2026-08-25·FREE·11 min read

Band Theory from the Ground Up — Why It Had to Be Silicon

What actually separates a conductor from an insulator? Starting from why atomic energy levels smear into bands, this piece builds up bandgaps, Fermi statistics and doping with no prior physics assumed — and lands on the real answer to 'why is every chip made of silicon', which turns out to be about the oxide, not the element.

#semiconductor #band-theory #bandgap #doping #silicon #hardware

ModalitytextTaskhardware

Between "conducts" and "doesn't"

Copper conducts electricity. Glass doesn't. School physics gives you two categories, conductors and insulators. Yet the inside of every computer is made of neither — it's made of semiconductors.

People often describe a semiconductor as "a material that conducts moderately well," and that framing sends you down the wrong path. The value isn't that its resistance sits in the middle; it's that the same piece of material can be switched between conducting and not conducting. If middling resistance were the point, pencil lead would do. The switching is what gives you a 0 and a 1, and therefore computation. Band theory is the way of seeing solids that makes that switching intelligible — and followed to the end, it also answers, surprisingly, why every chip on Earth is silicon.

The picture: a sold-out lower deck and an empty upper deck

Think of the energy states an electron can occupy as seats in a stadium. The Pauli exclusion principle says one seat holds one electron (two, counting spin), so electrons fill from the bottom up.

The crucial consequence is that nobody can move in a completely full row. It's a packed commuter train — if the seat next to you is taken, there's nowhere to shuffle to. Current means electrons moving, so a fully occupied row carries none. What matters isn't "are there lots of electrons" but "is there an empty seat next to them." That's the first counterintuitive hurdle.

Line up three materials in this picture:

Conductor (copper): the top row is only half full. Neighbours are empty, so a small push slides everyone along.
Insulator (glass): the lower deck is packed solid, the upper deck is empty — and the staircase between them is enormously tall. Electrons can't climb up, and downstairs they can't move.
Semiconductor (silicon): structurally identical to the insulator, except the staircase is only moderately tall. Ordinary room-temperature heat bounces a small fraction of electrons upstairs.

The packed lower deck is the valence band, the empty upper deck is the conduction band, and the height of the staircase is the bandgap. So insulators and semiconductors are not different kinds of thing — they differ only in how tall the staircase is. That sentence is the spine of everything below.

One more consequence: an electron that jumps upstairs leaves an empty seat behind. Electrons near that vacancy can now shuffle, so the vacancy itself appears to travel. Book it as a moving positive charge and you have the hole. A semiconductor always has two carriers to track.

Where the "bands" come from

An isolated atom has discrete allowed energies: a handful of shelves at fixed heights, nothing in between. Bring a second atom close and the two electron waves overlap; since Pauli forbids two identical states, one shelf splits into two. Three atoms, three shelves; four atoms, four. A cubic centimetre of solid silicon packs roughly 5×10²² atoms, so the shelf splits 5×10²² ways, the spacing becomes unresolvably fine, and what you see is a continuous band. That's all a band is.

So where does the gap come from? In a crystal the atoms sit in a regular repeating array, and electrons behave as waves, so certain wavelengths get reflected by that periodicity. A reflected wave can't propagate, so the energies corresponding to those wavelengths have no available seats at all. It's the same reasoning as X-ray diffraction off a crystal, applied to electron waves.

Bandgaps are quoted in electron volts (eV) — the energy an electron picks up crossing a one-volt difference. Representative values: germanium 0.66, silicon 1.12, gallium arsenide 1.42, silicon carbide (4H-SiC) about 3.3, diamond about 5.5, and SiO₂ — silicon's own oxide — about 9 eV. Bind silicon to oxygen and the gap grows nearly eightfold, turning it into a flat-out insulator. That fact does a lot of work later.

The mechanism: staircase height versus how many can climb it

The first equation gives the probability that a state is occupied — the Fermi–Dirac distribution.

f(E) = \frac{1}{1 + \exp\!\left(\dfrac{E - E_F}{k_B T}\right)}

(1)

$E$ is the energy of the state, $E_F$ is the Fermi level (the height at which occupancy is exactly 50%), $k_B$ is Boltzmann's constant, $T$ is absolute temperature. The product $k_B T$ measures how much punch room-temperature heat carries: at 300 K (about 27 °C) it is roughly 0.026 eV, or 26 meV.

In words, equation (1) says: states well below the Fermi level are essentially full, states well above it are essentially empty, and the width of the fuzzy boundary between them is set by $k_B T$ . Heat the material and the boundary blurs, so electrons appear in the upper states.

The shape is also, exactly, the sigmoid function from machine learning. Select sigmoid in the figure below.

FIG 1The sigmoid curve *is* the Fermi distribution. Read the horizontal axis as (E_F − E)/k_BT and the vertical axis as "probability this seat holds an electron": it passes through exactly 0.5 at x=0 (the Fermi level) and pins to 0 or 1 within a few tick marks. At room temperature that entire transition is about 26 meV wide

The second equation is the one that matters most. For a material with bandgap $E_g$ , the number of electrons that thermal energy alone lifts into the conduction band — the intrinsic carrier concentration $n_i$ — is:

n_i = \sqrt{N_c N_v}\,\exp\!\left(-\frac{E_g}{2 k_B T}\right)

(2)

$N_c$ and $N_v$ count the seats available in each band; they're material-specific and land around 10¹⁹ cm⁻³. The prefactor isn't the point. The point is that $E_g$ sits in the exponent, which says something blunt: raise the staircase and the number of electrons able to climb it doesn't drop a bit — it drops by orders of magnitude.

Feel it: exponents move in decades

Measured values at 300 K: germanium about 2×10¹³ per cm³, silicon about 1×10¹⁰, gallium arsenide about 2×10⁶. Moving the bandgap from 0.66 to 1.12 to 1.42 eV — barely a doubling — drops the carrier count by seven decades.

FIG 2Ignore the curve labels; watch only how the straight line and the exponential pull apart. Because the bandgap rides in the exponent, the difference between 1.12 eV and 5.5 eV is not "5×" but tens of decades. The log/linear toggle is the other half of the lesson — an exponential straightens out on a log axis, which is why semiconductor datasheets plot on log scales

Read the same equation along the temperature axis and you get a number engineers use daily: in silicon, $n_i$ roughly doubles for every ~10 °C. The current leaking through a transistor that is supposedly OFF comes from those carriers, so a hotter chip leaks more, leaking heats it further, and it leaks more again. That feedback is the physical content of "thermal runaway." MOSFETs from the Ground Up and The Physics of Power pick up the thread.

Checking it in code

import numpy as np

kT = 8.617e-5 * 300          # thermal energy at room temperature [eV] ≈ 0.0259

def n_i(Eg, Nc, Nv, kT=kT):
    """Intrinsic carrier concentration [cm^-3]. Nc, Nv = how many seats exist."""
    return np.sqrt(Nc * Nv) * np.exp(-Eg / (2 * kT))

for name, Eg, Nc, Nv in [("Ge",   0.66, 1.0e19, 6.0e18),
                         ("Si",   1.12, 2.8e19, 1.0e19),
                         ("GaAs", 1.42, 4.7e17, 7.0e18)]:
    print(f"{name:5s} Eg={Eg:.2f} eV  n_i={n_i(Eg, Nc, Nv):.1e} cm^-3")

# Ge    Eg=0.66 eV  n_i=2.2e+13 cm^-3
# Si    Eg=1.12 eV  n_i=6.5e+09 cm^-3
# GaAs  Eg=1.42 eV  n_i=2.1e+06 cm^-3

Measured values are about 2.4×10¹³, 1.0×10¹⁰ and 2.1×10⁶, so the toy calculation lands within a small factor. Exact agreement was never on offer — $N_c$ and $N_v$ are themselves approximations — but get the exponent right and the decade is right. That's the character of an exponential model.

Doping: one part per million moves six decades

Pure silicon packs 5×10²² atoms per cm³, of which a mere 10¹⁰ conduct. That's very nearly an insulator, and you can't build a circuit from it. Enter doping — deliberately contaminating the crystal.

Silicon has four outer electrons and bonds to four neighbours. Substitute phosphorus or arsenic, which have five, and four go into bonds while one is left over — a spare that needs only a few tens of meV to reach the conduction band, because the dopant provides a landing step just below the top instead of the full 1.12 eV staircase. Such an atom donates electrons — a donor — and electrons (negative) become the majority carrier, hence n-type. Substitute boron, with three outer electrons, and a bond comes up one short, leaving a vacancy that supplies holes: an acceptor, giving p-type.

The leverage is startling. Add phosphorus at one part per million of those 5×10²² atoms and you get 5×10¹⁶ donors per cm³ — roughly five million times the native 10¹⁰ carriers. In resistivity terms, a fall from the 10⁵ Ω·cm range to the 0.1 Ω·cm range: six decades, from one part per million of contamination. It's also why fabs are fanatical about cleanliness — an unintended impurity at the same concentration changes the material just as much.

At equilibrium, doped or not, $n p = n_i^{\,2}$ holds ( $n$ is electron concentration, $p$ is hole concentration). In words: push one up and the other falls by the same factor, because the product is fixed — the law of mass action. For the n-type silicon above, $p$ collapses to the 10³ range and holes are all but wiped out. That asymmetry is exactly what starts happening when n-type meets p-type, which is where pn junctions and diodes begin.

Why it had to be silicon

Historically, the first transistor was not silicon. The 1947 point-contact device at Bell Labs was germanium, as was Kilby's 1958 integrated circuit — germanium was easier to purify to high quality at the time and offered better mobility. So what flipped it?

1. The bandgap (0.66 vs 1.12 eV). Germanium carries thousands of times more intrinsic carriers than silicon, meaning thousands of times more leakage, growing exponentially with temperature. Fine on a lab bench, useless in an engine bay or in military hardware. Silicon holds up to junction temperatures in the 150 °C class.

2. Availability. Silicon is the second most abundant element in the Earth's crust after oxygen — functionally, it's sand. Germanium is scarce, recovered in small quantities as a by-product of other metal refining. Silicon's 1414 °C melting point matters too: it survives the high-temperature steps that came later.

3. The decider was the oxide. Expose silicon to hot oxygen and SiO₂ grows on the surface by itself. That film is a genuine insulator with a ~9 eV gap; it's hard, it doesn't dissolve in water, it forms a well-behaved interface with the silicon beneath — and it blocks dopant diffusion. Which unlocks a procedure: grow oxide → photolithographically open windows only where you want them → drive dopants in only there. That is the planar process, established by Jean Hoerni at Fairchild in 1959, and every integrated circuit since descends from it. Germanium's oxide, GeO₂, dissolves in water and supports none of this.

So the answer comes out backwards from what you'd expect. Silicon won not because silicon was the best material, but because silicon's oxide was. On electron mobility alone, gallium arsenide beats it; for emitting light, direct-gap GaAs and GaN win outright, while indirect-gap silicon is so bad at emitting photons that there is no silicon LED. Silicon became the foundation of mass production anyway, because it was the only material where the entire manufacturing process closed. A systems victory, not a materials one.

SiO₂'s reign wasn't absolute either: as scaling thinned the gate dielectric toward 1 nm, electrons began tunnelling through it, and from around 2007 high-permittivity hafnium-based oxides (high-k) took over the gate. What Moore's Law Actually Says follows that history.

How this gets used in practice

Who touches it, and when. Device and process engineers setting electrical targets for a new process. Power-electronics designers selecting and derating parts. Solar cell and LED designers choosing a material at all. Embedded hardware engineers land here too, the moment they commit to an operating temperature range.

The parameters you actually type. Doping isn't specified as a "concentration" — it's dose (total implanted ions, cm⁻²) plus implant energy (keV, which sets depth), and the spec isn't complete without the anneal (RTA) that follows. You verify with sheet resistance (Ω/sq, four-point probe) and a SIMS depth profile. On the datasheet side you're reading $E_g$ , $n_i$ , reverse leakage $I_R$ , junction temperature $T_j$ , and subthreshold swing (SS, mV/decade). Simulation happens in TCAD (Sentaurus, Silvaco); the circuit side lives in SPICE BSIM model parameters.

The traps that turn into incidents.

Leakage grows exponentially with temperature. Qualify only at room temperature and the part can leak an order of magnitude more at real in-enclosure temperature. Derating is not "leave some margin" — it's an exponential estimate.
The 60 mV/decade wall. Subthreshold swing cannot go below $\ln(10)\cdot k_B T/q \approx 60$ mV/decade at room temperature. That floor comes straight from the $k_B T$ in equation (1); no circuit trick beats it, which is why "just lower the supply voltage further" keeps getting rejected.
Thermal budget. Dopants keep diffusing through every subsequent hot step. Design against the as-implanted profile and the finished wafer won't match.
More doping is not simply faster. Higher doping buys carriers but costs mobility (impurity scattering) and lowers breakdown voltage. A low-resistance contact and a high-voltage drift layer want doping at opposite ends of the range.
Don't pick the wrong material. Indirect-gap silicon will never be a light emitter. Conversely, automotive inverters needing high blocking voltage and hot operation reach for wide-gap SiC (~~3.3 eV) or GaN (~~3.4 eV) — a wider gap means a stronger breakdown field and far less leakage when hot.

How it shows up in interviews and design reviews. "Why silicon?" — the answer is only complete if you reach the oxide and the process. "How does $n_i$ vary with temperature?" — exponentially, roughly doubling every 10 °C. "Where does 60 mV/decade come from?" — the $k_B T$ in the Fermi distribution. Those three reveal directly whether band theory is a tool you can use or a topic you once read about.

Summary

A band is what happens when atoms crowd together and one shelf splits many ways; a gap is the stripe of energies where electron waves get reflected by the lattice and no seats exist
Insulators and semiconductors don't differ in kind — only in staircase height
Carrier count depends exponentially on the gap, so temperature and material choice both move things by decades, not percentages
Doping shifts resistivity six decades using one part per million of impurity, and $np = n_i^2$ means whatever you gain in one carrier you lose in the other
Silicon won on the strength of SiO₂ — an improbably convenient oxide — not on the merits of the element itself

The natural next step is to bond n-type against p-type and put a gate on top: MOSFETs from the Ground Up connects the switching mechanism to the leakage we met here. For how this physics is actually imposed on a wafer, see How Chips Are Made.