USB spectrometer

Introduction

Optical spectroscopy extensively contributed to the development of physics in recent two centuries. For instance, it was used to determine the ion composition of distant stars, although this task had been enlisted several years before as one of questions which mankind can never answer. Moreover, the solar spectra revealed new, at that time unknown spectral lines, which were later resolved to come from helium. At the turn of the twentieth century, the absorption and emission spectra of various materials supported the need of building a new, quantum theory. Optical spectroscopy developed into an essential measurement method in analytic chemistry, astronomy and many other fields.

Most optical spectrometers that produce a spectral waveform spanning over a selected region of wavelengths may be divided by their principle of operation into 3 most important groups:

• Fourier-transform spectrometers use Michelson interferometer to measure how the interference signal oscillates when one of the mirrors moves. The interference function is later digitally processed to get the intensity spectrum. They can achieve very high spectral resolution and can measure even far-infrared spectra, assuming IR optics and detector are used.
• Spectrometers with a monochromator usually utilize rotating diffraction grating. They are simpler and less sensitive to environment, as no interference fringes are measured. The Czerny-Turner setup is often used, but other different setups may suit a particular purpose better. For example, to replace the rotating grating by a prism with moving mirror may be favourable for UV spectroscopy or when higher diffraction orders are to be eliminated.
• The latter setup can be modified to obtain a spectrometer with multi-channel detector and no moving components. The grating is then fixed and the spectrum is projected on a broadband detector, which is usually a CCD.

  The spectrometers in this category largely differ by sensitivity, resolution, application, size and price. For instance, the High Accuracy Radial Velocity Planet Searcher at La Silla Observatory in Chile is able to detect 1 m/s velocity difference in star's movement, enabling to indirectly search for extrasolar planets. On the other hand, spectrometers of this type can be made very compact and they are optimal for applications where resolution and sensitivity is not critical. % http://www.horiba.com/scientific/products/oem-mini-ccd-spectrometers/miniature-spectrographs/ Some models are available as a PCI card fitting into a computer.

However, even the price of these handheld spectrometers usually exceeds $ 1000, making them hardly affordable as a educational gadget for a high-school science practicals, for example. Moreover, the supplied proprietary software usually does not permit any modifications or scripting, which may be needed when automating routine work in research laboratory.

This research project focuses on building a simple, cheap and compact spectrometer meeting following objectives:

1. The spectral resolution (Δ λ) should be better than 5 nm.
2. No moving components should be used.
3. The device and supplied software should be easy to use so that anybody with basic knowledge of spectroscopy can use it immediately.
4. The supplied software must run on most major operating systems with dependence on freely available programs only.
5. Except for initial calibration, no further maintenance would be required. Each device built will store its own calibration data in its permanent memory to prevent the need for recalibration when another model is connected.
6. Communication with computer and electrical supply should be performed via single USB cable.

Theory

General dispersive spectrometer design

In this research project, a diffraction grating design with a CCD detector was used. This section describes how the light components are separated and delivered to the CCD.

In the horizontal plane, the incoming light must be spatially limited by a vertical slit. Then the spatial limitation is converted to angular limitation using a concave mirror or a convex lens. Ideally, we obtain a parallel beam of white light. Its width is limited either by the size of the mirror or lens used, of by the angle of the beam passing through the slit.

The light components of different wavelengths separate after being reflected into different angles by a diffraction grating. Finally, the angular separation might be converted back to spatial separation using another lens or mirror as drawn in Fig. 1a. However, this would require a big lens with diameter similar to the length of CCD. The spatial difference would be proportional to its focal length, probably resulting in a small spatial separation. Fortunately, the second lens is not necessary, as the first lens can be shifted several millimeters away from the slit to create a real image in arbitrary distance from the grating (Fig. 1b). The beam incident on the diffraction grating becomes slightly convergent when a single lens is used. This results in nonuniform angular difference even for single wavelength and the projected image of the slit becomes slightly blurred. However, for the sharp angle of convergence this effect can be neglected.

Fig. 1: Horizontal schematics and light paths

There are several other phenomena limiting the spectral resolution of this setup, but assuming an ideal diffraction grating is used, the nonzero width of the input slit dominates. % list them

In the vertical projection, a single point in the slit plane is projected onto another point on the sensor as shown at Fig. 2a. As the CCD sensor is very narrow, the spectrometer would be sensitive in only one point at the slit, not depending on the incident angle. This would be optimal if either a thin optical fiber or a point source are measured.

Fig. 2: Vertical schematics and light paths

However, it was expected that the incident light would mostly be cover some area on the slit and that the spectrometer should be sensitive at nonzero length along the slit. Therefore a cylindrical mirror was inserted near to the focal plane to concentrate each spectral line (image of the input slit) into a much shorter line at the detector (Fig. 1c, 1b).

If the light entered the slit at a single point and this point slightly moved in vertical axis, its image would move in vertical axis, too. When a cylindrical mirror is used, the point image would still be projected onto the CCD, albeit under different angle. The detector could not be put directly in the focal plane of this mirror, because the horizontal focus given by the lens must be maintained. Nonetheless, the sensitivity to the light entering the slit was substantially improved.

CCD detector

Overview

The charge coupled device, CCD, was originally designed not to be an optical detector, but a memory instead. It should work as an analog shift register, storing arbitrary charge in each of its cells and reading them sequentially. Soon it was discovered that illumination causes internal photoelectric effect in the silicon substrate and fills up each memory cell with additional charge proportional to the incident light intensity. Although CCD was rarely used as a memory, it found widespread application as a sensitive optical detector both in consumer electronics and in high-end scientific instruments such as astronomical telescopes.

CCD detectors may be divided into two groups: linear CCDs, where one row of pixels is shifted to the output, and matrix CCDs, where the bottom row is fully shifted out always when all columns are shifted down by one pixel. For this project, a linear CCD was chosen over a matrix CCD because only a linear detector several centimeters long was needed. Still, it would be possible to build a custom matrix CCD driver, too [ccd-matr]. Another option, a CMOS sensor, was not used because of worse availability.

Driving a linear CCD

Although the internal design of the particular components may differ, the basic principle of CCD operation should remain similar. In the CCD register, a silicon substrate is covered by a thin layer of insulator and the electrodes. When a voltage is applied to an electrode, the electrons get trapped under the layer of insulator. By changing the voltage distribution on the surface, the charge packet can be moved with negligible loss.

In a common two-phase driven CCD, there are two electrodes on the insulating layer over each pixel. The geometry of electrodes is asymmetrical, dragging the electron packet nearer to one end of electrode [peterson]. Therefore, a two-phase only signal is needed to transport the charge to the next pixel. The two phase crossing must be nearly symmetrical, which requires them to be switched at once (In one of the first designs, these signals were switched in two consequent instructions of the microcontroller and the delay of ca. 80 ns probably caused the CCD to malfunction).

The silicon surface can integrate the charge produced by internal photoelectric effect, but this is rarely used now. Instead, contemporary sensors contain an array of photodiodes, which are much more sensitive than the actual CCD register. When a transfer gate impulse comes, the charge is transferred to each corresponding pixel in the CCD register. This also eliminates blur of the image when the charges are shifted [nec3799dsheet]. At the end of the analog register, the charge is coupled out to a pre-amplifier.

Fig. 3: Simplified diagram of waveforms driving a linear CCD. Input levels are TTL, corresponding output ranges approximately from 2 to 7 V. Grey areas contain each pixel's output voltage.

CCD output handling

Depending on the particular design, the two CCD phases and additional one to three inputs need to be activated by an impulse so that each pixel is read. After correct pixel driving signals are input, the CCD shifts the pixel's charge to its internal preamplifier and the corresponding voltage signal occurs at one pin. Sometimes is the output voltage stored in a sample-and-hold circuit.

The CCD detector output shall depend on the illumination linearly, but the signal is inverted, at least in the CCD detectors considered, e. g., for the CCD used in this project (NEC μPD3799): Under no illumination, the output is approximately +7 V, whereas under full saturation the output drops to +2.4 V.

Fig. 4: One of the first measurements of the CCD output on an oscilloscope. On the screen, there are two consecutive readouts of CCD. The detector window was covered by four thin paper stripes. Note the end-of-register signal in the middle of the screen and the blooming which slightly overlaps the less illuminated areas.

The following list sums the most important sources of measurement error:

• Blooming occurs when excess charge concentrates in one cell and it leaks into neighbouring cells. It manifests as flat, fully saturated regions with sharp edges, effectively obscuring the signal in the vicinity of the intensively illuminated spot. Most modern CCDs address this by electronic design, but the simple linear CCDs used in this project do not. Overexposed areas must be avoided by either reducing the light intensity or integration time. For instance, the saturation exposure of the used linear CCD is lower than 0.5 lx s [nec3799dsheet].
• Charge volatility: The stored charge decays over time. This is expected to be caused by both thermal electrons and incident light. At room temperature, it was experimentally proven that the CCD must be read in ~ 0.5 s after gate transfer to avoid this. This presents a time constraint to the data acquisition and transmitting time. When the CCD is read in bursts, a sawtooth-like additive bias will occur.
• Oversaturation: When the CCD is illuminated by common daylight for a while (> 10 lx s), the chip becomes so oversaturated that several full read-outs are needed to drain the integrated charge. Otherwise the output voltage may drop even under the saturation voltage. Sometimes this manifested as a step transition from saturation of one colour component in the middle of CCD reading, even when the CCD has been in dark for the last integration period.
• Electrical noise can arise from the switching power supply or from other AC sources, including the USB signalling and ADC convertor noise.
• Thermal noise is caused by random thermal electron-hole pairs. It is probably lower than electronic noise in the sampling circuit used. Thermal noise be addressed by thermoelectrical cooling of the sensor, as a 20 'C temperature drop is expected to effectively suppress the thermal noise and conductivity. No cooling was used in this design because of rather high power requirements of a Peltier cell.

When a colour CCD is used, the red, green and blue signals must be averaged to obtain sensitivity covering the whole spectrum to be measured. This was done by inserting a 10 kΩ resistor at each channel and by connecting their second terminals together. The resulting spectral sensitivity is inhomogeneous, which was later addressed by amplitude calibration.

Description of the Atmega8 microcontroller

The proper CCD driving, analogue to digital conversion and consequent data transmission would require complicated circuitry if constructed from discrete components or TTL logic. Even such designs were constructed earlier, but most of these tasks can be easily handled by a common microcontroller. The most demanding task is the USB communication, which determines the minimum MCU computational power to approximately 12 MIPS at least. This topic is discussed later.

The microcontroller should be equipped by a sufficient internal analog-to-digital converter (ADC). Usage of an external ADC would present additional cost and complexity. Many contemporary microcontrollers have an internal 10-bit converter able to perform at least 10,000 samples per second, which is sufficient for this application.

Additionally, the need to avoid the mentioned sawtooth bias requires that the whole CCD is read at once and the spectral curve is stored in the memory until it is transmitted to the computer. Many microcontrollers come with fairly limited amount of internal RAM, which ranges from tens of bytes to 1 kB in the middle class. Taking into account that about 100 bytes are allocated by various program data and that each pixel occupies 10 bits, there remains room for approximately 720 pixels if 1 kB of RAM was used. This severely limits the theoretical spectral resolution of 5300 pixels. The real spectral resolution in this case is, however, more limited by imprecise optics, and therefore the 1 kB of RAM was considered sufficient.

If higher optical resolution was achieved and consequently bigger data buffer was appropriate, microcontrollers with 2 kB or 4 kB of RAM are available for approximately 5 Euro and 8 Euro, respectively. A cheaper solution would be an external RAM. As of 2010, the market is saturated by cheap parallel static RAM circuits with capacity in tens of kB. However, most of these circuits use more than 20 pins for addressing and data transmission, which would occupy all available pins of the cheaper microcontroller. The SRAM circuits communicating over the serial line are much less available, although there are plenty of rather cheap serial EEPROM circuits, which have limited number of write cycles. Possible candidates could be Microchip 23K640 and 23K256 models with 8 and 32 kB, respectively. Anyway, the use of external RAM would make the device more complicated and error prone. Therefore no external RAM was used in this design.

As 1 kB of RAM was considered sufficient, the Atmel Atmega8 microcontroller was found to fullfill all aforementioned requirements. Moreover, this MCU supports in system programming, which was very handy to develop and debug communication with both CCD and computer. The Atmega8-16PU version has 28 pins, of which 4 were occupied by DC supply, 2 by USB, 2 by oscillator, 4 by CCD driving signals and one for ADC input. There remained enough pins and capacity for a 7-segment display and simple user controls on the device, but later it was considered impractical and only the computer control was left. As for 2010, Atmega8-16PU circuit is still one of the most cost effective microcontrollers on the market, with price of about 1.5 Euro.

Communication interface

Comparison of different interfaces

• The serial port (RS232) can use 3 wires only, one wire for each direction and a ground. The data rate must be set on both sides. Serial line is often used in industrial applications and for older computer peripherals. The microcontroller used in this project has hardware circuit for asynchronous serial communication, which could make this option favourable. However, RS232 begins to disappear from newer computers and it requires the user to set the serial port options manually, too.
• The parallel port would probably require writing an own driver for data transfer and it would be too bulky for this project.
• The GPIB interface is often used to communicate with scientific instruments, but is rarely available at standard computers, let alone notebooks. Moreover, it shares the disadvantages of parallel port.
• The ethernet interface uses complicated, high-level protocol, which probably could not be processed by the microcontroller unit (MCU) and would increase the device complexity.

In the low-speed mode, the signalling speed is 1.5 Mb/s. This data rate can be processed by the MCU firmware, although the MCU has no hardware acceleration for USB communication.

For higher data rate in full-speed mode of 12 Mb/s, a dedicated USB to RS232 converter may be used. An example is the Future Technology FT232BM circuit. According to the datasheet, it should be able to communicate with computer via USB, handle all USB protocol messages and communicate with the MCU via serial port. In this case the maximum data rate over serial port should be 2 Mb/s. [ft232bldsheet] This may not seem as an advantage over firmware USB driver until one realizes that it is the USB protocol in the low-speed mode what presents the bottleneck, limiting the average data rate to less than one tenth of the signalling speed. Likewise, when the MCU communicates using the USART (Universal Synchronous and Asynchronous serial Receiver and Transmitter) subcircuit, less time is spent by the driver routines used to handle the messages. The USB-RS232 converter was not used here in order to keep the device simple.

Hardware for USB communication

As no clock signal is sent over USB and the USB specification requires 1.5 % maximum tolerance for the timing [usb], the MCU clock shall be timed precisely. Typically, the firmware drivers require oscillator whose frequency would be a multiple of 1.5 MHz. The minimum clock frequency is limited by the particular driver; mostly at least 8 processor steps per bit are required. A 12 MHz quartz crystal was used in this project.

Fig. 5: a) Pins on the USB B connector; b) Minimal circuit for connecting ATMEGA8 to USB.

The USB cable contains 4 wires, which are listed in the Table 1. The pins of the USB-B connector, used on the device side, are drawn in Fig. 5a. The logical one voltage should fall between 2.7 and 3.6 V [usb] , therefore the 5 V TTL output of the MCU must be limited by Zener diodes. Additionally, the D- line must have logical one when the MCU is reset [p. 141]{usb], which is ensured by the 1 kΩ pull-up resistor (Fig. 5b). ]

The USB datasheet [p. 128]{usb] has some stricter requirements, e. g. for the device resistance, but the circuit from Fig. 5 worked flawlessly. Similar schematics are used in many other projects.

Brief description of USB protocol

All data are sent in packets. The packet contains an initial sync sequence, the packet identification number and optionally payload or control data along with its checksum. Some packets transmit payload data, while other packets are used for handshake and identification only. Three packets, token, data and handshake, usually form a transaction [eckhardt-usb]. The USB protocol, mostly in a simplified, yet working form, is handled fully by the firmware driver.

After being requested by the host, each USB device identifies itself by a vendor number and an USB class. Most devices fall into the human interface device or mass storage classes. Each class is handled slightly differently by the computer. The human interface device class was chosen due to its simplicity and widespread universal drivers [usb]. USB protocol is rather complicated and its more detailed description is beyond the scope of this document.

There are several software implementations of the USB protocol for AVR microcontrollers. For example, Igor Češko [cesko-usb] and Jan Smrž [smrz-usb] implemented their own. This project utilizes the V-USB driver by Christian Starkjohan [obdev-usb] because of its good documentation and numerous practical examples. The driver is licensed under GNU GPL for noncommercial use.

All these drivers use a hardware interrupt, which is triggered by incoming USB packet. At 1.5 Mbit/s signalling speed and 12 MHz main oscillator, only 8 clock cycles remain to receive, convert and store each bit. This requires that the microcontroller completely halts all another operation when any USB packet is received or transmitted. Besides, the USB protocol has the master-slave topology, and accordingly the MCU stops to receive or transmit packets asynchronously with its operation, which has to be taken into account when programming real-time tasks.

Solution

Hardware

Optical setup

The optical components holders were made of a standard aluminium L-profile 25 mm high, 2 mm thick and with 12 mm long foot. The beam path axis was chosen to be 14 mm above the base surface. This provided optimal vertical room for the beam, maintaining enough room between the light path and the base for bolts etc.

• The input slit was made of two razor blade segments, 15 mm long. One of the segments was glued onto the inner side of the L-profile to cover half of the input hole. Its sharp edge was adjusted to be perpendicular to the base. The second segment was clamped next to it by a short screw with a spring bushing. This makes adjusting the slit width possible. In the final setup, the slit width of 50 μm -- 100 μm appeared to be appropriate.
• The collimating lens had 14 mm diameter, focal length f = 30 mm, plano-convex geometry and standard anti-reflex coating. (Generally, any common small lens of similar f may be used.) It was glued at the edges to a hole of similar diameter, keeping its axis at the beam axis. At the base of the L-profile, a notch instead of a hole was grinded to enable focusing.
• The reflective diffraction grating was attached nearly perpendicular to the incident beam. For the first prototype, a 20×20 mm segment cut from the outer edge of a blank CD-R was used. Notes to this solution are discussed later. The grating was rotated so that the most desired spectral range in the first diffraction order is projected on the CCD.
• The cylindrical mirror was obtained from a scrap photocopier. It is a polished metal rectangle, 42×15 mm in size. Its focal length is approximately 30 mm. The mirror was attached to a metal chip, which had a groove at the bottom. The groove fitted to two spikes in the bottom of the L-profile. At the top, an M3 screw allows to adjust the vertical angle of the mirror.
• The CCD detector was glued to a low profile PMMA slab. Its pins were shortened not to obscure any passing light and soldered to the controller circuit. The slab had a notch for the bolt.
  The output signal from three colour channels was averaged by three 10 kΩ resistors near the CCD and it was connected using a thin shielded cable to eliminate noise from the digital wires. Between the supply voltage and ground, a 100 nF capacitor was inserted.

Fig. 6: Detail of the CCD mount and the adjustable cylindrical mirror

The optical setup was mounted (by M3 bolts and nuts) to a 2 mm thick aluminium plate, of approximately 70×90 mm in size. It provided a decently rigid, yet light base. There remained 6 degrees of freedom, that had to be precisely adjusted to obtain effective and narrow spectral lines projected on the CCD.

1. The input slit width: The second razor segment should limit the light entrance to an uniform, thin line.
2. The longitudinal position of the collimating lens should focus the slit onto the CCD in the horizontal plane.
3. The vertical grooves orientation of the diffraction grating should be parallel to the slit.
4. The horizontal angle of the diffraction grating: The whole desired spectral range should be projected onto the CCD.
5. The cylindrical mirror and CCD should be perpendicular to the beam coming from diffraction grating, which illuminates the center of CCD.
6. The vertical angle of cylindrical mirror and the corresponding position of the CCD detector: Along with previous setting, it should focus the spectral lines to illuminate the narrow CCD line the most effectively.

The coarse setting was done visually using the light from a standard fluorescent lamp. The fine tuning of the lens position was performed when the device was connected to computer and the spectral lines were optimised to be as narrow as possible.

Where to obtain an optical grating?

Future models can use a professional grating, which can cost 50 Euro or more [edmund-gratings] . This price is not prohibitive, however it would spoil one of the objective of minimizing the device's cost. Maybe several gratings could be replicated after one professional grating is bought.

Alternatively, there is a proof that a reasonable diffraction grating may be manufactured using a nearly home-built ruling engine, which was accomplished by Brian Manning even without the contemporary availability of MCUs, lasers and stepping motors [manning]. Anyway, the precision requirements for such ruling engine are very tight. To achieve 1 nm spectrometer resolution at the 500 nm wavelength, the RMS error of groove spacing must be at most few nanometers. This limits the groove spacing error to less than 10 diameters of silver atoms!

For the first prototype, the CD-R slab worked quite well. The cheapest CD-R had unexpectedly low stray light and no apparent Rowland ghosts coming from low-frequency modulation of the grating. Apart from non-optimal efficiency and fragile metallic coating on the surface, the biggest problem was caused by the grooves curvature. A monochromatic beam limited by the slit is projected into a crescent shape in the focal plane of the collimating lens. Instead, a vertical line is required, as the cylindrical mirror focuses the beam in the vertical plane.

Microcontroller circuit

The CCD driving signals use the TTL levels, hence the CCD inputs may be connected directly to the MCU pins. For the used NEC μPD3799 CCD detector four signal wires are needed: the phase 1 and phase 2 signals, the RB (reset gate) and the TG (transfer gate) signals. When another CCD is used, additional driving signals may have to be added, e. g., for NEC μPD3797, additional CLB (reset feed-through level clamp clock) negative impulse must be included after the RB impulse. According to the particular datasheet, both the firmware routines and hardware wiring can be easily modified to accomplish this. Note that one MCU clock takes only 83 ns and some minimum pulse length and delay must be maintained according to the CCD datasheet.

As was noted in the section on CCD output, the output signal of CCD is too high to be directly connected to the analog-to-digital converter (ADC) of the MCU. The output must be linearly inverted, amplified and shifted before passing to the ADC, which measures in the range of 0 to 5 V. There is an older project of Paul Stoffregen, who accomplished this transformation using five operational amplifiers, of which two act as a sample-and-hold circuits, storing the two voltage levels of dark and saturated signal [pjrc]. A circuit of two potentiometers and only one operational amplifier was used here instead, as both the dark and the saturated voltage levels both prove to be relatively stable.

If only one colour component becomes saturated, the output signal gets distorted and therefore unusable. Accordingly, the signal range should be stretched so that full saturation of single colour component corresponds to full saturation of the signal entering the ADC. In other words, the output signal must be amplified with gain enough to prevent the saturation of any colour component in CCD. The circuit with inverting preamplifier and microcontroller is shown in Fig. 7.

Fig. 7: The microcontroller circuit. At its left side it is connected to the supply circuit, at right to the CCD detector.

The operational amplifier supply voltage has to be 12 V, which presents a risk of destroying the ADC circuit by excess voltage, as the maximum input voltage of ADC is limited to 5.3 V in this case [atmega-dsheet]. In order to prevent this, a 10 kΩ series resistor and a diode with low voltage drop limit the ADC input voltage, as shown at Fig. 7.

Supply circuit

For the voltage conversion, a boost converter integrated circuit MC34063 was used. In short, the circuit senses the output voltage at the CII (comparator inverted input). When the voltage drops below 1.25 V, the SC (switch collector) shorts to ground, letting the current through the L₁ coil and R_sc grow. The CLS (current limit sense) voltage drops as the current increases and when the voltage drop exceeds approximately 300 mV, the SC pin closes. The energy stored in the L₁ coil creates an impulse of higher voltage that charges the C₂ capacitor. The C₁ capacitor determines the time interval of operation.

Fig. 8: The supply circuit. At its left side it is connected to the computer via USB cable.

The output voltage is given by the voltage divider parallel to the C₂. The selected resistances of 25 kΩ and 2 kΩ set the minimum C₂ voltage to 16 or 17 V. This voltage is limited by the LM7812 and LM7805 stabilizers to 12 V and 5 V, respectively. Reducing the C₂ voltage would increase efficiency, but risk of negative ripple in the 12 V stabilized output would be present due to the voltage drop at LM7812.

The USB signalling wires D+, D- are connected according to the Fig. 5.

Spectrometer casing

These two bolts, as well as whole optical mount, were not connected to the USB ground to prevent accidental short-circuit or ground loop.

Although the matt inner side of the black plastic box has low reflection on its own, the cylindrical mirror and CCD were additionally covered by a folded sheet of black paper to minimize the stray light coming directly from the input slit or any tiny hole left in the plastic case.

Microcontroller firmware

Full microcontroller firmware is described in the original PDF document attached.

Computer software

The software in computer was split into two programs:

• the backend ensures communication with the spectrometer and transforms the data according to the wavelength and amplitude calibration,
• and the frontend, which provides graphical user interface and, calling the backend repeatedly, presents the measured spectra in real time.

Command line backend for USB communication

It was developed and tested under Linux 2.6, but it is designed to compile and run under Windows and other operating systems, too. Under Linux, the backend depends on libusb, on Windows it depends on the Driver Development Kit (DDK), which should be shipped along with MinGW compiler [obdev-usb].

The full code is included in the attached CD and it will not be listed here. The commandline backend accepts several commands. The first parameter determines the action performed and optional parameters are in square brackets:

1. set_measure [integration time] starts the measurement, integration time ranges from 1 to approx. 200
2. get_spectrum receives, transforms and writes out the spectrum to the stdout
3. write_calib_ampli lambda1:amplitude1,[lambda2:amplitude2,[...]] write amplitude calibration to the EEPROM
4. write_calib_lambda pixel1:lambda1,[pixel2:lambda2,[...]] write wavelength calibration
5. read_calib_ampli [--short] reads amplitude calibration from the EEPROM
6. read_calib_lambda [--short] write wavelength calibration

The calibration write commands use the second parameter to set the calibration points in the EEPROM. Each calibration point occupies 4 bytes in the EEPROM. See the README file for more information.

For wavelength calibration, pairs of pixel number and corresponding wavelength in picometers have to be input in the format of pairs of integers related by colon and separated by comma. First two bytes are the integer position of the pixel. Following two bytes are the wavelength divided by 32 pm. Therefore, the wavelength of 2²¹= 2,097,152 pm is encoded as the maximum integer value 2¹⁶= 65,536.

For amplitude (i. e. sensitivity) calibration, the pairs of wavelength and corresponding coefficient are used, as the sensitivity curve mostly depends not on the optical geometry, but rather on the averaged spectral sensitivity of the three CCD channels. If the projected image moves slightly, only the wavelength calibration would have to be repeated. The wavelength and the coefficient are calculated similarly as above.

Data preprocessing according to calibration

Then, the calculated wavelength is input to the second spline interpolation, which calculates the sensitivity coefficient for the given wavelength and divides the measured data by this coefficient.

The spectrum is output to stdout with two columns of tab-separated numbers, and usually this output is redirected into a file. This form of communication does not cause much holdup, because the file is stored in the disk cache and possibly does not get written to the disk until it is read by frontend and rewritten by new measurement.

Python module backend

The functions provided by the backend module are nearly the same, except for not existing equivalent for the get_spectrum parameter. The SetMeasure function not only starts the measurement, but also receives the spectrum, performs calibration transformation and return it as two lists of wavelenghts and corresponding values.

Graphical user interface

The gtk libraries are a widget toolkit often used to create multiplatform graphical applications. The gtk libraries are commonly installed on Linux desktops by default, under Windows, one has to download and install them. Both python and gtk are released under free licenses compatible with GPL [python-lic], [gtk-lic]. Thanks to this, all the software equipment of the spectrometer can be distributed without any license costs.

The graphical user interface provides only basic functions as for now (see Fig. 9). It allows to switch between manual spectral measurement and automatic measurement, which runs the spectrometer repeatedly when the previous spectrum was acquired and displayed.

Fig. 9:A screenshot of the graphical user interface, showing the orange-red fluorescence of a highlighter pen illuminated by a green laser pointer, whose reflection manifests as the narrow peak at 532 nm. Note also the weak component at ~ 810 nm coming from the laser diode pumping the Nd:YVO₄ crystal.

The Sensitivity slider at the top sets the integration time to be two on the power of the value set. The most practical values are between 2 and 5.

Averaging can be set either to the moving average of 2 to 64 waveforms, or to average infinitely to suppress the sampling noise.

There is an option to calibrate the amplitude, which is the most important for measurement of a transmittance T(λ) of a sample. The calibration curves are stored in the computer memory. The spectrometer should be recalibrated always when the light source changes. It is recommended that after several minutes of CCD warm up, the source is turned off and the dark signal calibration curve I_D(λ) is set, then the source is turned on again without the sample and the full signal curve I_F(λ) is set. Long averaging time improves accuracy. When the amplitude calibration is applied to the measured signal I_M(λ), the transmission T(λ) is calculated the following way:

Thanks to the interactive interface of the matplotlib package, the graph may be zoomed and panned, as well as easily exported into a bitmap file.

Calibration

The calibration procedure consists of two steps, that shall be done in the following order: the wavelength and the amplitude calibration.

For the wavelength calibration, a common compact fluorescent lamp (CFL) may be used. Such lamp contains diluted inert gas and mercury vapor and at the inner surface, a thin fluorescent layer of europium and terbium salts is sputtered. The output spectrum consists of many spectral lines with different intensity and width, of which the most intensive are listed in the Table 2.

The spectrum was first acquired without calibration, then the peaks were easily identified and corresponding calibration points were saved. As noted above, the backend uses spline interpolation/extrapolation and accordingly even for three distant enough calibration points it was possible to obtain reasonable spectral accuracy.

The amplitude calibration requires a source of a broad and known spectrum. Possible candidates are a halogen bulb or common daylight. Similarly as for the wavelength calibration, several calibration points shall be calculated from the shape of uncalibrated spectrum. The value of each point is obtained as a ratio of the measured intensity to the expected real intensity of the source.

It must be noted that the amplitude calibration does not take into account the linear background. This background depends on CCD temperature and sensitivity. Therefore, the background transforms to a curve that may not be easily visually subtracted without previous observation of its shape for zero input light.

Additionally, it was observed that the spectrometer sensitivity varies by several percent from pixel to pixel. This can be accounted to irregularities of the cylindrical mirror or the CCD detector. Unfortunately, the corresponding compensation curves would not fit into the internal EEPROM of the microcontroller. Both unwanted effects must be compensated in the graphical user interface as the amplitude calibration curves for a particular device and ligth source.

Results

In this section, several measured spectra will be presented to illustrate the performance and shortcomings of the current design.

In the Figure 10, the spectra of a starting 15 W IKEA compact fluorescent lamp is shown. This lamp was used to calibrate the spectrometer. The spectrum has been averaged from 16 measurements to suppress the noise and no amplitude calibration was used here.

Fig. 10:Emission of a compact fluorescent lamp, 5 s after cold start (black) and ~ 60 s later (blue). The horizontal axis is the wavelength in nanometers.

Initially, the argon filling emits mostly in the near infrared region. The slow start is probably caused by slow evaporation of the mercury filling, which manifests as growth of the spectral lines in visible region. Note that on the contrary, some lines intensity drops during the start.

With longer averaging and proper zoom, several other lines may be resolved, such as the weak violet mercury line at 406 nm.

Both the blue mercury line at 436 nm and the infrared argon line at 811 nm are emitted by diluted gas and are therefore narrow enough to assess the spectral resolution. The spectrometer measured their width at half maximum as 5 to 8 nm, which can be considered to be the maximum spectral resolution of the device. However, with proper zoom, the green peak at the 544 to 546 nm has a shallow, yet stable notch in its center. This may be accounted either to the nonuniformity of CCD response, or to the fine resolution of spectra. (Both is possible as the convolution function is probably not gaussian, but has a sharp peak instead.)

Fig. 11: a) Not normalized transmission through a 5 mm cuvette of pure water (black line); of water with ~ 10 μmol/l solution of KMnO₄ (red line) and of water with ~ 100 μmol/l solution thereof (green line). The horizontal axis is the wavelength in nanometers. b) The simple setup for transmission measurement; the white LED and the light bulb complement to a broader spectrum.
a) b)

The Figure 11 depicts three spectral curves measured by passing an incandescent bulb light through a 10 mm long cuvette with pure water and two different concentrations of potassium permanganate (KMnO₄). In low concentration, the very strong absorption in the green spectral region causes the pink tint of the solution. In higher concentration, broader spectral region from 450 to 650 nm gets absorbed. This explains the visible hue difference of different concentrations. The amplitude growth in unabsorbed regions is caused by the cuvette displacement.

In this measurement, the amplitude calibration did not correct the pixel-to-pixel nonuniformity, although the spectrum was averaged from 16 measurements again. For practical application, the proper calibration for the particular source and fixed cuvette position is important.

Conclusion

In the above sections, a concept of a cheap, compact digital spectrometer was presented. If a CCD detector from an old scanner is used, the components cost less than 20 Euro. (A new suitable CCD can be obtained for under 10 Euro if a bigger package is requested.) The work per each next piece can be expected to be less than 5 hours, since the first prototype is tested and both the software and firmware are ready.

In the Results section it was shown that such device works with sensitivity and resolution acceptable for many tasks both in education and research. The graphical user interface makes the device relatively easy to work with.

The main conclusion of this work is that it is possible to build a relatively usable, cheap and portable digital spectrometer.

Fig. 12: Front view of the assembled spectrometer

Appendix

• [bhagwat] Amit Bhagwat and Ankur Kumar. CCDImager – the use of the Atmel Mega32 microcontroller for imaging, 2007. URL link.
• [cesko-usb] Igor Ceško. Implementation USB into microcontroller: Igorplug-USB. URL link.
• [atmega-dsheet] Atmel Corporation. ATmega8(L) data sheet, 2003. URL link.
• [nec3799dsheet] NEC Corporation. μPD3799 data sheet, 1999. URL link.
• [eckhardt-usb] Kamil Eckhardt. USB - univerzální sériová sběrnice, 2002. http://home.zcu.cz/~eckhardt/popis.html
• [6] Filip Dominec et al. Programujeme jednočipy, czech textbook at Wikibooks, 2009. URL link.
• [7] Python Software Foundation. Python license, 2010. URL link.
• [edmund-gratings] Edmund Optics Inc. Reflective ruled diffraction gratings, 2010. URL link.
• [ft232bldsheet] Future Technology Devices Intl. Ltd. datasheetarchive.com/. FT232BL datasheet, 2005. 33
• [manning0] Brian Manning. A grating ruling engine. Scientific American, 4:232, April 1975.
• [peterson1] Courtney Peterson. How It Works: The Charged-Coupled Device, or CCD, 2001. URL link.
• [12] Jan Smrž. Implementace USB rozhraní AVR mikrokontrolérem. Master thesis, CVUT, Praha, 2008. URL link.
• [obdev-usb] Christian Starkjohann. Virtual USB port for AVR microcontrollers, 2010. URL link.
• [pjrc] Paul Stoffregen. CCD Array Reader Project, 1992. URL link.
• [gtk-lic] The GTK+ Team. The gtk+ license, 2008. URL link.
• [usb] The USB Implementers Forum Inc. (USB-IF). Universal Serial Bus Revision 2.0 specification, 2010. URL link.
• [cfl-spectrum] Wikipedia user Deglr6328. Fluorescent lighting spectrum, 2009. URL link