Insulators that do something
Most of the electronics in a phone is not semiconductor at all. Next to every chip sit dozens of tiny ceramic capacitors, around a thousand in a modern smartphone and several thousand in a car, made of hundreds of layers of a barium titanate ceramic a micrometre thick. The phone's radio filters are thin films of aluminium nitride that turn electrical signals into sound waves and back. A watch keeps time because a sliver of quartz vibrates 32 768 times a second when an electric field pushes it. A hospital ultrasound probe sends sound into the body from an array of lead zirconate titanate elements. And inside every transistor of the chip, the insulating layer that lets the gate control the channel is now hafnium oxide, a material few people had heard of in 2000.
All of these are dielectrics: insulators whose interest lies in how their charges shift, without flowing, when an electric field is applied. This lesson follows that shift from its four microscopic mechanisms to the capacitors it makes possible; then to the crystals whose shape changes when they polarise (piezoelectrics) and those whose polarisation persists and can be switched (ferroelectrics); and finally to the gate dielectrics where materials science and transistor scaling meet.
How an insulator polarises
Put an insulator between two charged plates. No current flows, but the charges inside shift slightly: positive ones towards the negative plate, negative ones towards the positive. Each shift makes a tiny dipole, and the dipoles together partly cancel the field. The material is polarised. There are four ways it can happen, and they differ enormously in speed.
- Electronic polarisation. Every atom's electron cloud shifts slightly relative to its nucleus. Electrons are light, so this responds up to ultraviolet frequencies, around Hz. Every material has it; in diamond and silicon it is the only mechanism.
- Ionic polarisation. In an ionic crystal, cations and anions are displaced in opposite directions, stretching the bonds. Ions are thousands of times heavier than electrons, so this follows the field only up to infrared frequencies of about Hz, where the lattice vibrates.
- Orientational polarisation. Molecules with permanent dipoles, such as water, rotate towards the field against the jostling of thermal motion. Rotating a whole molecule is slow: in liquid water it takes about ten picoseconds, so the effect fades in the microwave range.
- Interfacial (space-charge) polarisation. Mobile ions or electrons pile up at grain boundaries, electrodes or phase boundaries. It can be huge but is slow, typically gone above kilohertz to megahertz frequencies.
Permittivity, loss and the symmetry of crystals
Definition— polarisation, susceptibility, relative permittivity
The polarisation is the electric dipole moment per unit volume, in C/m². In a linear, isotropic dielectric , where is the electric susceptibility. The electric displacement is , with relative permittivity (dielectric constant) .
Definition— complex permittivity, loss tangent
Under an alternating field the polarisation lags behind, which is written as a complex permittivity . The real part stores energy; the imaginary part dissipates it as heat. The loss tangent is .
Definition— Debye relaxation
Dipoles that relax towards equilibrium with a single time constant give
where is the static value and the value after the dipoles have dropped out. The loss peaks at , where it equals .
Definition— piezoelectric, pyroelectric, ferroelectric
A piezoelectric material polarises when stressed (the direct effect) and strains when an electric field is applied (the converse effect), linearly: and , with piezoelectric coefficients in pC/N (equivalently pm/V). A pyroelectric material has a spontaneous polarisation that changes with temperature. A ferroelectric is a pyroelectric whose spontaneous polarisation can be reversed by an applied field. Above its Curie temperature a ferroelectric becomes paraelectric.
These definitions are nested, and symmetry decides which a crystal can be. If a crystal has a centre of symmetry, inverting every coordinate leaves it unchanged; but inversion reverses a polarisation while leaving a stress unchanged, so a linear coupling between them must vanish. Of the 32 crystallographic point groups, 21 lack a centre of symmetry, and 20 of those are piezoelectric (the cubic group 432 is the exception, by accident of its other symmetries). Ten of the 20 are polar, with a unique direction that a spontaneous polarisation can point along: these are pyroelectric. Ferroelectrics are those polar crystals in which the polarisation can be switched, which is a matter of energy, not symmetry alone.
Polarisation, computed
The program does two things. First it splits the static permittivity of several solids and of water into its electronic part, from the optical refractive index, and the rest. Then it computes the Debye relaxation of water, whose molecules' rotation dominates its permittivity, and how deep microwaves penetrate it at frequencies used for heating.
import numpy as np
# 1. How much of the static permittivity is electronic? At optical frequencies only the electrons
# can follow the field, so eps_optical = n^2. Rounded room-temperature literature values.
data = [("diamond", 2.42, 5.7), ("silicon (n at 1.55 um)", 3.48, 11.7), ("silica glass", 1.458, 3.8),
("sodium chloride", 1.544, 5.9), ("magnesium oxide", 1.736, 9.8), ("water", 1.333, 78.4),
("barium titanate ceramic", 2.4, 2000)]
print("material n n^2 static eps_r electronic share")
for name, n, es in data:
print(f"{name:24} {n:5.3f} {n * n:5.2f} {es:9.1f} {100 * min(n * n / es, 1):5.1f} %")
# 2. Orientational polarisation of water: a Debye relaxation, eps = eps_inf + (eps_s - eps_inf) / (1 + i w tau)
eps_s, eps_inf, tau = 78.4, 5.2, 8.3e-12 # water at 25 C, rounded literature values
c = 2.998e8
print(f"\nwater: loss peak at f = 1/(2 pi tau) = {1 / (2 * np.pi * tau) / 1e9:.1f} GHz")
print(" f/GHz eps' eps'' power penetration depth")
for f in [0.0001, 0.915, 2.45, 5.8, 19.2, 100, 1000]:
w = 2 * np.pi * f * 1e9
eps = eps_inf + (eps_s - eps_inf) / (1 + 1j * w * tau)
e1, e2 = eps.real, -eps.imag
n_complex = np.sqrt(e1 - 1j * e2) # complex refractive index n - i kappa
kappa = -n_complex.imag
depth = c / (2 * w * kappa) if kappa > 0 else float("inf") # intensity falls by e over this distance
depth_text = f"{depth * 1000:9.3g} mm" if depth < 1e3 else " very large"
print(f"{f:9.4f} {e1:7.2f} {e2:7.2f} {depth_text}")Diamond and silicon are entirely electronic: matches the static permittivity. Silica, salt and magnesia get half or more of their permittivity from ions moving. Water's permittivity of 78 is 98 per cent orientational, and barium titanate's thousands come almost entirely from the ions of its ferroelectric lattice, as we will see.
Water's loss peaks at 19.2 GHz. A domestic microwave oven runs at 2.45 GHz, well below the peak, and that is a good compromise: at the peak, the power would be absorbed within half a millimetre of the surface, and food would burn outside while staying cold inside. At 2.45 GHz the computed penetration depth is about 19 mm, and industrial heaters at 915 MHz reach further still.
Permittivity and capacitor dielectrics
A parallel-plate capacitor of area and dielectric thickness has capacitance and stores energy . To pack more capacitance into less volume, increase , make the layer thinner, and stack many layers in parallel. Each choice has a price: a thinner layer sees a higher field for the same voltage and may break down, and very high permittivities usually come from ferroelectric materials whose permittivity changes with temperature, voltage and age.
Example— How many layers does a 10 µF ceramic capacitor need?
A common multilayer ceramic capacitor (MLCC) measures 1.6 mm × 0.8 mm (the "0603" size). Suppose the overlapping area of its interleaved nickel electrodes is 1.2 mm × 0.6 mm, the barium titanate layers between them are 1.0 µm thick, and their relative permittivity is 3000. One layer gives
Layers connect in parallel, so 10 µF needs about 523 layers, each with its electrode, stacked in a part less than a millimetre high. The same 10 µF with polypropylene film () 5 µm thick would need a quarter of a square metre of film for every microfarad, rolled into a cylinder, which is why film capacitors are large but are chosen for high voltages and low losses.
Engineers sort capacitor dielectrics by what they trade away.
| Dielectric | (approx.) | Strengths | Weaknesses |
|---|---|---|---|
| Polypropylene film | 2.2 | very low loss, high breakdown field, self-healing | bulky |
| Calcium zirconate-based ceramic ("C0G", Class 1) | 20–100 | stable with temperature and voltage | modest capacitance |
| Barium titanate-based ceramic (X7R, X5R; Class 2) | 1000–5000 | huge capacitance per volume | capacitance falls with DC bias, temperature and time |
| Anodic Al₂O₃ (electrolytic) | about 9 | very thin oxide grown electrochemically, large area | polarised; liquid electrolyte ages |
| Anodic Ta₂O₅ (tantalum) | about 27 | compact, stable | polarised; failure can be violent if misused |
The Class 2 ceramics' weaknesses all have one cause: barium titanate is ferroelectric, and its enormous permittivity comes from domains that respond easily to small fields. A DC voltage across the capacitor pins those domains, and the small-signal capacitance can fall to a fraction of its nominal value. Designers read the manufacturer's DC-bias curves, not just the number printed on the reel. Large capacitors can also hold a dangerous charge long after the power is removed, which is why equipment with them includes bleed resistors and why technicians are trained to discharge them before working.
Piezoelectricity: sensors, actuators and ultrasound
Pierre and Jacques Curie discovered piezoelectricity in 1880 by squeezing crystals of quartz, tourmaline and Rochelle salt and measuring the charge on their faces. The first practical use came during the First World War, when Paul Langevin built quartz ultrasonic transmitters to detect submarines. Today the main families are single crystals (quartz, lithium niobate), ceramics (lead zirconate titanate, PZT), polymers (polyvinylidene fluoride, PVDF) and thin films (aluminium nitride, AlN, and its scandium-alloyed form).
A piezoelectric element is also a capacitor, which explains a surprise in the numbers below. The program squeezes discs of four materials with the same force, then computes a stack actuator's stroke and the design of a medical ultrasound transducer.
import numpy as np
eps0 = 8.854e-12
# 1. Direct effect: squeeze a disc 10 mm across and 1 mm thick with 10 N along its axis.
# Charge Q = d33 F; the disc is a capacitor C = eps0 eps_r A / t, so the open-circuit voltage is Q / C.
# Rounded, representative coefficients: d (pC/N) and relative permittivity.
A, t, F = np.pi * 0.005**2, 1e-3, 10.0
print("material d/(pC/N) eps_r charge/nC C/pF voltage/V g = d/(eps0 eps_r) in V m/N")
for name, d, er in [("quartz (d11)", 2.3, 4.5), ("PVDF film", 25, 12), ("aluminium nitride", 5.5, 9),
("soft PZT ceramic", 400, 1800)]:
Q = d * 1e-12 * F
C = eps0 * er * A / t
print(f"{name:18} {d:8.1f} {er:6.1f} {Q * 1e9:8.3f} {C * 1e12:7.1f} {Q / C:8.2f} {d * 1e-12 / (eps0 * er):.4f}")
# 2. Converse effect: a multilayer PZT stack actuator, 100 layers each 0.1 mm thick, driven at 100 V
d33 = 400e-12
print(f"\nstack of 100 layers at 100 V: extension {100 * d33 * 100 * 1e6:.1f} um "
f"(field per layer {100 / 0.1e-3 / 1e6:.1f} MV/m)")
# 3. A medical ultrasound transducer: a PZT plate resonates in thickness when t = half a wavelength
v_pzt, rho_pzt = 4350.0, 7750.0 # longitudinal sound speed (m/s) and density (kg/m^3), representative
Z_pzt = rho_pzt * v_pzt / 1e6 # acoustic impedance in MRayl
Z_tissue = 1.54 # soft tissue, MRayl
for f0 in [2.5e6, 5e6, 10e6]:
print(f"{f0 / 1e6:4.1f} MHz needs a plate {v_pzt / (2 * f0) * 1e3:.3f} mm thick")
T_direct = 4 * Z_pzt * Z_tissue / (Z_pzt + Z_tissue) ** 2
print(f"\nZ(PZT) = {Z_pzt:.1f} MRayl, Z(tissue) = {Z_tissue} MRayl")
print(f"energy transmitted straight from PZT into tissue: {100 * T_direct:.1f} %")
# A quarter-wave matching layer of impedance sqrt(Z_pzt Z_tissue) transforms the load, like an
# antireflection coating. Transmission against frequency, from the transmission-line formula.
Zm = np.sqrt(Z_pzt * Z_tissue)
print(f"ideal matching layer impedance: {Zm:.2f} MRayl")
print(" f/f0 transmitted (with layer)")
for r in [0.5, 0.7, 0.85, 1.0, 1.15, 1.3, 1.5]:
kd = np.pi / 2 * r # phase thickness of the layer at this frequency
Zin = Zm * (Z_tissue + 1j * Zm * np.tan(kd)) / (Zm + 1j * Z_tissue * np.tan(kd))
R = abs((Zin - Z_pzt) / (Zin + Z_pzt)) ** 2
print(f" {r:4.2f} {100 * (1 - R):5.1f} %")The surprise is in the voltage column. PZT produces 170 times more charge than quartz for the same force (4 nC against 0.023 nC), yet a lower open-circuit voltage, 3.2 V against 7.4 V, because its permittivity is 400 times larger and its capacitance correspondingly bigger. Which matters depends on the job. A charge amplifier reading an accelerometer wants charge, so PZT wins. A sensor read by a voltage amplifier, or a flexible PVDF film detecting a footstep, wants a large voltage coefficient , where PVDF's 0.24 V m/N beats PZT's 0.025 tenfold. Quartz is chosen for something else entirely: its coupling is weak, but its elastic properties hardly change with temperature in certain crystal cuts, which makes it the standard frequency reference for watches and electronics.
The converse effect gives motion measured in nanometres per volt: a 100-layer stack moves 4 µm at 100 V, with great force and in microseconds. Such stacks drive fuel injectors, inkjet print heads, the scanners of atomic force microscopes and the fine-positioning stages of chip lithography.
Ultrasound. A PZT plate rings at the frequency where its thickness is half a wavelength of sound, so a 5 MHz probe uses plates 0.44 mm thick, cut into hundreds of narrow elements that are fired with small time delays to steer and focus the beam. The problem is acoustic impedance : PZT's is about 34 MRayl, soft tissue's 1.5, and a sound wave crossing such a mismatch is mostly reflected, so only 17 per cent of the energy would get in. The cure is the acoustic version of an anti-reflection coating: a layer a quarter of a wavelength thick with impedance MRayl, which transmits everything at the design frequency and more than three-quarters within ±15 per cent of it. Real probes use two or three matching layers, of particle-filled epoxies, to widen the band.
Ferroelectrics and hysteresis
In 1920 Joseph Valasek at the University of Minnesota found that Rochelle salt's polarisation, plotted against the applied field, traced a loop just like the magnetisation of iron, and the name "ferroelectric" followed by analogy (no iron is involved). The breakthrough material came during the Second World War, when groups in the United States, Britain, Japan and the Soviet Union independently found the extraordinary permittivity of barium titanate, BaTiO₃.
Barium titanate has the perovskite structure: barium ions at the corners of a cube, oxygen ions at the face centres, and a small titanium ion in the octahedral cage at the centre. Above 120 °C or so the cube is perfect and the crystal is paraelectric. Below it, the titanium ions shift off-centre by about a hundredth of a nanometre, all in the same direction within a region, the cell stretches slightly into a tetragonal shape, and each cell carries a dipole: a spontaneous polarisation of about 0.26 C/m² (26 µC/cm²) at room temperature.
Above the Curie temperature the permittivity follows the Curie–Weiss law, , with a Curie constant K for barium titanate: 50 K above the permittivity is about 3000, and it grows without limit as the transition approaches. The crystal is becoming "soft" against the displacement that is about to freeze in, and that softness is where the high permittivity of capacitor ceramics comes from.
Landau theory describes this with a free energy expanded in the polarisation, , where changes sign at . Below the energy has two wells, at , and a uniform crystal can only switch when the applied field tilts the landscape so far that its well disappears. The first part of the program computes that intrinsic coercive field for barium titanate. The second simulates what real materials do: switch region by region. It treats a film as many small regions, each flipping up at its own field and down at its own field , with the switching fields spread around an average, which is the Preisach model of hysteresis. The numbers are illustrative values of the size measured in lead zirconate titanate films.
import numpy as np
# 1. Landau theory of a uniform (single-domain) ferroelectric below its Curie temperature.
# Free energy per volume F = a P^2 / 2 + b P^4 / 4 - E P with a < 0. Equilibrium: E = a P + b P^3.
# The polarisation can only flip when its minimum disappears, at E = (2 / 3 sqrt 3) |a| Ps.
eps0 = 8.854e-12
C_curie = 1.5e5 # Curie constant of barium titanate, K (rounded)
Ps = 0.26 # spontaneous polarisation of barium titanate at room temperature, C/m^2
for dT in [10, 50, 100]:
a = -dT / (eps0 * C_curie) # a = (T - T0) / (eps0 C), in m/F
Ec = 2 / (3 * np.sqrt(3)) * abs(a) * Ps
print(f"T0 - T = {dT:3} K: Landau coercive field {Ec / 1e5:6.1f} kV/cm")
print("measured in barium titanate crystals: of order 1 kV/cm\n")
def trap(y, x):
"""Trapezoid-rule integral of y over x."""
return float(np.sum((y[1:] + y[:-1]) * np.diff(x)) / 2)
# 2. A real film or ceramic switches region by region. Model it as many small regions ("hysterons"),
# each flipping up at a field alpha and down at a field beta < alpha: a Preisach model.
# Illustrative values of the size found in lead zirconate titanate films.
Ps_film = 30.0 # uC/cm^2
Ec_mean = 50.0 # kV/cm
chi_lin = 0.04 # reversible (non-switching) response, uC/cm^2 per kV/cm (eps_r about 450)
hc = np.linspace(0.3, 1.7, 29) * Ec_mean # half-width of each hysteron
hb = np.linspace(-1.2, 1.2, 25) * Ec_mean # its bias (internal field from defects)
HC, HB = np.meshgrid(hc, hb)
weight = np.exp(-((HC - Ec_mean) / (0.3 * Ec_mean)) ** 2 / 2) * np.exp(-(HB / (0.4 * Ec_mean)) ** 2 / 2)
weight /= weight.sum()
alpha, beta = HB + HC, HB - HC
state = -np.ones_like(HC) # start fully poled down
E_path = np.concatenate([np.linspace(0, 150, 151), np.linspace(150, -150, 301), np.linspace(-150, 150, 301)])
P_path = []
for E in E_path:
state[E >= alpha] = 1.0
state[E <= beta] = -1.0
P_path.append(Ps_film * np.sum(weight * state) + chi_lin * E)
P_path = np.array(P_path)
# Read the loop from its last full sweep (the second half of the path, -150 -> +150 and the branch before it)
down, up = slice(151, 452), slice(452, 753)
Pr_up = np.interp(0, E_path[up], P_path[up])
Pr_down = np.interp(0, E_path[down][::-1], P_path[down][::-1])
Ec_plus = np.interp(0, P_path[up], E_path[up])
Ec_minus = np.interp(0, P_path[down][::-1], E_path[down][::-1])
area = -(trap(P_path[down], E_path[down]) + trap(P_path[up], E_path[up])) # loop area, uC/cm^2 x kV/cm
energy = area * 1e-2 * 1e5 # J/m^3 per cycle (uC/cm^2 = 1e-2 C/m^2, kV/cm = 1e5 V/m)
print(f"remanent polarisation: +{Pr_down:.1f} / {Pr_up:.1f} uC/cm^2")
print(f"coercive field: {Ec_minus:.1f} / +{Ec_plus:.1f} kV/cm")
print(f"polarisation at +150 kV/cm: {P_path[150]:.1f} uC/cm^2")
print(f"energy lost per cycle (loop area): {energy / 1e6:.2f} MJ/m^3")
print(f" for a memory capacitor 1 um x 1 um x 100 nm: {energy * 1e-19 * 1e12:.2f} pJ per full cycle")
# ASCII P-E loop: rows are P from +40 down to -40 uC/cm^2, columns are E from -150 to +150 kV/cm
rows, cols = 13, 41
grid = [[" "] * cols for _ in range(rows)]
for E, P in zip(E_path[151:], P_path[151:]):
c = int(round((E + 150) / 300 * (cols - 1)))
r = int(round((40 - P) / 80 * (rows - 1)))
if 0 <= r < rows:
grid[r][c] = "*"
for r in range(rows):
if grid[r][cols // 2] == " ":
grid[r][cols // 2] = "|"
print("\n" + "\n".join("".join(row) for row in grid))
print("-150 kV/cm" + " " * 21 + "+150 kV/cm")The loop shows the three numbers that characterise a ferroelectric: the remanent polarisation (about 29 µC/cm² here) that stays when the field is removed, which is what makes ferroelectrics memories; the coercive field (about 48 kV/cm) needed to reverse it; and the loop area, 6 MJ/m³ per cycle here, which is lost as heat and limits how fast and how often bulk ferroelectrics can be driven. For a memory capacitor one micrometre square and 100 nm thick, a full cycle costs about 0.6 pJ.
Domains and poling. A ferroelectric crystal or grain splits into domains of different polarisation directions, which reduces the electrostatic energy of its surfaces, just as magnets form magnetic domains (Magnetic materials). A freshly sintered PZT ceramic, with its randomly oriented grains and domains, has no net polarisation and is not piezoelectric. It becomes useful only after poling: a strong field applied at elevated temperature aligns the domains as far as the grains allow, and the material keeps that alignment. Heating a poled part above its Curie temperature, or too close to it, undoes the poling, which is why piezoelectric sensors carry maximum temperature ratings well below .
Measuring a loop is simple in principle. Charles Sawyer and Charles Tower's circuit of 1930 puts a large reference capacitor in series with the sample: the reference's voltage is proportional to the charge, and so to , while the drive voltage gives . Modern ferroelectric testers integrate the current instead, with the same idea.
High-k dielectrics in transistors
A transistor's gate controls the channel through a thin insulator, and the charge it can induce is the capacitance per area, . For forty years that insulator was silicon dioxide, grown by oxidising the silicon itself, with a nearly perfect interface. As transistors shrank, the oxide had to thin with them, reaching about 1.2 nm, four or five atomic layers, by the mid-2000s. Then quantum mechanics intervened: electrons tunnel straight through a barrier that thin, and the gate leakage grows roughly tenfold for every 0.2 nm removed.
The way out was to keep the capacitance and make the layer physically thicker, with a higher permittivity. Engineers compare dielectrics by their equivalent oxide thickness, : the thickness of SiO₂ that would give the same capacitance. But permittivity is not the only requirement. As John Robertson summarised in 2004, the material must also be stable in contact with silicon at processing temperatures, form a clean interface, and, critically, present a barrier of more than about 1 eV to electrons and holes in the silicon, or they will simply flow over it. Permittivity and band gap pull in opposite directions: the large permittivities come from soft, polarisable lattices, which tend to have small gaps.
The program compares candidate oxides at the same EOT of 1.0 nm using the simplest tunnelling estimate, transmission with through a rectangular barrier of height , taking the same tunnelling mass for every oxide.
import numpy as np
hbar, me, q = 1.054571817e-34, 9.1093837015e-31, 1.602176634e-19
# Approximate values compiled by J. Robertson (2004): relative permittivity, band gap (eV),
# conduction-band offset to silicon (eV), i.e. the barrier an electron must tunnel through.
oxides = {"SiO2": (3.9, 9.0, 3.2), "Al2O3": (9, 8.8, 2.8), "HfO2": (25, 5.8, 1.4),
"ZrO2": (25, 5.8, 1.5), "La2O3": (30, 6.0, 2.3), "Ta2O5": (22, 4.4, 0.35),
"TiO2": (80, 3.5, 0.0)}
EOT = 1.0e-9 # target: the capacitance of 1.0 nm of SiO2
m_eff = 0.3 # a single tunnelling mass for every oxide: a crude simplification
def wkb_exponent(barrier_eV, thickness):
"""2 kappa t for a rectangular barrier; transmission ~ exp(-2 kappa t)."""
kappa = np.sqrt(2 * m_eff * me * barrier_eV * q) / hbar
return 2 * kappa * thickness
print(f"every film below has the capacitance of {EOT * 1e9:.1f} nm of SiO2: "
f"{3.9 * 8.854e-12 / EOT * 100:.2f} uF/cm^2")
ref = wkb_exponent(3.2, EOT)
print("oxide eps_r gap/eV barrier/eV thickness/nm tunnelling relative to SiO2")
for name, (er, gap, cbo) in oxides.items():
t = EOT * er / 3.9 # physical thickness with the same capacitance
if cbo <= 0:
print(f"{name:6} {er:5.1f} {gap:5.1f} {cbo:5.2f} {t * 1e9:6.2f} no barrier: electrons flow straight over")
continue
rel = np.exp(-(wkb_exponent(cbo, t) - ref))
print(f"{name:6} {er:5.1f} {gap:5.1f} {cbo:5.2f} {t * 1e9:6.2f} {rel:9.1e}")
# Real stacks keep a thin SiO2 layer next to the silicon. Capacitances in series add like resistors in
# parallel, so EOTs simply add; WKB exponents through successive layers add too.
print("\nSiO2 interlayer + HfO2 EOT/nm tunnelling relative to 1.0 nm SiO2")
for t_il, t_hk in [(0.5e-9, 2.0e-9), (0.5e-9, 3.2e-9), (0.8e-9, 1.5e-9)]:
eot = t_il + t_hk * 3.9 / 25
rel = np.exp(-(wkb_exponent(3.2, t_il) + wkb_exponent(1.4, t_hk) - ref))
print(f" {t_il * 1e9:3.1f} nm + {t_hk * 1e9:3.1f} nm {eot * 1e9:5.2f} {rel:9.1e}")The table reads like a selection process. Alumina is stable and has a high barrier, but its permittivity of 9 buys little thickness. Tantalum oxide's barrier for electrons is only 0.35 eV, and titanium dioxide, with , has none at all: its conduction band lines up with silicon's, so electrons would pour over it. Lanthanum oxide looks best on paper but absorbs water and reacts with silicon. Hafnium and zirconium oxides, with permittivities around 20–25 and barriers of about 1.5 eV, are the compromise, and hafnium oxide also survives the high-temperature steps of chip making in contact with silicon better than zirconium oxide. Even with a 0.5 nm silicon dioxide layer kept at the interface for its clean bonding to silicon, a stack with the same capacitance as 1.0 nm of SiO₂ leaks orders of magnitude less.
Treat the tunnelling ratios as directions, not values: a single tunnelling mass, sharp barriers and no defects make the predicted gains far larger than anything measured, because real stacks also leak through traps in the film. What the estimate gets right is the ranking, and why a high barrier counts as much as a high permittivity.
Intel introduced hafnium-based high-k dielectrics with metal gates in its 45 nm process in 2007, and the rest of the industry followed; the metal gate was needed because polysilicon gates interact badly with hafnium oxide. Hafnium-based stacks are now in essentially every advanced logic transistor, including the fin and gate-all-around devices of Lumenary EE's transistor scaling.
Then came a surprise. In 2011 Tim Böscke and colleagues at Qimonda and NaMLab in Dresden reported that thin films of hafnium oxide doped with silicon could be ferroelectric, in a non-centrosymmetric orthorhombic phase that forms in films a few nanometres to tens of nanometres thick. A material already qualified for every fab could now store a bit in its polarisation. As of 2026 ferroelectric hafnia (often alloyed with zirconium) is being developed for ferroelectric field-effect transistors and capacitors, embedded memory, and in-memory computing, with the open questions being endurance, wake-up and imprint effects, and variability in tiny devices.
The professional view
Making MLCCs. A slurry of fine barium titanate powder is cast into tapes a micrometre or so thick; nickel electrode patterns are screen-printed on them; hundreds of sheets are stacked, laminated, diced and fired so that ceramic and metal sinter together. Because nickel would oxidise in air, firing is done in a reducing atmosphere, and the titanate is doped (for example with rare-earth and manganese oxides) so that it stays insulating after the oxygen-poor firing, a piece of defect chemistry that made cheap base-metal MLCCs possible. Major producers include Murata, TDK and Taiyo Yuden in Japan, Samsung Electro-Mechanics in South Korea and Yageo in Taiwan. The furnaces and powders belong to industry; the processing principles are those of Processing and heat treatment.
Piezoelectrics. PZT contains lead, and lead-free alternatives (potassium sodium niobate, bismuth sodium titanate and barium titanate-based ceramics) are a major research and regulatory topic; as of 2026 PZT remains in use under exemptions in rules restricting hazardous substances, and lead-free ceramics are entering some applications. In phones, aluminium nitride and scandium-alloyed AlN films form the bulk acoustic wave filters that separate radio bands; thin-film lithium niobate is being developed for electro-optic modulators (Lumenary EE's photonics).
Dielectric data. Measured permittivities and losses across frequency come from impedance analysers at low frequencies, resonant cavities and network analysers at microwave frequencies, and optical methods above; the methods are standardised by bodies such as the IEC and ASTM. Computed dielectric and piezoelectric tensors for thousands of crystals are available from the Materials Project, with the caveat that density functional theory captures electronic and ionic contributions but not domain motion.
Common misconceptions
Misconception
"Ferroelectrics contain iron." The name is an analogy with ferromagnetism: both show hysteresis, domains and a Curie temperature. Barium titanate, PZT and hafnia contain no iron.
Misconception
"The dielectric constant is a constant." It depends on frequency (as mechanisms drop out), temperature (especially near a Curie temperature), applied DC field (strongly, in ferroelectric ceramics) and time (Class 2 capacitors lose a few per cent per decade of time as domains settle).
Misconception
"A higher permittivity always makes a better dielectric." A capacitor or gate dielectric also needs a high breakdown field, low loss, stability and, for gates, a large band offset. Titanium dioxide has a permittivity twenty times that of silica and is useless as a gate insulator on silicon.
Misconception
"Any crystal will generate a voltage if you squeeze it." Only crystals without a centre of symmetry can be piezoelectric. A ceramic of randomly oriented grains, even of a piezoelectric material, shows no net effect until it is poled.
Misconception
"Piezoelectric floor tiles could power a city." A footstep does only a few joules of mechanical work on a floor, and a piezoelectric tile converts a small fraction of that. Harvesters are useful for powering wireless sensors, not for supplying grids.
Exercises
1. Dielectric and ferroelectric materials
Warm-upHumanWhich of these, as described, will produce a net charge on their faces when squeezed? Select all that apply.
2. Dielectric and ferroelectric materials
CoreHumanA gate stack has 0.4 nm of SiO₂ () next to the silicon, then 3.0 nm of hafnium oxide with . What is its equivalent oxide thickness, in picometres (to the nearest whole number)?
3. Dielectric and ferroelectric materials
CoreHumanA PZT plate in which longitudinal sound travels at 4350 m/s is to resonate in its thickness mode at 7.5 MHz. How thick should it be, in micrometres?
4. Dielectric and ferroelectric materials
CoreHumanWater's dielectric loss peaks near 19 GHz. Why do microwave ovens heat food at 2.45 GHz instead?
5. Dielectric and ferroelectric materials
CoreHumanTitanium dioxide has a relative permittivity of about 80, more than three times hafnium oxide's. Why is it not used as the gate dielectric on silicon?
6. Dielectric and ferroelectric materials
CoreComputationalWrite debye(f, eps_s, eps_inf, tau) returning the complex relative permittivity of a single Debye relaxation at frequency f (in Hz), and loss_peak(tau) returning the frequency (in Hz) at which is largest.
Use with . With Python's complex numbers, has a negative imaginary part, so the result automatically has the form .
7. Dielectric and ferroelectric materials
StretchComputationalIn Landau theory a uniform ferroelectric in a field satisfies (with , below the Curie temperature). Write landau_ps_ec(a, b) returning the pair (Ps, Ec): the spontaneous polarisation (where , ) and the coercive field, the largest field the negatively polarised state can withstand, which is the local maximum of on the branch .
You may solve it with calculus or numerically; the tests accept either to a relative accuracy of .
8. Dielectric and ferroelectric materials
Warm-upHumanA freshly sintered PZT ceramic is not piezoelectric. What one-word name is given to the treatment that makes it piezoelectric, by applying a strong field at elevated temperature to align its domains?
Further reading and resources
- Piezo-electric and allied phenomena in Rochelle salt · J. Valasek · 1921Physical Review. The first report of ferroelectric hysteresis, in a crystal of Rochelle salt.
- High dielectric constant oxides · J. Robertson · 2004European Physical Journal Applied Physics. The requirements a gate dielectric must meet, with the permittivities, gaps and band offsets used in this lesson.
- A 45nm logic technology with high-k + metal gate transistors, strained silicon, 9 Cu interconnect layers, 193nm dry patterning, and 100% Pb-free packaging · K. Mistry and others (Intel) · 2007IEEE International Electron Devices Meeting. The first high-volume transistors with hafnium-based gate dielectrics.
- Ferroelectricity in hafnium oxide thin films · T. S. Böscke, J. Müller, D. Bräuhaus, U. Schröder and U. Böttger · 2011Applied Physics Letters. The discovery that thin doped hafnia, already in every chip, can be ferroelectric.
- IEEE Standard on Piezoelectricity (ANSI/IEEE Std 176) · IEEEThe conventions for piezoelectric coefficients and their measurement.
- DoITPoMS · University of CambridgeFree teaching packages, including dielectric materials, ferroelectrics and piezoelectric materials.