ProjekinAja. Embedded Systems Stage 1 · Foundations Lesson 1.4
● Stage 1 · Lesson 1.4

Passive Components for Real: Resistors, Capacitors, Inductors

A 10 µF capacitor is not 10 µF. It is 10 µF at 0 V, 25 °C, on a good day.

⏱ ~28 min read 🧱 The parts you thought were simple 🔋 Failure class: component physics → stability

🎯 By the end you'll understand

Why a resistor's tolerance is the least interesting number on its datasheet, and what temperature coefficient, power derating, and voltage coefficient actually cost you. What the three characters in X7R, X5R, C0G, and Y5V mean, and why a 10 µF 0402 ceramic can deliver 3 µF in your circuit while measuring perfectly on the bench. ESR and ESL as the reason a capacitor stops being a capacitor above a few megahertz, and why "one big electrolytic" and "one small ceramic" are two different components doing two different jobs. Inductor saturation current, DCR, and the two incompatible ways vendors define Isat. And how every one of these turns into a Beacon part number with a reason attached.

KEY INSIGHTThe value printed on the part is a measurement condition, not a property

The board worked at 2.5 V input and squealed at 4.2 V. Nothing in the circuit was voltage-dependent — except the output capacitor, which had quietly become a third of itself.

A colleague brings you a buck regulator that misbehaves. At a low input voltage it is clean. Charge the battery to 4.2 V and the output develops a slow, irregular wobble and the inductor makes an audible noise. The schematic is copied from the datasheet's reference design. The layout was reviewed. The part numbers match the recommended list — 10 µF on the output, exactly as specified. Then someone puts an LCR meter on the removed capacitor with a DC bias applied and finds 2.9 µF. The recommended 10 µF was never on the board, because the 10 µF was a 6.3 V-rated X5R in an 0402 package, and at 3.3 V of bias that dielectric gives back about a third of its nameplate. The circuit was operating outside the regulator's specified output-capacitance range from the first day, and the only reason it worked at all was that nobody had charged the battery yet.

🔑 Hold onto this

Every passive has a nameplate value that is true under exactly one set of conditions, and a real value that depends on voltage, temperature, frequency, age, and mechanical stress. Class 2 ceramics (X5R, X7R, Y5V) lose capacitance under DC bias — a large fraction of it, sometimes most of it — because their dielectric is ferroelectric and the applied field reduces its permittivity. Class 1 ceramics (C0G/NP0) do not, at the cost of being roughly a thousand times less dense. Resistors drift with temperature and, in high-value thick-film parts, with voltage. Inductors stop being inductors above their saturation current, abruptly. The engineering skill is not memorising these effects; it is knowing which of them your circuit is sensitive to, and derating on purpose rather than by accident.

CORE CONCEPTResistors: five numbers, and only one of them is on the silkscreen

A surface-mount resistor is a thin or thick film of resistive material printed on a ceramic substrate, laser-trimmed to value, and covered in glass. That construction gives you five specifications that matter, and they are not equally important in the way beginners assume.

1. Tolerance — usually the number you should care least about

Tolerance is the guaranteed spread at 25 °C, at time of manufacture, at low applied voltage. The standard values come from the E-series: E24 gives 24 values per decade and pairs with 5 % parts, E96 gives 96 values per decade and pairs with 1 %. In modern practice the price difference between 5 % and 1 % thick-film chip resistors is negligible — often the same reel cost — so most designs simply use 1 % throughout and stop thinking about it. That is the right call, and it is also why tolerance is rarely the thing that bites. A pull-up resistor does not care about 1 %. An LED ballast does not care. A voltage divider ratio for an ADC cares a little, and a current-sense shunt cares a lot, but those two you specify deliberately anyway.

The exception worth naming: tolerance on a ratio matters more than tolerance on a value. Beacon's battery divider uses two 1 MΩ resistors and the firmware multiplies by exactly 2. If both are 1 % and drift independently, the ratio error can approach 1 % — about 37 mV on a 3.7 V cell, which on the flat middle of a LiPo discharge curve is a meaningful chunk of the state-of-charge estimate. Using a matched pair or a resistor network fixes it; so does a one-point calibration in production test (Lesson 6.3). Deciding which of those you can afford is a real engineering decision that starts with a tolerance number.

2. Power rating — a thermal statement with a temperature attached

Lesson 1.2 established the mechanism: power is heat, and a package can only shed so much. The datasheet number (0402 = 63 mW or 100 mW depending on the series, 0603 = 100 mW, 0805 = 125 mW, 1206 = 250 mW are the common values) is specified at a stated ambient — usually 70 °C — and comes with a derating curve that falls linearly to zero at the maximum film temperature, typically 155 °C. Above 70 °C ambient you get less. On a dense board with hot neighbours, the local ambient is not the room's ambient. The working rule stays: design to 50 % of the rating, and treat anything above 70 % as a thermal design problem that needs a bigger package or a plan.

3. Temperature coefficient — the drift you never see on the bench

Resistance changes with temperature, quoted in parts per million per degree Celsius. Standard thick-film chip resistors are commonly ±100 ppm/°C, sometimes ±200 ppm/°C for high values; thin-film parts reach ±25 ppm/°C and precision thin film ±10 ppm/°C or better. Do the arithmetic before dismissing it: a ±100 ppm/°C resistor over Beacon's 0–45 °C environment (R9) drifts by 100 × 45 = 4500 ppm = 0.45 %. That is comparable to the 1 % tolerance you paid for, which is the point — buying a 0.1 % resistor and pairing it with a 200 ppm/°C tempco buys you nothing outside a climate-controlled room.

Tempco matters most where a resistance sets an absolute quantity: a current-sense shunt, an oscillator's RC, a precision divider, a bias network for an analog sensor. It matters least where the resistance sets a ratio between two parts that share a temperature — a divider made from two resistors of the same series on the same board tracks itself and largely cancels.

4. Voltage coefficient — the one nobody mentions

Thick-film resistors are granular: conduction happens through a percolating network of conductive particles in glass. Apply a strong electric field and that network changes slightly, so resistance depends on applied voltage. The specification is a voltage coefficient of resistance, commonly around −100 ppm/V or worse for high-value thick-film parts, and negligible for thin film. At Beacon's 3.7 V it is invisible. In a 1 MΩ divider sensing a 400 V DC bus, it is a systematic error of several tenths of a percent that no calibration at low voltage will find. Store this one; it will save you an afternoon exactly once in your career.

5. Pulse and ESD withstand — where thick film really fails

A thick-film chip resistor is a very thin serpentine track. A short, high-energy pulse — an ESD strike, an inrush surge, a hot-plug event — deposits its energy in that track faster than the ceramic can conduct it away, and the film opens or drifts permanently upward. This is why series resistors sitting on external-facing lines (Beacon's 22 Ω on the USB D± pair, per the canon) should be specified as pulse-withstanding or anti-surge types where the vendor offers them, and why the resistor closest to a connector is the one that mysteriously reads 30 % high after a field return. Wirewound and metal-film through-hole parts, and thin-film chips with specific pulse ratings, do much better here.

💡 The reading order for a resistor datasheet

Power rating and its derating curve first (does it survive?), then temperature coefficient (does it stay put?), then tolerance (is it close enough today?), then the maximum working voltage (a 0402 is typically rated 50 V, and a high-value divider across a high rail can exceed it long before it exceeds its power rating), and last the pulse rating if the part touches the outside world. Notice that "tolerance", the number in the part's name, is fourth.

DEFINITIONCeramics: the three characters that decide everything

A multilayer ceramic capacitor (MLCC) is a stack of interleaved metal electrodes separated by thin ceramic layers, fired into a monolithic block. Capacitance is C = ε₀εr · A · N / d, where N is the number of layers, A the overlap area, d the layer thickness, and εr the dielectric's relative permittivity. Every trade-off in this section falls out of that equation.

Class 1 versus Class 2

Class 1 dielectrics (C0G, also written NP0) use paraelectric ceramics with modest permittivity — εr in the tens. They are linear: capacitance does not depend meaningfully on applied voltage, they have very low loss, they do not age, and their temperature coefficient is specified as 0 ± 30 ppm/°C. They are, in every electrical sense, the capacitor you were taught about in physics class. The price is density. Because εr is small, a C0G part of a given value is enormous compared with its Class 2 equivalent: an 0402 C0G tops out in the low nanofarads, an 0603 around 10 nF, an 0805 in the tens of nanofarads. You cannot buy a 10 µF C0G in a package that fits on Beacon.

Class 2 dielectrics (X7R, X5R, Y5V and relatives) are based on barium titanate, a ferroelectric with εr in the thousands. That is where the density comes from — and every unpleasant behaviour in this lesson is the price of that permittivity. Ferroelectric ceramics have domains that reorient under an applied field; once they are aligned, the material's incremental permittivity falls. That single physical fact produces DC-bias derating, temperature dependence, aging, dielectric absorption, and the piezoelectric "singing" that makes ceramic capacitors audible under an AC bias.

Decoding the EIA three-character code

For Class 2 parts the code is read as low-temperature limit, high-temperature limit, and the allowed capacitance change over that range:

Character 1 — low tempCharacter 2 — high tempCharacter 3 — change over range
X = −55 °C5 = +85 °CP = ±10 %
Y = −30 °C6 = +105 °CR = ±15 %
Z = +10 °C7 = +125 °CS = ±22 %
8 = +150 °CV = +22 % / −82 %

So X7R is ±15 % from −55 to +125 °C. X5R is ±15 % from −55 to +85 °C — the same tolerance over a narrower range, which is why X5R is cheaper and denser and why it is the default in consumer products. Y5V is +22 %/−82 % from −30 to +85 °C, which is a specification so wide that it is close to a confession: at the top of its range a Y5V part may retain less than a fifth of its marked value. Y5V exists because it is cheap and because some applications genuinely do not care. Yours probably does.

Read the third character as what it is: a guaranteed band, not a typical. And read it as applying to temperature only. The DC-bias effect is not in the code at all. A part can be perfectly within its X7R specification and still deliver 40 % of its value in your circuit, because bias derating is documented in a curve, not in the part number.

✗ The single most expensive misreading in this lesson

"X7R is ±15 %, so my 10 µF is at worst 8.5 µF." No. The ±15 % is the temperature band, measured at a small AC test signal with no DC bias. Add the initial tolerance (±10 % for a K-code part, ±20 % for M), add the DC-bias derating (which can exceed −70 % for a dense small-package part at its rated voltage), add aging (a few percent per decade of hours), and the worst-case effective capacitance of a nominally 10 µF part can be under 2 µF. These effects multiply; they do not average out, and they are not independent random variables you can root-sum-square away, because bias derating is systematic — every unit on every board has it.

DC-bias derating, the mechanism

To make an MLCC dense you make the dielectric layers thin. A 10 µF 6.3 V part in an 0402 case has layers on the order of a micrometre. Apply 3.3 V across a one-micrometre layer and the field is roughly 3.3 MV/m — an enormous field, well into the region where barium titanate's domains saturate. The higher the field, the lower the incremental permittivity, and the lower the capacitance the circuit actually experiences.

Three design levers follow directly, and they are the only three you have:

What you cannot do is buy a better brand. All manufacturers of Class 2 MLCCs face the same physics; the curves differ in detail, not in kind. What differs is disclosure — Murata, TDK, Samsung and KEMET all publish bias curves and free simulation tools (SimSurfing, TDK's characteristic viewer, KEMET's K-SIM), and using them for the exact part number you intend to buy is a ten-minute task that removes this entire failure mode from your project. The widget below teaches the shape; the vendor's tool gives you the number for the part.

why thin layers derate 10 µF 6.3 V 0402 layers ≈ 1 µm 3.3 V → ≈ 3.3 MV/m keeps ≈ 30 % 10 µF 16 V 1206 layers much thicker 3.3 V → weak field keeps ≈ 80 % same nameplate, same tolerance code, two-and-a-half times the capacitance in circuit the shape you must expect applied DC bias → C C0G X7R, big case X5R, dense 0402 the curve is monotonic and systematic — every unit does this it is not a tolerance and it does not average away across a batch
Left: the same capacitance built two ways. Density costs field strength, and field strength costs permittivity. Right: the qualitative shape of the DC-bias curve for three dielectric choices. The percentages shown are representative of published vendor curves for parts of this class — always pull the curve for the exact part number you intend to order.

The other Class 2 behaviours, briefly and honestly

HOW IT WORKSESR, ESL, and why a capacitor stops being a capacitor

An ideal capacitor's impedance falls forever as frequency rises: |Z| = 1/(2πfC). A real one is that capacitance in series with a resistance (ESR — electrode and termination losses plus dielectric loss) and an inductance (ESL — the loop the current takes through the electrodes, terminations, pads, traces, and vias). The result is a V-shaped impedance curve: capacitive on the way down, resistive at the bottom, inductive on the way up. The bottom of the V is the self-resonant frequency, fSRF = 1 / (2π√(L·C)), and above it the part is an inductor with a capacitor in the way.

frequency (log) → |Z| 100 kHz 5 MHz 50 MHz 500 MHz 1 µF 0603 100 nF 0402 f_SRF capacitive: |Z| = 1/(2πfC) ESR floor inductive: |Z| = 2πf·L this is where decoupling stops working above resonance the smaller part wins — but the loop you solder it into sets most of the L
The impedance V. The capacitive slope is set by capacitance, the floor by ESR, the rising slope by ESL. The smaller 100 nF part resonates higher and therefore stays useful further up in frequency. Shape is representative; exact positions depend on the part and, above all, on the mounting.

The arithmetic, once

An 0402 MLCC's own ESL is roughly 0.4–0.6 nH; an 0603 roughly 0.8–1 nH. But the loop that matters includes the pads, the trace to the via, the via barrel, the plane, and the return — the mounting inductance — which on an ordinary layout is 1–2 nH, comfortably more than the part. So take L ≈ 1.5 nH as a realistic total for a well-mounted 0402:

This is the origin of the classic advice to place a 100 nF and a 1 nF in parallel: the small one covers the frequencies where the big one has become inductive. It is also the origin of a long-running argument among people who lay out boards for a living, and the argument is worth understanding because it teaches you how to read this kind of advice.

🧪 What experienced engineers actually argue about

Two capacitors of different values in parallel do not simply cover each other's gaps. Between the small part's resonance and the large part's resonance, the large part is inductive and the small part is still capacitive, and that inductor-plus-capacitor forms a parallel resonant tank — an anti-resonance where the combined impedance is higher than either part alone. The classic 100 nF + 1 nF pair has a real impedance peak between them. Signal-integrity practitioners (Bogatin and Sandler have both written extensively on this) generally argue for many identical capacitors rather than a decade-spread ladder, because identical parts resonate together and their ESRs damp the peak, and because on a modern board the mounting inductance dominates the part's own inductance anyway. The counter-argument is that anti-resonance peaks are damped enough by real ESR and plane losses that the ladder is harmless and the extra value costs nothing. For Beacon — a low-power BLE tag with no high-speed digital bus — this argument is academic, and the practical rule is: one 100 nF as close as possible to each supply pin with the shortest possible via to the ground plane, plus bulk. But you should know why the rule you were given is contested, so that when you build something with a fast bus you know which literature to read.

ESR: when you want it low, and when you need it high

Ceramics have very low ESR — single-digit milliohms is common. Aluminium electrolytics have hundreds of milliohms to ohms, tantalum capacitors tens to hundreds, polymer electrolytics tens. Low ESR sounds strictly better, and for decoupling it is: less voltage developed across the capacitor for a given current spike, and less self-heating from ripple current.

But not always. Some older linear regulators and some control loops were compensated assuming a specific ESR range in the output capacitor — the capacitor's ESR provides a zero that stabilises the loop. Replace that tantalum with a "better" ceramic and the zero disappears and the regulator oscillates. Modern parts are usually explicitly ceramic-stable and say so on the front page, but if you ever find a datasheet with a stability plot whose axes are ESR and load current, that is what you are looking at, and the plot is not decoration. Read it before you improve anything.

The other place ESR earns its keep is ripple-current heating in power supplies: P = Iripple,rms² × ESR, dissipated inside the capacitor, which for an electrolytic sets its life. Electrolytic lifetime roughly halves for every 10 °C of internal temperature rise, so ESR is a lifetime specification wearing an impedance costume.

📖 The division of labour

Ceramic (MLCC): tiny, cheap, very low ESR and ESL, no wear-out mechanism, but derates hard under bias and cracks under flex. Job: high-frequency decoupling and anything under a few tens of microfarads. Aluminium electrolytic: enormous capacitance per rupiah, high ESR and ESL, dries out with time and heat (a wear-out mechanism you must design a lifetime around). Job: bulk energy storage and low-frequency ripple. Tantalum / polymer: dense, stable capacitance with no bias derating, moderate ESR; tantalum fails short and can ignite if abused, so it needs voltage derating of 2× or more. Job: bulk where the board is small. Film: stable, lossless, no derating, physically large. Job: precision analog and mains-adjacent circuits. Beacon uses ceramics for everything, because at 25 units at 35 × 25 mm nothing else fits — and that decision is exactly what makes this lesson's derating arithmetic mandatory rather than optional.

SEE ITCapacitor DC-bias derater — pick a part, watch the circuit change its mind

Choose a dielectric, a package, a voltage rating, a nominal value, and the DC bias your circuit will actually apply. The bench reports the effective capacitance after bias, temperature, aging, and worst-case initial tolerance, then judges it against the job you assigned it. The model is fitted to the shape of published vendor curves and is illustrative in its exact percentages — for a real design, pull the curve for the exact part number from the vendor's own tool. What is not illustrative is the direction, the magnitude, and the fact that the effect is systematic.

Interactive · Effective capacitance bench

Try to build a compliant 10 µF buck output capacitor in an 0402. Then find out what it costs to build a compliant one.

10 µF
3.3 V
Nominal
10.0 µF
Effective, typical unit
3.0 µF
Fraction retained
30 %
Worst-case unit
2.7 µF

Effective capacitance against applied DC bias, from zero to the part's rated voltage, at the selected temperature and age. The flat grey line is the nameplate value you specified in the BOM. The gap between the two lines is the part of your design that exists only on paper.

Things to try, in this order. Start at the defaults — 10 µF, X5R, 0402, 6.3 V, 3.3 V bias — and read the buck verdict: the part is not merely derated, it is outside the range the regulator's datasheet asks for, and it got there without a single component failing. Now fix it three different ways and compare the costs. Move to an 0805 and the effective value roughly doubles, at the cost of board area that Beacon's 35 × 25 mm (R8) does not have much of. Move to a 16 V rating in the same 0402 and watch what happens: the nominal 10 µF is not even available at 16 V in an 0402, and the bench says so — the part you specified does not exist. Go back to 0402 6.3 V and put two 10 µF in parallel: the bench models one part, but you can read the answer off it, because two identical derated parts give you twice a third, not two-thirds of twenty. Then switch the dielectric to C0G and watch the availability collapse; switch to Y5V and watch what happens when you also move the temperature to 85 °C. Finally, set the role to the RC filter and discover that only one dielectric on the list can hold ±5 %, and it cannot hold microfarads.

CORE CONCEPTInductors: the part that fails by ceasing to exist

An inductor stores energy in a magnetic field: E = ½LI². In a switching regulator it is the element that makes the whole thing work, moving energy in packets from input to output. Three numbers on its datasheet decide whether your regulator works, and the third one is defined differently by different vendors, which is a preview of Lesson 1.6.

Inductance, and the frequency it was measured at

Inductance is specified at a stated test frequency and a stated (usually zero) DC current. Ferrite-cored parts hold their value well below saturation; powdered-iron and composite cores "soft-saturate", losing inductance gradually as current rises, which some designers prefer because the failure is graceful.

Saturation current (Isat) — a definition, not a limit

As current rises, the core's magnetic domains align. Past a point, adding current adds almost no flux, so dL/dI collapses and the part stops behaving like an inductor: the current, which had been ramping linearly, turns sharply upward toward whatever the circuit's resistance allows. In a buck converter that means the switch current spikes far beyond the design value in a fraction of a microsecond, and the current-limit comparator has to catch it. Sometimes it does.

The trap is the definition. Isat is specified as the current at which inductance has dropped by a stated percentage — and vendors variously use 10 %, 20 %, or 30 %. A part with "Isat = 1.5 A (−30 %)" and one with "Isat = 1.5 A (−10 %)" are not the same part, and the second is considerably better. Two datasheets, two definitions, one BOM line. You must read the footnote, and if you are comparing parts you must normalise them to the same criterion using the curves.

Rated current (Irms) and DCR — the thermal specification

Separately from saturation there is a heating limit: the DC winding resistance (DCR) dissipates I²·DCR, and the rated current is the current that produces a stated temperature rise, conventionally 40 °C. That is also a definition, not a limit, and it says nothing about saturation. A part can be thermally fine and magnetically saturated, or magnetically fine and cooking. The design current must be below both. And note the peak-versus-average distinction: saturation cares about the instantaneous peak, which in a buck is the DC output current plus half the ripple current; heating cares about the RMS.

💡 Sizing a buck inductor in four lines

Ripple current ΔIL = Vout · (1 − Vout/Vin) / (L · fsw). Aim for ΔIL at roughly 30 % of the maximum output current. Peak current Ipk = Iout(max) + ΔIL/2, and that is the number to compare against Isat — with margin, because Isat falls with temperature. Then check I²·DCR against the efficiency you promised. Worked for Beacon's buck at 3.7 V in, 3.3 V out, L = 4.7 µH, fsw = 1.5 MHz: ΔIL = 3.3 × (1 − 0.892) / (4.7 µH × 1.5 MHz) ≈ 51 mA. Against Beacon's 40 mA active load that ripple is larger than the load itself — which is normal and fine for a light-load converter, and it is exactly why the regulator spends most of its life in a pulse-skipping mode rather than continuous conduction. Lesson 4.2 does this properly.

Two more practical notes. Shielded versus unshielded: an unshielded inductor radiates its field into everything nearby, and "nearby" on a 35 × 25 mm board includes an antenna. Shielded parts cost more and are what you use next to RF. Audible noise: inductors sing for the same reason capacitors do — magnetostriction and winding movement at audio-band switching frequencies. If your regulator drops into a burst mode whose repetition rate lands between 1 and 15 kHz under some load, your product whistles, and customers notice.

BEACONBeacon's passive BOM, with a reason attached to every part

Beacon's canon fixes the active parts; this lesson fixes the passives around them. Here is the reasoning, in the order the decisions get made, using the canon's numbers.

1

The ADC hold capacitor: 100 nF, and it must be X7R or better

Lesson 1.3 sized this at 100 nF to sit 10,000 : 1 above the ADC's sampling capacitor while settling through 500 kΩ in 250 ms. The bias across it is at most 2.1 V (half of a full 4.2 V cell), so pick a 16 V or 25 V rating in an 0402 and the derating is a few percent — the widget shows about 12 % at 16 V, less at 25 V. Now the honest part: the firmware's 250 ms wait was derived from a capacitance. If procurement substitutes a Y5V, that 100 nF becomes perhaps 30 nF at 45 °C, τ falls from 50 ms to 15 ms, the wait is now wildly conservative — and the sampling droop over 64 averaged conversions triples. Neither symptom is dramatic; the battery reading is simply a bit off, in a way that varies with temperature, and nobody finds it for a year. Specify the dielectric on the BOM line, not just the value.

2

Decoupling at the ESP32-C3-MINI-1: 100 nF per supply pin, plus 10 µF bulk

The module's own datasheet and reference design specify the decoupling; follow it rather than inventing your own. What this lesson adds is the mounting discipline: the 100 nF goes on the same side as the pin, with its ground pad connected to the plane by its own via, and the loop from pin to capacitor to via to plane kept as short as the footprint allows. That loop is 1–2 nH and it dominates the capacitor's own 0.5 nH, so a longer trace to a "shared" ground via can double the effective inductance and halve the frequency above which the decoupling stops working. The bulk 10 µF exists to supply the BLE burst — tens of milliamps stepping on in microseconds — without the LDO's or buck's loop having to react instantly. Its DC-bias derating matters here in a benign direction: 10 µF that behaves as 3 µF still supplies far more charge than the burst needs, which is why bulk decoupling is the one place you can be sloppy about this.

3

The buck output capacitor (Stage 4): the one that must be checked, not assumed

The canon replaces the breadboard's XC6220 LDO with a TPS62840 buck in Lesson 4.2, and a buck's datasheet specifies an output capacitance range that its control loop was designed around. That specification is on effective capacitance. If you place a 10 µF 6.3 V X5R 0402 and the circuit sees 3 µF at 3.3 V, you are outside the specified range, and the symptoms are the ugly kind: a control loop behaving differently from its simulation, larger output ripple, an audible or intermittent instability that appears at some load currents and battery voltages and not others. This is the lesson's opening scene, and it is entirely preventable by opening the vendor's bias curve once. The fix on a 35 × 25 mm board is usually two parallel capacitors or a higher voltage rating in the same footprint, not a bigger package.

4

Resistors: 1 % thick film everywhere, with three exceptions

The two 1 MΩ divider resistors want tolerance and tempco matching more than absolute accuracy — same series, same package, placed adjacently so they share a temperature. The 22 Ω series resistors on the USB D± pair (canon) sit on an external connector and should be pulse-rated parts, because that is the line that meets a person's finger and a cheap charger. The WS2812B-2020 needs no ballast at all — it carries its own constant-current drivers — so the only resistor near it is the small series part on the data line, chosen for edge quality rather than for current. Its high-side switch, however, is a power question, and Lesson 1.2's 50 % rule decides that package. Everything else on Beacon is a pull-up, a pull-down, or a series damping resistor, and 1 % 0402 thick film is correct and costs almost nothing at 25 units.

beacon_adc.cthe constant that is secretly a component specification
/* Battery-sense settling time.
 * Derived from Rth = R1 || R2 = 500 kOhm and C_hold = 100 nF: tau = 50 ms, 5*tau = 250 ms.
 * This number is only correct while C_hold is an X7R (or better) with >= 16 V rating.
 * A Y5V substitution at 45 C would make the real tau roughly 15 ms and would triple
 * the droop across 64 averaged conversions. Do not change the BOM line without
 * changing this constant -- and vice versa. */
#define BEACON_ADC_VBAT         ADC1_CHANNEL_4   // GPIO4 (canon pin map)
#define BEACON_VBAT_SETTLE_MS   250
#define BEACON_VBAT_AVG_SAMPLES 16

static int beacon_vbat_mv(void)
{
    vdiv_enable(true);                            // PMOS connects the divider
    light_sleep_ms(BEACON_VBAT_SETTLE_MS);         // hide the wait, ~130 uA not 40 mA
    int raw = adc_read_avg(BEACON_ADC_VBAT, BEACON_VBAT_AVG_SAMPLES);
    vdiv_enable(false);
    return adc_raw_to_mv(raw) * 2;                  // exactly 2 because R1 == R2
}
🔑 The BOM line is part of the design

Three of the four decisions above are invisible in the schematic. A schematic says "100 nF" and "10 µF"; it does not say X7R, 16 V, 0402, and it certainly does not say "the firmware's timing constant depends on this". Beacon's BOM therefore carries dielectric, voltage rating, and package on every capacitor line, and a one-line reason on the four lines where the reason is not obvious. That discipline is what makes a second-source substitution during a shortage a five-minute decision instead of a three-week field failure — and Stage 6's cost work (6.8) depends on being able to tell which lines are negotiable.

WEAKNESSWhere this bites

✗ Verifying capacitance with an LCR meter and concluding the part is fine

A handheld LCR meter measures with a small AC signal and no DC bias. The 10 µF 0402 that gives 3 µF in your circuit will measure close to 10 µF on that meter, every time, on every unit. You will conclude the parts are good and go looking for the fault elsewhere — which is exactly how the opening scene consumed two days. To measure what the circuit sees you need a meter with a DC bias source (many benchtop LCR meters have one, or you can add an external bias tee), or you infer it from circuit behaviour: measure the actual output ripple, or step the load and measure the droop, and back out the capacitance. Better still, do not measure at all: read the vendor's published bias curve for the part number before you place it.

✗ Treating derating effects as tolerances that average out

There is a strong instinct to root-sum-square independent errors, and it is correct for genuinely independent random variables. DC-bias derating is not one. It is a deterministic function of the applied voltage that every unit of that part number obeys, so it applies fully to every board you build — a 100 % systematic shift, not a distribution. The same is true of the temperature term and of aging. Only the initial tolerance is a distribution. The correct arithmetic is therefore: multiply the systematic factors, then apply the tolerance to the result — which is what the widget does when it reports a worst-case unit. Getting this wrong is how a design "passes on paper" and then fails on the majority of units rather than a few.

✗ Assuming the reference design's part list is a specification

Reference designs list values, and their BOMs list the part numbers the evaluation board was built with — usually generous packages and voltage ratings, because an evaluation board has room. When you shrink the layout onto a 35 × 25 mm product and swap the 0805 for an 0402 "because it is the same 10 µF", you have quietly changed a specified parameter of the regulator's control loop. The same trap catches inductors: an evaluation board's 4.7 µH shielded part with 1.5 A saturation becomes a smaller, cheaper 4.7 µH with 0.6 A saturation, and the converter is fine at light load and current-limits on the burst. The value is the least specific thing about a passive component. Package, rating, dielectric or core, and the vendor's own curves are the specification.

KEY INSIGHTHow this connects

The through-line of this lesson is a habit, not a fact: when you write a value on a schematic, ask what conditions make it true and whether your circuit satisfies them. That habit turns into a datasheet-reading skill in 1.6, into a measurement skill in 1.7, and eventually into the reason Beacon's Rev A only has three defects instead of eight.

GLOSSARYKey terms introduced

Class 1 / Class 2 dielectric
Class 1 (C0G, NP0) uses low-permittivity paraelectric ceramic: linear, stable, no bias derating, no aging, low density. Class 2 (X7R, X5R, Y5V) uses ferroelectric barium titanate: high density, and hence bias derating, temperature dependence, aging, and piezoelectric behaviour.
DC-bias derating
The systematic loss of capacitance in a Class 2 MLCC when a DC voltage is applied, caused by the field reducing the dielectric's incremental permittivity. Worse for thinner layers, so worse for denser packages and lower voltage ratings. Not a tolerance — every unit does it.
EIA temperature code
The three characters (e.g. X7R) giving the low-temperature limit, high-temperature limit, and allowed capacitance change over that range, measured with no DC bias applied.
ESR (equivalent series resistance)
The capacitor's internal loss resistance. Sets the floor of the impedance curve, the heating from ripple current (P = Irms²·ESR), and — for some regulator loops — the compensation zero.
ESL (equivalent series inductance)
The capacitor's series inductance, dominated in practice by the mounting loop (pads, trace, via, plane) rather than by the part itself. Sets the rising side of the impedance V.
Self-resonant frequency (fSRF)
f = 1/(2π√(L·C)), where a capacitor's impedance is at its minimum and equal to ESR. Above it the part behaves as an inductor and no longer decouples.
Anti-resonance
The impedance peak formed between two paralleled capacitors of different values, where the larger part is already inductive and the smaller is still capacitive. The reason "one big plus one small" is contested advice rather than a law.
Saturation current (Isat)
The current at which an inductor's inductance has fallen by a stated percentage — 10 %, 20 %, or 30 % depending on the vendor. Compare against peak current, not average, and normalise the definition before comparing two parts.

ACTIVE RECALLCheck your understanding

Answer in your own words first, then reveal.

1A regulator datasheet asks for 10 µF on its output. You place a 10 µF ±10 % 6.3 V X5R 0402 on a 3.3 V rail. Walk through every factor between the nameplate and what the loop sees, and say which of them are systematic.›

Four factors, applied multiplicatively. DC bias at 3.3 V on a 6.3 V-rated dense 0402: representative curves for this class of part leave roughly 30 % — systematic, identical on every board. Temperature: X5R is ±15 % to 85 °C, so at the top of Beacon's 45 °C range expect a few percent more loss — systematic. Aging: a few percent per decade of hours from reflow, so several percent gone within a year — systematic. Initial tolerance: ±10 % for a K-code part — this one is the only genuine distribution. Multiply: 10 µF × 0.30 × 0.96 × 0.94 ≈ 2.7 µF typical, and about 2.4 µF on a unit at the bottom of its tolerance. The regulator was designed for 10 µF and is seeing roughly a quarter of it. The critical point is that three of the four are systematic, so this is not a yield problem affecting a few boards — it is the design, on every board.

2Your colleague proposes replacing the 100 nF decoupling capacitors with 1 µF "for more margin", in the same 0402 package and the same 16 V rating. Argue both sides, then decide.›

For: at low frequencies 1 µF has ten times lower impedance, so the bulk-supply job improves, and one part number replaces two. Against: the mounting inductance is unchanged at roughly 1.5 nH, so fSRF falls by √10 — from about 13 MHz to about 4 MHz. Above 4 MHz the part is an inductor, and above about 13 MHz it is a worse inductor than the 100 nF would have been at the same frequency only because there is more capacitance in series with it doing nothing useful. If the thing you were decoupling has energy at tens of megahertz, the swap makes the high-frequency job worse while making the low-frequency job better. Also, 1 µF in an 0402 at 16 V is a denser part than 100 nF at 16 V, so it derates more under the same 3.3 V bias. Decision for Beacon: keep 100 nF at the pin and add bulk separately where there is room. Beacon has no fast digital bus, so in truth either choice works — which is the honest answer, and knowing that it is the honest answer is worth more than a rule.

3Two 4.7 µH inductors: part A says Isat = 1.2 A at −30 % and DCR = 90 mΩ; part B says Isat = 0.9 A at −10 % and DCR = 150 mΩ. Beacon's converter has a 51 mA ripple and a 40 mA active load but a 700 mA peak during a hypothetical Wi-Fi OTA burst. Which do you choose, and what else do you need?›

First normalise: the two Isat figures are not comparable as printed. Part B's 0.9 A is where it has lost only 10 %; its −30 % point is higher, plausibly above part A's 1.2 A. To compare them properly you need the L-versus-I curves, not the headline numbers — and if only one vendor publishes a curve, that is itself information about which vendor to trust. Second, compute the peak: 700 mA + 51/2 ≈ 726 mA during the OTA burst, which is inside both parts' headline ratings but with less margin than it looks, because Isat falls with temperature and both are quoted at 25 °C. Third, DCR: at 726 mA, part A dissipates 0.726² × 0.09 = 47 mW and part B 79 mW, which matters for efficiency during OTA but is irrelevant for the 40 mA steady state. Fourth — and this is the point of the question — you also need to know whether the part is shielded, because it sits on a 35 × 25 mm board with a 2.4 GHz antenna at one edge (R8, canon 5.7). An unshielded part that wins on every number in the table can still be the wrong choice.

4A field unit reports a battery voltage about 6 % lower than its neighbours, and the discrepancy grows in the afternoons. The divider resistors measure correctly. What is your first hypothesis from this lesson, and how would you confirm it without a site visit?›

The hypothesis: the hold capacitor on that unit is not the specified X7R — a Y5V or an out-of-spec substitution — so its capacitance falls sharply as the enclosure warms in the afternoon. That does two things. The settling time constant shrinks, which alone would make the reading better, not worse. But the droop across the averaged conversions grows in proportion to Cs/Chold, and it is that droop which pulls the reading down, more strongly at higher temperature. A competing hypothesis worth holding is a leaky PMOS or a higher pin leakage at temperature, which produces a DC offset through Rth (Lesson 1.3) and also worsens with heat. To distinguish them without visiting: Beacon's firmware records the raw ADC value, the number of samples averaged, and the achieved wait (Lesson 1.3), and it reports temperature from the SHT40 in the same payload. Ask the unit for a reading at 1 sample instead of 16. If the discrepancy largely disappears, it is droop, and therefore the capacitor. If it persists unchanged, it is a DC leakage path, and therefore the switch or the pin. One firmware command, no travel — which is why those three fields are in the payload.

Next: a MOSFET is a resistor you set with the gate — and in between, a heater →