How a Phone Knows Up From Down: The MEMS Accelerometer Explained

01 How a glass slab knows which way is up

A smartphone has no eyes, no inner ear, and no plumb line, yet the instant you rotate it in your hand the display flips from portrait to landscape with a small physical certainty that feels almost reflexive. The organ responsible is a chip smaller than a grain of rice: a MEMS accelerometer, a microelectromechanical system in which actual moving machine parts — beams, springs, a suspended mass — are carved out of a sliver of silicon using the same photolithography that builds the phone's processor.

The first surprise about this chip is what it actually measures. An accelerometer does not report velocity, and it does not report motion in any everyday sense. As the Wikipedia summary of the device puts it, an accelerometer measures proper acceleration — the acceleration of an object relative to an observer in free fall. A phone lying flat on a table, apparently motionless, reports a steady acceleration of about 9.81 m/s² pointed straight up. A phone dropped from your hand reports exactly zero while it falls. That inversion — still means one g, falling means nothing — is the seed of everything the sensor does, and of the deepest problem it poses.

02 A spring and a mass, carved out of silicon

Strip the packaging away and the physics is antique: a proof mass suspended on springs, the same arrangement Galileo would recognise from a weight hanging on a wire. When the phone accelerates, the silicon frame moves and the proof mass briefly lags behind, stretching or compressing its spring tethers until the spring force catches up. Hooke's law does the rest: the displacement of the mass is proportional to the force applied to it, and the force is proportional to acceleration. Measure how far the mass moved and you have measured how hard the phone accelerated.

The engineering feat is not the principle but the manufacturing. In a consumer MEMS accelerometer, the proof mass, its folded spring legs, and the anchor points are all etched from a single layer of monocrystalline silicon, freeing a mechanical structure that would sit comfortably inside the cross-section of a human hair. The whole micromachine is a few hundred micrometres across, weighs essentially nothing, and — because it is a chip — costs a couple of dollars and never needs a bearing lubricated. In the engineerguy video that accompanies this article, Bill Hammack cracks one open on camera to show exactly this lattice of springs, rendered at die scale.

Modern parts squeeze astonishing economy from the design. Bosch Sensortec's BMA400, a current-generation consumer accelerometer, draws 14.5 microamps at maximum performance, 5.8 microamps in a typical use case, and 3.5 microamps in its low-power mode — a micromechanical machine that runs for years on a battery, precisely because the moving parts are so small that almost no energy is needed to interrogate them.

03 Capacitive readout: measuring trillionths of a metre

A displaced proof mass is useless until something can measure the displacement, and no ruler works at this scale. The trick is capacitance. Beside the moving mass, the chip carries fixed electrodes; the mass and electrodes together form tiny parallel-plate capacitors whose capacitance depends on the gap between them. Move the mass by even a fraction of the width of its gap and the capacitance shifts by a measurable amount. Interrogating those plates with a high-frequency signal turns a displacement far below the wavelength of visible light into a clean electrical number.

The sensitivity is the part that borders on the implausible. The capacitive gaps in consumer parts sit on the order of a micrometre, and the electronics can resolve mass motions of a few trillionths of a metre — displacements smaller than the diameter of an atom. That extreme resolution is why the readout is done differentially: the mass sits between paired electrodes, and the circuit measures the difference between two capacitances, cancelling drift, noise, and temperature effects that would swamp a single-ended measurement.

What the user gets at the end of this chain is a digitised axis reading. The resolution of that number is a deliberate design choice, not a law of nature. Analog Devices' ADXL345 — for years the canonical smartphone accelerometer — offers selectable full-scale ranges of ±2, ±4, ±8, and ±16 g. Set it to ±2 g and one least-significant bit of the output is worth 3.9 milli-g; set it to ±16 g to survive rougher handling and each LSB is worth 31.2 milli-g. Sensitivity is traded against headroom, eightfold, on command.

04 Gravity versus acceleration: Einstein's equivalence problem

Here is the philosophical snag buried in the chip. Einstein's equivalence principle says that locally, no experiment can distinguish uniform acceleration from gravity — and the accelerometer is that experiment run in hardware, so it cannot tell the difference either. Tilt the phone and the gravity vector, which never changes, appears in the sensor's frame as a changing acceleration along its axes. Shake the phone and genuine acceleration lands on the very same axes. One number, two entirely different physical causes.

This is why the resting phone reports one g pointed upward. The table is pushing up on the phone hard enough to cancel gravity; the chip, measuring proper acceleration, sees that push — not the gravity it cannot see. In free fall, with nothing pushing, it reads zero. The practical consequence for software is that the raw stream from a single accelerometer is always an ambiguous blend of tilt and true motion. To untangle it, a phone needs a second opinion.

05 Drift, and the sensor-fusion fix

That second opinion is the MEMS gyroscope, a sibling chip that measures rotation rate, typically by sensing the Coriolis force on a vibrating structure. In principle, integrating the gyro's output over time yields orientation. In practice, tiny rate errors accumulate without bound: a fraction of a degree of drift per minute compounds into nonsense within hours. The accelerometer drifts too, but in a different currency — its long-term average is anchored to gravity, so its tilt reading is slow but absolute, while the gyro is instant but unmoored.

Sensor fusion is the algorithm that marries their strengths: the gyro supplies fast, smooth short-term rotation, the accelerometer keeps pulling the estimate back toward the gravity reference, and a magnetometer — a compass — adds an absolute heading so the fused orientation stops yaw drift around the vertical. Phones run complementary or Kalman-style filters that blend the three continuously, which is why your display rotates crisply yet never slowly winds up tilted. It is a small, perpetual negotiation between three liars, each checked against the others.

06 What one chip enables: steps, falls, and crash detection

The payoff for all this microscale physics is scattered through daily life. The obvious one is screen rotation, which is just the fused gravity vector crossing a tilt threshold. Step counting comes from the same data: the rhythmic two-peak signature of a walking stride is so characteristic that a filtered accelerometer can tally steps and, with stride-length estimates, feed the health app its distances. Fitness trackers, sleep monitors, and the tap-to-wake gesture all read from the same stream.

The steeper end of the curve is safety. Fall detection on modern watches watches for the sharp spike-and-free-fall signature of a person hitting the ground, then the sustained stillness afterwards, and calls for help if the wearer does not respond. Optical image stabilisation leans on the gyro to counter handshake faster than any motor could. And car-crash detection on recent phones waits for the one pattern a pocket device should never see from a human body: decelerations of many g sustained across a collision, cross-checked against the sudden stop. A silicon spring, displaced by trillionths of a metre, is the witness.