OpenBench

Build log

Science Shouldn’t Cost Thousands: Building an Under-£10 Light Detector for an Open-Source Plate Reader

Part 1: circuit design and characterisation, the first of many home-built instruments

This is the first write-up in an ongoing project: building real laboratory instruments at home from scratch, with a soldering iron, a 3D printer and parts that cost a few pounds. Each instrument starts by studying how the professional version works, then works downwards to something any student could build.

Abstract

This project started in an A-level Biology lesson a few weeks ago. I used a colorimeter for the first time, for a quantitative Benedict's test of glucose concentration, and the results were poor because the school's colorimeters were so inaccurate. I wanted to understand why, so I decided to build one myself and measure exactly how good it is.

Absorbance microplate readers are central to biology and biochemistry, but commercial instruments cost thousands of pounds, which keeps them out of most schools and home labs. As the first step towards an open-source, 3D-printed plate reader, I designed, built and characterised the light-detection circuit at its core.

Each of two identical channels uses a BPW34 silicon photodiode, a transimpedance amplifier built on one MCP6002 op-amp, and a 16-bit ADS1115 analogue-to-digital converter read by an Arduino Nano. One channel measures light that has passed through the sample. The other watches the light source directly, copying the double-beam design of laboratory spectrophotometers so that changes in source brightness cancel out.

On a breadboard, the amplified chain held steady to within 0.4 mV in constant light, only about three of the converter's 0.125 mV steps. The two channels agreed to within 3.2% under the same light, and their ratio moved only in single converter steps (0.46%, about ±0.002 absorbance units). On the final soldered board, both channels responded independently to light, and the dark output was 3.2 to 4.1 mV, unchanged by room lighting, which identifies it as the op-amp's output floor.

The detector parts cost under £10. Part 2 will cover the 3D-printed optical enclosure and the first calibration against a laboratory spectrophotometer.

Block diagram of the double-beam signal chain: LED to cuvette to photodiode to amplifier A to the ADS1115 converter to the Nano, with a second photodiode and amplifier B reading the LED directly
The double-beam signal chain. Only the sample beam passes through the cuvette; both channels are read at the same moment. The source in this build is a white LED with a red gel filter — the diagram also names the WS2812B, which was the alternative I considered.

My mission

I run Overengineered Sim Racing Solutions (oesimracing.com), designing, 3D printing and building affordable sim racing accessories myself. Reviewer Laurence Dusoswa tested my gear and rev indicator and concluded: “For roughly £45 / €50, I highly recommend it.”

Now I’m applying the same principle where it matters more. In sim racing, high prices shut people out of a hobby; in science, they decide who gets to do real experiments. My mission is to build lab instruments any student can afford, test them rigorously, and publish how to make them.

Introduction

A microplate reader shines light through up to 96 small samples and measures how much is absorbed. That single number drives a huge range of everyday lab work: protein and DNA assays, enzyme kinetics, bacterial growth curves and drug screening. The physics behind it is the Beer–Lambert law, which says absorbance rises in proportion to the concentration of the absorbing substance.1

A = log10(I0 ÷ I) = ε c l

Here I₀ is the light through a blank, I is the light through the sample, ε is the molar absorption coefficient, c is the concentration and l is the path length.

The problem is cost. A commercial absorbance plate reader typically costs several thousand pounds, so most schools never have one, and students meet these experiments only on paper. Open hardware has already shown another way. The OpenFlexure Microscope brought lab-grade, motorised microscopy to 3D printers,2 and a team at the University of Pennsylvania built an open-source plate reader for under $3,500.3 That is impressive, but it is still far beyond a student budget.

I want to be fair here: commercial plate readers aren’t overpriced. They’re used in hospitals and drug development, where a wrong number can mean a wrong diagnosis. Every one has to hit tight tolerances, carry certified calibration and pass regulatory checks, and I understand why that costs thousands.

But that standard doesn’t help a student who just wants to see Beer–Lambert work with their own eyes. I’m not trying to build something a hospital would trust. I’m trying to build something a school can afford, and then prove, with real data, how good it actually is. If the choice is between my £10 detector and nothing, I know which one teaches you more.

Before this project, I built an OpenFlexure microscope myself. It taught me how much a well-designed open instrument can do, but I was following a published instruction manual written by experts. This plate reader is different: there is no manual. Every design decision, from choosing the photodiode to the layout of the soldered board, is my own, and so is every mistake recorded in the discussion below.

I felt the gap between real and school equipment first-hand. In my first quantitative Benedict's test in A-level Biology, the school colorimeters gave readings too inaccurate to make a usable calibration curve. Nobody in the room could say whether the fault was the instrument, the method or us. Building my own detector, and measuring its performance at every stage, was my way of finding out what a colorimeter needs to get right.

My long-term aim is a plate reader that anyone with a 3D printer can build for under £100. I started by studying how real instruments work and then worked downwards to parts I could afford, solder and print. Running my sim racing business (see My mission) meant I already had the printer, the soldering iron and practice at turning an Arduino prototype into a finished product people pay for. This project points those same skills at a harder problem, where the result has to be measurably accurate, not just work.

This paper, Part 1, covers only the electronics: the circuit that turns light into an accurate number. It is the part that decides whether the whole instrument can be trusted. My specific aims were to:

  1. Build a light detector with enough resolution to measure absorbance up to about 2 (1% of the light getting through).
  2. Add a reference channel so that drift in the light source cancels out, as in a double-beam spectrophotometer.
  3. Measure the detector's noise, matching and dark signal honestly, and record what went wrong along the way.

I am an A-level student, not a trained engineer, so I have tried to report the mistakes as clearly as the results.

Design: working down from a laboratory instrument

I designed the circuit by taking each part of a laboratory spectrophotometer and finding the cheapest part that does the same job without giving up accuracy.

FunctionLaboratory spectrophotometerThis detector
Light sourceTungsten or xenon lamp with a monochromatorWhite LED with a red gel filter
DetectorSilicon photodiode or photomultiplier tubeBPW34 silicon PIN photodiode
Current-to-voltage conversionLow-noise transimpedance amplifierMCP6002 op-amp with a 100 kΩ feedback resistor
Digitisation16 to 24-bit converterADS1115 16-bit converter
Drift correctionReference beam (double-beam design)Second identical photodiode channel
Control and outputEmbedded computerArduino Nano over USB

Why a photodiode and not an LDR. Light-dependent resistors are the usual cheap choice, but they respond slowly, drift with temperature and are not linear in light intensity. A photodiode produces a current almost exactly proportional to the light falling on it, across many orders of magnitude. That linearity is what Beer–Lambert measurements depend on. The BPW34 has a large 7.5 mm² sensitive area and costs about 50p.4

Turning a tiny current into a voltage. The photodiode's current is tiny: 0.3 to 3.5 µA in room light. A transimpedance amplifier holds the photodiode at 0 V and forces its current through a feedback resistor, Rf, giving an output voltage of:

Vout = Iphoto × Rf

With Rf = 100 kΩ, each microamp becomes 0.1 V. A 10 nF capacitor (Cf) across the resistor removes fast noise, such as flicker from mains lighting. It limits the bandwidth to about 160 Hz, which is still far faster than any absorbance measurement needs.

fc = 1 ÷ (2π Rf Cf) = 1 ÷ (2π × 100 kΩ × 10 nF) ≈ 159 Hz

I chose the MCP6002 because it runs from a single 5 V supply, its inputs draw almost no current, and one chip holds two amplifiers: one for each channel.5

Why 16 bits matter. Absorbance is logarithmic, so the signal shrinks very quickly. At A = 1 only 10% of the light gets through, and at A = 2 only 1%. The Nano's built-in 10-bit converter has steps of about 4.9 mV, so a dark sample falls into just a few steps. The ADS1115 has steps of 0.125 mV at its ±4.096 V range, which is 40 times finer.6

Sample (blank reads 0.25 V)Nano built-in converter (4.9 mV steps)ADS1115 (0.125 mV steps)
A = 1 (25 mV left)about 5 steps, roughly ±0.09 Aabout 200 steps, roughly ±0.002 A
A = 2 (2.5 mV left)0 or 1 step, not measurableabout 20 steps, measurable

The double-beam idea. An LED's brightness drifts as it warms up. If only the sample beam were measured, that drift would look like a change in absorbance. Laboratory double-beam instruments solve this by measuring a reference beam at the same moment. I did the same: a second photodiode sees the LED directly and never through the sample, and the instrument reports the ratio of the two. Subtracting the dark readings and dividing by the same ratio for a blank gives:

A = −log10( [(S − Sdark) ÷ (R − Rdark)] ÷ [(S0 − Sdark) ÷ (R0 − Rdark)] )

Here S and R are the sample and reference voltages, and S₀ and R₀ are the same readings taken with a blank. A brightness change multiplies S and R equally, so it cancels in the ratio. The two channels are identical apart from the cuvette, so anything that changes both equally, like LED brightness, cancels in the ratio S ÷ R.

Materials and methods

The detector was built in five stages between 16 and 18 October 2026. Each stage was tested before the next was added, so that any fault could be traced to one change. All parts were ordered from AliExpress on 23 September 2026.

Components
PartRoleKey specification
BPW34 PIN photodiode (×2)Light detector, sample and reference7.5 mm² sensitive area, peak response about 900 nm
MCP6002 dual op-ampTwo transimpedance amplifiersSingle 5 V supply, rail-to-rail output
100 kΩ resistor (×2)Feedback resistor RfSets the gain: 0.1 V per µA
10 nF ceramic capacitor (×2)Feedback capacitor Cf159 Hz bandwidth, removes mains flicker
100 nF ceramic capacitorSupply decoupling at the op-ampAbsorbs fast spikes on the 5 V line
ADS1115 module16-bit converter, 4 inputs±4.096 V range, 0.125 mV steps, I²C address 0x48
Arduino Nano (CH340 clone)Control, maths, USB output5 V, ATmega328P
5 × 7 cm double-sided perfboard, DIP-8 socketPermanent boardSocket lets the op-amp be swapped

Stage 1: identifying the photodiode. The BPW34's clear package made its anode and cathode impossible to tell apart by eye. I identified them by connecting the photodiode with a 100 kΩ load resistor to the Nano's analogue input A0 and checking which orientation gave a reading that rose with light. I used the Nano's 1.1 V internal reference and averaged 20 readings for finer steps.

Close-up of a BPW34 photodiode in its clear plastic package with two bent legs
The BPW34's clear package: the anode and cathode cannot be told apart by eye.
A photodiode and resistor on a breadboard wired to an Arduino Nano for the orientation test
Stage 1: the photodiode on the Nano's A0 input with a 100 kΩ load resistor.

Stages 2 and 3: the amplifier chain on a breadboard. I then built one transimpedance amplifier on a breadboard. The photodiode's cathode went to the op-amp's inverting input (pin 2) and its anode to ground. Rf and Cf were connected in parallel between pins 1 and 2, and pin 3 was grounded. The output (pin 1) went to the ADS1115's input A0.

A breadboard carrying an Arduino Nano, an op-amp and an ADS1115 module, wired with jumper leads
Stage 3: the amplifier chain, the Nano and the ADS1115 on the breadboard.
Close-up of the ADS1115 converter module on the breadboard with its I²C leads attached
The ADS1115 module, wired to the Nano over I²C.

Stage 4: the reference channel. The MCP6002's second amplifier became the reference channel. It was wired exactly like the first: Rf and Cf between pins 6 and 7, pin 5 grounded, the reference photodiode's cathode to pin 6, and the output (pin 7) to the ADS1115's input A1.

Stage 5: the soldered perfboard. To make the circuit permanent, I transferred it to a 5 × 7 cm double-sided perfboard. I planned the layout on a photo of the board with the parts placed, then soldered connections as bridges along rows and insulated wires underneath. A solder line along one row became the ground bus. The op-amp sits in a socket so that a damaged chip can be swapped in seconds.

The perfboard with the Arduino Nano, the ADS1115 module and the op-amp socket placed on it
The circuit transferred to a 5 × 7 cm perfboard, with the op-amp in a socket.
The same board photograph annotated with coloured lines marking the 5 V, ground and signal links
The layout planned on a photo of the board before soldering.

During soldering, the converter's clock (SCL) and data (SDA) lines ended up on the Nano's A4 and A5 pins the wrong way round. The Nano's hardware I²C only works on fixed pins, so instead of desoldering I wrote a small software I²C driver that generates the signals on any two pins. It keeps the same function names as the standard Adafruit library, so the rest of the code did not change.

Firmware. The Nano reads both channels alternately (sample, reference, sample, reference), so both see the same moment of light. It averages 16 readings per channel, converts the counts to volts, and prints the sample voltage, reference voltage, photocurrent and their ratio over USB. In the full colorimeter firmware, the LED is switched off before every measurement and a fresh dark reading is subtracted.

Tests.

  1. Steady-light stability: about 15 consecutive readings of one channel in constant room light.
  2. Channel matching: both photodiodes side by side under the same light, recording the ratio.
  3. Independence (torch test): a phone torch shone on one photodiode at a time. Only that channel should rise.
  4. Dark test: both photodiodes covered with the room lights on, then off. A true dark reading should not change.

Results

The amplified chain was about 700 times steadier than the Nano's own input, and the soldered two-channel board passed all three of its tests.

Stage and testResultWhat it shows
1. Nano A0 with nothing connected0.52 to 0.81 V, wanderingThe baseline: a floating input picks up noise
1. Photodiode on A0, covered by a finger, then room lightabout 0.15 V, then about 0.23 VThe photodiode works; skin lets some red and infrared light through
3. Amplifier chain, room light and phone torch0.25 V (2.5 µA), 0.35 V (3.5 µA)The gain matches Vout = I × Rf
3. Steady-light stability, about 15 readings0.0271 to 0.0275 V (0.4 mV spread)Steady to about 3 converter steps
4. Channel matching, same lightSample 0.026 V, reference 0.027 V, ratio 0.968The channels agree within 3.2%
4. Ratio wobbleSteps of exactly 0.0046 (0.9630, 0.9676, 0.9721)One converter step at 26 mV, about ±0.002 absorbance
4. Torch on the reference onlyReference 0.033 to 0.040 V, sample steady at 0.026 VNo crosstalk between channels
5. Soldered board, dim evening lightSample 6.4 to 9.3 mV, reference 12.3 to 19.0 mVBoth soldered channels are live
5. Soldered board, torch on each in turnOnly the lit channel roseChannels independent after soldering
5. Dark test, lights on, then offSample 3.2 to 4.1 mV, reference 3.4 to 4.1 mV, unchangedA dark offset that is not light

Stage 1: the baseline. With nothing connected, the Nano's analogue input wandered between 0.52 and 0.81 V. This is not a fault. An unconnected input picks up electrical noise from its surroundings, and it showed how unsuitable the bare input is for small signals. With the photodiode and load resistor connected, room light gave about 0.23 V. A finger over the photodiode only lowered this to 0.15 V, because red and near-infrared light pass through skin and the BPW34 is most sensitive in the near-infrared. Pulse oximeters work on the same principle.

Serial output listing photodiode voltages between 0.52 and 0.81 volts with nothing connected
The Stage 1 baseline: the Nano's input with nothing connected, wandering between 0.52 and 0.81 V.

Stage 3: the amplified chain. In room light the amplifier gave 0.25 V, and a phone torch gave 0.35 V. These correspond to photocurrents of 2.5 and 3.5 µA, as expected from Vout = I × Rf. In a later, dimmer capture, readings ranged from 0.0269 to 0.0350 V. Over about 15 consecutive readings in steady light, the output stayed between 0.0271 and 0.0275 V.

That 0.4 mV spread is about three steps of the converter. Compared with the 290 mV wander of the bare input, it is about 700 times steadier. The comparison is not perfectly like for like, since one is a floating input and the other a connected circuit, but it shows the scale of the improvement. Even at only 0.27 µA of photocurrent, the signal sat about 220 converter steps above zero.

Serial log of output voltages holding between 0.0288 and 0.0294 volts
Before the torch: the chain sits at 0.0288 to 0.0294 V, about 0.29 µA.
Serial log of output voltages climbing from 0.0309 to about 0.0350 volts and holding there
Torch on: the output climbs to 0.0345 to 0.0350 V and holds while the light does.
Serial log of output voltages settling between 0.0269 and 0.0275 volts
Torch off: it settles at 0.0271 to 0.0275 V — the 0.4 mV spread quoted above, about three converter steps.

Stage 4: two channels. With both photodiodes under the same room light, the sample channel read 0.026 V and the reference 0.027 V, a ratio of 0.968. The photodiodes therefore agree within 3.2%, and taking a blank reading removes this difference. The ratio moved between 0.9630, 0.9676 and 0.9721: steps of exactly 0.0046. At 26 mV, one converter step is 0.48% of the signal, so these jumps are single steps of the converter, not noise in the circuit. In absorbance, they are about ±0.002.

Shining a torch on the reference raised it from 0.033 to 0.040 V while the sample stayed at about 0.026 V, so the channels do not interfere with each other.

Stage 5: the soldered board. After the faults described below were fixed, both channels worked on the soldered board. In dim evening light, the sample read 6.4 to 9.3 mV and the reference 12.3 to 19.0 mV. They differ because room light reached the two photodiodes unevenly. A torch on each photodiode in turn raised only that channel.

The finished perfboard with the Nano and converter attached, standing on a wooden table
Both channels live on the soldered board.

In the dark test, both channels read 3.2 to 4.1 mV. Switching the room lights off made no difference, so this offset is not light leaking in. It is the op-amp's output floor: on a single 5 V supply, the MCP6002's output cannot quite reach 0 V, and both amplifiers share one chip. Over the run, the offset crept by less than 1 mV, consistent with the chip warming up after switch-on. Once the LED is in place, the signal will be 0.5 to 3.5 V, so this offset is only 0.1 to 0.7% of it, and the firmware subtracts a fresh dark reading before every measurement anyway.

One dark reading came out as a perfect sample-to-reference ratio of 1.0000. It looked too good to be true, and it was. Both channels sat on the same whole number of converter steps (about 33, or 4.1 mV), so the ratio compared one offset with an identical one. At that signal level, one step either way would change the ratio by 3%. A ratio only means something once both channels carry real light.

Discussion

What the numbers mean. The most important result is that, in steady light, the detector's resolution is limited by the converter's step size, not by noise from the circuit. That is the best a 16-bit design can do. At a typical blank signal of 0.5 to 3.5 V, one converter step is between 0.004% and 0.025% of the signal, so the detector should resolve absorbance changes far smaller than 0.001. The ±0.002 figure from Stage 4 applies only to the very dim room light used in testing (26 mV).

This is also my first answer to the question I started with. A colorimeter's accuracy depends on how much signal reaches the detector and how finely it is measured. In my circuit, both can now be checked with numbers.

Faults found and fixed. Most of my time went on finding faults, not on building. I have listed every significant one, because each taught me a rule I now follow.

#StageSymptomCauseFix and lesson
11Readings of 0.52 to 0.81 V with nothing connectedA floating input acts like an aerialA 100 kΩ resistor to ground. Always measure a “nothing connected” baseline first.
210 V in lightPhotodiode the wrong way roundSwapped the legs. Reversing it is harmless; it just reads zero.
31Upload stuck at “Uploading…”Clone Nano needs a different bootloader settingWorked through a checklist of causes until the upload completed.
43Converter not respondingOne Nano header pin (A5) had never been solderedSoldered it. Inspect every joint before building.
53Feedback resistor touching other chip pinsPins 1 and 2 are only one hole apartStood the resistor upright. Mount short spans vertically.
63Wiring looked impossible in a photoAn angled photo distorted rows and columnsMapped from a straight-down photo by column number.
73Blank screen, then “ADS1115 not found”The converter sat on the unpowered half of a split breadboard railJoined the rails and made the firmware report its progress and time out instead of freezing.
83“USB device not recognised”A short between 5 V and ground after rewiringRebuilt the power wiring one wire at a time. Use red only for 5 V and black only for ground.
94Reference channel stuck at about 0.9 mVAmplifier B pinned at 0 V, most likely by the old pin 6–7 tie-off jumper or a reversed photodiodeChecked both; the channel then read 24 to 27 mV.
105Both soldered channels read 0.000 VThe 5 V bus was never joined, so the op-amp had no power; its ground pin was also floatingSee below. When every channel fails the same way, check what they share.
A straight-down photograph of the breadboard showing the op-amp, resistor and jumper leads by column
Fault 6: an angled photo distorted the rows, so the wiring was mapped from a straight-down shot by column number.
Serial output repeating “Booting… looking for ADS1115” and “ADS1115 not found: check VDD, GND, SDA->A4, SCL->A5, ADDR->GND”
Fault 7: the firmware now reports its progress and times out instead of freezing.

Fault 10 was the hardest. The board looked finished, yet both channels read exactly zero whichever way round the photodiodes were fitted. That pointed to something the two channels share: power and ground. I had no multimeter, so I wrote a short sketch that turned a spare Nano input into a voltmeter and probed the board point by point. I also audited photos of the solder side hole by hole.

The soldered perfboard photographed from the side, annotated with red lines tracing the supply and ground links
The board as first soldered, with the supply and ground links traced.

The audit found four problems. Two supply wires ended on neighbouring pads that were never bridged, so the op-amp had no 5 V. Its ground pin sat on an isolated blob of solder. The decoupling capacitor was not on the 5 V line, and one planned ground link was missing. I spotted the unbridged 5 V junction myself after the first photo review had marked it only as “unconfirmed”. Once these were fixed, both channels worked first time.

Limitations.

  • No controlled light source yet. All tests used room light or a phone torch. Matching and stability need repeating with the LED in a light-tight holder (Part 2).
  • The ratio under changing light is untested. I showed the channels are independent, but not yet that their ratio stays constant while the brightness changes. That is the real test of the double-beam design.
  • Small samples. Stability was judged from about 15 readings. Part 2 will use 30 or more repeats and report a standard deviation and a limit of detection.
  • No multimeter. Several faults would have been found in minutes with one. It is the first thing I would buy.
  • Dark offset. The 3 to 4 mV op-amp floor is subtracted in software, but it limits how small a signal can be measured. A dual-supply or precision op-amp would remove it, and the socket makes that an easy upgrade to test.

Conclusion and what comes next

I have built a two-channel, double-beam light detector for about £8 in parts. In steady light, it is limited by its converter rather than by circuit noise. Its two channels agree within 3.2% and work independently, and its dark offset is understood and corrected in software. That is a sound foundation for a colorimeter, and later a plate reader.

The bigger lesson was about method. Changing one thing at a time, recording a baseline before trusting a reading, and asking what a strange number actually means found every fault in this build. When a result looked perfect, like the dark ratio of 1.0000, it was worth asking why.

Part 2 will cover:

  1. The 3D-printed, light-tight optical block: cuvette holder, LED, red filter and both photodiodes in fixed positions, with the circuit board standing upright beside it.
  2. Tests of repeatability, linearity and stray light, including the double-beam ratio under changing light.
  3. A calibration curve with a food dye, compared against a laboratory spectrophotometer.
  4. A rerun of the quantitative Benedict's test that started this project, measured on both my colorimeter and the school's.

After that, the same detector design will be copied into eight channels, one per row of a 96-well plate, with a moving 3D-printed stage. The colorimeter is instrument number one; it will not be the last.

Sim racing taught me that a product built carefully at a fair price can earn trust from people who have used the expensive alternatives. Science is a harder audience, because an instrument has to prove its accuracy with data, not just impress. That is what Part 2 is for.

What it cost

Prices are what I paid on AliExpress in September 2026, divided by the pack size. The Arduino Nano came out of my workshop, so its line is what one costs to buy rather than what this build spent. The complete first order, including parts for later stages (multiplexers, temperature sensors, a red filter, a 96-well plate and syringes), came to £41.72.

PartQuantityPack priceCost in this detector
BPW34 photodiode2£5.02 for 10£1.00
MCP6002 dual op-amp1£1.59 for 5£0.32
ADS1115 converter module1£3.10 for 2£1.55
5 × 7 cm double-sided perfboard1£2.22 for 5£0.44
DIP-8 socket, resistors, capacitors, wireassortedfrom kitsabout £0.30
Arduino Nano (CH340 clone)1£4.50£4.50
Totalabout £8.10

Sources