Avalanche Photodiode (APD)

Impact ionisation — gain inside the semiconductor

A regular PIN photodiode converts each absorbed photon into one electron-hole pair, and the photogenerated carrier traverses the depletion region under a modest reverse bias to produce a photocurrent. One photon in, one electron out. Quantum efficiency might be 80 percent; responsivity at 1550 nm is roughly 1 A/W.

An avalanche photodiode does the same thing — and then keeps going. By raising the reverse bias to within a few volts of the device's breakdown voltage, the electric field in a thin "multiplication region" near the p-n junction climbs to 10⁵ to 5×10⁵ V/cm. At those field strengths the photogenerated electron is accelerated to kinetic energies above the semiconductor bandgap before its next phonon-scattering event. When the energetic electron collides with a lattice atom, it transfers enough energy to ionise the atom — creating a new electron-hole pair from a previously bound valence electron. The new carriers are then accelerated by the same field and can trigger further ionisations.

Photon → 1 primary e-h pair → impact ionisation cascade
                                    ▼
                  e⁻ → → →   |     |
                              e⁻    h⁺
                              │     │
                              ▼     ▼
                    e⁻ ─ e⁻ ─ e⁻ ─ e⁻ ─ e⁻ ──→ contact
                            h⁺   h⁺   h⁺ ──→ contact

  Multiplication factor M = (total carriers out) / (initial pair)
  Typical: M = 10 to 100 for telecom APDs
           M = 1000 for low-light detectors
           M → ∞ above breakdown (SPAD / Geiger mode)

The number of secondary pairs produced per primary photoelectron is the multiplication factor M. M is determined by the bias voltage, the temperature, and the device geometry. For a high-quality silicon APD running at 0.9 × V_br at 25 °C, M = 50 is comfortably reproducible; for an InGaAs APD at 1550 nm, M = 10-20 is typical because the higher k-factor (ionisation-rate ratio) makes higher gains too noisy. M is one of the two key APD figures of merit along with the noise floor.

Why internal gain beats external amplification

Every photonic receiver has a noise floor set by the first electronic amplifier stage following the photodiode. The amplifier adds voltage and current noise that cannot be reduced below the device-physics limits of its input transistors. A typical 10 GHz transimpedance amplifier has an input-referred current noise of a few pA/√Hz; an equivalent PIN photodiode plus that amplifier has a noise-equivalent power (NEP) on the order of 10⁻¹² W/√Hz at 1550 nm and 10 GHz bandwidth.

An APD bypasses that limit by amplifying the signal inside the photodiode, before the amplifier ever sees it. If the photocurrent is multiplied by M, the amplifier sees M times the signal but its own noise is unchanged. The effective NEP — which is what determines the minimum detectable optical power — improves by roughly M, down to 10⁻¹⁴ W/√Hz for telecom-grade APDs.

PIN + TIA:    Signal = I_ph                Noise = i_n,TIA          → NEP ~ 10⁻¹² W/√Hz
APD + TIA:    Signal = M · I_ph            Noise = i_n,TIA          → NEP ~ 10⁻¹⁴ W/√Hz
                                            (plus √F(M) excess noise term)

The catch is that multiplication itself is statistical — each photoelectron does not produce exactly M secondary pairs but a random number with mean M and significant variance. This shot-noise excess is captured by the excess noise factor F(M), defined so that the noise current squared is 2qI_ph M² F(M) Δf. F(M) > 1 always, but is well-controlled in silicon (F ≈ M^0.3 at moderate gain) and worse in InGaAs (F ≈ M^0.7). The optimum operating M maximises signal-to-noise ratio = M · I_ph / √(2qI_ph M² F(M) Δf + i_n,TIA²) — typically M_opt ≈ 50-100 for silicon, M_opt ≈ 10-20 for InGaAs.

Device structure — SAM and SACM layers

Modern telecom and LIDAR APDs use Separate Absorption and Multiplication (SAM) structures: one layer of narrow-bandgap material absorbs the light efficiently at low field; a separate adjacent layer of wider-bandgap material has the high field that does the multiplying. This separates the optical-absorption job from the electrical-multiplication job, which lets each layer be optimised independently.

SAM-structured InGaAs/InP APD (telecom, 1550 nm)

           anode (p++)                   ← top contact
              │
           InP charge layer              ← controls field in InGaAs
           │
   high E  InP multiplication layer      ← impact ionisation here
              │
   low E   InGaAs absorption layer       ← photon absorption here
              │
           InP buffer
              │
           InP substrate (n++)           ← bottom contact

   Operation: photon absorbed in InGaAs creates primary e-h pair;
   electron drifts up into InP multiplication layer where the field
   is 10⁵ V/cm and triggers cascade; secondary carriers reach contacts.

The reason SAM exists is twofold. First, the bandgap of InP (1.35 eV) is too large to absorb 1550 nm photons (0.8 eV), so you need narrow-gap InGaAs for the absorption — but InGaAs has a much higher dark current than InP would, so you don't want to multiply in InGaAs. Second, the ionisation-rate ratio in InP is far more favourable than in InGaAs, giving lower F(M). SAM puts each job in the right material.

Some modern designs add a Separate Absorption, Charge, and Multiplication (SACM) structure that inserts a thin "charge layer" between absorption and multiplication, allowing precise control of the field in each region by adjusting the doping. SACM is the current state-of-the-art for telecom InGaAs APDs (Hamamatsu, Discovery Semiconductors, Thorlabs).

Worked example — SNR comparison at 100 nW input

Consider a 1550 nm receiver detecting a 100 nW (-40 dBm) optical signal at 1 GHz bandwidth, with a 1 A/W photodiode (PIN or APD).

Signal photocurrent: I_ph = 100 nW × 1 A/W = 100 nA

Case 1: PIN photodiode
  Shot noise:    i_shot² = 2 · q · I_ph · Δf = 2 · 1.6e-19 · 100e-9 · 1e9 = 3.2e-17 A²
                 i_shot  = 5.7 nA RMS
  TIA noise:     i_TIA   = 3 pA/√Hz · √1 GHz = 95 nA RMS
  Total noise:   ≈ 95 nA (TIA-dominated)
  SNR:           100 nA / 95 nA = 1.05 (≈ 0.4 dB)
  → marginal — not enough margin for low BER

Case 2: APD with M = 30, F(M) = 4
  Multiplied signal:  M · I_ph = 30 × 100 nA = 3 μA
  APD shot noise:     2qI_ph M²F Δf = 2 · 1.6e-19 · 100e-9 · 900 · 4 · 1e9 = 1.15e-13 A²
                      = 0.34 μA RMS
  TIA noise:          95 nA RMS (unchanged)
  Total noise:        √(0.34² + 0.095²) ≈ 0.35 μA RMS
  SNR:                3 μA / 0.35 μA = 8.6 (≈ 9.3 dB)
  → 9 dB improvement → 10× margin → comfortable BER < 10⁻¹²

Same photocurrent, same TIA, same wavelength — the APD gives 9 dB better SNR. That nine-decibel improvement is the difference between a 40 km fibre span and a 60 km span, between a 100 m LIDAR and a 200 m LIDAR, between a flawed quantum-key-distribution link and a working one. Internal gain is the receiver designer's most powerful lever.

PIN vs APD vs SPAD — when each is right

Property	PIN photodiode	APD (linear mode)	SPAD (Geiger mode)
Operating bias	5-20 V reverse	30-400 V (just below V_br)	3-10 V above V_br
Gain	1	10-100 typical, up to 1000	10⁶ (digital pulse)
Output type	Analog photocurrent	Analog photocurrent × M	Digital pulse per photon
NEP at 1 GHz	10⁻¹² W/√Hz	10⁻¹⁴ W/√Hz	not applicable (photon-counting)
Single-photon sensitivity	No	No (analog)	Yes (one photon = one pulse)
Bandwidth	1 MHz – 100 GHz	1 MHz – 50 GHz	Limited by dead time (~10-100 ns)
Stability requirement	Loose	Tight (gain temp-coeff)	Moderate (above-breakdown)
Typical use	Short-range fibre, photography, solar cells	Long-haul fibre, LIDAR, X-ray	Quantum communication, FLIM, dToF LIDAR

Silicon vs InGaAs — wavelength dictates the choice

Material	Wavelength range	k-factor	V_br typical	M_opt	Where used
Silicon	300-1100 nm	0.005-0.04	100-400 V	50-150	905 nm LIDAR, visible, X-ray scintillator coupling
InGaAs / InP	900-1700 nm	0.3-0.5	20-70 V	10-20	1310/1550 nm fibre, 1550 nm LIDAR, quantum networking
Germanium	800-1700 nm	0.7-1.0	20-40 V	5-10	Legacy 1300 nm receivers, niche IR sensing
InGaAsP	1100-1600 nm	0.3-0.6	30-70 V	10-20	Tunable telecom receivers
HgCdTe	2-12 μm	~0	10-40 V	500-1000	Mid-infrared sensing, astronomy

Silicon is the workhorse below 1.1 μm with extraordinarily low k-factor — meaning silicon APDs can run very high gain (M up to 150) before excess noise destroys SNR. This is what makes 905 nm silicon APDs the preferred choice for mid-range automotive LIDAR (Luminar, Innoviz, Aeva): the gain margin compensates for atmospheric attenuation. InGaAs takes over above 1.1 μm because silicon is transparent there — InGaAs has a higher k-factor and worse F(M), so usable M is smaller, but the wavelengths it covers are exactly the telecom and eye-safe bands where every long-haul fibre system needs detectors.

SPAD and silicon photomultipliers

Operating an APD above its breakdown voltage puts it in Geiger mode: any single photoelectron triggers a self-sustaining avalanche that drives the device current to milliamps within nanoseconds, producing a digital pulse rather than analog gain. The avalanche is then quenched (current dropped to allow recombination) by an external circuit before the device can detect another photon. The dead-time between pulses is tens to hundreds of nanoseconds.

SPADs are photon-counting detectors. Each pulse is one detected photon; the count rate is the photon arrival rate (within saturation limits). They're used in:

• Direct-time-of-flight LIDAR. Apple's iPhone Pro LiDAR module uses a SPAD array sensor (Sony IMX556) that detects single returning photons and records their arrival time. The Sony IMX series is now dominant in automotive dToF.
• Quantum key distribution. InGaAs SPADs (ID Quantique, Aurea Technology, Princeton Lightwave) detect single photons at 1550 nm for QKD over fibre. Cooling to -40 °C plus active gating reduces dark counts below 10² per second.
• Fluorescence-lifetime imaging microscopy (FLIM). SPAD arrays with picosecond timing detect single fluorescent-photon emissions and bin them by arrival time for lifetime histograms.
• Time-correlated single-photon counting (TCSPC). The technique behind LIDAR ranging, FLIM, and many quantum-optics experiments. PicoQuant's HydraHarp uses SPAD inputs at picosecond resolution.

A silicon photomultiplier (SiPM) is an integrated array of hundreds to thousands of small SPADs, all connected to a common output. Each microcell fires independently when hit; the total output current is the sum across the array. SiPMs have replaced photomultiplier tubes in many applications (PET scanners, gamma cameras, calorimeters) because they're solid-state, magnetic-field-tolerant, and lower-voltage. Hamamatsu MPPC and SensL FBK are the dominant SiPM suppliers.

Where APDs ship today

• Long-haul fibre receivers. Every 10G, 100G, 400G, and 800G coherent or direct-detect receiver above the 80 km span uses an InGaAs APD as front-end detector. Ciena, Cisco, Huawei, and Nokia DWDM long-haul interfaces all incorporate Hamamatsu G14858 / Thorlabs APD400C / Discovery DSC-50S series.
• Automotive LIDAR. 905 nm silicon APDs (linear mode) in Luminar Iris, Innoviz One, Aeva 4D LIDAR. 1550 nm InGaAs APDs in eye-safer systems like Luminar Halo. SPAD arrays (Sony IMX series) in lower-cost short-range solid-state LIDARs.
• Optical time-domain reflectometers (OTDR). Field instruments that test optical fibres for breaks and attenuation by injecting a laser pulse and detecting back-scattered light. The detector is invariably an InGaAs APD because the back-scattered signal is 70 dB weaker than the input pulse.
• X-ray spectroscopy and medical imaging. Silicon APDs paired with scintillators (LYSO, CsI(Tl), BGO) detect single X-ray photons in PET scanners, gamma cameras, and luggage scanners. Hamamatsu S8664 is the canonical part. APD-coupled scintillators replaced bulkier photomultiplier tubes in 2010s-era PET designs.
• Quantum communication. InGaAs SPADs detect single photons at 1550 nm for QKD; commercial systems from ID Quantique (Geneva) and Toshiba (Cambridge) all use cryogenically cooled SPAD detectors with gated bias for dark-count suppression.
• Gamma-ray astronomy. SiPM arrays in the H.E.S.S., MAGIC, and CTA imaging atmospheric Cherenkov telescopes detect the few-nanosecond Cherenkov flashes from cosmic ray air showers. Hamamatsu MPPC arrays at -40 °C give the sensitivity needed.
• Laser rangefinders. Military and surveying laser rangefinders use 1550 nm InGaAs APDs at km-class ranges; sport laser rangefinders use 905 nm silicon APDs at the 1000-yard golf-rangefinder class.

Pitfalls and failure modes

• Excess bias driving runaway. Gain M rises exponentially as the bias approaches V_br. A small over-voltage transient can drive M past the device's safe-operation limit, producing currents that locally heat the multiplication region, which lowers V_br, which raises M further — thermal runaway. Modern APDs include integrated thermistors and the bias supply has tight current limiting and overvoltage clamps; cheap designs without these die randomly.
• Temperature coefficient of V_br. Silicon APDs have a V_br temperature coefficient of about 0.1 V/°C; InGaAs APDs are similar. A 25 °C ambient swing changes V_br by 2.5 V, which changes M by 30-50 percent at constant bias. Mitigation: temperature-stabilise the device with a TEC (thermo-electric cooler), or use a bias controller that tracks V_br via gain calibration.
• Dark current at high temperature. Dark current doubles roughly every 10 °C in silicon and every 8 °C in InGaAs. At 60 °C ambient an InGaAs APD's dark current may be 10× its 25 °C spec, and it gets multiplied by M just like signal. The receiver SNR collapses. Mitigation: TEC cooling to 0 °C or -20 °C in extended-temperature applications.
• Quenching circuit too slow in SPADs. If a SPAD avalanche is not quenched within the device's safe-current-time product, the multiplication region overheats and the device drifts in V_br or fails outright. Passive quenching (a 100-500 kΩ series resistor) is simple but slow (50-200 ns); active quenching (a fast comparator and switch) gets dead-time below 20 ns but is more complex and adds noise.
• Afterpulsing in SPADs. Charge trapped in defects during an avalanche can re-trigger the device with a few microseconds delay — afterpulsing. Cryogenic cooling reduces it; gated operation (only enabling detection in known time windows) eliminates it. Quantum-key-distribution systems gate their SPADs to a few ns around expected photon arrival times.
• Optical crosstalk in SiPM arrays. Each microcell avalanche emits a few photons that can trigger neighbours, leading to correlated multi-cell events that look like a brighter pulse than there really was. Mitigation: optical trenches between microcells (Hamamatsu's TSV-Trench technology), passive optical coatings, or correction algorithms in firmware.

A short history — from McKay to single-photon LIDAR

Avalanche multiplication in semiconductor diodes was first observed and modelled by Kenneth McKay at Bell Labs in 1954-1955 — the same year and the same lab that produced the first silicon solar cell. McKay's measurements of impact ionisation rates in silicon and germanium established the parameter that bears his name (the McKay theory of avalanche multiplication). The first practical avalanche photodiodes appeared in the early 1960s for nuclear-particle detection.

Telecom-grade InGaAs APDs for fibre receivers were developed through the 1980s as long-haul fibre systems pushed toward 1.3 and 1.55 μm wavelengths where silicon was blind. By the early 1990s SAM-structured InGaAs/InP APDs were the standard front-end of every transoceanic fibre receiver. Silicon APDs in the same era found their footing in nuclear medicine (PET scanners) and laser rangefinding.

The 2010s saw SPADs move from physics labs into automotive and consumer products. Sony's stacked-BSI CMOS SPAD-array sensors (IMX series, starting with the IMX458 in 2018) put millions of SPAD pixels into LIDAR modules and smartphones. Apple's 2020 iPhone 12 Pro LiDAR uses an evolved Sony SPAD-array sensor for dToF depth mapping. SiPMs replaced photomultiplier tubes in nuclear-medicine imaging through the 2010s-2020s as cost dropped 10× and reliability improved. The frontier today is single-photon-sensitive sensor arrays for quantum networking, LIDAR, and biomedical imaging at scale.