🔬 Build 05 · AMO Lab Operations Handbook

AMO Lab Operations Handbook

A field guide for the operations that make an AMO experiment work: align light, prepare atoms, drive RF and microwaves, lock lasers, debug failures, and turn lab intuition into numbers.

Use it as an operations handbook

Start from the lab problem, then open the matching drawer.

AMO technique pages are easiest to use when organized by what you are trying to do at the table, not by component names. These tabs now map to common lab actions.

01Align light

Fiber coupling, AOM geometry, polarimetry, beam delivery, and the diagnostics that reveal bad modes.

02Prepare atoms

Optical pumping, stretched states, repump logic, and the selection rules that decide where population goes.

03Drive transitions

RF antennas, microwaves, impedance thinking, safety, and what power actually reaches the atoms.

04Stabilize and optimize

Laser locks, control scripts, Bayesian optimization, QuTiP models, and failure patterns worth remembering.

1 · Optical Fiber Coupling
Mode-matching visualiser — beam waist animated into fiber
Almost every laser beam in an AMO experiment is delivered to the optical table via optical fiber. This is not merely a convenience: fibers spatially filter the beam, mechanically decouple the laser table from the experiment table, and allow laser sources to be swapped without realigning downstream optics. The practical consequence is that fiber coupling efficiency determines beam quality at the atom.

Three fiber types appear in a typical lab:
  • Single-mode (SM), supports only the fundamental HE₁₁ mode; any higher-order input content is rejected. Essential wherever spatial coherence matters.
  • Polarisation-maintaining (PM), SM fiber with stress rods that introduce birefringence; preserves the input polarisation when aligned to the principal axis. Every beam requiring a defined polarisation at the atom must travel on PM fiber.
  • Multimode (MM), large core, easy to couple; used for wavemeter pick-offs, diagnostics, and detection paths where polarisation is unimportant.
Mode-matching: the key formula + calculator

A lens of focal length f converts an input Gaussian beam of 1/e² radius w_in into a focused waist:

$$w_0 \approx \frac{\lambda f}{\pi w_{\rm in}}$$

The coupling condition is w₀ = w_f, where w_f = MFD/2 is the fiber mode radius. Solving for the required focal length:

$$f = \frac{\pi w_{\rm in} w_f}{\lambda}$$

Trade-off: shorter f → tighter focus, higher peak efficiency, stricter alignment tolerances. Longer f → more robust over temperature drifts.

Target: 50–80% coupling into SM/PM fiber; >80% into MM. PM fiber extinction ratio target: >20 dB (P∥/P⊥ > 100), meaning <1% power in the wrong polarisation axis.

🔧 Mode-matching calculator

Formula: f = π w_in w_f / λ  ·  Verify with VNA or throughput measurement

PM fiber alignment procedure
  1. Use a PBS to prepare clean linear polarisation upstream.
  2. Rotate a HWP after the PBS to align the polarisation to the fiber's keyed (slow/fast) axis.
  3. Optimise coupling for maximum throughput.
  4. Measure extinction ratio (ER) at the output through an analyser PBS.
  5. Iterate HWP angle to maximise ER.
Target ER > 20 dB. Residual polarisation impurity (leakage into the orthogonal axis) is the dominant limit on optical-pumping fidelity and on the σ⁺–σ⁻ purity of gray-molasses imaging beams.
📦 Thorlabs, SM & PM Fibers
SM-780HP (780 nm, MFD 5.0 μm) · PM780-HP (PM, 780 nm) · P1-630A-FC (630 nm)
📦 Thorlabs, Coupling Lenses
C220TMD (f=11 mm) · C230TMD (f=4.51 mm) · C240TMD (f=8 mm)
2 · Double-Pass Acousto-Optic Modulators (AOMs)
Cat's-eye double-pass AOM (Fig. 3). Laser → PBS → lenses → AOM: 0th order blocked, +1st order (ν₀+f) deflected at Bragg angle θ_B → QWP → cat's-eye lens ↔ mirror → retroreflects → AOM again (+f) → lens → PBS. After the double QWP pass the beam is H-pol; PBS deflects it to FC. Output: ν₀+2f, direction unchanged, frequency-agile.
An AOM diffracts light off a moving acoustic grating, shifting the optical frequency by f_RF while deflecting the beam by the Bragg angle. The problem with a single pass: the Bragg angle depends on f_RF, so tuning the frequency also steers the beam. This coupling is unacceptable for precision experiments.

The solution is double-passing: retroreflect the first-order beam back through the same crystal. The frequency shift doubles to 2 f_RF and the angular deviations cancel exactly, the output beam direction is independent of f_RF. Every tunable beam in the experiment uses a double-pass AOM.
Cat's-eye retroreflector geometry

The most compact double-pass geometry:

  1. AOM crystal → first-order beam deflected by Bragg angle
  2. Lens (f = one focal length from crystal) collimates and focuses
  3. QWP + mirror at one focal length from lens, retroreflects beam
  4. QWP traversed twice → polarisation rotated 90°
  5. Second pass through AOM → frequency shifted by another f_RF
  6. PBS transmits the 2f_RF output (orthogonal polarisation) and rejects zero-order

The cat's-eye geometry makes retroreflection insensitive to mirror tilt, which is why it is preferred over simpler flat-mirror configurations.

Typical performance
~90%
Single-pass eff.
~80%
Double-pass eff.
40–50 dB
Extinction (AOM off)
Longitudinal vs. shear-wave AOMs
PropertyLongitudinalShear-wave
Acoustic modeAtoms ∥ propagationAtoms ⊥ propagation
Sound velocity~4–6 km/s~1–2 km/s
Deflection angleLarger (AOD use)Smaller
RF bandwidthBroaderNarrower
Diffraction eff.Lower at peakHigher at peak
In the lab: Cooling/imaging beams → shear-wave AOM. Tweezer array generation → longitudinal AOM (AOD), because large deflection range is needed to position hundreds of tweezers.

🔧 AOM switching-time calculator

Formula: t_rise = d_beam / v_s  ·  Focus tighter to get faster switching

📦 AOM vendors
Gooch & Housego · AA Opto-Electronic (MT series) · Isomet
📦 EOM vendors (fast switching)
Qubig · EOSpace · Jenoptik
3 · Polarimetry, Stokes Parameters & the Poincaré Sphere
Polarisation purity is one of the most critical and least forgiving parameters in AMO experiments. A 1% admixture of the wrong circular component in an optical-pumping beam can reduce state-preparation fidelity from >99% to <95%. A few percent of linear polarisation in a σ⁺–σ⁻ gray-molasses beam opens decoherence pathways. Measuring polarisation quantitatively, not just qualitatively, is essential.

Stokes parameters

The complete polarisation state is described by four intensity measurements:

ParameterMeaning
S₀ = I_totalTotal power
S₁ = I_H − I_VHorizontal vs. vertical linear
S₂ = I₊₄₅ − I₋₄₅Diagonal linear
S₃ = I_RCP − I_LCPRight vs. left circular
$$P = \frac{\sqrt{S_1^2 + S_2^2 + S_3^2}}{S_0}$$

P = degree of polarisation (1 = fully polarised)

Rotating-QWP measurement method

Place a QWP on a rotation stage before a PBS analyser. As QWP angle θ is swept, transmitted intensity follows:

$$I(\theta) = \tfrac{1}{2}(A + B\sin 2\theta + C\cos 4\theta + D\sin 4\theta)$$ $S_0 = A - C \;\mid\; S_1 = 2C \;\mid\; S_2 = 2D \;\mid\; S_3 = B$

Use 50–100 evenly spaced angles for reliable results.

QWP retardance calibration

$$\delta = 2\arccos\!\left(\sqrt{\frac{I_{\rm min}}{I_{\rm max}}}\right)$$

Measure between two PBS ports with linear input. Any deviation from 90° is corrected in the fitting procedure. Calibrate every QWP in optical-pumping and imaging paths.

Poincaré Sphere: RCP (σ⁺) sits at the north pole (S₃ = +1), LCP (σ⁻) at the south pole (S₃ = −1). All linear states lie on the equator. Elliptical states are on the surface. A pure state has |S| = S₀; partially polarised light has |S| < S₀.
4 · Laser Frequency Stabilisation
🔐

Laser frequency stabilisation (SAS, beat-note, and PDH cavity locking) is covered in full on the dedicated Laser Locking page, including interactive error-signal simulators, a lock-hierarchy diagram, and cavity design calculators.

Laser Locking page →
Optical pumping drives an atom into a specific Zeeman sublevel via repeated photon absorption and spontaneous emission cycles. In a tweezer experiment, its purposes are twofold:
  1. State preparation, put the atom in a single, well-defined |F, m_F⟩ state before any coherent manipulation.
  2. Detection preparation, define the initial condition for state-selective fluorescence imaging.

Imperfect optical pumping is a direct source of systematic error in lifetime measurements, qubit state detection, and gate fidelity.

Cesium: pumping to |F=4, m_F=+4⟩

Scheme: σ⁺-polarised light driving F=4 → F′=4 has no allowed absorption for an atom already in m_F=+4 (which would require Δm_F=+1, but no m′_F=+5 exists in the excited state). That state is dark; all other m_F sublevels are continuously depopulated until the atom accumulates in m_F=+4.

Repumper: A beam resonant with F=3→F′=4 returns any population that decayed to F=3.

Why the magnetic-field direction matters: The quantisation axis is defined by the local B-field, not the laser k-vector. If the bias field is misaligned with the beam, the σ⁺ in the lab frame decomposes into σ⁺, π, and σ⁻ in the atom's frame. Apply a bias field of ~6 G along the optical axis of the pumping beam.

Diagnostic: the depumping-ratio test

  1. Drive F=4→F′=4 without the repumper, atoms in F=4 scatter photons and heat out of the trap.
  2. Under aligned B-field: atoms in the dark state |4,+4⟩ survive for ~1 ms (off-resonant scattering only).
  3. Under deliberately misaligned B-field (~45°): σ⁺ acquires σ⁻ component → atoms depumped in ~10 μs.
  4. Target depumping ratio > 100 (aligned/misaligned survival times).

Adjustment knobs: laser frequency (exact line centre), QWP orientation (polarisation purity), bias field direction. Together these achieve >99% pumping fidelity.

Quick diagnosis checklist:
✅ ER of pumping beam fiber > 20 dB
✅ Bias field coil along pump beam axis
✅ Pump laser on F=4→F′=4 (not F=3→F′=4)
✅ Repumper on F=3→F′=4
✅ Depumping ratio > 100
✅ Pump pulse duration > 5 × (1/Γ_scatter)
📦 Bias coil drivers
Wavelength Electronics low-noise current sources · itech programmable supplies
📦 Key optics
Thorlabs PBS cubes · Thorlabs QWPs · Thorlabs HWPs

Lithium: D2-line optical pumping

Challenge: The ⁶Li D2 excited-state hyperfine splitting is only ~5 MHz, comparable to Γ. Different F′ components overlap, making it impossible to address a single excited hyperfine level cleanly. In practice, a combination of F=3/2→F′=3/2 and F=1/2→F′=3/2 light is used.

Procedure

  1. Coarse beam alignment to MOT on the diagonal camera.
  2. Fine alignment on a single trapped atom.
  3. Verify resonance by scanning laser frequency over the atom-loss signal.
  4. Characterise fidelity with the depumping-ratio test (same as Cs).
Advantage: Ground-state hyperfine splitting of 228 MHz ≫ D2 linewidth → state-selective probing is clean and free of cross-talk.
Practical note: In a dual-species (Li+Cs) setup, the Li pumping beams co-propagate with the Cs beams; beam geometry and PM-fiber infrastructure are shared, so only minor frequency and alignment adjustments are needed once Cs pumping is optimised.
📄 References · Gehm (2003), Properties of ⁶Li (NCSU) · Steck, Cs D-line data
Coherent control of atomic hyperfine states requires delivering oscillating magnetic fields at precise frequencies, 228 MHz for ⁶Li ground-state splitting, ~76 MHz in the Paschen-Back (high-field) regime, 9.2 GHz for Cs. The engineering challenge is delivering enough B-field amplitude at the atom while fitting inside an existing vacuum apparatus. Safety-critical high-current electronics (Feshbach coils) demand dedicated hardware interlocks that are independent of the computer control system.

Loop antenna for Li hyperfine transitions

A single-turn loop radiates primarily through its magnetic dipole field. The radiation resistance of a small loop of area A at frequency f is:

$$R_{\rm rad} \approx 20\pi^2\left(\frac{2\pi A}{\lambda^2}\right)^2 \;\Omega$$

For any loop that fits near a vacuum cell, R_rad ≪ 50 Ω. Efficient power delivery from a 50 Ω source therefore requires an impedance-matching network.

Three approaches tested (with VNA)

  1. Capacitive loading, series or parallel capacitor shifts resonance. Result: parallel ~47 pF proved most effective for 76 MHz, compact geometry.
  2. Transmission-line stub matching, moves impedance on Smith chart. Works in principle, but spurious resonances can complicate things.
  3. Discrete LC networks, more design freedom but more components.

Achieved: reflection minimum ~10 dB, sufficient B-field at the atom with ~100 W amplifier.

Lessons learned

  • Lead length (connector to loop) contributes parasitic inductance at 100 MHz, non-negligible.
  • Simulate with SimSmith before building every iteration.
  • The resonant frequency scales inversely with circumference at fixed inductance.

🔧 Loop antenna quick estimate

Approximate formulas, always verify with VNA measurement.

📦 Tools & instruments
Siglent SVA1015X VNA · SimSmith (free Smith chart simulator) · Mini-Circuits RF amplifiers

Feshbach coil safety interlocks

Feshbach coils produce fields of order 1000 G by carrying large DC currents. Sudden current interruption generates inductive voltage spikes that can damage power supplies and coils.

Hardware interlock logic (essential rules)

  • Monitor: coil temperature, current level, supply voltage
  • On threshold exceeded: ramp down smoothly, do not switch off abruptly
  • All interlock logic implemented in relay hardware, independent of computer control
  • Interlock circuit must be untriggerable by software bugs
⚠️ Never bypass a hardware interlock for convenience. Inductive spikes from sudden coil switching have destroyed power supplies and cracked coil epoxy in labs worldwide. The hardware interlock is not optional. A software interlock fails if the computer crashes. A relay cannot be bypassed by software.
📦 High-current supplies
Kepco BOP bipolar · iSeg precision current sources · AMETEK Programmable Power
Optimising an ultracold-atom experiment means navigating a high-dimensional continuous parameter space where each measurement takes seconds to minutes. A MOT has ≥5 knobs; a tweezer experiment adds trap depth, cooling polarisation, B-field direction. Naïve grid search over 5 variables × 10 points = 10⁵ trials, months of run time.

How it works

Gaussian-Process Regression (GPR): a non-parametric Bayesian model that maintains a probabilistic map of the response surface (e.g. atom survival vs. laser detuning + power). At each step it returns a predicted value and an uncertainty; unexplored regions have high uncertainty.

Acquisition function: selects the next measurement point by trading off exploitation (sample near current optimum) and exploration (reduce uncertainty in poorly sampled regions).

Acquisition functionWhen to use
Expected Improvement (EI)Default; works well near optimum
Upper Confidence Bound (UCB)When signal is absent, need broad search
Probability of Improvement (PI)Conservative; avoids risk

Workflow

  1. Collect 5–20 initial points from Latin hypercube or random design
  2. Train GPR model; inspect posterior mean + uncertainty
  3. Select next point by maximising EI/UCB
  4. Run experiment, append data, retrain, repeat
Convergence: 30–100 evaluations for 4–8 dimensional spaces (vs. thousands for grid search).
📦 Software
scikit-learn GPR · BayesianOptimization (Python) · Meta Ax platform · Frazier (2018), BO tutorial
GP posterior demo, atom survival vs. cooling detuning

Grey dotted = true function · Purple band = GP ±2σ · Yellow = observations · Green dashed = next query point

The master-equation simulations in the Li and Cs chapters are implemented in QAtomTweezer, a purpose-built Python library on top of QuTiP. The key idea is to treat a single atom with full hyperfine structure coupled to a 1D harmonic oscillator representing one motional axis of the tweezer.

Physical model

$$\hat{H} = \hat{H}_{\rm internal} \otimes \hat{H}_{\rm HO}$$
  • H_internal: all hyperfine and Zeeman sublevels (typically 12–24 states for D1 of an alkali)
  • H_HO: harmonic oscillator, truncated at N_HO = 10–20 Fock states
  • Total dimension: d_int × N_HO (e.g. 12 × 12 = 144 for Li D1)

Key feature: the full matrix exponential is used for the recoil operator R̂ = e^(iη(â+â†)) (not the Lamb-Dicke expansion), so the code is valid beyond the strict Lamb-Dicke regime.

3 validation cross-checks

  1. Fock distribution P(n), fit to Boltzmann to extract T_eff
  2. Excited-state fraction p_e, compare to measured photon rate
  3. Temperature minimum, verify location in 2D parameter scans

Code structure

# Entry points:
QAtomTweezer.py
QAtomTweezer_SingleLevel.py

# Main callable:
SteadyStateTweezer(
    x,      # [δ1, δ2, Ω1, Ω2, φ1, φ2]
    wh,     # trap freq in units of Γ
    Nh,     # HO truncation
    atom,   # AtomSettings object
    eta,    # Lamb-Dicke parameter
    pol,    # polarisation config
)
# Returns: ⟨n⟩, P(n), p_e

Performance

1–5 s
per steady-state eval (144×144)
10–30 min
2D scan 50×50 (joblib)
📦 Software stack
QuTiP documentation · joblib (parallelism) · Johansson et al. (2013), QuTiP 2 paper
AMO experiments for Li–Cs tweezer work require four distinct laser wavelengths, each with its own technical demands. All are produced by external-cavity diode lasers (ECDLs) in Littrow configuration, except the 1064 nm tweezer which uses a commercial Nd:YAG/fiber laser.
External-Cavity Diode Lasers (ECDLs), general principles

A bare semiconductor diode lases on multiple longitudinal modes. Two tuning mechanisms in a bare diode:

  • Injection current: ∂ν/∂I ~ 1–3 GHz/A (fast but noisy)
  • Temperature: ∂ν/∂T ~ −20 to −40 GHz/K (slow, hysteretic)

Neither alone provides the narrow linewidth (<100 kHz) or mode-hop-free tuning range (>1 GHz) required.

The Littrow ECDL solution

  • Holographic grating at Littrow angle feeds first-order diffraction back into the diode
  • Selects one longitudinal mode of the extended cavity
  • Grating angle (PZT-tuned) + injection current → mode-hop-free tuning over 1–2 GHz
  • AR coating on front facet suppresses internal Fabry-Pérot resonances
~100 kHz
Free-running linewidth
1–2 GHz
Mode-hop-free scan
~30 MHz/V
PZT tuning

Common failure modes

  • Mode hops: usually from temperature drift; cure with better temperature control
  • Reduced output: check AR coating; diodes degrade with age and excessive current
  • Multiple modes: grating feedback misaligned; realign while monitoring on wavemeter
  • Linewidth broadening: current noise from noisy driver; use low-noise current source

Alignment tips

  1. Set temperature for approximate target wavelength
  2. Coarsely align grating to first-order feedback with IR card
  3. Monitor wavelength on wavemeter; find single-mode region
  4. Maximise mode-hop-free range by co-scanning PZT + current (feed-forward)
📦 ECDL vendors
Toptica DL Pro · Sacher LION · MOGLabs CEL · RIO Photonics (DFB)
671 nm , Lithium D-lines (⁶Li / ⁷Li)
Transitions

D1 (670.992 nm) and D2 (670.977 nm)

Role in experiment

MOT (D2), Zeeman slower (D2), Λ-GM cooling and tweezer imaging (D1)

Linewidth

Γ/2π = 5.87 MHz (D1 and D2 nearly identical)

Key challenge

D1 and D2 lines separated by only ~10 GHz, both needed simultaneously. D2 laser offset-locked to D1. Lightest alkali: single-recoil scale ≈3.5 μK, so each scatter cycle heats strongly unless cooling is simultaneous.

Locking strategy

D1: SAS locked to vapour cell. D2: beat-note locked to D1 (Vescent D2-135).

Laser source

Toptica or home-built ECDL at 671 nm; TA amplifier often needed for MOT power

📦 Diode sources
Thorlabs L671P200 · Eagleyard Photonics · Laser Components
852 nm , Cesium D2 (¹³³Cs)
Transitions

D2 (852.347 nm); D1 at 894.6 nm also used in some labs

Role

MOT, fluorescence imaging, repumper

Linewidth

Γ/2π = 5.23 MHz

Key challenge

Large hyperfine splitting (9.193 GHz) means cooler and repumper must be offset by ~9.2 GHz; use an AOM chain or separate laser with beat lock.

Locking strategy

SAS locked directly to Cs D2 line in vapour cell.

685 nm , Cesium quadrupole transition (5D₅/₂)
Transitions

6S₁/₂ → 5D₅/₂ (electric quadrupole, Γ/2π ≈ 117.6 kHz)

Role

Narrow-line sideband cooling of Cs in tweezer; excited-state lifetime measurement

Linewidth

Γ/2π = 117.6 kHz  (resolved sidebands at typical trap frequency)

Key challenge

Forbidden E2 transition → I_sat ≈ 2.3 W/cm² (much higher than D-lines). No SAS possible. Requires PDH lock to ULE cavity for <1 kHz linewidth. Astigmatism correction needed (prism pair before cavity).

Locking strategy

PDH locked to ULE cavity (L=77.5 mm, F≈15 000, linewidth ~100 kHz). Laser linewidth ~1 kHz.

📦 Diode source
Thorlabs L685P010 (AR-coated front facet strongly preferred for stable ECDL operation)
1064 nm , Optical tweezer trap (Li + Cs)
Transitions

Far-detuned (no resonant absorption); acts as conservative dipole trap potential

Role

Creates the optical tweezer potential; all atoms trapped in the 1064 nm focus

What matters here

Intensity noise, not frequency noise. Intensity noise at trap frequencies (kHz) causes parametric heating.

RIN target

< −130 dBc/Hz at trap sidebands. Needs intensity stabilisation (AOM servo on pick-off PD).

Locking strategy

Free-running (stable Nd:YAG or fiber laser); intensity servo via AOM feedback.

📦 Laser sources
Coherent Mephisto · NKT Photonics Koheras · Azurlight fiber amplifier · Thorlabs Nd:YAG

Tacit knowledge that isn't in any textbook — collected from years of debugging in cold-atom and tweezer labs. Each section covers a specific experimental area.

⚠️
AOM double-pass: retroreflect off the AOM crystal, not the fiber coupler. The retro mirror must send the beam back through the AOM at exactly the same transverse position. If your retro mirror is after a fiber launch, you lose the double-pass frequency shift — the AOM shifts the beam twice only when both passes hit the active aperture.

Diagnostic: block the retro and check for a single-pass diffracted spot. Restore retro and verify the RF power needed for 80% efficiency drops by roughly half compared to single-pass.
⚠️
AOM efficiency drops if the beam is not focused inside the crystal. The acoustic wave has a finite Bragg acceptance angle. A collimated 3 mm beam has a much smaller angular spread than one focused to 100 μm — tight focus gives better efficiency for high-frequency AOMs (>200 MHz) because the Bragg condition is satisfied over a wider range of angles.

Rule of thumb: focus to a spot size where the Rayleigh range ≈ acoustic transit length (~5–10 mm for most crystal heights).
⚠️
EOM Vπ is not constant across drive frequencies. At 9.2 GHz (Cs hyperfine), a resonant EOM has dramatically lower Vπ than at 20 MHz (PDH). Driving an EOM off-resonance or above its bandwidth produces almost no modulation — always check the EOM spec sheet for the resonant frequency and 3 dB bandwidth.
⚠️
PM fiber after EOM scrambles your polarization state. An EOM requires linearly polarized input aligned to its crystal axis. If you launch into PM fiber before the EOM, check that the PM fast axis aligns to the EOM crystal axis, not just to some arbitrary horizontal. A quarter-wave rotated PM fiber coupling completely kills the EOM modulation depth.
⚠️
Coupling efficiency >80% but PER is only 17 dB? Your beam is elliptical or astigmatic. Polarization extinction ratio (PER) in a PM fiber requires the input beam to be perfectly linearly polarized and Gaussian. An astigmatic beam (common after an AOM) couples light into both PM axes — producing high efficiency but poor PER. Add a cylindrical lens pair to correct astigmatism before coupling into PM fiber.
⚠️
Power instability through PM fiber: check the patch cable stress, not the laser. A bent PM fiber or a clamp pressing on the cable applies stress birefringence, rotating the polarization axis and beating with the PM alignment. Slow power drifts often trace to a fiber cable that rests against a moving part of the table (shutter actuator, beam dump, cooling water line) rather than the laser itself.
⚠️
Mode-matching to a fiber: optimize tip/tilt first, then z-position of the lens. Counter-intuitive but true: the lens transverse position sets the input angle, and the lens z-position sets the mode waist at the fiber face. Iterate tip/tilt → z-lens → tip/tilt. Trying to optimize all three simultaneously gives slow convergence.
⚠️
Don't clean fiber connectors with compressed air — use IPA and a reel cleaner. Compressed air blows particles onto the ferrule face. Use a single-use reel cleaner (Cletop-S) for APC/PC connectors. Inspect with a fiber scope before mating — a single 1-μm particle on the 9-μm core is enough to degrade coupling efficiency from 85% to 50%.
⚠️
SAS crossover resonances are not real transitions — don't lock to them by accident. A crossover peak appears halfway between two real transitions that share a common ground state. It's the tallest feature on the SAS spectrum (coherent population transfer), so it's tempting to lock there. But it's 212.5 MHz below the 87Rb F=2→F'=3 cycling transition, not 0 MHz.

Verification: compare your SAS spectrum peak positions against published Rb/Cs hyperfine intervals. For Rb87 D2: F=2→{1,2,3} = 72, 157, 267 MHz above F=2→F'=1; crossovers fall at 36, 115, 212 MHz.
⚠️
PDH error signal polarity depends on which sideband is higher in frequency. If you flip the LO phase by 180° or swap the mixer IF/RF ports, the error signal flips sign — and your PI controller drives the laser away from resonance instead of toward it. Always verify locking direction: give the laser a small push (tap the table) and confirm the lock corrects, not amplifies, the perturbation.
⚠️
Servo bump tells you where your loop bandwidth is — not where you want it. The characteristic bump in the in-loop noise PSD at the servo bandwidth is not an artifact to filter away — it's diagnostic. Flattening it by lowering gain reduces bandwidth and worsens noise. Instead, improve the phase margin by adding a lead compensator or reducing cable length in the feedback path.
⚠️
High-finesse cavity drift: check temperature, not the laser. A 10 cm Zerodur spacer expands by ~0.5 nm/°C (CTE ~2×10-8/K). A 1 mK temperature drift shifts the resonance by ~2 MHz at 780 nm. If your PDH-locked laser frequency drifts slowly over minutes, the cavity is thermally drifting — improve vacuum can temperature stabilization before touching laser current.
⚠️
Quantization axis must be defined before the pump beam arrives. If the bias magnetic field is not on (or is lower than the stray-field level ~100 mG), the atom has no preferred axis. The σ+ photon pumps into different mF levels depending on local stray field direction. Always verify the bias field is on and stable before running the OP sequence.
⚠️
Re-pumping power too high scrambles your state preparation. The repump beam (F=1→F'=2 for Rb87) should be weak enough to recycle dark atoms without broadening the resonance for the main pump. Too much repump power heats the sample by preventing atoms from dwelling in the F=1 dark state, and can repump atoms out of the target mF state.
⚠️
OP efficiency check: use a short probe pulse, not absorption imaging. To verify mF state purity, do a quick Stern-Gerlach separation with a field gradient before imaging. Absorption imaging along the same OP beam axis can misrepresent state purity because the imaging beam itself pumps atoms during exposure. A short (10 μs) probe pulse after Stern-Gerlach separation resolves individual mF populations cleanly.
⚠️
Intensity noise heats your atoms. Frequency noise does not (for far-detuned traps). A 1064 nm tweezer far from any transition is an intensity-noise-limited trap. The trap depth U₀ ∝ I, so fractional intensity noise δI/I directly converts to fractional trap-depth noise, driving parametric heating at 2ωtrap. Frequency noise at 1064 nm has no first-order effect on the trap depth (it only shifts absolute energy by the tiny light-shift gradient).

Action: use a low-noise laser source at 1064 nm and servo the intensity with an AOM + photodiode + PI loop. Target RIN < −140 dBc/Hz at 2ωtrap (~2×100 kHz = 200 kHz).
⚠️
SLM-generated tweezer arrays: check diffraction efficiency, not just spot count. An SLM can project 100 spots, but if your hologram algorithm doesn't compensate for the sinc envelope of the SLM pixel pitch, outer spots in the array are 30–40% dimmer than central ones. Run a weighted Gerchberg-Saxton (WGS) algorithm that measures spot intensities in a feedback loop and reweights the target amplitudes accordingly.
⚠️
Evaporation ramp that works in a crossed ODT fails in a single-beam tweezer. A crossed ODT has equal radial and axial trapping in two dimensions. A single-beam tweezer has radial frequency ωr >> axial ωz — atoms escape axially first as you ramp down power. Design the evaporation ramp by monitoring axial size with in-situ imaging, not just total atom number.
⚠️
Camera background subtraction: take dark frames at the same temperature as your science frames. EMCCD and sCMOS dark current is strongly temperature-dependent (~2× per 6–7°C). If your camera warms up between your background run and your data run, the dark floor changes, creating a systematic offset in atom number. Let the camera thermalize for at least 30 minutes and take background frames immediately before (or interleaved with) science shots.
⚠️
Single-atom fluorescence: saturation parameter determines photon budget, not just pulse duration. At high saturation s >> 1, the scattering rate plateaus at Γ/2. Going from s=10 to s=100 only gives you ~10% more photons while increasing the recoil heating by 10×. The optimal single-atom imaging uses s ≈ 1–5 with a short pulse, not maximal power.

Rule of thumb: for Rb87 D2 with NA=0.5 objective, s=2 gives ~400 photons in 1 ms with ~8 recoil kicks — sufficient for single-atom detection with EMCCD EM gain ~300.
⚠️
Charge smear on EMCCD: don't readout during the fluorescence pulse. EMCCD vertical shift (frame transfer) takes ~1 ms. If your shutter timing overlaps with the shift, every row smears fluorescence into adjacent rows, broadening PSFs and reducing per-pixel SNR. Use a mechanical shutter or an AOM to gate the light, not just the camera trigger.
⚠️
Calibrate your RF π-pulse time at the same bias field you use for science. The Zeeman shift moves the transition by ~0.7 MHz/G for Rb87 |F=1,mF=−1⟩→|F=2,mF=0⟩. If you calibrate the π-pulse at 1 G and run science at 3 G, you're off by 1.4 MHz — and instead of a π-pulse you're doing a Ramsey dephasing experiment. Always recalibrate ΩR and the resonance frequency when you change the bias field.
⚠️
Ground loop in the RF circuit kills coherence time. An RF ground loop between your signal generator, amplifier, and antenna creates a circulating current that modulates the local magnetic field at 50/60 Hz and its harmonics. Use a single-point ground, float the amplifier chassis on rubber feet, and check T₂* with a Ramsey sequence: if you see T₂* oscillate at 50 Hz, you have a ground loop.
⚠️
Avoid antenna resonances: design for broadband, not peak efficiency. A resonant antenna at 6.835 GHz (Rb87 clock transition) is efficient but narrow-band. Any drift in the antenna resonance (thermal expansion of coax, SMA connector loosening) shifts the coupling and changes your effective Rabi frequency. A broadband impedance-matched antenna (s11 < −15 dB over ±100 MHz) is more robust for day-to-day reproducibility.
⚠️
Eddy currents from fast field switching limit how quickly you can apply a Feshbach field. A large metallic vacuum chamber acts as a shorted transformer: when you ramp the Feshbach coil current rapidly (>1 A/ms), induced eddy currents oppose the change and the actual field at the atoms lags behind the coil current by the chamber's L/R time constant (typically 1–10 ms). Always measure the actual field delay with RF spectroscopy before assuming the field follows your current waveform.
⚠️
Stray-field zeroing: do it with atoms at the operating temperature, not room temperature. Mu-metal shielding has temperature-dependent permeability. The field zero you find at room temperature can shift by 20–50 mG when the chamber bakes out or the surrounding equipment heats up during a long experimental run. Zero the field with atoms by minimizing the Ramsey fringe decay rate at the exact operating conditions.
⚠️
MOT coil polarity: anti-Helmholtz for trapping, Helmholtz for bias field — never mix up. Anti-Helmholtz coils produce a gradient (zero at center, ∇B ≠ 0) — correct for MOT. Helmholtz produces a uniform field. If you accidentally run anti-Helmholtz during optical pumping, the atoms experience a spatially varying quantization axis and your OP efficiency drops to ~50% at best. Double-check the relay/H-bridge state before every sequence that switches between gradient and bias.
🔬
Hood Lab note: Most of the pitfalls above were discovered the hard way — by running a beautiful experiment for three days and then realizing the state preparation was wrong. The most common root cause is not verifying assumptions: always ask "how do I know the field is on?", "how do I know the fiber is PM-aligned?", "how do I know the AOM is double-passing correctly?" Build a short diagnostic sequence for each subsystem and run it at the start of every session.

📚 References & Further Reading

Primary Source & Textbooks

Metcalf & van der Straten, Laser Cooling and Trapping (1999)

Foot, Atomic Physics (OUP, 2005)

Saleh & Teich, Fundamentals of Photonics (Wiley)

Pozar, Microwave Engineering (Wiley, 4th ed.)

Key Papers
📄 Black (2001), PDH locking tutorial (Am. J. Phys.) 📄 Johansson et al. (2013), QuTiP 2 📄 Kaufman & Ni (2021), Tweezer arrays review (Nat. Phys.) 📄 Frazier (2018), Bayesian optimisation tutorial
Vendor App Notes & Tools
🔗 Thorlabs fiber coupling guide 🔗 AA Opto AOM selection guide 🔗 Vescent laser locking app note 🔗 SimSmith, Smith chart tool 🔗 QuTiP documentation 🔗 scikit-learn GP regression