Machines you can't see, carved into silicon
We usually think of a chip as something purely electronic — no moving parts, just electrons sloshing through frozen geometry. A microelectromechanical system breaks that picture wide open. Using the very same photolithography that prints transistors, engineers etch *actual mechanical structures* into silicon: springs, hinges, beams, plates and combs, all of them micrometres across. They bend, twist and vibrate. A whole tiny machine — too small to see, with features finer than a red blood cell — sits on the same die as the electronics that listen to it.
Why bother carving machines this small? Because at the micron scale, physics is generous. A beam a few microns thick is absurdly stiff for its size yet feather-light, so it responds to the faintest push almost instantly and survives being dropped on the floor. Thousands of identical units come off one wafer, so a sensor that once cost hundreds of dollars and filled a shoebox now costs cents and hides under a grain of rice. That collapse in size and price is exactly why a $200 phone can carry the kind of inertial sensing that used to guide missiles.
The accelerometer: a proof mass on tiny springs
Picture a small weight hanging from springs inside a sealed box. Shove the box left and the weight, by inertia, lags behind — it appears to swing right relative to the walls. Measure how far it swings and you have measured the acceleration. That is the entire idea of an accelerometer, and a MEMS version shrinks it to the size of a dust mote. The weight is called the proof mass: a slab of silicon suspended by flexible silicon springs so it can slide a little when the chip accelerates.
How do you read a motion of mere nanometres? With capacitance. The proof mass carries a row of comb-like fingers that interleave with fixed fingers anchored to the frame, forming a pair of parallel-plate capacitors. When the mass slides, one gap narrows while the other widens; one capacitance rises while the other falls. Recall that for parallel plates C = εA/d — capacitance is inversely proportional to the gap d, so even a nanometre of motion produces a measurable change. The catch: that change is on the order of femtofarads (10⁻¹⁵ F), a thousand times smaller than the stray capacitance of the wire reading it.
Top view of a differential capacitive accelerometer
fixed fingers (anchored) fixed fingers
▐ ▐ ▐ ▐ ▐ ▐ ▐ ▐
─┼───┼───┼───┼── proof mass (slides ←→) ──┼───┼───┼─
▐ ▐ ▐ ▐ ◄═══ springs ═══► ▐ ▐ ▐ ▐
C1 (gap d1) C2 (gap d2)
At rest: d1 = d2 -> C1 = C2 -> output = 0
Accelerate →: mass lags, d1 shrinks, d2 grows
C1 > C2 -> ΔC = C1 - C2 ∝ acceleration
Differential trick: reading C1 - C2 cancels temperature drift and
common-mode noise that hit BOTH capacitors equally.A femtofarad wiggle is far too small to send anywhere, so the first electronics on the die sit *right next to* the springs: a low-noise charge amplifier — essentially the op-amp from rung 1 wired to convert tiny charge changes into a respectable voltage swing. This is the deepest reason MEMS and microelectronics share a chip: the signal is so frail it must be amplified *before* it travels even a millimetre, or stray pickup would swamp it.
The gyroscope: catching the Coriolis ghost
An accelerometer feels straight-line shoves but is blind to *rotation*: spin a phone flat on a turntable and the proof mass barely cares. To sense turning, you need a MEMS gyroscope, and it exploits a subtler effect. Try walking in a straight line toward the centre of a spinning merry-go-round — a sideways force you never asked for shoves you off course. That phantom push is the Coriolis force, and it appears whenever something moves *within* a rotating frame.
A MEMS gyroscope builds its own little rotating-frame experiment. It keeps a proof mass vibrating rapidly back and forth along one axis — the *drive* motion — using electrostatic comb drives, like a tuning fork humming at tens of kilohertz. As long as the chip isn't turning, that's all it does. But the instant the chip rotates about a perpendicular axis, the Coriolis force kicks the vibrating mass *sideways*, perpendicular to its drive motion. That sideways nudge — the *sense* motion — is proportional to the rotation rate, and it's read with the very same comb-capacitor trick the accelerometer used.
MEMS gyroscope = a driven mass + the Coriolis force
Drive (always on): mass oscillates ◄═══════► along X
v (velocity)
Chip rotates about Z at rate Ω
│
▼ Coriolis force F = -2 m (Ω × v)
Sense (the signal): mass kicked ▲ along Y
│ amplitude ∝ Ω
▼
drive frequency ~ tens of kHz · sense amplitude → ΔC → voltage
No rotation (Ω = 0) -> no sideways motion -> output = 0The IMU: six senses, one motion story
Each sensor alone has a fatal weakness. The accelerometer knows which way is down (thanks to that 1 g vector) but is hopelessly jittery — every footstep and table-tap shows up as acceleration noise. The gyroscope is smooth and fast but drifts away over seconds. Put a three-axis accelerometer and a three-axis gyroscope on one chip and you get an inertial measurement unit (IMU): six degrees of freedom, three of acceleration and three of rotation, reporting your full motion many hundreds of times a second.
The magic isn't the parts — it's the sensor fusion that combines them. Each one covers the other's blind spot: the gyro provides the fast, smooth, short-term answer, while the accelerometer's long-term average of 'down' is used to gently pull the gyro's drift back into line. A filter (a complementary filter, or the famous Kalman filter) blends them continuously, trusting the gyro over short windows and the accelerometer over long ones. The result is an orientation estimate that is both responsive *and* stable — neither sensor could give that alone.
- Drone stabilisation — the IMU reports tilt and turn rate hundreds of times a second so the flight controller can correct a gust before you ever see it wobble.
- Phone & game controllers — screen rotation, step counting, and the motion you swing into a tennis game all come from reading the IMU.
- Dead reckoning — when GPS drops out (a tunnel, indoors), the IMU keeps estimating where you moved by integrating acceleration and rotation — until drift forces a fresh fix.
- Image stabilisation — a camera reads its own gyro to predict hand-shake and shift the lens or crop the frame to cancel it, mid-exposure.
Pressure sensors and strain gauges: feeling a squeeze
Not every physical quantity is motion. A MEMS pressure sensor is built around a wafer-thin silicon diaphragm — a drumhead a few microns thick stretched over a sealed vacuum cavity. Air pressure on one side pushes the diaphragm inward, bowing it by a fraction of a micron. Read how far it bows and you've read the pressure. There are two common ways to read it: capacitively, exactly like the accelerometer (the bowing diaphragm changes a gap), or by detecting the *stretch* in the silicon itself — which leads us straight to the strain gauge.
A strain gauge turns deformation into resistance. Stretch a wire and it gets longer and thinner, so its resistance rises (R = ρL/A); squash it and resistance falls. Glue such an element to a beam, a bridge girder, or a pressure diaphragm and its resistance now tracks the strain. Silicon does this far more strongly than metal through the piezoresistive effect — squeezing the crystal changes how easily electrons flow — giving MEMS gauges a sensitivity (the *gauge factor*) of 100 or more, against about 2 for a metal-foil gauge.
Reading a strain gauge with a Wheatstone bridge
Vexc
│
┌────┴────┐
R1 Rg ← strain gauge (R changes with strain)
│ │
Vo- ● ● Vo+ Vout = Vo+ - Vo-
│ │
R2 R3
└────┬────┘
│
GND
Balanced (no strain): R1/R2 = Rg/R3 -> Vout = 0
Under strain: Rg shifts a little -> small Vout appears
ΔR/R is often < 0.1%, so Vout is only millivolts riding on volts.
-> needs an instrumentation amplifier, then an ADC.The universal chain: from a whisper to a number
Step back and every sensor in this guide — and almost every sensor you will ever meet — collapses into the same four-stage pipeline. A transducer turns a physical quantity into a frail electrical one. An amplifier blows that whisper up to a usable size. An analog-to-digital converter freezes it into a number. And the number lands in a register where firmware can finally use it. Learn this chain once and you understand the front end of a phone, a Mars rover, an ECG machine, and a kitchen scale all at once.
The universal sensor signal chain
┌──────────────┐ μV–mV ┌──────────────┐ V ┌──────────┐ bits
│ TRANSDUCER ├─────────►│ AMPLIFIER ├──────►│ ADC ├──────► number
│ (MEMS, gauge)│ tiny │ (op-amp / │ clean │ │ to MCU
└──────┬───────┘ analog │ instr amp) │ swing └────┬─────┘ register
│ └──────────────┘ │
physical world digital world
accel · rotation · pressure · strain a value firmware can read
Same four boxes for EVERY sensor. Only the leftmost box changes.
(Often an anti-aliasing filter sits between the amplifier and the ADC.)Two pieces of this chain you already own from rung 1. The amplifier is built from the op-amp — for the bridge and biopotential sensors it's usually the differential-input instrumentation amplifier, chosen precisely because it amplifies the *difference* between two wires while ignoring noise common to both. And the ADC is the same converter that gave you V_ref / 2ⁿ step sizes and the Nyquist sampling rule. Sensors don't introduce new electronics so much as point the electronics you know at the physical world.