Atomic Force Microscopy: Feeling Atoms With a Needle

The instrument: a spring, a needle and a laser

An atomic force microscope contains no lens. Its entire optical train exists only to read the bending of one micromachined spring: a cantilever, a diving board of single-crystal silicon or low-stress silicon nitride ~100–200 µm long and 0.5–5 µm thick — sub-micron for the softest levers, because k goes as t³ — carrying an etched pyramid on its free end. The pyramid terminates in an apex of ~2–10 nm radius, so the last few atoms of the tip do all the imaging.

• Approach. A piezoelectric scanner brings sample and tip together in steps far smaller than a wavelength of light, until the lever registers a force.
• Bend. The lever is a Hooke's-law spring, F = k·Δz. At k ≈ 0.1 N/m a deflection of 0.1 nm is only 10 pN — comparable to the force that ruptures a single hydrogen bond.
• Read. A ~670 nm laser diode reflects off the gold-coated back of the lever onto a four-quadrant photodiode ~3 cm away. The normalised vertical difference, (top − bottom)/(sum), is the deflection; the left–right difference reports twisting, which is how lateral-force (friction) imaging works.
• Hold. A proportional–integral loop compares that signal with a setpoint and drives the Z electrode of a lead zirconate titanate (PZT) tube to keep the force constant.
• Raster. Quadrant electrodes on the same tube sweep X and Y. The Z correction voltage, plotted against X and Y, is the image.

Nothing in that chain requires photons or electrons to touch the specimen. AFM reports a mechanical force, which is why it images an insulating polymer or a living bacterium in buffer as readily as a gold film.

The force curve: dispersion in, Pauli repulsion out

Within about a nanometre of the surface the Lennard-Jones picture takes over. Two neutral atoms attract through London dispersion as −C/d⁶ and repel as +C′/d¹² once their electron clouds overlap, because the Pauli exclusion principle forces overlapping electrons into higher orbitals. A tip is not one atom, so the pairwise −1/d⁶ term must be integrated: for a sphere of radius R above a flat half-space the Hamaker result is F_vdW ≈ −A R/(6d²), with the Hamaker constant A of order 10⁻¹⁹ J for silicon across vacuum.

Worked example. Take R = 5 nm and d = 0.5 nm. Then F_vdW ≈ (10⁻¹⁹ J × 5×10⁻⁹ m) / (6 × (5×10⁻¹⁰ m)²) ≈ 3×10⁻¹⁰ N, about 0.3 nN. That is thirty times the 10 pN a 0.1 nm bend represents, which is why the feedback loop, not the operator's hand, has to govern the approach.

The same expression explains the AFM's most notorious instability. Differentiating gives a force gradient ∂F/∂z ≈ AR/(3d³) ≈ 1.3 N/m at that separation, and when the gradient exceeds the lever's own stiffness the equilibrium vanishes and the tip snaps into contact. A 0.1 N/m lever is more than ten times too soft to sit stably in the attractive well, so working there needs either a far stiffer sensor or an oscillating mode. Push closer still and the curve turns over into the repulsive wall, where the contrast becomes almost purely Pauli repulsion — the regime that produces textbook molecular images.

Reading 0.1 nanometres: the optical lever and the noise floor

Detecting a 0.1 nm deflection is the real trick, and it is solved by geometry. A cantilever loaded at its free end does not merely translate; it tilts. For an end-loaded beam of length L the slope at the tip is θ = 3Δz/(2L), and a reflected ray turns by twice the mirror's tilt, so the spot walks 2θ×D across a detector a distance D away.

Worked example. With L = 100 µm and Δz = 0.1 nm, θ = 3(0.1 nm)/(2 × 100 µm) = 1.5 µrad. Doubling on reflection gives 3 µrad, and over D = 3 cm the spot moves 3×10⁻⁶ × 0.03 m = 9×10⁻⁸ m ≈ 90 nm. That is the celebrated optical lever gain ~1000× — 0.1 nm of tip motion becomes ~100 nm of spot shift on a photodiode ~3 cm away, the scheme Meyer and Amer brought to the AFM in 1988. Ninety nanometres of beam walk across a millimetre-scale photodiode is a trivial electrical measurement, so an ordinary detector delivers vertical resolution ~0.1 nm (1 Å).

The limit is not the optics but Brownian motion. Equipartition gives ½k⟨z²⟩ = ½k_BT, so a 0.1 N/m lever at 300 K rattles with an rms amplitude of √(4.1×10⁻²¹ J / 0.1 N/m) ≈ 0.20 nm — twice the resolution being claimed. The escape is bandwidth: nearly all of that thermal energy sits in a narrow peak at the lever's resonance, and the off-resonance noise density is a fraction of a picometre per √Hz, so restricting the loop to roughly a kilohertz leaves a few picometres of noise and the ångström is real. That same restriction sets the frame rate — ~1 scan line per second, so a 512-line image takes ~9 min.

Contact, tapping, and the water film that ruins everything

In contact mode the loop holds deflection constant with a soft lever (k ≈ 0.1 N/m) and the tip ploughs along inside the repulsive wall. In ambient air this is brutal, because every surface carries several monolayers of adsorbed water. A meniscus condenses around the tip and pulls it down with F ≈ 4πRγcosθ: for R = 10 nm and γ = 72 mN/m that is about 9 nN, and a blunted tip with a thicker film pushes it into the ~10–100 nN range. Ten nanonewtons borne on a Hertzian contact under two nanometres across is a mean pressure of several gigapascals, which is why contact-mode AFM in air sweeps DNA and adsorbed protein off the substrate rather than imaging it.

Tapping mode, commercialised by Digital Instruments in 1993, refuses to stay in contact. The lever is driven near ~300 kHz at 20–100 nm free amplitude, so each cycle it punches through the water film, touches for well under a microsecond, and a stiff spring rips it back out. The feedback variable becomes the oscillation amplitude, and the phase lag between drive and response maps energy dissipation — stiffness, adhesion, composition — alongside topography.

Worked example. A rectangular silicon lever 125 µm × 35 µm × 4 µm has k = Ewt³/4L³ = (169 GPa × 35 µm × (4 µm)³)/(4 × (125 µm)³) ≈ 48 N/m, matching the ~40 N/m of a tapping lever, and f₀ ≈ 0.162(t/L²)√(E/ρ) ≈ 350 kHz. Note the t³: a 10% error in etched thickness is a 33% error in k, which is why nominal catalogue spring constants are never trusted for quantitative force work.

In liquid there is no meniscus at all, so imaging can run at tens of piconewtons — the reason live-cell and single-molecule AFM is done in buffer. At the other extreme, frequency-modulation AFM in ultra-high vacuum uses a stiff quartz sensor (the qPlus tuning fork, k ≈ 1800 N/m) at sub-ångström amplitude and tracks the resonance shift Δf ≈ −(f₀/2k)⟨∂F/∂z⟩.

Calibration, standards and how an AFM is specified

An AFM produces volts, not nanometres, so two calibrations stand between the instrument and a number. The deflection sensitivity in nm per volt is the slope of the hard-contact region of a force curve taken against something incompressible, such as sapphire. The spring constant comes from one of three standard methods: thermal noise (Hutter and Bechhoefer, 1993), which fits a Lorentzian to the first resonance and applies equipartition; the Sader method (1999), using plan-view dimensions with the resonance and quality factor measured in air; or Cleveland's added-mass method. Two of them agreeing to 10–15% is the usual acceptance criterion.

Scanner calibration is a separate problem. Open-loop PZT tubes show 10–15% hysteresis plus creep lasting minutes after a step, so metrology-grade instruments run closed-loop against capacitive or strain-gauge sensors. Z is calibrated against traceable step-height artifacts spanning roughly 8 nm to 1 µm, X and Y against pitch gratings of a few microns period. ISO 11952 covers dimensional calibration of scanning probe microscopes, areal roughness parameters (S_a, S_q, S_z) are defined by ISO 25178, and ASTM E2382 is the reference guide to scanner- and tip-related artifacts.

The tip should be characterised rather than assumed: imaging a deliberately spiky reference sample and running blind tip reconstruction returns an upper-bound estimate of the apex. A tip that began at ~2–10 nm can double its radius within a single frame on a hard sample, and nothing in the image says so.

From a foil lever in 1986 to pentacene in 2009

Gerd Binnig and Heinrich Rohrer built the scanning tunnelling microscope at IBM Zürich in 1981 and shared the 1986 Nobel Prize in Physics for it — but the STM demands a conducting sample, which excludes most of chemistry. In the same year Binnig, Quate and Gerber published the atomic force microscope in Phys. Rev. Lett. 56, 930 (1986). Their prototype was a strip of gold foil with a fragment of diamond glued to it, read by a second STM tip behind the lever, and it demonstrated in air a lateral resolution of 30 Å with a vertical resolution better than 1 Å. It became a product only when that STM readout gave way to the optical lever (Meyer and Amer, 1988), which needs no conductive cantilever, no vacuum and no second feedback loop. Tapping mode followed in 1993 and opened up soft matter; in 1995 Franz Giessibl showed genuine atomic resolution on the Si(111)-7×7 reconstruction by frequency-modulation AFM in ultra-high vacuum.

The defining image came in 2009, when Leo Gross and colleagues picked a single carbon monoxide molecule up onto the apex: pentacene's carbon rings resolved with a CO-terminated tip at 5 K (IBM Zürich, 2009). The CO is chemically inert and presents a single-atom oxygen terminus that resists vertical compression, and the constant-height contrast is essentially pure Pauli repulsion. Later work showed the crisp lines between atoms are sharpened by lateral tilting of the CO — a caution against reading such pictures as literal bond-order maps.

Two other threads matter: single-molecule force spectroscopy, which arrived with Rief and co-workers in 1997 pulling titin's immunoglobulin domains open one at a time near 200 pN, and Toshio Ando's high-speed AFM, which filmed myosin V walking along an actin filament in 2010 at roughly a tenth of a second per frame. Binnig, Gerber and Quate received the 2016 Kavli Prize in Nanoscience for the AFM.

How it lies to you, and what it is not

Every AFM image is a dilation of the surface by the tip, and the tip is never a point. A hemispherical particle of radius r imaged with an apex of radius R appears about 4√(Rr) wide: a particle of 5 nm radius scanned with a 10 nm tip images ~28 nm across, nearly three times its true 10 nm diameter. Its height is still correct. The governing rule of AFM interpretation is that heights are quantitative and lateral widths are not.

The other classic artifacts announce themselves once you know them. A chipped apex produces a double tip, repeating every feature at a fixed offset. A piezo tube sweeps on an arc, so a flat sample images as a shallow dome; the flattening that removes that bow also removes real long-wavelength topography — the commonest way to manufacture a misleadingly flat result. Thermal drift of ~0.1 nm/s shears a nine-minute frame by tens of nanometres. And overlaying trace with retrace is the standard diagnostic: if they disagree, the gains are too low or the tip is dragging.

The biggest conceptual trap is the word atomic. Lattice-resolved contact-mode images of mica or graphite in air are averages over a contact many atoms wide: they reproduce the periodicity beautifully but never show a point defect, which is the real test. Likewise, measured DNA heights of 0.5–1.5 nm rather than the 2.0 nm B-form diameter reflect tip compression and mica-surface electrostatics, not a thinner molecule.

Finally, the look-alikes set out in the table above: STM needs a conductor and maps the local density of states rather than shape, SEM needs vacuum and offers no calibrated Z axis, and cryo-EM or crystallography average a molecule's interior over millions of copies. AFM touches only the outside of one object — but in buffer, at room temperature, and it can watch that object move.

Atomic force microscopy against the techniques it is most often confused with

Technique	Quantity measured	Sample requirement	Resolution and practical limits
AFM	Force between tip apex and surface: van der Waals attraction, then Pauli repulsion	Almost anything flat enough — metal, glass, polymer, DNA, live cells in buffer; air, liquid or vacuum	Vertical ~0.1 nm (1 Å); lateral set by the tip, ~2–10 nm in air, atomic in UHV at 5 K. ~1 scan line per second, so a 512-line image takes ~9 min
STM	Electron tunnelling current, I ∝ exp(−2κd) with κ ≈ 1.1 Å⁻¹	Must be electrically conducting; best in ultra-high vacuum	Routine atomic resolution, but the image maps the local density of states, not topography
SEM	Secondary and backscattered electrons from a focused beam	Vacuum; insulators usually need a conductive sputter coating	~1 nm lateral with enormous depth of field, but no calibrated height axis
Cryo-EM / X-ray crystallography	Electrons or X-rays scattered through the whole specimen	A vitrified thin film, or a single crystal	Atomic coordinates of the interior, averaged over millions of copies; no single-molecule, real-time surface data
Stylus profilometer	Deflection of a diamond stylus dragged under load	Robust surfaces only	Tip radius ~1–10 µm at stylus loads of ~0.01–1 mN (ISO 3274 nominally 0.75 mN); the AFM is its nanoscale descendant