trade-off between retention and switching voltage, and the depolarising field, screening and strain boundary conditions decide whether the polar phase survives at all in a thin film.
The commercially consequential branch is doped hafnia, usually Hf0.5Zr0.5O2 (HZO). Its attraction is that it can be crystallised at temperatures compatible with CMOS back-end-of-line processing and scales to thin films, so it can be added on top of finished logic as embedded non-volatile or computational memory. Its problem is that the ferroelectric orthorhombic phase is metastable: it must be stabilised against the non-polar monoclinic ground state, and the sources show that oxygen-vacancy concentration and the resulting chemical strain are a controlling variable in ~7 nm HZO nanoparticles. Because crystallisation kinetics are so sensitive, device-to-device non-uniformity across a wafer is the stated barrier to high-volume manufacturing, to the point that groups are applying unsupervised learning (PCA plus K-means) to intra-die variation to predict the behaviour of unseen dies. Process work is aimed at the same problem: visible nanosecond laser annealing of a Si3N4/TiN/HZO stack that mimics a CMOS geometry shows a sharp threshold in pulse energy density for crystallisation, with the ferroelectric phase fraction depending on film thickness.
A second branch is layered van der Waals ferroelectrics, where the polar axis can lie in-plane and the material can be stacked with semiconductors without lattice matching. A SnSe/alpha-In2Se3 p-n heterojunction has been used to build a ferroelectric junction field-effect transistor memory with 0.1 pA reverse leakage, a diode ideality factor of 1.95, 900 meV of ferroelectric barrier modulation and a 1.8 V memory window. NbOI2 and its NbOI2-xClx alloys are being developed as both a tunable electronic material and a photonic one, with composition-dependent second-harmonic generation and, in slab waveguides, giant optical spin-splitting via the optical spin Hall effect over tens of micrometres. Machine-learning screening of the two-dimensional ABC2X6 family has produced multiferroic candidates such as monolayer AuCrP2S6, with a ferromagnetic ground state, 7.46 pC/m out-of-plane polarisation and a ~130 meV per formula unit switching barrier, plus a bulk-photovoltaic readout scheme for four-state memory.
Underneath both branches, the field is re-tooling its theory and its measurement hygiene. A unified definition of ferroelectricity based on switchable polarisation differences between energetically equivalent states, rather than polar space-group membership, has been used to screen synthesised materials and identify new quantum ferroelectrics including Ba3I6 and Cs2PdC2. Finite-temperature behaviour can now be predicted from ground-state properties alone, without the fitted temperature-dependent parameters that Landau-Ginzburg treatments require, with quantitative agreement for PbTiO3, SrTiO3 and KTaO3. On the sceptical side, series capacitance artefacts can produce apparent polarisation hysteresis loops even under the double-wave method, and loops of that kind appear in the published literature; and a 25-year-old prediction of room-temperature out-of-plane ferroelectricity in compressively strained SrTiO3 has been re-examined at up to -3% strain, where butterfly capacitance-voltage hysteresis and domain-like piezoresponse were found to be inconsistent with a stable ferroelectric state.
Viability (3/5)
TLDR: Real switching in real devices, but phase metastability, wafer-scale variability and false-positive measurements all remain live.
Drivers (3/5)
Supply side is well evidenced. In a single month the sources show new fabrication routes (visible-laser BEOL crystallisation of HZO; chemical vapour transport growth of NbOI2-xClx alloys; bulk metathesis synthesis of wurtzite-derived Mg3WN4, with the reaction starting near 380 C and completing by 600 C, motivated partly by ferroelectric switching), new screening machinery (high-throughput identification of conventional and quantum ferroelectrics under a unified definition; ML screening of 2D ABC2X6 multiferroics), and new design tools (a GPU-accelerated full-vector polarisation solver replacing CPU solvers with simplified electrostatics and reduced-dimensional polarisation fields; ground-state-only prediction of finite-temperature ferroelectric properties). That is a healthy pipeline of independent groups and complementary capability.
Demand side is thinner in this corpus. The only explicit statement is that AI workloads are driving exponentially growing demand for non-volatile and computational memory, given as motivation in an abstract rather than as market evidence. No source here contains foundry roadmaps, product announcements, capacity commitments or pricing. The score is capped at 3 for that reason: the push is demonstrable, the pull is plausible but unmeasured in the supplied material.
TLDR: Strong and diverse supply-side push across many independent groups; demand is asserted rather than quantified in these sources.
Novelty (3/5)
Where the sources make comparisons, they are explicit and the margins are meaningful. The finite-temperature framework improves on Landau-Ginzburg by removing the need for prior knowledge of critical behaviour and fitted temperature-dependent parameters, and reproduces measured dielectric and ferroelectric properties across full temperature ranges for PbTiO3, SrTiO3 and KTaO3. The GPU solver is positioned against CPU solvers that use simplified electrostatics and reduced-dimensional polarisation representations, which limit the domain structures they can capture. Visible-wavelength laser annealing is set against prior UV and IR work, exploiting TiN absorption in a CMOS-like stack. The unified definition of ferroelectricity is a genuine conceptual widening: polar space-group criteria miss ferroelectricity in non-polar space groups, and dropping that criterion expands the accessible material set.
At device level the ferroelectric junction FET memory is described as a fundamentally different device rather than an increment, with 900 meV of barrier modulation and near-ideal junction behaviour, and the AuCrP2S6 proposal solves a named problem in multiferroic memory, the lack of non-destructive readout, using photocurrent sign and spin channel as two independent probes. What is missing is any benchmark against the technologies these would displace: no endurance, retention, energy per bit or density comparison against existing memory appears in these sources, so the novelty cannot be scored above contested.
TLDR: Clear, specific gains over named prior methods in theory, process and device architecture; almost no head-to-head comparison against incumbent memory technologies.
Diffusion (2/5)
The most credible adoption path, HZO embedded memory, is explicitly held up by fabrication-induced performance variability, called a formidable barrier to high-volume manufacturing, with device-to-device non-uniformity making wafer-scale performance unpredictable. Using unsupervised learning to predict unseen dies is a mitigation, not a fix; the underlying cause is crystallisation kinetics that depend on oxygen vacancy concentration and annealing conditions and on pulse energy density thresholds tied to film thickness.
Other barriers are visible across the corpus. Several effects are not room-temperature or not robust: charged domain wall conductivity in BaTiO3 falls by orders of magnitude below about 5 C as the wall fragments into micron-scale segments at the tetragonal-orthorhombic transition, and oxygen-vacancy-doped EuTiO3-delta reaches a maximum Curie temperature of only about 11 K at n approximately 10^21 cm^-3. Van der Waals devices are exfoliated-stack demonstrations with no manufacturability evidence here. And the measurement-artefact problem plus the strained-SrTiO3 reassessment mean adopters cannot take literature claims at face value without independent verification, which itself slows transfer.
TLDR: Manufacturing variability is the acknowledged blocker for the near-term HZO path, and the exotic branches are single-flake demonstrations.
Impact (4/5)
The functional breadth here is unusual for one material class. In memory, HZO targets embedded non-volatile and computational memory directly in the CMOS back end, and vdW junction transistor memory demonstrates a 1.8 V window with picoamp leakage. In magnetoelectronics, electric-field control of the anomalous Hall effect has been shown at room temperature in Py/BaTiO3/La0.7Sr0.3MnO3 heterostructures on MgO and LaAlO3, and polar distortions have been observed coexisting with altermagnetic order in MnTe at atomic resolution, both routes to voltage-controlled spintronics. In photonics, NbOI2 offers record optical nonlinearities and dielectric anisotropy with sub-micrometre spin-orbit control for on-chip polarisation handling and beam steering. Reconfigurable conductive channels written optically into an insulator add a further device class.
The score is 4 rather than 5 because none of these sources quantifies economic value, addressable market or system-level gain. The judgement rests on the number of independent, room-temperature-capable device functions demonstrated from one materials family, not on any financial evidence.
TLDR: If the manufacturing problem is solved, ferroelectrics touch memory, logic-adjacent compute, photonics and magnetoelectronics from a single materials base.
Timing Soon (2-5yr)
The HZO work in this corpus is not asking whether the material is ferroelectric; it is asking how to make it uniform across a wafer and how to crystallise it within a BEOL thermal budget. That is a late-stage problem, which places the embedded-memory application in the two-to-five-year band on process maturity grounds, though these sources give no product timeline.
Everything else is earlier. The 2D multiferroic candidates are computational predictions awaiting synthesis, the newly identified quantum ferroelectrics are screening hits selected for experimental viability rather than demonstrated devices, and domain wall electronics still loses its conductive channel across an ordinary phase transition near 5 C. For context on how slow this can be: room-temperature out-of-plane ferroelectricity in compressively strained SrTiO3 was predicted 25 years ago and is still not established.
TLDR: HZO is at the manufacturing-yield stage now; the van der Waals and quantum-ferroelectric branches are a decade out.
Overrated or underrated? Fairly rated
Ferroelectrics are neither hype nor a lab curiosity. The hafnia branch has a specific, industrially legible problem (crystallisation-kinetics-driven device-to-device variability) and specific, industrially legible attempts to solve it (defect-chemistry control of the orthorhombic phase, threshold-mapped laser annealing, statistical screening of dies). That is what a technology looks like when it is close but not yet clean.
Where expectations should be discounted is the exploratory tail. The corpus itself supplies the discount: series-capacitance artefacts can manufacture convincing hysteresis loops that appear in published work, and a flagship strained-perovskite target turns out to show polarity from inhomogeneous internal fields rather than a stable switchable state. Treat any new claimed ferroelectric, especially thin-film or 2D, as unconfirmed until strain-based (S-E) or structural evidence accompanies the polarisation loop. Correspondingly, the ML-screened multiferroic and quantum-ferroelectric candidates should be read as a widened search space, not as devices.
Prediction
By 31 December 2028, no independently reproduced demonstration of stable room-temperature out-of-plane ferroelectric switching in compressively strained SrTiO3 capacitors will have been published, and published wafer-scale HZO capacitor studies will still report device-to-device variability large enough to require statistical or machine-learning screening rather than deterministic process control.
Evidence base
- HZO is positioned as a strong embedded non-volatile memory candidate on BEOL compatibility and scalability, but sensitive crystallisation kinetics produce device-to-device non-uniformity described as a formidable barrier to high-volume manufacturing; PCA and K-means clustering are used to predict performance of unseen dies (4 May 2026).
- Visible nanosecond laser pulses crystallise HZO in a Si3N4/TiN/HZO stack that mimics CMOS device geometry, with a sharp threshold in pulse energy density and phase fraction dependent on film thickness; prior work used UV or IR light (5 May 2026).
- In ~7 nm Hf0.5Zr0.5O2 nanoparticles the fraction of orthorhombic phase depends on oxygen vacancy concentration set by annealing atmosphere, with Landau-Ginzburg-Devonshire calculations attributing stabilisation to vacancy-induced chemical strain (5 May 2026).
- A SnSe / alpha-In2Se3 van der Waals p-n heterojunction memory shows 0.1 pA reverse leakage, diode ideality factor 1.95, 900 meV ferroelectric barrier modulation and a 1.8 V memory window (8 May 2026).
- Series non-ferroelectric capacitance can produce apparent polarisation hysteresis loops even under the double-wave method, and such suspicious loops are identified in published papers; a commercial ceramic capacitor’s D-E loops are shown to be apparent rather than ferroelectric (14 May 2026).
- Compressively strained SrTiO3 capacitors at up to -3% strain, with tetragonality comparable to bulk PbTiO3, show butterfly capacitance-voltage hysteresis and domain-like piezoresponse yet behaviour inconsistent with a stable ferroelectric state, 25 years after the effect was predicted (29 May 2026).
- Room-temperature electric-field control of the anomalous Hall effect is demonstrated in epitaxial Py/BaTiO3/La0.7Sr0.3MnO3 heterostructures on MgO and LaAlO3 substrates (2 June 2026).
Open questions
- Can HZO device-to-device variability be reduced by direct process control (oxygen partial pressure, laser energy density window, film thickness uniformity) to the point that per-die statistical screening becomes unnecessary, or is screening a permanent feature of the technology?
- Do the ML-screened candidates, in particular AuCrP2S6 with its 7.46 pC/m polarisation and ~130 meV/f.u. switching barrier, survive attempted synthesis and show the predicted four-state behaviour with bulk-photovoltaic readout?
- How many published thin-film and 2D ferroelectricity claims survive re-measurement using piezoelectric strain loops rather than polarisation loops alone, given that series capacitance can mimic hysteresis even under the double-wave method?
- Can charged domain wall conductivity in a perovskite be made robust across phase transitions and ordinary temperature swings, or does the fragmentation observed at the tetragonal-orthorhombic transition rule out domain wall interconnect?
Assessment drafted 2026-08-31 from up to 18 KB sources using the technology-scorecard framework; scores are a draft read pending review.