A personal view of the development of the FTS-14 Spectrometer
Raul Curbelo
BIO-RAD Digilab Div.
68 Mazzeo Drive, Randolph, MA 02368
(Note: Click here to view a photo gallery related to the FTS-14's development.)
To have a perspective of the factors that led us to the development of the FTS-14, we need to look back at the two decades before 1970. A combination of science and technology became available during the 1960's as the result of the advances from the late 40’s through the 50’s. The development of the 296 interferometer in 1968, and the availability of low cost minicomputers, made it possible for us to develop the FTS-14.
At Bell Labs, research efforts to improve communication systems led to fundamental advances in communications theory. We call it information theory now. Nyquist and Shannon made the more important contributions.
Fellgett at the University of Cambridge, demonstrated in his 1951 thesis, the advantage of multiplexed spectroscopy, while trying to improve the sensitivity of his astronomical infrared measurements.
The effort to reduce the size of electronic components led the research on semiconductors by Schockley and others at Bell Labs, to the invention of the transistor in 1948 and to usable devices in 1951. The search for improved yield and reliability of transistors at Fairchild, led Noyce to the invention of integrated circuits.
The scientific advances in communications before 1960, in conjunction with the development of the transistor, led to an expansion of related technologies beyond the communications applications, and led to applications in all sorts of commercial instrumentation. These included solid state detectors, better computers and faster analog electronics including better analog to digital converters. The publication of the FFT algorithm by Cooley and Tukey in 1965, made practical the Fourier transform of large arrays in the slow computers of the time.
Larry Mertz pioneered the application of the multiplexed advantage in infrared spectroscopy, first with a FIR polarization interferometer built in 1954 at Baird Associates. At Block Associates, Mertz championed rapid-scan interferometers and data processing methods that were the basis for the I3 and I4 interferometers developed at Block in the early 60’s. Rapid-scan interferometers minimize the effects of scintillation noise, are less sensitive to external vibrations, reduce the effect of slow drift, and can be implemented in a compact instrument without a chopper. The mechanical and optical engineering for these interferometers, were done by N. Young, and the electronics by Gerry Wyntjes.
The physicist Mertz understood where in the interferogram was the spectral information and the importance of the phase information in the interferogram and developed an algorithm that provides the best corrections to the resulting spectrum when the needed constrains are met. His book, "Transformations in Optics", includes the key information theory aspects of interferogram spectral estimation.
In 1960 you could get a spectrum from an interferogram by taking punched cards to a main frame computer. This was an expensive operation. It took one hour of CPU time to transform a 1K interferogram.
For getting a spectrum in the laboratory, Mertz first used an audio tape recorder playing a tape loop with the recorded interferogram into an audio wave analyzer with the output going to a chart recorder. Here he retained the multiplexed advantage in the measurement but forfeited it in the data processing, although not entirely, because he played back the tape at a faster rate. He tried other interferogram storage methods, including mechanical switched capacitor filters and magneto-strictive delay lines. The storage capacity was too small for practical use, and the implementation difficulties were substantial.
In 1963, Block Associates bought a used Recomp computer. The memory and the registers were a magnetic drum with 4K words. Cycle time was in the order of a millisecond. Mertz programmed the Fourier transform in machine language. A 1K points transform took four hours. In 1965 the FFT in the same computer took 24 minutes.
In 1964 at Block Engineering we developed a complete spectrometer system that would coherently add (the term coadd was coined by Myron Block) successive interferograms, and play them back at a faster rate to a built-in wave analyzer. This was the Block Coadder, the prototype was built with Digital Equipment Corporation (DEC) modules and the production units used the first low cost commercial IC's, packaged in transistor cans.
For the Coadder, the core memory stack had 1K of 16-bit words and a 10 m sec cycle time. It used a 10 bit A/D; we built the rest of the electronics. The digital circuits were 24 cards. This system became the Model 200 Spectrometer. The resolution was limited to 10 cm-1 in the mid-infrared (MIR), and it was primarily an emission spectrometer, without a built-in infrared source. About 30 Model 200 systems were sold. Manfred Low of Rutgers University was one of the early users.
The HeNe laser was invented in 1961. Commercial lasers in '63 were 1 meter long and needed a large RF power supply. By 1966 these had evolved to a foot long with a DC supply, when we first used a laser in an interferometer for the NASA expedition to observe the solar eclipse in the South Atlantic.
The processes developed to fabricate semiconductors were adopted to produce new detectors, including pyroelectric detectors first developed by Marconi in the UK. A commercial pyroelectric bolometer available in 1969 (DTGS) had good sensitivity (D* >3e7) and a frequency response beyond 1KHz that allowed rapid scan in the MIR, as opposed to the thermistor bolometers, which had a cutoff at about 200 Hz.
Transistors and integrated circuits started a revolutionary trend of improving capability and performance, and shrinking the size and cost of electronics that continues today. The effect of this trend was most visible in the evolution of computers.
In the early 60's, the first commercial interactive minicomputer, the PDP1, filled a room with 6-foot high 19" racks. Each 18-bit register was 8" of rack panel made-up of 18 modules (cards), one per bit, built with discrete transistors. Transistors were $12 and it took 12 transistors to make a flip-flop. With 4K words of memory, it sold for more than $100,000.
By 1965 the PDP8 was a 12-bit machine. Originally designed as
a controller, it was the first RISC machine, had
3 registers and 4 logic instructions including addition, all in a 4-foot
high table top rack, that sold for $18,000.
The Varian 620i was a slow 18-bit machine, still too expensive for a spectrometer.
In 1969 the Data General introduced the Nova minicomputer in a 5 ½"-high box. It used mid-scale integrated circuits, up to four flip flops per IC. The CPU fit in two 15" x 15" boards. It had four registers and seven instructions.
For us, at Block Engineering, another milestone was a NASA contract in 1968 to build an interferometer for emission measurements, to operate primarily in the NIR, with 10X more resolution and larger optics (10 times the area) than the interferometers built at Block to date. In this interferometer, the support mechanism for the moving mirror needed a longer travel for the resolution, and had to have small enough tilt for interferometer operation at 1 micron with the larger mirrors, and smooth motion in the micron scale with no stickiness. These requirements led us to the air bearing. The tilt specification required that the center of support be fixed with respect to the center of gravity of the moving system. With the help of John Hero of Dover Instruments, we developed a linear bearing that had less than one arc second of tilt over its travel.
The air bearing with hovercraft type pads had been developed earlier. We chose the cylindrical form for easy manufacture with the tooling used to make rotating thrust bearings. A rotation constraint was needed, and the solutions generated a sequence of difficulties that proved to be long lasting.
The linear motor had to move a 0.5 Kg mass more than 10 mm and have small velocity errors. The solution was to use the magnet and voice coil for the largest loudspeaker built locally. An accelerometer was used to reduce the effect of large external vibrations.
The controller for the interferometer was built mostly with technology developed earlier, extended for the new mechanical and optical requirements.
All the system specs were met or exceeded, and resulted in a system that was manufacturable. We got 0.5 cm-1 resolution that was comparable to that of the better analytical dispersive spectrometers of the day, and the sensitivity was orders of magnitude better.
The commercial version of this interferometer was named the 296. The 296 interferometer had two Michelson interferometers sharing the moving mirror. The reference interferometer used a portion of the rear surface of the moving mirror to generate two reference interferograms, from a laser and a white light source respectively. Wyntjes, using a neon bulb for the monochromatic source, had developed this technique in 1962. He used two identical interferometer cubes assembled back to back, with the moving mirrors connected by a rigid rod through the common linear motor. The same mechanical arrangement was used in the 1966 eclipse interferometer.
The 296 interferometer casting was made massive in search of stability, but we had departed from the cube topology used by Mertz and Young. The open V was designed for easy beamsplitter change, but worked as a chimney for the heated air and we could see the effect of the rising air bubbles in the output signal. A roof on top of the V reduced the problem to an acceptable level. Moreover, the thermostated interferometer resulted in a temperature gradient in the 1" thick aluminum fixed mirror support, until we found the right location for the heating pads.
The performance of the 296 interferometer and the availability of 16-bit minicomputers at a reasonable price made it obvious that a commercial FTIR spectrometer could be built to compete with the better dispersive spectrometers available.
The FTS-14 optical head was designed to emulate the dual-beam dispersive spectrometers we hoped to replace. This was a single-beam spectrometer, and the sample and reference beams had to be measured in sequence. The foci in the sample compartment were near the exit ports, as in the dispersive spectrometers, so we could use the existing accessories for dispersive spectrometers. The resolution was controlled with an aperture at the focus of the source paraboloid, which determined the solid angle through the interferometer, and it used all the luminosity advantage of the interferometer. The location of the sample after the interferometer eliminated the sample radiation from the measured spectrum, which could be the only source of something like the stray light in dispersive spectrometers, and at the same time minimized sample heating because the visible radiation of the source with a MIR beamsplitter was stopped by the beamsplitter coating. But now the user did not have a visible beam to aid in positioning samples. A source of visible light (a surplus collimator) was added to aid in the placement of samples and alignment of sample compartment accessories.
Two prototypes were built in early 1969 on large Formica tables. Don Graham of CSPI help us with the programming of a Varian 620i to collect the interferograms and compute the Fourier transform. These instruments were exhibited at a conference at the Ohio State University (OSU), and later at the Anaheim Conference Center. Tom Dunn, then President of Digilab, was the driving force in marketing.
The optical head controller evolved from the 296 controller, adding computer control of the interferometer scan, the mirror flippers, and the resolution aperture, and monitoring of internal diagnostic signals.
The production data system was designed around the Nova. The Nova CPU had cycle time of 2.6 microseconds (0.38 MHz!).
The system was first built with a minimum of 8K words of magnetic core memory (made up of 4K boards), of which 4K were used for programs and the rest used for data. This restricted the resolution in MIR to 2 cm-1 for a single-beam spectrum. Soon 8K boards became available allowing 32K memory in the same chassis.
For 0.5 cm-1 resolution we resorted to optical filters that allowed undersampling without aliasing. The list price of the system was about $74,000. About 25 core systems were sold.
We described the system at the Aspen International Conference in March 1970.
The FTS-14 system was a large instrument. The optical head was on a freestanding cabinet 38" wide by 45" high. The computer, and the optional high-speed paper tape reader and punch and later the disk, occupied the middle bay of the data system cabinet. This cabinet was big, 63" wide x 36" high, with the Teletype (TTY) and digital plotter on top. The spectrometer interface was built in a vertical rack, in the right-hand bay of this cabinet. It included the 15-bit A/D converter, the digital interface between the Nova and the optical head, the plotter and the TTY, and the hardware multiply and divide needed for acceptable computation times. A 2 cm-1 single-beam spectrum was computed in 30 seconds.
The Analogic A/D converter was a five or six small-card assembly on its own back plane, with a 15’ signal cable from the spectrometer. Think about the noise pickup, which had a substantial effect on the S/N of the result. The first improvement was a low pass filter at the input of the A/D. The actual I/O interface was 19 cards. The Nova bus was extended 15' to reach the I/O rack!
In a second row, we had the plotter and paper tape interfaces. The lower row was the signed multiply and divide with 16 cards. This was a very low volume production; therefore the I/O rack was wired-wrapped manually.
We got some of the first Novas shipped by DG. They had no software at all. To start the system, you had to enter a four-instruction bootstrap loader program using the front panel switches in the Nova. With that program in memory, you loaded a binary loader with the paper tape reader, and then you could load your programs. This procedure remained true after we developed our spectrometer programs, because the memory was not retained on power off.
Charlie Foskett developed the software for the early systems with part-time help from Brough Turner. The software was developed in the target system, the one and only, with an assembler, that showed up soon after delivery of the first computers. This was a standard two-pass assembler; therefore, after you punched the source code tape, you passed it through the reader twice, all at ten characters per second.
No operating system was available. Foskett built a 400-word supervisor program to connect the different tasks, which was used until the RDOS operating system became available (~1974).
The data file size was specified beforehand and assigned to absolute addresses in memory. Data was lost when a different file size was specified.
The only user interface was the TTY and its built-in paper tape reader and punch. By this time the TTY was a very reliable machine except for the mechanical tape reader.
A debugger allowed insertion of break points (up to a maximum of 4) where the program halted and the TTY printed the contents of the 4 accumulators. Typing an address would return the content of that location. To this date, Foskett remembers 054, the address of the I/O errors.
The FFT was coded in mixed radix (four/two) to optimize performance. Block floating point was used to maintain precision with the 16-bit word. Double precision software was introduced in 1974.
The IR Executive was built with three letter commands, so three command or parameters could be packed in a 16-bit word. The sign bit identified commands or parameters.
An important feature was the plotting of the frequency axis with the spectrum to take advantage of the precision of the laser reference, as opposed to the common practice in dispersive systems, which used a chart recorder with preprinted wave number scale. A less welcome feature of the plotter was the sampled appearance of the spectrum (a staircase of 200 steps per inch), that "everybody" knew was not real. It took some time for users to understand the information content, and it took faster computers that allowed enough interpolation for a better appearance of the digital plot.
Another problem we had was that the ratio of the sample and the reference spectra sometimes exceeded unity. Mrs. Sadtler educated me that all transmission values had to be less than 100%. She demonstrated running a spectrum in a dispersive spectrometer of a sample she may have seen before, because at a certain point in the plot, she moved her hand and placed her thumb to prevent the pen from going over 100%.
In mid 1970 with the addition of a 128K disk, the FTS-14 could achieve full MIR coverage at 0.5 cm-1 without under-sampling.
Brough Turner, who had joined us full time, implemented the software. Brough developed the programs that coadded interferograms to the disk and did the FFT in place in the disk. All this development had to be repeated for three different disks in a period of two years. These small disks (12" high by 30" deep in a 19" rack) were not very reliable then. The progress in memory capacity of the disk, and the increased flexibility of the operating system came at the expense of extra overhead, which slowed things down. This process still continues today, but we have always been saved by faster silicon.
The software continued to grow, providing more data processing features. The number of three letter commands and parameters grew with each release of the software.
The spectrometer resolution and sensitivity improved, leading to the development of a family of spectrometers, but that is another story.