Skip to content
LumenaryAcademy
Functional 75 min

Semiconductor materials

Intrinsic and doped carriers computed against temperature, mobility, the purest material industry makes, and why power electronics and lighting moved to GaN and SiC.

Code tested · 6AI-draftedAwaiting review by a materials scientist

Your guides: Chloe · The Materials Scientist, Margaret · The Tester, Gideon · The Failure Analyst

MotivationH

One atom in ten million

A silicon wafer for a microprocessor starts as quartz sand and ends as the most perfect large object people make. It is a single crystal 300 mm across, without a single dislocation, and its unintended impurities are counted in parts per trillion. Then, deliberately, impurities are added back: a typical starting wafer has about one boron atom for every 37 million silicon atoms, and the transistors built on it have regions with one phosphorus or arsenic atom per hundred. Those tiny additions change the conductivity by factors of a million.

That is the defining trick of semiconductors. In Electronic structure and bands we saw that a semiconductor has a full valence band, an empty conduction band and a gap small enough that heat lifts a few electrons across it. Because the number of those intrinsic carriers is so small, a few added impurity atoms can supply far more, and decide whether the current is carried by electrons or by holes. This lesson does the counting properly, as a function of temperature; follows silicon from sand to wafer; and explains why the newer power electronics in electric cars and data centres, and almost all lighting, have moved from silicon to compound semiconductors such as silicon carbide and gallium nitride.

The devices themselves, diodes and transistors, belong to Lumenary EE's semiconductors and diodes and transistors; the chemistry of the group-13, -14 and -15 elements is in the solid state and main-group chemistry. Here we look at the materials.

ComputationC

Intrinsic carriers

In a pure semiconductor every electron in the conduction band leaves a hole in the valence band, so the electron and hole densities are equal: n=p=nin = p = n_i, the intrinsic carrier density. To compute it, count states and multiply by their occupation.

Near the bottom of the conduction band the band looks like free electrons with an effective mass me∗m_e^*, with the E\sqrt E density of states. With the Fermi level several kBTk_BT below the band, the Fermi–Dirac distribution is close to the Boltzmann factor, and integrating gives

n=Nc e−(Ec−EF)/kBT,p=Nv e−(EF−Ev)/kBT,Nc=2(2πme∗kBTh2)3/2,n = N_c\,e^{-(E_c - E_F)/k_BT}, \qquad p = N_v\,e^{-(E_F - E_v)/k_BT}, \qquad N_{c} = 2\left(\frac{2\pi m_{e}^*k_BT}{h^2}\right)^{3/2},

and similarly for NvN_v with the hole mass. The effective densities of states NcN_c and NvN_v are a few times 101910^{19} cm⁻³ for silicon at 300 K. Multiplying nn and pp eliminates the unknown Fermi level:

np=NcNv e−Eg/kBT=ni2,ni=NcNv e−Eg/2kBT.np = N_cN_v\,e^{-E_g/k_BT} = n_i^2, \qquad n_i = \sqrt{N_cN_v}\,e^{-E_g/2k_BT}.

The product np=ni2np = n_i^2 holds whether or not the material is doped, as long as it is in equilibrium; it is the semiconductor version of the law of mass action. The program evaluates nin_i for five semiconductors over a range of temperatures.

python
import numpy as np

kB = 8.617333e-5            # Boltzmann constant, eV/K
N300 = 2.5094e19            # 2 (2 pi m_e kB T / h^2)^(3/2) at 300 K, in cm^-3

def n_intrinsic(Eg, me_dos, mh_dos, T):
    """Intrinsic carrier density (cm^-3): n_i = sqrt(Nc Nv) exp(-Eg / 2kT), Boltzmann statistics."""
    Nc = N300 * me_dos**1.5 * (T / 300) ** 1.5
    Nv = N300 * mh_dos**1.5 * (T / 300) ** 1.5
    return np.sqrt(Nc * Nv) * np.exp(-Eg / (2 * kB * T))

# Room-temperature gaps (eV) and density-of-states effective masses (units of m_e), rounded literature values;
# the masses of the wide-gap materials are less certain, which hardly matters next to the exponential.
materials = {"Ge": (0.66, 0.55, 0.37), "Si": (1.12, 1.09, 1.15), "GaAs": (1.42, 0.067, 0.48),
             "4H-SiC": (3.26, 0.77, 1.0), "GaN": (3.39, 0.2, 1.5)}

temps = [300, 400, 500, 700, 1000]
print("n_i in cm^-3 (gap held at its 300 K value)")
print("material " + "".join(f"{T:>11} K" for T in temps))
for name, (Eg, me, mh) in materials.items():
    print(f"{name:8} " + "".join(f"{n_intrinsic(Eg, me, mh, T):13.2e}" for T in temps))

# A device stops working when intrinsic carriers swamp its doping. At what temperature does
# n_i reach 1e15 cm^-3, a typical doping of a high-voltage device's thick, lightly doped layer?
print("\ntemperature at which n_i reaches 1e15 cm^-3:")
for name, (Eg, me, mh) in materials.items():
    lo, hi = 200.0, 3000.0
    for _ in range(60):                           # bisection: n_i rises steadily with T
        mid = 0.5 * (lo + hi)
        if n_intrinsic(Eg, me, mh, mid) < 1e15:
            lo = mid
        else:
            hi = mid
    print(f"  {name:7} {lo:6.0f} K  ({lo - 273.15:5.0f} C)")
Runs entirely in your browser, in an isolated worker. Nothing is sent to a server.

At room temperature silicon has about 1.2×10101.2 \times 10^{10} intrinsic carriers per cubic centimetre (careful measurements give close to 1.0×10101.0 \times 10^{10}), germanium about 2×10132 \times 10^{13}, gallium arsenide about 2×1062 \times 10^6. Silicon carbide and gallium nitride have, in effect, none: around 10−810^{-8} and 10−1010^{-10} per cubic centimetre, which is about one carrier per hundred cubic metres of silicon carbide, and fewer still in gallium nitride. Raising the temperature from 300 to 500 K multiplies silicon's nin_i by more than ten thousand.

The second table explains a practical limit. A device works only while its deliberately added carriers outnumber the intrinsic ones. Germanium reaches that point near 130 °C, and germanium transistors were usually limited to well below 100 °C, one reason silicon replaced germanium in the 1960s. Silicon reaches ni=1015n_i = 10^{15} cm⁻³ near 300 °C, and in practice silicon power devices are rated to about 150–175 °C. Silicon carbide and gallium nitride would not reach that point until well over 1000 °C, so for them the limits come from packaging, contacts and oxides instead.

DefinitionsH

Donors, acceptors and the carriers they supply

Definition— donor, acceptor, n-type, p-type

A donor is an impurity with one more valence electron than the atom it replaces (phosphorus, arsenic or antimony in silicon). The extra electron is weakly bound and easily released into the conduction band, leaving a fixed positive ion. An acceptor has one fewer (boron, aluminium, gallium in silicon); it easily captures an electron from the valence band, leaving a mobile hole and a fixed negative ion. A semiconductor whose majority carriers are electrons is n-type; one whose majority carriers are holes is p-type. Electrons in p-type material and holes in n-type material are minority carriers.

Definition— extrinsic semiconductor, charge neutrality

A semiconductor whose carriers come mainly from impurities is extrinsic. In equilibrium the material is electrically neutral:

n+NA−=p+ND+,n + N_A^- = p + N_D^+,

where ND+N_D^+ and NA−N_A^- are the densities of ionised donors and acceptors. Together with np=ni2np = n_i^2 and the occupation statistics of the impurity levels, this fixes the Fermi level and every carrier density.

How weakly is the donor's electron bound? A phosphorus ion in silicon, with its extra electron, looks like a hydrogen atom with two changes: the electron has the effective mass m∗m^* instead of mem_e, and the attraction is screened by the crystal's relative permittivity εr\varepsilon_r. Hydrogen's 13.6 eV binding energy and 0.0529 nm Bohr radius become

Ed=13.6 eV×m∗/meεr2,a∗=0.0529 nm×εrm∗/me.E_d = 13.6\ \text{eV}\times\frac{m^*/m_e}{\varepsilon_r^2}, \qquad a^* = 0.0529\ \text{nm}\times\frac{\varepsilon_r}{m^*/m_e}.

For silicon (m∗=0.26 mem^* = 0.26\,m_e, εr=11.7\varepsilon_r = 11.7) that gives about 26 meV and an orbit 2.4 nm across, spread over hundreds of silicon atoms. The measured values (45 meV for phosphorus, 54 meV for arsenic, 39 meV for antimony, and 45 meV for the boron acceptor) are larger, because close to the ion the screened-hydrogen picture fails and each element's core matters. For gallium arsenide, with its small effective mass, the model gives 5.5 meV and an orbit 10 nm across, close to the measured binding of common shallow donors there. Either way the binding is comparable with kBT=26k_BT = 26 meV at room temperature, so almost all such impurities are ionised.

The program below solves charge neutrality for silicon with 101610^{16} phosphorus atoms per cubic centimetre, from 30 K to 800 K, by bisection on the Fermi level.

python
import numpy as np

kB = 8.617333e-5     # eV/K

# Hydrogen model of a shallow donor: binding energy 13.606 eV * (m*/m_e) / eps_r^2, orbit radius 0.0529 nm * eps_r / (m*/m_e)
for host, m_cond, eps in [("Si", 0.26, 11.7), ("GaAs", 0.067, 12.9)]:
    print(f"{host:5}: donor binding energy {13.606 * m_cond / eps**2 * 1000:5.1f} meV, "
          f"orbit radius {0.0529 * eps / m_cond:5.2f} nm")

# Silicon doped with 1e16 phosphorus atoms per cm^3: solve charge neutrality n + 0 = p + N_D+ for the Fermi level
Eg, Ed = 1.12, 0.045          # gap and donor binding energy (eV); energies measured up from the valence band
Nd = 1e16                     # cm^-3
def NcNv(T):
    s = (T / 300) ** 1.5
    return 2.62e19 * s, 2.69e19 * s     # 2.5094e19 * m^1.5 for m_e,dos = 1.09 and m_h,dos = 1.15

def carriers(T):
    Nc, Nv = NcNv(T)
    kT = kB * T
    def excess(EF):           # positive charge minus negative charge
        n = Nc * np.exp((EF - Eg) / kT)
        p = Nv * np.exp(-EF / kT)
        Nd_plus = Nd / (1 + 2 * np.exp((EF - (Eg - Ed)) / kT))   # spin degeneracy 2 of the neutral donor
        return p + Nd_plus - n, n, p, Nd_plus
    lo, hi = -0.5, Eg + 0.5
    for _ in range(200):      # bisection: the excess falls as E_F rises
        mid = 0.5 * (lo + hi)
        if excess(mid)[0] > 0:
            lo = mid
        else:
            hi = mid
    _, n, p, Ndp = excess(lo)
    return lo, n, p, Ndp

print("\n  T/K   E_c - E_F/eV    n/cm^-3     p/cm^-3   donors ionised")
rows = []
for T in [30, 50, 75, 100, 150, 200, 300, 400, 500, 600, 700, 800]:
    EF, n, p, Ndp = carriers(T)
    rows.append((T, n))
    p_text = f"{p:9.2e}" if p >= 1 else "      < 1"
    print(f"{T:6}   {Eg - EF:8.3f}      {n:9.2e}   {p_text}     {100 * Ndp / Nd:5.1f} %")

print("\nlog10(n) against 1000/T: freeze-out, saturation, intrinsic")
for T, n in rows:
    bar = int(round((np.log10(n) - 12) * 8))
    print(f" {1000 / T:6.2f} | " + "*" * max(bar, 0) + f"  {T} K")
Runs entirely in your browser, in an isolated worker. Nothing is sent to a server.

The output shows the three regimes every semiconductor engineer learns to recognise.

  • Freeze-out (below about 100 K). There is not enough thermal energy to ionise all the donors: at 50 K only 5 per cent have given up their electron. Silicon electronics for very cold environments, such as the control chips for quantum computers, must be designed with this in mind.
  • Saturation or the extrinsic range (about 150 to 500 K). Essentially every donor is ionised and n≈ND=1016n \approx N_D = 10^{16} cm⁻³, independent of temperature. At 300 K the holes number only ni2/n≈104n_i^2/n \approx 10^4 per cubic centimetre: a trillion times fewer than the electrons. This flat region is where devices are meant to work.
  • Intrinsic (above about 600 K here). Carriers lifted across the gap outnumber the donors' electrons, nn and pp rise together, and the doping no longer controls anything.
ComputationC

Mobility, resistivity and the wafer specification

Carriers in an electric field E\mathcal E accelerate, scatter off lattice vibrations and impurities, and settle into an average drift velocity v=μEv = \mu\mathcal E. The mobility μ\mu, in cm²/(V s), measures how freely they move; a smaller effective mass and less scattering mean higher mobility. The conductivity is

σ=q (nμn+p μp),ρ=1/σ.\sigma = q\,(n\mu_n + p\,\mu_p), \qquad \rho = 1/\sigma.

In lightly doped silicon at room temperature, electrons have μn≈1400\mu_n \approx 1400 cm²/(V s) and holes about 470. Adding dopants adds ionised impurities that scatter carriers, so mobility falls as doping rises. Device engineers use empirical fits to measurements; the program uses the main terms of the fit by Giorgio Masetti, Maurizio Severi and Sandro Solmi (1983) to compute resistivity against doping, then models how dopants spread along a growing crystal, which we will need in the next section.

python
import numpy as np

q = 1.602176634e-19   # C

def mobility(N, carrier):
    """Room-temperature mobility in silicon (cm^2/(V s)) against dopant density N (cm^-3):
    the main terms of the empirical fit of Masetti, Severi and Solmi (1983)."""
    if carrier == "electron":            # electrons in phosphorus-doped silicon
        mu_min, mu_max, N_ref, alpha = 68.5, 1414.0, 9.20e16, 0.711
    else:                                # holes in boron-doped silicon
        mu_min, mu_max, N_ref, alpha = 44.9, 470.5, 2.23e17, 0.719
    return mu_min + (mu_max - mu_min) / (1 + (N / N_ref) ** alpha)

print(" N/cm^-3    mu_n    rho (n-type, ohm cm)    mu_p    rho (p-type, ohm cm)")
for N in [1e13, 1e14, 1e15, 1e16, 1e17, 1e18, 1e19]:
    mn, mp = mobility(N, "electron"), mobility(N, "hole")
    print(f"{N:8.0e}  {mn:6.0f}   {1 / (q * N * mn):12.4g}          {mp:6.0f}   {1 / (q * N * mp):12.4g}")

# What doping gives a 10 ohm cm p-type wafer? Solve rho(N) = 10 by bisection on log N.
lo, hi = 12.0, 19.0
for _ in range(60):
    mid = 0.5 * (lo + hi)
    rho = 1 / (q * 10**mid * mobility(10**mid, "hole"))
    lo, hi = (mid, hi) if rho > 10 else (lo, mid)
print(f"\n10 ohm cm p-type needs about {10**lo:.2e} boron atoms per cm^3, "
      f"one per {5.0e22 / 10**lo:.1e} silicon atoms")

# Czochralski growth: the crystal rejects or accepts each impurity with distribution coefficient k = C_solid / C_liquid.
# Scheil: after a fraction f of the melt has frozen, the crystal growing has C_s = k C0 (1 - f)^(k - 1).
k_values = {"boron": 0.8, "phosphorus": 0.35, "gallium": 0.008, "iron": 8e-6}
fractions = [0.0, 0.25, 0.5, 0.75, 0.9]
print("\ndopant concentration along the ingot, relative to the first slice (seed end)")
print("element    k        " + "".join(f"  f={f:<5}" for f in fractions))
for el, k in k_values.items():
    rel = [(1 - f) ** (k - 1) for f in fractions]
    print(f"{el:10} {k:<8g} " + "".join(f"{r:8.2f}" for r in rel))
Runs entirely in your browser, in an isolated worker. Nothing is sent to a server.

Resistivity is how wafers are bought and sold. A common substrate for logic chips is p-type at about 10 Ω cm, which the program shows needs 1.4×10151.4 \times 10^{15} boron atoms per cubic centimetre: one boron atom for every 37 million silicon atoms. Over the range from 101310^{13} to 101910^{19} cm⁻³, resistivity falls by about five orders of magnitude, less than the six orders by which the doping rises, because mobility falls as the doping increases.

ExplanationH

From sand to wafer

Purifying silicon. Quartz (silicon dioxide) is reduced with carbon in large electric arc furnaces to metallurgical-grade silicon, about 98–99 per cent pure; most of it goes into aluminium alloys and silicones. For solar cells and chips it is converted to a volatile chlorosilane, purified by repeated distillation, and decomposed back onto heated silicon rods in what is called the Siemens process. The resulting polysilicon is 99.9999 per cent pure or better for solar use and 99.999999999 per cent ("eleven nines") for electronics. Silicon has 5.0×10225.0 \times 10^{22} atoms per cubic centimetre, so even nine nines leaves 5×10135 \times 10^{13} impurity atoms per cubic centimetre; the transition metals among them, which create states deep in the gap that destroy carrier lifetimes, must be kept far lower still. These are large industrial chemical plants handling flammable and corrosive gases under strict regulation, a reminder of why such processes stay in industry.

Growing the crystal. Almost all silicon for chips is grown by the method Jan Czochralski described in 1918, when he was studying how fast metals crystallise. Polysilicon and a measured amount of dopant are melted (silicon melts at 1414 °C) in a silica crucible under flowing argon. A small seed crystal, cut to the desired orientation, usually (100), is dipped into the melt and slowly withdrawn while seed and crucible rotate in opposite directions. Atoms freeze onto the seed in its orientation, and the crystal grows downwards into the melt.

pull up slowlypull rodseedneckshouldercrystal(rotating)molten siliconabout 1420 °Csilica cruciblegraphite holder,heaterheaterargon flowcrucible counter-rotates
A Czochralski crystal puller (schematic, not to scale). The seed is dipped into molten silicon and slowly raised while rotating; the thin neck grown first lets dislocations escape at its surface, after which the crystal widens through the shoulder into a cylindrical body. Heaters around a graphite holder keep the silica crucible's melt just above the melting point, and argon sweeps away silicon monoxide evaporating from the melt.

The first few centimetres are deliberately grown very thin, a few millimetres across, at a fast pull rate. This Dash neck, introduced by William Dash at General Electric in the late 1950s, lets dislocations from the seed, created by the thermal shock of dipping it, glide out to the surface and disappear; everything grown after it is dislocation-free. Then the diameter is increased through a shoulder to the full size and held constant by adjusting the pull rate and heater power, with the diameter monitored optically. A modern 300 mm ingot can be around two metres long and weigh several hundred kilograms. It is ground to a precise diameter, sliced with diamond-coated wire saws, and the slices are lapped, etched and polished to wafers 775 µm thick (the thickness set by the SEMI M1 standard for 300 mm wafers) and flat to within tens of nanometres.

Dopants are not frozen in evenly. When a melt freezes, each impurity divides itself between solid and liquid in a ratio set by the phase diagram, the distribution coefficient k=Cs/Clk = C_s/C_l (see Phase diagrams). With kk below 1 the crystal takes up less impurity than the melt holds, so the rejected impurity accumulates in the shrinking melt, and later parts of the ingot get more. If the melt stays well mixed, the Scheil equation gives the concentration in the crystal after a fraction ff has frozen:

Cs(f)=k C0 (1−f)k−1.C_s(f) = k\,C_0\,(1 - f)^{k - 1}.

The table from the last program shows the consequences. Boron, with k=0.8k = 0.8, varies by only 58 per cent from one end of the ingot to the point where 90 per cent has frozen, so p-type ingots yield many wafers within a resistivity specification. Phosphorus (k=0.35k = 0.35) varies more than fourfold. Gallium (k=0.008k = 0.008) varies almost tenfold, yet many solar-cell makers switched from boron to gallium doping around 2020, because boron pairs with the oxygen dissolved from the crucible and the pairs degrade cells under light. Iron, with kk near 10−510^{-5}, is almost completely rejected: a crystal grown from slightly contaminated melt is a hundred thousand times cleaner than the melt. The same rejection drives zone refining, the method that first made germanium pure enough for transistors.

That crucible oxygen, around 101710^{17} to 101810^{18} atoms per cubic centimetre, is not all bad: controlled heat treatments make it form tiny precipitates deep in the wafer that trap stray metal atoms away from the surface where transistors are built, a process called internal gettering. When oxygen cannot be tolerated, for high-voltage devices and radiation detectors, crystals are grown by the float-zone method instead, melting a narrow zone that moves along a polysilicon rod held without any crucible. Dopants are then often introduced as described in Diffusion, or by neutron irradiation, which turns a precise fraction of silicon-30 atoms into phosphorus.

ExplanationH

Compound semiconductors: GaAs, GaN and SiC

Silicon dominates because it is abundant, grows huge perfect crystals, and forms an excellent native oxide. But its gap is indirect and fairly small, and its electrons are not especially mobile. For three jobs, other materials win.

Fast and radio-frequency electronics: gallium arsenide. Electrons in GaAs have an effective mass of 0.067 me0.067\,m_e and a mobility around 8000 cm²/(V s) in pure material, several times silicon's. GaAs heterojunction bipolar transistors amplify the radio signals in most mobile phones, and its direct gap makes it an efficient emitter of infrared light, used in remote controls, optical links and the vertical-cavity lasers of face-recognition modules. Its crystals are harder to grow than silicon's because arsenic is volatile: the melt is covered by a layer of molten boron oxide, or the crystal is frozen slowly in a sealed vertical boat.

Lighting: gallium nitride. GaN's direct gap of 3.4 eV lies in the near ultraviolet, and alloying with indium nitride brings emission down to blue and green. For decades nobody could grow GaN crystals of usable quality or make it p-type. In the late 1980s Isamu Akasaki and Hiroshi Amano at Nagoya grew good GaN on sapphire using a thin aluminium nitride buffer layer, and activated magnesium acceptors with an electron beam; Shuji Nakamura at Nichia found that thermal annealing activated them too, and made the first bright blue LEDs in 1993. All three shared the 2014 Nobel Prize in Physics. A blue LED coated with a yellow phosphor makes white light, now the most widely sold kind of lamp. Remarkably, GaN LEDs work even though the films contain 10810^8 to 10910^9 dislocations per square centimetre, which would ruin a GaAs or silicon device. The materials of LEDs and lasers are developed further in Optical materials.

Power: silicon carbide and gallium nitride. A power transistor must block high voltage when off and conduct with low resistance when on. The voltage is held by a thick, lightly doped drift layer, and the electric field in it must stay below the critical field EcE_c at which carriers gain enough energy to knock out more carriers in an avalanche. A wider gap means a larger EcE_c: roughly 0.3 MV/cm in silicon, about 2.5 in 4H-SiC and 3.3 in GaN. A higher critical field lets the layer be thinner and more heavily doped, and its on-resistance falls as 1/(εμEc3)1/(\varepsilon\mu E_c^3), a combination Jayant Baliga proposed as a figure of merit in 1982.

python
eps0 = 8.854e-14      # F/cm
q = 1.602e-19         # C

# Rounded, commonly quoted room-temperature values: gap (eV), relative permittivity,
# electron mobility (cm^2/(V s)), critical field for breakdown (V/cm). The critical field depends on doping.
mats = {"Si": (1.12, 11.7, 1350, 0.3e6), "4H-SiC": (3.26, 9.7, 900, 2.5e6),
        "GaN": (3.39, 9.0, 1000, 3.3e6), "beta-Ga2O3": (4.8, 10.0, 200, 8e6)}

VB = 1200.0           # blocking voltage of the device, V
print(f"drift layer of a unipolar switch blocking {VB:.0f} V (ideal one-sided junction, punch-through ignored)")
print("material      eps*mu*Ec^3 vs Si   thickness/um   doping/cm^-3   R_on,sp/(mOhm cm^2)")
ref = None
for name, (Eg, er, mu, Ec) in mats.items():
    eps = er * eps0
    bfom = eps * mu * Ec**3
    ref = ref or bfom
    W = 2 * VB / Ec                      # triangular field profile peaking at Ec
    Nd = eps * Ec**2 / (2 * q * VB)      # doping that gives that profile
    Ron = W / (q * mu * Nd)              # = 4 VB^2 / (eps mu Ec^3)
    print(f"{name:12}  {bfom / ref:12.0f}       {W * 1e4:9.1f}     {Nd:12.2e}     {Ron * 1e3:10.3f}")
Runs entirely in your browser, in an isolated worker. Nothing is sent to a server.

To block 1200 V, a silicon drift layer must be 80 µm thick and doped at only 2.4×10142.4 \times 10^{14} cm⁻³, with a resistance of about 150 mΩ cm². A silicon carbide layer does the same job 9.6 µm thick, doped sixty times more heavily, with about three hundred times less resistance. That is why the traction inverters of many electric vehicles, especially those with 800 V batteries, use SiC transistors, and why GaN transistors, built on silicon wafers for cost, now fill compact phone chargers and data-centre power supplies at 650 V and below (see Lumenary EE's power electronics). Beta-gallium oxide, with a 4.8 eV gap, promises more still and, unlike GaN and SiC, can be grown from a melt; but it conducts heat poorly and cannot yet be made p-type, so as of 2026 it remains mostly in research and early products.

SiC brings a second advantage: it conducts heat about three times better than silicon. Its crystals are grown not from a melt (at ordinary pressures it decomposes before it melts) but by physical vapour transport, condensing vapour onto a seed above about 2000 °C. Defects that once crippled it, such as hollow-core "micropipe" dislocations, have been engineered down, and as of 2026 production is moving from 150 mm to 200 mm wafers.

ExplanationH

Heterostructures and band engineering

The deepest idea in compound semiconductors is that you need not accept the band structure nature gives you. Grow one semiconductor on another, atomic layer by atomic layer, and at the interface the band edges jump. Those jumps, the band offsets, can confine electrons, guide them and separate them from the impurities that scatter them.

Two conditions make it work. First, the crystals must fit: if their lattice parameters differ, the upper layer is strained and, beyond a critical thickness, relieves the strain with misfit dislocations. Gallium arsenide (0.5653 nm) and aluminium arsenide (0.5661 nm) differ by only 0.14 per cent, so the alloy AlxGa1−xAs\text{Al}_x\text{Ga}_{1-x}\text{As} can be grown on GaAs at any composition. For other pairs, the alloy composition is chosen to match: by Vegard's law an alloy's lattice parameter is close to the composition-weighted average of its ends, which is how In0.53Ga0.47As\text{In}_{0.53}\text{Ga}_{0.47}\text{As} is matched to indium phosphide for the photodetectors of fibre-optic links. Second, the growth must be precise: molecular beam epitaxy (MBE) and metal–organic chemical vapour deposition (MOCVD) deposit crystals one atomic layer at a time.

0.540.580.620.66lattice parameter (nm)012band gap (eV)SiGeGaPAlPGaAsAlAsZnSeInPInAsGaSbInSbAlGaAsInGaAs
Band gap against lattice parameter for common semiconductors at room temperature. Filled circles have direct gaps, open circles indirect. Materials on the same vertical line can be grown on each other without strain: AlGaAs alloys lie almost exactly above GaAs, and the InGaAs alloy curve (which bows below a straight line) crosses the lattice parameter of InP at a gap of about 0.74 eV, the composition In₀.₅₃Ga₀.₄₇As.

The applications followed one after another.

  • The double-heterostructure laser. In 1963 Herbert Kroemer, and independently Zhores Alferov and Rudolf Kazarinov, proposed sandwiching a narrow-gap layer between wider-gap layers, so that electrons and holes are trapped together where they recombine and the light is guided by the higher refractive index of the narrow-gap layer. In 1970 Alferov's group in Leningrad, and Izuo Hayashi and Morton Panish at Bell Labs, made such lasers run continuously at room temperature. Every fibre-optic transmitter and DVD laser descends from them; Alferov and Kroemer shared the 2000 Nobel Prize in Physics.
  • Quantum wells. Make the narrow-gap layer only a few nanometres thick and the carriers in it occupy discrete levels, like a particle in a box, whose energies are set by the thickness. Laser wavelengths can then be tuned by geometry.
  • Modulation doping and the HEMT. In 1978 Raymond Dingle, Horst Störmer, Arthur Gossard and William Wiegmann at Bell Labs put the donors in the AlGaAs and left the GaAs undoped. Electrons fall from the donors into the GaAs, where they form a two-dimensional sheet, far from the ions that would scatter them. Takashi Mimura's group at Fujitsu turned this into the high-electron-mobility transistor in 1980, used ever since in satellite receivers and radio telescopes. In gallium nitride the same kind of electron sheet forms at an AlGaN/GaN interface without any doping at all, induced by the materials' built-in electrical polarisation, and it is the basis of GaN power and radio transistors.
  • Strained silicon. Even silicon chips use band engineering: since the early 2000s, transistors have had silicon–germanium regions that stretch or compress the channel, changing the band structure to raise mobility.
AlGaAsGaAsAlGaAsconduction bandvalence bandΔEc ≈ 0.24 eVΔEv ≈ 0.13 eV1.42 eV1.80 eVlightelectron levelhole level
Energy bands across a quantum well: a thin layer of GaAs between layers of Al₀.₃Ga₀.₇As. The wider gap of the alloy (about 1.80 eV against 1.42 eV) is split between a conduction-band step of about 0.24 eV and a valence-band step of about 0.13 eV. Electrons and holes are both trapped in the GaAs, where they occupy discrete levels (dashed) and recombine to emit light.
ExplanationH

The professional view

Wafer supply. A handful of companies grow most of the world's 300 mm silicon, among them Shin-Etsu Handotai and SUMCO in Japan, GlobalWafers in Taiwan, Siltronic in Germany and SK Siltron in South Korea. Specifications run to dozens of parameters: resistivity and its variation, oxygen and carbon content, flatness, particles, and metal contamination counted in units of 101010^{10} atoms per square centimetre. Silicon-on-insulator wafers, with a thin crystalline layer on an oxide, are made largely by the Smart Cut process from the French company Soitec and its licensees. The industry's standards are set through SEMI.

Compound substrates and epitaxy. SiC substrates come from producers in the United States, Europe, Japan and, increasingly, China; GaN for LEDs and power is mostly grown as thin films on sapphire, SiC or silicon in MOCVD reactors from makers in Germany, the United States and China. Device performance is often limited not by the material's ideal properties but by its defects: basal-plane dislocations in SiC, carbon and other traps in GaN, and the quality of interfaces with gate oxides.

Where research is going (as of 2026). Ultrawide-gap materials (gallium oxide, aluminium nitride, diamond) for higher voltages; GaN on large silicon wafers for lower cost; gallium and germanium supply, after export controls introduced in 2023 made their concentration in a few countries visible; and two-dimensional semiconductors such as MoS₂ as possible channels for future transistors, discussed in Two-dimensional and quantum materials and Lumenary EE's transistor scaling.

Common misconceptionsH

Common misconceptions

Misconception

"n-type silicon is negatively charged." Every electron a donor gives to the conduction band leaves a positively charged donor ion behind. Doped material is electrically neutral; "n" and "p" name the majority carrier.

Misconception

"Doping adds carriers, so more is always better." Heavier doping lowers mobility, shortens minority-carrier lifetimes and, above about 101910^{19} cm⁻³, makes silicon degenerate, behaving more like a poor metal. Each region of a device is doped for its job: lightly where voltage is blocked, heavily where contacts are made.

Misconception

"Purer is always better, so intrinsic silicon is the ideal." Intrinsic silicon's conductivity changes about tenfold for every 30 kelvin near room temperature. Doping makes the carrier density stable, which is what a circuit needs.

Misconception

"Wide-gap semiconductors are simply better than silicon." They are better for high voltage, high temperature and short wavelengths. For dense logic, silicon's huge perfect wafers, cheap processing and superb oxide are still unmatched, and SiC and GaN wafers remain far more expensive and defective.

Exercises

Exercises

1. Semiconductor materials

Warm-upHuman

Which of these impurities, substituting for silicon atoms, make silicon n-type? Select all that apply.

Choose all that apply

2. Semiconductor materials

CoreHuman

Use the hydrogen model to estimate the binding energy of a shallow donor in gallium nitride, taking the electron effective mass as 0.20 me0.20\,m_e and the relative permittivity as εr=9.5\varepsilon_r = 9.5. Give the answer in meV, to the nearest whole number. (Hydrogen's binding energy is 13.606 eV.)

3. Semiconductor materials

CoreHuman

The lattice parameters of GaAs, InAs and InP are 0.56533 nm, 0.60583 nm and 0.58687 nm. Using Vegard's law (the lattice parameter of InxGa1−xAs\text{In}_x\text{Ga}_{1-x}\text{As} is (1−x) aGaAs+x aInAs(1-x)\,a_{\text{GaAs}} + x\,a_{\text{InAs}}), what indium fraction xx makes the alloy lattice-matched to InP? Give xx as a percentage, to the nearest whole number.

4. Semiconductor materials

CoreHuman

Why can a silicon carbide power transistor have a far lower on-resistance than a silicon one rated for the same blocking voltage?

Choose one

5. Semiconductor materials

CoreComputational

In the saturation range every dopant is ionised. Charge neutrality n+NA=p+NDn + N_A = p + N_D together with np=ni2np = n_i^2 then gives a quadratic for nn.

Write electron_density(Nd, Na, ni) returning the electron density (same units as the inputs) for a semiconductor containing both donors Nd and acceptors Na, with intrinsic density ni. It must work for n-type, p-type, compensated and intrinsic material, and must stay accurate when ni is ten orders of magnitude smaller than the doping.

Runs entirely in your browser, in an isolated worker. Nothing is sent to a server.

6. Semiconductor materials

StretchComputational

A Czochralski ingot is grown from a well-mixed melt, so by the Scheil equation the dopant concentration in the crystal, relative to the first slice, is (1−f)k−1(1-f)^{k-1} after a fraction ff of the melt has frozen. A wafer is within specification if its dopant concentration is no more than a factor (1+tol)(1 + \text{tol}) above that of the first slice.

Write usable_fraction(k, tol) returning the fraction of the ingot, measured from the seed end, that is within specification, for a distribution coefficient 0<k<10 < k < 1.

Runs entirely in your browser, in an isolated worker. Nothing is sent to a server.

7. Semiconductor materials

Warm-upHuman

Name the crystal-growth method in which a seed crystal is dipped into a melt and slowly pulled out while rotating, used for almost all silicon wafers. (Give the inventor's surname, as the method is named.)

Further reading and resources