Measurement and Analysis of 1/f Noise in Uncooled Microbolometers by Jason T. Timpe Submitted to the Department of Electrical Engineering and Computer Science in Partial Fulfillment of the Requirements for the Degree Master of Engineering in Electrical Engineering and Computer Science at the Massachusetts Institute of Technology May 22, 2000 Copyright 2000 Jason T. Timpe. All rights reserved. The author hereby grants to M.I.T. permission to reproduce and distribute publicly paper and electronic copies of this thesis and to grant others the right to do so. Author DepArtment of Electrical Engineering and Computer Science May 22, 2000 Certified by_ igH Qing Hu TOsis Spervisor Accepted by A Irthur C. Smith Chairman, Department Committee on Graduate Theses MASSACHUSETTS INSTITUTE OF TECHNOLOGY JUL 3 12002 LIBRARIES 1 Measurement and Analysis of 1/f Noise in Uncooled Microbolometers by Jason T. Timpe Submitted to the Department of Electrical Engineering and Computer Science May 22, 2000 In Partial Fulfillment of the Requirements for the Degree of Master of Engineering in Electrical Engineering and Computer Science ABSTRACT A method for measuring the 1/f noise in bolometers was developed that would be most conducive to a production environment. Several experiments were performed to discover how best to reduce the 1/f noise through processing changes. A model was developed to predict the performance of an infrared camera based on the 1/f noise measurement and other measurements made on the unpackaged wafers. Thesis Supervisor: Qing Hu Title: Associate Professor of Electrical Engineering and Computer Science 2 Chapter 1 Introduction Infrared (IR) imaging systems have the potential to make a dramatic impact on our way of life. They already perform several useful functions for military applications including weapons sights and targeting systems. In the commercial market, night-vision systems could be put into cars and planes to allow drivers and pilots to navigate more safely at night. Fire fighters could use them to see through smoke and to identify hazardous floors and walls. Unfortunately, the cost of producing and maintaining cryogenically cooled IR imaging systems has prevented them from achieving this potential. Only recently have room-temperature microbolometers opened up the possibility of high-performance, lowcost IR imaging systems. However, as with any new technology, microbolometer IR focal plane arrays are a long way from reaching their theoretical limits of performance. It is important that manufacturers continue to study how to increase the sensitivity and reduce the noise of these devices. 3 Previous studies indicate that the noise in an IR imaging system is dominated by the 1/f noise in the microbolometers themselves. The goal of this thesis was to develop an efficient way to measure 1/f noise, to use this measurement technique to perform experiments that might indicate a way to reduce the noise, and to prove that this measurement was valid by developing a model that could predict the performance of an IR imaging system. Chapter 2 describes the 1/f measurement system and the improvements made to make the system fit better into a production environment. It includes background on 1/f noise. Chapter 3 describes the experiments performed in an attempt to find a way to reduce the 1/f noise. It includes a description of how the microbolometers work and the results of the experiments as well as suggestions for further experiments. Chapter 4 describes the model used to predict system performance and the results of comparing these predictions to actual performance. It includes a description of the rest of the IR imaging system. 4 Chapter 2 1/f Measurement System 2.1 Overview It is extremely helpful, from a production standpoint, for noise measurements to be made as early in the manufacturing process as possible. This way potentially bad wafers or die can be removed from the line before more time and money is spent building a product that will not perform up to specifications. Furthermore, it is easier to identify both the causes of and the solutions to problems when the measurement is done at the detector level since the effects from the signal processor and the readout circuitry are not included. Finally, it decreases turnaround time for experiments since parts do not have to be packaged into systems before they can be tested. It is also important, from a production standpoint, for a test station to be both efficient and easy to operate. A test station that has a long test time or requires large amounts of operator intervention is a waste of time and money. It is best to make a test station as automatic as possible. However, there can often be a tradeoff between 5 automation and reliability unless self-checks are implemented so that the computer can handle unusual situations. A 1/f noise test station was constructed based on the procedure described by Lentz [1]. Several improvements were made in an attempt to make the station more productive. 2.2 1/f Noise 1/f noise or low frequency noise is distinguished by a power density spectrum (PSD) that is proportional to 1/f. This means that the spectral density of the noise increases without limit as the frequency decreases. 1/f noise is ubiquitous, appearing in everything from transistors and resistors to the fluctuations of a membrane potential in a biological system. The 1/f noise in the detectors can be observed as a voltage fluctuation, but it is actually due to a change in resistance. This means that a change in the bias voltage across the detector causes an equivalent change in the magnitude of the 1/f noise voltage. This fact is important for separating the 1/f noise from the other types of noise in the detector. It also suggests some interesting things about the source of the 1/f noise. The resistance of the detector is determined by: R = Wt 1 (2.1) where R is resistance, p is resistivity, w is width, t is thickness and 1 is length. Since the physical size of the detector cannot fluctuate that much, the change in resistance must be due to a change in resistivity. The resistivity of a semiconductor is given by: 6 p = 1 1 q( uMnn+ p,p) (2.2) where q is the charge of an electron, n and p are the number of negative and positive carriers respectively, and p is the mobility of the carriers. Since the charge of an electron is a physical constant, the 1/f noise must be due to fluctuations in either mobility or the number of carriers. The exact mechanism that causes such fluctuations is unknown but they may be due to traps and other defects in the material. The question is where such defects occur and how they can be removed. 2.3 Other Noise Sources There are other kinds of noise in the microbolometers besides I/f noise. Like every resistive element, they have Johnson noise. At thermal equilibrium the random motion of charge carriers in a resistive element generates a random electrical voltage across the element. This noise is white, which means that its PSD is flat across all bandwidths. Johnson noise is dependent on resistance and temperature (because an increase in temperature causes an increase in the mean kinetic energy of the carriers), but not on bias voltage. Another type of noise is thermal fluctuation noise. This is due to the fluctuations in temperature of the detector due to radiative exchange with the background. This is also a white noise although it is band limited by the thermal time constant of the detector. Finally there is noise due to the drift of the microbolometer temperature over time. These low-frequency artifacts show up as a 1/f 2 PSD. This noise is particularly 7 troublesome because it can overwhelm the 1/f noise if care is not taken to reduce this component. 2.4 Test Circuit To measure the 1/f noise it is necessary to have a very low-noise amplifier that will operate at low frequencies. The original circuit designed by Lentz is an 8-stage parallel bridge circuit. Two of the eight stages are shown in Figure 2-1. The device under test is labeled R in the circuit diagram. Rc is the resistor network shown in Figure 2-2 in series with two 10 kM wire-wound mechanical potentiometers. During testing, its value is adjusted so that it is 10 times the value of R. The entire circuit is placed in a test box and connected via BNC connectors to the device under test and to the rest of the test equipment as shown in Figure 2-3. 1k LM399 sg monitor -: biaswk Figure Vs 500k=Rc Ra= I k +9V(2) 04 ~~~hihi io Rb=lk +10 Rd -9V(3) - MOut 10k + -9V(3) Figure 2-1: Two stages of 1/f noise test circuit. 8 .. 50k -77 62.5k _/ LI L2 LO 50k 20k -_ 10k 12.5k S2 SS S6 Figure 2-2: Resistor network for adjustable Rc. High HP 3478A Multimeter Bolometer Low Model 113 Test Input A Out Box Pre-amp and Filter Bias Monitor Out High Node I HP 3458A Multimeter HP 3561A Dynamic Signal Analyzer 7 Figure 2-3: 1/f noise test station setup. The test station operates as follows. The bias is adjusted via a 10 ko wire-wound mechanical potentiometer until the bridge is biased at .41 V. This means that there is about .04 V across the detector itself. This voltage is high enough to allow the resistance to be measured but low enough that heating effects are relatively insignificant. After the bridge is balanced, the resistance of the detector can be calculated by measuring the 9 current through the 'hi node'. The resistance is measured so that all parts can be tested at the same power level rather than at the same voltage level. The voltage is then adjusted to bias the detector at 1 ptW. The voltage necessary for this can be calculated using the following equation: (2.3) Vmas, = II-.Viu W - R, The spectrum analyzer then averages 16 periodograms of 400 frequency points each from 0.1 Hz to 40 Hz. Each periodogram is a noisy estimate of the PSD of interest, thus averaging them improves the estimate. This estimated PSD contains all of the noise sources described above as well as noise from the test station itself. These components must be separated from each other to give an accurate measure of the 1/f noise. The first step is to isolate the 1/f noise from the white noise and 1/f 2 noise. This can be done by fitting the data to a curve of the form: S=a 2 +b +b 2 (2.4) where f is frequency. Once this is done, the parameter b gives the 1/f noise voltage at 1 Hz. However this value still contains the noise from the test station as well as that of the detector. Ideally, the test box would have no 1/f noise, however this is not true in practice. In order to separate the detector noise from the test box noise, the measurement is repeated at bias levels of % ptW and 0 pW. The % piW test serves as a check on the test system. If the system is working properly the 1/f noise voltage from the detector at % ptW should be 2of the noise voltage at 1 pW. At 0 pW, since the detector is unbiased, it should have only Johnson noise. Therefore, there should be no 1/f noise from the 10 detector, so any 1/f noise measured at this bias is due to the test station alone. This means that the 1/f noise voltage can be calculated as follows: V = b 1 w -b , (2.5) where V. is the noise voltage. For modeling purposes, it would be better to have a noise figure of merit that was independent of bias. This is done by dividing V,, by Vb.. This VN/V can then be used to calculate the noise voltage at whatever bias the system is running at. 2.5 Test Station Improvements There were several problems with the test station designed by Lentz, particularly from a production standpoint. Several improvements were made to the test station that increased reliability and efficiency. Lentz describes many environmental sources of noise including air currents, light, and EML Placing the DUT within a light tight enclosure solved most of these problems. A light tight enclosure is a large metal box that can be closed to prevent light from entering. This also removes any air currents that could be caused by people walking past the test station or other activity in the lab. Grounding the casing would also reduce EMI. The test box is run using batteries to prevent ground loops, however there were still occasions when spurious noise signals suggested EMI. In the original circuit, two BNC connectors were used to connect to the device under test. One was attached to the 'hi' node and one to the 'low' node. Since the 'low' node is at ground, this BNC can be eliminated. This improved the reliability of the test setup by removing another ground loop. It is also helpful to keep the BNC connectors as short as possible. 11 Despite these improvements to the testing environment, some spurious signals did appear occasionally. An example of this is shown in Figures 2-4 and 2-5. This plots two different tests performed on the same pixel. On the second plot there is a large peak that interrupts the normally smooth curve. This could be due to vibrations caused by other machinery that was running in the lab. The best solution to this problem would be to place the setup on an isolation table that would reduce these vibrations. However, since there was none available at the time, this theory could not be tested. Figure 2-4: Typical Noise Plot 1.80E-07 1.60E-07 1.40E-07 1.20E-07 e 1.OOE-07 - 8.OOE-08 6.OOE-08 4.OOE-08 2.OOE-08 O.OOE+00 0.1 1 10 100 Frequency (H) Another problem with the test circuit is that it was designed to measure detectors with a resistance smaller than 50 kO. Unfortunately, detectors occasionally have a larger resistance than this. To fix this problem the resistor network shown in Figure 2-2 was 12 Figure 2-5: Spurious Noise Signal 2.00E-07 1.80E-07 1.60E-07 1.40E-07 1.20E-07 CL I.OOE-07 8.OOE-08 6.OOE-08 4.OOE-08 2.00E-08 - 0.OOE+00 0.1 1 10 100 Frequency (Hz) changed to that shown in Figure 2-6. This not only made the circuit more robust, it also made the switching pattern more straightforward. This would make it easier for technicians to run the test station. 200k 20k 100k 100k 100k 20k 20k 20k Figure 2-6: New resistor network. 13 Much of the original test station required manual operation. Balancing the bridge, adjusting the bias, and even operating the dynamic analyzer were all done manually. A few circuit changes and a software program remedy this situation. The first step was to write a program that could operate the dynamic analyzer over a GPIB interface. This reduces a complicated measurement device down to a simple point and click user interface and removes the necessity of having an operator who understands how the dynamic analyzer works. The next step was to replace the potentiometer that adjusts the bias with a circuit that uses a digital potentiometer to perform the same function. This circuit is shown in Figure 2-7. With this circuit in place, the software can now control the bias of the circuit through a digital I/O card. This means that the majority of the test can now be performed automatically. Only the resistance measurement and the balancing of the bridge need to be performed by the operator. This makes the test station much more efficient since the operator can now be freed up to perform other tasks while the test is running, and need not constantly monitor the testing. Furthermore, since resistance is usually measured earlier in production, the software was configured to allow the user to input the resistance directly rather than measuring it. This means that the only operator intervention is the balancing of the resistors. The original test station also performed data collection and analysis in two different steps. This means that the operator of the test station would not be able to get any real time feedback about whether or not the test station appeared to be working properly. For example, the 1/f noise at 1 ptW should be about twice that at 14 pW. If this +9V 2) + 5V 2k 00 lo p 0 +15V +9V(2) .01 U 619kk 00 15 p 10k 10k Figure 2-7: Digital Potentiometer Bias Circuit is not true, it is likely that there was some problem in one or the other of the measurements. For example, bumping the table could cause a large jump in the noise, or could cause the probes to slide off the pads. Furthermore, there are times when the curve-fitting algorithm chooses negative coefficients. This also indicates a problem with the data, most likely a low frequency artifact that is exaggerating the 1 /fQ noise. With the original test station, it is impossible to determine this until all of the data has been collected and then analyzed. This means that time is wasted collecting data that is erroneous. This problem is also solved in software. The computer that performs the data collection can also perform the analysis of the data. This means that information can be 15 provided after each set of 16 periodograms. Specifically, the coefficients from the curve fitting can be displayed along with plots of both the original data and the approximation. This information can then be used, either by the operator in a manual setting, or by the computer in an automatic one, to determine whether or not it is worthwhile to take the next set of data or whether there is some problem with the data. A final improvement that could be made is to replace the two mechanical 10 kQ potentiometers in the bridge with digital potentiometers. This would allow the entire circuit to be placed under computer control, and fully automate the test. This is probably the most risky of the changes made to the circuit since the noise measurement is extremely sensitive to noise in the bridge resistors. The original reason for using a switched resistor network rather than a large potentiometer was because such potentiometers had too much noise. Unfortunately, time constraints prevented this change from being implemented. The problem with computer control is that the computer can only handle situations that are preprogrammed. Unlike a human operator, it cannot adapt to unusual situations. It is important, therefore, to make the computer program as robust as possible, so that it can handle typical problems that may arise. One typical problem that can arise is due to artifacts that show up as 1/f, noise. Such spurious noise can drown out the 1/f noise of interest resulting in erroneous data or even no data at all. This happens most frequently when the measurement is taken too soon after a change in bias. Probably because of the changing temperature of the detector, measurements taken immediately after the bias changed have large 1/f noise. 16 This problem was overcome by implementing a one-minute delay between the change in bias and the beginning of the measurements. Another problem results from averaging the periodograms. If one of the periodograms is much different from the others, this can throw off the average and change the measurement. To prevent this, the dynamic analyzer is set in single auto range mode. This means that it sets its range at the beginning of each measurement and then rejects any periodogram that is outside of this range. Unfortunately, the range could be set too low, so that too many of the periodograms are rejected. To prevent this, a time limit was placed on the measurement. Typically it takes a little less than three minutes for the dynamic analyzer to complete an average of sixteen periodograms. The software has a time limit of four minutes, after which it will record that it timed-out, and will begin the measurement again. Spurious noise signals are easiest to spot if the 1/f noise is not proportional to the bias voltage. To take advantage of this, the software can compare the noise voltage at 1 pW and at ptW. If the later is not approximately half the former, there is a problem and the measurement should be repeated. All instances where the computer assumes that data is bad should be logged along with the assumed erroneous data. This way a human operator can examine the data afterwards in an attempt to find what caused the problems. 2.6 Conclusions The improvements made to the test station drastically reduced the amount of user intervention required in the test station without reducing the validity of the 17 measurements. Further improvements in this circuit would help little. Changing the adjustable bridge resistance to digital pots will allow the computer to take over the entire test, but since the user is already required to set up the test, there is little benefit gained from this step. The next large step in improving the efficiency of the test procedure will come from an ability to probe all of the test pixels on a single wafer in parallel. This will remove the necessity of the user having to change the probes after every test and will allow an entire wafer to be run without user intervention. 18 Chapter 3 1/f Noise Experiments 3.0 Overview With a reliable test station, there are a number of experiments that can be done in an attempt to discover the source of the 1/f noise and how to reduce it. According to Sze [41 the origin of 1/f noise in most semiconductor devices is due to the surface effect and carrier recombination at traps. It is possible that the 1/f noise in the microbolometers is due to similar effects. This immediately suggests a path of experimentation. First, testing parts with different thicknesses can determine whether or not it is a surface effect. Secondly, parts can be annealed in different atmospheres and at different temperatures. Annealing has a dramatic improvement on the 1/f noise of other devices and may prove equally effective here. 19 3.1 Microbolometer Overview A bolometer is a resistor with a high thermal coefficient of resistance (TCR). The microbolometers in these experiments consist of a thin layer of vanadium oxide (VO") between encapsulating layers of silicon nitride (Si 3N4). The VOx is the temperature sensitive material. The incoming infrared radiation strikes the microbolometer, and the energy from this radiation heats up the VOx causing a change in resistance. This resistance change is related to the power of the radiation and so to the infrared energy being emitted by the scene the imager is looking at. A short voltage pulse across the microbolometer measures this change in resistance through an integration capacitor. Thus an effective measurement is made of the infrared radiation being emitted by the scene. Ideally, all of the energy from the incoming radiation would be used to heat up the vanadium oxide. In reality, some of the energy is lost through the thermal connection to the substrate. To minimize this effect, the microbolometer is suspended above the substrate on two thin metal legs. Other losses include those due to imperfect optics and the fact that the microbolometer does not cover the entire pixel area. On every production die there are eight different test pixels that can be tested. Pixels in the actual array cannot be tested because they do not have the metal contacts and so there is nowhere to connect the probes. These eight test pixels are of various geometries and may or may not be suspended above the substrate. There were two main types of pixels used in this testing, called F2 and F2L pixels. Both are suspended above the substrate. The F2 pixels are approximately 15 pLm long and 38 pm wide. The F2L pixels are approximately 27 pm long and 19 pm wide. It 20 is important, when comparing different pixel shapes to take these differences into account. The F2L pixel geometry is most closely matched to the geometry of the pixels in the array. 3.2 Surface and Bulk Effects The first experiment run on the microbolometers was to vary the thickness of the vanadium oxide and determine what effect, if any this had on the 1/f noise. This experiment would determine whether or not the phenomenon that causes the 1/f noise is a surface effect or a bulk effect. The difference between a surface effect and a bulk effect is as follows. In the case of a bulk effect, the phenomenon is evenly distributed throughout the volume of the material. This means that the equation for the noise can be written as V XOVlp (3.1) Since the length and width of the microbolometers is the same for all pixels of a particular geometry, the noise voltage is inversely proportional to the square root of the thickness. The bulk effect equation can be understood by thinking of the detector as a noisy resistor. When the thickness is doubled, it is the same as putting two equivalent noisy resistors in parallel. This means that the equivalent circuit shown in Figure 3-1 can represent the thicker detector. Adding the resistances in parallel gives an equivalent of 2R. Adding the noise currents gives and equivalent of I*12. This means that the equivalent noise voltage is Vn/42. 21 I = ER R R ) R/2 I= E/R I = sqrt(2)*Et/R R/2 XK E = Et/sqrt(2) Figure 3-1: Equivalent Circuit for Bulk Effect A surface effect occurs when the phenomenon that causes the noise is concentrated near the surface of the detector. In this case, the equivalent circuit is that shown in Figure 3-2, where RB is a noiseless resistor that represents the bulk and Rs is a noisy resistor in parallel that represents the surface. In this case, doubling the thickness of the detector cuts the resistance in half while the noise remains constant. This means that the noise voltage is inversely proportional to the thickness of the detector. This assumes that the resistivity of the surface layer and the bulk layer are the same and that 22 the surface effects extend at most to a depth equivalent to the thickness of the original detector. X I=E/Rs RB+Rs= R I= Et/Rs R/2 R/2 E =Et/2 Figure 3-2: Equivalent Circuit for Surface Effect Because of the different behaviors of the surface and bulk effects, it should be possible to determine whether or not the effect is bulk or surface by increasing the thickness. This is important because it will determine how effective it is to reduce the 1/f noise by increasing the thickness. Increasing the thickness has some negative effects such as increasing the thermal time constant. Thus, it is important to characterize the 23 benefits properly so designers can make the proper choice when developing the next generation of pixel. 3.3 Thickness Experiments The thickness variation experiment was run on four wafers from two different lots, two from each lot. These wafers were run through the standard process with the rest of the lot, except at the vanadium deposition step. At this step, the experimental wafers had a deposition time of 60 minutes rather than the standard 40 minutes. Since the deposition time controls the amount of vanadium oxide deposited on the wafer, the thickness should be proportional to time. This means that the thickness of the vanadium should be 1.5 times greater in the experimental wafers. After processing the wafers were measured using the 1/f measurement station described earlier. Both F2 and F2L pixels were tested. Unfortunately, the resistance was greater than 50 kQ on some of the F2L pixels, so the test station was inadequate for their measurement . Initially the measurements from the experimental wafers were compared to wafers from the same lot that had been run at the standard deposition time. Using wafers from the same lot should minimize the number of uncontrolled variables, since there could be lot to lot variation. The lot comparison is shown in Table 3-12. 1Not all of the test station improvements described in Chapter 2 had been made at the time this experiment was done. 2 A complete listing of all of the data from the thickness experiments is shown in Appendix A. Only those measurements within 3a were considered for the statistical measurements. This is to prevent atypical pixels from influencing the statistics. 24 Lot 99.1 99.1 104.1 104.1 99.1 104.1 104.1 Part F2 F2 F2 F2 F2L F2L F2L Dep. Time Mean VnN 4.79E-07 40 min 2.89E-07 60 min 4.23E-07 40 min 60 min 3.23E-07 3.38E-07 60 min 4.46E-07 40 min 60 min 3.30E-07 Std. Dev. 1.15E-07 3.08E-08 4.24E-08 7.39E-08 7.56E-08 1.1OE-07 7.60E-08 Table 3-1: Thickness Experiment Lots The lot comparison data should support one of the above theories of 1/f noise, either the surface model or the bulk model. If it is assumed that the measurement of the standard wafers is accurate, then it is possible to predict the behavior of the experimental wafers. The mean of the measurement should be reduced by a factor of either 1.5 or 41.5, due to the surface model and the bulk model respectively. These predictions, along with the errors of the predictions are shown in Table 3-2. Lot Part 99.1 104.1 104.1 F2 F2 F2L Mean Bulk Bulk Err Surface Surface Err 2.89E-07 3.91E-07 -35.34% 3.19E-07 3.23E-07 3.45E-07 -6.87% 2.82E-07 3.30E-07 3.64E-07 -10.43% 2.97E-07 -10.51% 12.74% 9.84% Table 3-2 Bulk and Surface Predictions and Errors These results seem to indicate a surface effect, but are rather unsatisfying, particularly since the results from the F2 pixels from Lot 104.1 seem to indicate a bulk effect. In addition, due to the large resistance of the F2L pixels from Lot 99.1, there is no data for that lot. The error in the predictions could be due to an error in the models, but it seems more likely that the problem lies in the amount of data taken. More measurements would increase the statistical certainty of the predictions. 25 The wafers were packaged into systems so that comparisons could be made at the system level. This meant that no further measurements could be taken on the wafers. To get further data for the comparisons, it was possible to test other wafers that came from different lots, but were produced around the same time as the lots of interest. This has the advantage of giving a greater sampling, but it also adds in the variables from lot to lot variation. In this case, we are comparing all F2 pixels run at the standard deposition time to those run at the longer deposition time, and likewise with the F2L pixels. These comparisons as well as the model predictions are shown in Table 3-3. Part Dep. Time Mean F2 40 min 4.83E-07 F2 F2L F2L 60 min 40 min 60 min 3.06E-07 4.96E-07 3.32E-07 Bulk Model Bulk Err Surface Surface Err 3.94E-07 -28.93% 3.22E-07 -5.27% 4.05E-07 -22.19% 3.31E-07 0.23% Table 3-3: Model Comparison with All Lots These results clearly show a strong support of the surface model, particularly with the F2L pixels. The F2L pixels are those that most closely resemble the pixels that are in the actual array, so it seems fairly certain that the improvements seen at the detector level will be borne out at the system level. This potential will be explored further in Chapter 4. Although there was not time to make any more fully functional wafers with thickness variation, an experimental lot was made that had thickness variations between .5 times and 2 times the typical thickness. There was no intention to package these wafers, nor were the pixels suspended, so the measurements done could not be confirmed at the system level. The plot of the variation of Vn/V with vanadium thickness is shown in Figures 31 and 3-2. In Figure 3-1 a linear approximation is plotted. The closer the data points are 26 to this line, the greater the chance of a surface effect. In Figure 3-2 a square root approximation is plotted. The closer the data points are to this line, the greater the chance of a bulk effect. It is difficult to tell just by looking at the plots, which of the graphs most closely approximates the data, but Table 3-4 shows that the square-root approximation is actually closer to the measured data points. This is unfortunate since it actually goes against what we measured earlier. This seeming contradiction can be explained by realizing that it is possible for the surface to extend throughout the whole bulk. In the IX to 2X range, where the first experiments were done, the high noise region of the vanadium is in fact smaller than the entire thickness, so the noise goes down proportionally to the thickness, just as a surface effect should. However, at thicknesses less than IX, the high noise region extends throughout the entire thickness of the material. Thus, the noise looks like a bulk effect, and the noise goes as the square root of the thickness. Figure 3-1: Data Points and Surface Model 1.60E-06 1.40E-06 1.20E-06 1.OOE-06 8.OOE-07 6.OOE-07 4.OOE-07 - 2.OOE-07 - O.OOE+00 -- 0 I 0.5 __ _ I - I 1 1.5 VOx Thickness(X) 27 - 2 I 2.5 Figure 3-2: Data Points and Bulk Model 1.60E-06 - 1.40E-06 1.20E-06 - 1.OOE-06 8.OOE-07 6.OOE-07 4.OOE-07 - 2.OOE-07 0.OOE+00 10 0.5 T 1. 1 1.5 - 2 VOx Thickness(X) IX Resistivity F2L Pixels Approximation Error VnN VOx Thickness 4.88% 1.19E-06 1.25E-06 0.5 4.73% 1.19E-06 1.25E-06 0.5 15.52% 1.02E-06 8.81 E-07 5.90% 1.02E-06 9.61E-07 12.33% 8.46E-07 9.66E-07 1.5 12.18% 7.55E-07 8.46E-07 1.5 6.43% 7.22E-07 6.75E-07 2 8.85% Sqrt Thickness 0.707106781 0.707106781 1 1 1.224744871 1.224744871 1.414213562 Approximation Error Vn/V 3.12% 1.21 E-06 1.25E-06 2.97% 1.25E-06 1.21E-06 13.25% 8.81 E-07 9.98E-07 3.82% 9.98E-07 9.61 E-07 13.62% 8.34E-07 9.66E-07 10.54% 8.34E-07 7.55E-07 3.53% 7.22E-07 6.96E-07 7.26% Table 3-4: Bulk vs. Surface Model 28 2.5 In other words, it appears that the thickness of the high noise region is almost exactly that of a standard pixel, approximately 600 angstroms. In the cases where the pixel is thicker than IX, there is an additional low noise region that causes the 1/f noise to go down proportionally to the volume. If the pixel is made thinner, however, there is no low noise region at all, and so the noise goes down to proportionally to the square root of the volume because the high noise region extends throughout the entire bulk. The fact that the standard thickness lies right at the crux of these two regions could also explain variations in the noise measurements in production wafers. If some production technique can slightly vary the thickness of the surface layer, it could produce occasional good pixels when the surface is made thinner and there is a low noise region. Furthermore, since the thickness of the vanadium slightly varies from lot to lot, some of these wafers could have exceeded the thickness of the high noise region. 3.4 Annealing Experiments The characterization of the phenomenon that causes the I/f noise in the microbolometers is extremely useful, but it is still necessary to find a way to reduce the I/f noise. As seen above, the I/f noise is inversely proportional to the thickness of the detector, so increasing the thickness can reduce the 1/f noise. There is, however, a practical limit to this solution. First of all, increasing the volume of vanadium oxide increases the thermal mass of the detector, thereby increasing the thermal time constant. This has the undesired effect of slowing down the response of the detector. Furthermore, increasing the thickness increases the thermal conductivity. This means more of the energy will be conducted to the substrate, thus decreasing the sensitivity. 29 Figure 3-3: Histogram of Baking Experiments Pre mean = 1.13e-6 Pre std = 2.38e-7 Post mean = 9.44e-7 Post std = 2.84e-8 Twice mean = 9.349-7 Twice std = 3.93e-8 N Post-Baking 0 Pre-Baking M Twice-Baked 12 10 8 4) Cr U. 6 4 2 I- t 0 8.95E-07 9.90E-07 9.43E-07 1.04E-06 1.09E-06 1.13E-06 More VnN Figure 3-4 Plot by Part 2.50E-06 2.OOE-06 X Pre Baking + Post Baking 0 Twice Baked X - x 1.50E-06 xX XXX X Xt 'p 1.OOE-06 - X X X off*~ * .0 5.OOE-07 0.OOE+00 0 5 15 10 Part index 30 20 25 The fact that the 1/f noise seems to be a surface effect supports the theory that it is caused by traps. Traps are usually due to impurity atoms or lattice defects. Since the surface of a material undergoes a rougher treatment than the bulk, it seems likely that most impurities and defects would be concentrated at the surface. It is possible that annealing may removes some of these defects. The thermal energy will give the molecules in the material enough energy to move around and fix the defects. In other devices, it has been shown that low-temperature hydrogen annealing can remove most of the interface traps [4]. Because of the previous successes with other devices and annealing, an experiment was tried in which the wafers were annealed in a vacuum at 250* for one hour. They were tested both before and after the annealing. The results of this experiment are charted in Figure 3-33. It appears that there is some improvement in the 1/f noise after the annealing, however the overlap in the histograms makes it uncertain whether the improvement is real or merely an anomaly due to changes in the measurement environment. One encouraging fact is that there is less variation after baking. Each part is plotted separately in Figure 3-4. Here it is apparent that every pixel showed some improvement, although this improvement is not uniform. This could be due to the fact that some pixels had a greater number of defects to begin with, and yet there is a limit to the improvements that can be made with the annealing. Thus, annealing appears to bring all pixels down to the same level of noise. These results are encouraging, but the improvement is not nearly large enough. 31 The next step was to anneal the wafers for even longer. The same wafers were annealed for 2500 for another six hours. These are the 'twice baked' wafers shown in the above figures. It appears that the additional annealing had little or no effect on the pixels. It appears that the initial annealing brought the pixels to some plateau of noise level. 3.5 Conclusions As with any experiment, the results of both the thickness and annealing experiments were mixed. Certainly the thickness experiment seems to strongly indicate a surface effect, but it would be better to have more data points. Further experiments at greater thickness extremes, both lower and higher, could help to confirm the theory that the surface effect begins somewhere around IX thickness. The annealing experiment shows some improvement, but not enough to be satisfying. It definitely seems to show that at least part of the 1/f noise is due to traps in the surface. Annealing brings those pixels with a large number of surface defects back into range. Thus annealing can improve reliability and uniformity. The next step would be to try annealing at higher temperatures and in different atmospheres. Unfortunately, it will be difficult to reach higher temperatures without damaging the readout circuitry underneath the pixels. The aluminum metalization in the CMOS can be damaged at temperatures much higher than this. One possibility is to use a rapid thermal anneal. Because the detectors are thermally isolated from the substrate, this may heat the detectors to a high temperature while keeping the readout circuitry safe. It is also 3 A complete listing of all of the data from the baking experiments is shown in Appendix B. 32 unfortunate that the equipment was not available to anneal the wafers in different atmospheres, at least not at the temperatures of interest. 33 34 Chapter 4 System Modeling 4.0 Overview The results of the experiments described in the previous chapter are interesting, but they are only useful if they actually have some relation to the performance of the entire system. It would be helpful if the 1/f noise measurements could be used to predict the performance of the system. In this way, bad die can be removed from the line early on, since it can be predicted whether or not their system performance will be adequate. Furthermore, noise reduction experiments will take less time since they can be tested at the wafer level, rather than at the system level. Finally, a good model is a valuable design tool, since hypothetical designs can be modeled to determine their performance before time and money is spent actually building them. And different parameter curves can be plotted to help designers determine the relative effects of different variables. A good model can be developed from the theoretical performance of the system, but it can only be proven through experimental evidence. That is, the predictions of the 35 model must be borne out by actual system performance. No model can ever be perfect, since it is impossible to take into account all of the variables that can affect system performance. Throughout the time of these experiments, a model was developed and continuously modified as new results became available. 4.1 Model The model contained a number of sections. The first section consisted of an area where pixel parameters could be input. Some of these are physical constants or material properties that the designer has no control over. Both the electrical and thermal properties of the material are important. The material property of the greatest interest for this thesis is the 1/f noise figure of merit, labeled 'Vnove,'. This is the 1/f noise parameter that is measured by the test station. Hopefully, with this value as well as other measured values input, the model will accurately reflect the performance of the system. The pixel layout parameters also need to be input. These are the parameters over which the designer has the most control. They are determined by the size and shape of the pixel itself. For the most part they are limited by the layout technology available to the designer. They can also be limited by mechanical considerations since the bolometer is essentially a bridge. Making the parameters too extreme can cause bridges to collapse or to become detached during processing. The model also needs the ROIC parameters. These describe the behavior of the readout circuitry and its effects on the noise. One of the most important of these is Vbias. This is the amplitude of the bias pulse sent across the detector to measure the change in resistance. This is typically .7 V, but may be higher than this. As described in Chapter 3, 36 1/f noise varies with bias. Therefore, it is important to match the modeled bias voltage to the voltage that the system was tested at. The other ROIC parameters that are important are those involved in the integration of the current caused by the bias pulse, and the A/D conversion of this voltage. Finally, the other parts of the system are included in system parameters. Most important of these are the frequencies of interest. The type of test that the system will undergo determines the frequency band that the model must look at. The acceptance test used in this case looks at two different kinds of noise, spatial and temporal. The temporal noise is the variation of the pixels over a number of frames. The spatial noise is the variation of the pixels from the neighboring pixels. Since only one pixel is modeled, these tests must be converted to equivalent bandwidths. The temporal noise is considered to be the noise between 1 Hz and the upper limit set by the frame rate. The spatial noise is considered to be the noise between .01 Hz and 1 Hz. Any other bandwidths could be used in the model. For example, a lower frequency could be used for to model the amount of noise that would be observed by a human eye. The first important value that needs to be calculated is the effective fill factor. The fill factor is that percentage of the pixel area that is actually covered by the bolometer and therefore absorbing radiation. The atmosphere, the optics, the absorptance of the vanadium oxide, and the physical size of the detector all play a part in how efficient the detector is at absorbing the incoming radiation. Since many of these factors vary with wavelength, they can be entered as vectors of values within the wavelengths of interest (7 - 14 pm) and then a weighted average is taken. It is important when using the model as a design tool to edit this section carefully. Clearly, altering the 37 size of the pixel changes the fill factor, but changing the thickness can have an effect on the absorptance and thereby change the fill factor as well. The noise sources from the read out circuitry need to be measured independently and entered as factors into the model in a 'Noise Measurements/Estimates' section. It is in this section that one would normalize the 1/f noise with volume if the surface effect is assumed, or with the square root of the volume if the bulk effect is assumed. Throughout the time of using the model, the bulk effect was assumed since this would provide a more pessimistic estimate of the noise. The thermal time constant of the detector needs to be calculated from the thermal properties of the materials that make up the pixel. This is an important parameter because it determines the response time of the system to changes in the scene. A rapid thermal time constant can prevent unwanted visual effects when panning the imager across a scene, or when there is a rapidly moving object in the field of view. The thermal conductance and thermal capacitance of the pixel is calculated using the physical parameters of the metal, silicon nitride, and vanadium oxide, as well as the structure of the pixel. These two parameters than provide the thermal time constant. Like electrical conductivity, thermal conductivity is proportional to the thickness and the width and inversely proportional to the length of the material. However it is important to realize that the two legs of the bridge are electrically in series, while they are thermally in parallel. It is also important to consider that energy can be lost through radiation of the pixel. Next the responsivity of the pixel in counts per degree Kelvin needs to be calculated. This must take into account not only the efficiency of the pixel in absorbing 38 energy and converting this energy into a resistance change, but also the effectiveness of the readout circuitry in converting that resistance change in to a digital signal. The calculation proceeds as follows: 1) eff*size*I/4FA2 is the power of the incoming radiation 2) Dividing this by G gives the temperature change when this power is converted from radiative to thermal energy. 3) Multiplying by TCR and R then gives the change in resistance due to the incoming radiation. 4) Dividing by R then gives the fractional change in resistance. This is proportional to the fractional change in the current through the detector. 5) Multiplying by the original current is Vbias/R, gives the fractional change in the current through the detector. 6) The current than charges up the integration capacitor. Multiplying by Tint/Cap gives the change in the voltage on the capacitor. 7) Dividing by the slope of the A/D converter gives the responsivity of the pixel. Although the resistance of the pixel can be calculated from the sheet resistance of the VOx, the model will put a lower limit on the resistance based on the WeidemannFranz ratio. This is the ratio between the electrical and thermal conductivity of a material. Essentially, if a material has a specific thermal conductivity, there is a maximum electrical conductivity. The ideal material for the legs of the bridge would be something that had infinite electrical conductivity and zero thermal conductivity. Weidemann-Franz shows not only that this is impossible, but also exactly how close it is possible to get to such an ideal. 39 Another important calculation is the pixel TCR. This is not the same as the TCR of the vanadium oxide, because the legs are a part of the pixel. In other words, the legs contribute to the total resistance of the pixel, but have a very low TCR compared to the VOx. The effect of the legs on the TCR can be minimized by keeping the resistance of the legs small compared to the resistance of the bridge. Finally, the noise in various bandwidths is calculated. Each of these sections are the same, with only the bandwidths changed. First a high and low alias point is calculated. Some practical limit has to be set on the bandwidth of the 1/f noise so that the model can calculate the amount of noise that is alliased back into the bandwidth of interest. In this case it was assumed that there was no noise after 10/2nr, where t is the thermal time constant. Next, the thermal fluctuation noise is calculated. This noise is white, but it is bandlimited by the thermal time constant, and so the aliasing effects have to be taken into account. It is also useful to calculate the radiation limited NETD. This is the ideal noise. The noise of the pixel will be equal to this value when all other noise is removed except for the thermal fluctuation noise caused by radiative exchange between the pixel and its environment. This is what is known as background limited. Johnson noise is white, so it only has to be limited by the bandwidth of interest. The 1/f noise is calculated in two sections. One calculates the aliased 1/f noise, and another calculates the 1/f noise without aliasing. Then these are added in quadrature to give the total 1/f noise. Note that the 1/f noise parameter must be properly scaled with the square root of the volume because of the assumption that the 1/f noise is a bulk effect. 40 The remaining terms are ROIC noise and are mostly measured independently during ROIC design and development. All of the noise terms are then added in quadrature to give a total noise. 4.2 Model Confirmation Ideally, the model would accurately predict the behavior of any given system based on measurements made on the die. Practically, there are not only variations in the measurements that add error, but also problems with the model that cause inaccuracies in the predictions. It is important then to determine how reliable the model is. A number of production die were tested over the time of these experiments. These die were eventually packaged into systems, and this gave an ideal opportunity to test the model. Several parameters were used as input into the model. The most important of these was the Vn/V, since this would help to confirm not only the model, but the test station as well. The bias voltage was of course important, since some systems are run at higher biases than others and this can have a dramatic effect on the 1/f noise. The mean resistance of the detectors in the array is measured at wafer probe, so this was used as a resistance rather than calculating the resistance from the sheet resistance of the VOx. 41 Figure 4-1: Responsivity Check 80.00 75.00 70.00 65.00 60.00 55.00 - 50.00 - 45.00 40.00 35.00 30.00 30 35 40 45 50 55 60 Measured Responsivity (counts/dog C) 65 70 75 80 The model was run using the input parameters above and the results were compared to the measurements made at the system test station. The comparison is shown in Figures 4-1, and 4-2, and in Table 4-1. Figure 4-1 shows modeled versus measured responsivities. Ideally, all of the data points would lie on the line, and in fact they are quite close, except for a few outliers. Table 4-1 shows that the average error is about 3%. Of course the ± errors are canceling each other out, but even the absolute values of the errors average out to 9%. 42 Figure 4-2: TNETD Check 120 100 -I 40 20 0 0.00 20.00 40.00 60.00 80.00 100.00 120.00 Measured TNETD (mK) Figure 4-2 shows modeled versus measured temporal NETD. Again, the data points appear to be quite close to the ideal line. And in fact, the Table shows that these predictions are even better. The average magnitude of the error is only 6.8%. These results are quite promising. They show that the model, despite being based on only one pixel seems quite capable of predicting system performance, at least for responsivity and TNETD. The problem with this conclusion is that it is based on only one type of pixel, without any major difference that will cause changes in noise levels that the model should be able to predict. Fortunately, this experiment was made using the wafers from the original thickness experiment. These wafers have 1.5 times the VOx and so should have a significant difference in noise from the standard production wafers. The results from this 43 Bias Meas R. VnN 46.8 5.84E-07 35.5 8.06E-07 45.3 4.49E-07 24 1.06E-06 31.1 1.17E-06 31.1 1.22E-06 31.1 1.15E-06 26.6 9.92E-07 26.6 1.09E-06 28.7 1.06E-06 28.7 1.14E-06 26.6 1.08E-06 28.7 1.09E-06 24.8 9.93E-07 23.3 1.02E-06 24.8 1.03E-06 24.8 9.84E-07 23.3 1.01E-06 21.9 9.84E-07 23.3 1.04E-06 Mean 28.85 9.96E-07 (X) 1.50 1.25 1.25 1.00 1.00 1.00 1.00 1.00 1.00 1.25 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.06 Model Reasp 71.12 40.57 Responsivity 56.35 51.8 51.87 58.55 49.55 52.13 50.48 51.13 50.16 52.29 53.56 49.55 49.36 57.35 55.57 42.00 57.98 50.01 59.66 55.13 56.40 51.37 62.12 56.28 42.28 55.37 58.40 55.12 58.27 58.97 61.20 59.67 62.79 55.73 58.54 54.21 58.9 60.38 58.06 53.99 Error TNETD 26.21% 67.44 -21.69% 82.48 -19.03% 78.59 -0.98% 78.35 0.92% 87.60 14.45% 85.80 9.20% 87.73 10.31% 90.06 2.41% 90.68 18.80% 90.84 5.09% 85.51 -14.68% 90.45 12.17% 92.17 1.83% 82.66 -0.81% 79.96 -0.46% 76.54 8.78% 88.14 3.90% 83.32 -1.18% 76.40 8.14% 81.80 3.17% 83.83 Model TNETD 54.02 88.61 78.92 83.89 98.9 87.3 89.3 83.29 Error -19.90% 7.43% 94 84.82 88.61 102.81 85.32 82 88.73 85.96 83.65 82.87 84.09 83.74 85.54 0.42% 7.07% 12.90% 1.75% 1.79% -7.52% 3.66% -6.63% 3.63% 13.67% -7.43% -0.80% 10.97% 12.31% -5.09% -0.54% 10.07% 2.37% 2.01% Table 4-1: Model and Measurement Results experiment are shown in Figures 4-3, and 4-4 and in Table 4-2. Again the results are very good with the average magnitude of the error at 6.49% for TNETD and 5.50% for responsivity. The only unusual thing is that almost all of the modeled TNETD results are below what was measured. This seems to indicate that there is some source of noise that the model is not taking into account. 44 Figure 4-3: Responsivity Check, Thicker Wafers 80 70 60 50 I V. 40 30 20 10 0 0 10 50 40 30 20 60 70 80 Measured Responsivity (counts/deg C) Figure 4-4: TNETD Check, Thicker Wafers 70 60 00000 50 a I- 40 w zI- 30 20 10 0 0 10 20 40 30 Measured TNETO (mK) 45 50 60 70 Mean Model Meas R. Vn/V Bias (X) 27 26.2 26 28.1 4.88E-07 4.25E-07 3.57E-07 3.81E-07 26.825 4.13E-07 Model 1.25 1 1 1.25 Resp. 67.57 55.32 53.12 53.85 Resp 70.06 54.57 61.51 54.46 Error TNETD TNETD Error 3.69% 54.16 51.86 -4.25% -1.34% 62.81 59.29 -5.60% 15.81% 65.74 55.85 -15.040% 1.15% 63.99 63.31 -1.06% 1.125 57.465 60.15 4.83% 61.675 57.5775 -6.49% Table 4-2: Thicker Wafer Model and Measurements 4.3 Conclusions The model seems to work quite well. It seems to accurately predict system responsivity and TNETD. It is also capable of predicting these values when the design of the pixel is changed. Certainly improvements could be made to the model. The results are not shown here, but the modeled results for SNETD are dramatically different from what is measured at the system test station. This could be due to certain effects caused by the shutter of the camera that are not taken into account by the model. Despite this deficiency, the model is still a powerful design tool. 46 Chapter 5 Conclusions The results of this work provide a number of powerful tools for the manufacturers and designers of infrared imaging systems based on uncooled microbolometer technology. There is now a more efficient, highly automated system for measuring 1/f noise at the wafer level. Furthermore, there is a greater understanding of what can be done to reduce this noise so that it no longer dominates the system. Most importantly, there is a model that can be used to predict performance based on many different design parameters. This is a valuable design tool for any engineer. It allows an easy way to try out new designs. Plus it can be used to develop parametric curves that will show relative effects of various parameters. Finally, it can be used to develop intuition about how the imaging system and microbolometers function. Certainly there is further work that needs to be done to bring uncooled infrared systems closer to the theoretical limits of their performance. There is still no clear understanding of what exactly causes the 1/f noise, nor, more importantly, how to 47 improve it. The model is an excellent tool, but still cannot adequately predict spatial noise. The test station is highly automated, but still can only test one pixel every twenty minutes, and with 128 test pixels on a wafer, this hardly allows for thorough testing of every wafer. Despite these challenges, the utility of uncooled technology and the wide variety of applications open to infrared solutions is sure to encourage engineers to continue to investigate these detectors and push the envelope of performance. 48 Appendix A VOx Thickness Experiment Measurements Lot 99.1 99.1 99.1 99.1 99.1 99.1 99.1 99.1 99.1 99.1 99.1 99.1 99.1 99.1 99.1 99.1 99.1 99.1 99.1 99.1 99.1 99.1 99.1 99.1 99.1 99.1 99.1 99.1 104.1 104.1 104.1 104.1 104.1 Wafer Die Part Dep. Vhigh 25 25 22.4 22.4 22.4 22.4 22.5 22.5 21.5 21.5 21.5 21.5 40 40 40 0.163 0.082 9.16E-08 5.53E-07 5.06E-08 5.83E-07 0.152 6.78E-08 4.31 E-07 40 21.5 40 40 0.076 0.152 0.076 0.152 0.076 0.147 0.074 0.147 0.074 0.147 0.074 0.091 0.045 0.098 3.76E-08 6.72E-08 3.96E-08 7.21 E-08 3.63E-08 2.85E-08 1.60E-08 6.76E-08 3.24E-08 6.69E-08 3.07E-08 8.51 E-08 3.47E-08 4.58E-08 2.59E-08 7.31 E-08 4.41 E-08 7.42E-08 Meas R. 21.5 21.5 21.5 21.5 21.5 21.4 21.4 21.4 21.4 21.5 21.5 21.4 21.4 21 21 20.1 20.1 18.8 18.8 19 49 Time 40 40 40 40 40 40 40 40 40 40 40 40 40 0.049 0.147 40 0.074 40 40 0.147 0.074 0.147 0.074 40 40 40 40 40 40 40 40 40 0.147 0.074 0.147 0.074 0.147 0.074 40 0.136 0.068 40 0.14 b VnN 4.39E-07 4.27E-07 4.69E-07 4.70E-07 4.59E-07 1.84E-07 1.80E-07 4.56E-07 4.21 E-07 4.52E-07 4.03E-07 9.25E-07 7.18E-07 4.49E-07 4.61 E-07 4.95E-07 5.87E-07 4.97E-07 4.64E-08 6.03E-07 9.41 E-08 6.37E-07 4.51E-08 6.OOE-07 6.47E-08 4.37E-07 3.35E-08 4.40E-07 7.83E-08 5.31 E-07 4.05E-08 5.41 E-07 6.09E-08 4.13E-07 2.88E-08 3.84E-07 5.17E-08 3.74E-07 2.80E-08 3.90E-07 5.60E-08 3.88E-07 Lot 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 Wafer Die Part Meas R. F2 F2 F2 F2 F2 F2 F2 F2 F2 F2 F2 F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L 19 20 20 20.4 Dep. Time 40 40 40 40 20.4 20.2 20.2 40 40 20.4 20.4 40 20 20 19.6 19.6 19.6 19.6 18.6 18.6 18.6 18.6 40 40 40 19.6 40 40 19.6 46.8 46.8 20.7 20.7 20.7 20.7 20.7 20.7 21.9 21.9 21.9 21.9 40 40 40 40 40 40 40 40 40 40 40 40 40 40 40 40 40 40 40 40 Vhigh b VnN 0.07 0.143 0.072 0.143 2.91 E-08 6.44E-08 3.23E-08 7.32E-08 3.70E-08 6.40E-08 3.17E-08 6.74E-08 3.47E-08 5.62E-08 2.98E-08 5.1 OE-08 2.85E-08 5.57E-08 2.81 E-08 4.94E-08 2.63E-08 4.76E-08 2.27E-08 6.28E-08 3.09E-08 8.26E-08 4.45E-08 5.89E-08 2.85E-08 6.31 E-08 3.32E-08 6.62E-08 3.65E-08 1.30E-07 6.36E-08 7.80E-08 4.33E-08 6.51 E-08 3.07E-08 7.25E-08 3.49E-08 6.66E-08 3.38E-08 5.94E-08 3.28E-08 9.99E-08 5.35E-08 6.64E-08 3.15E-08 3.68E-07 4.42E-07 4.13E-07 5.07E-07 4.93E-07 4.42E-07 4.20E-07 4.67E-07 4.64E-07 3.92E-07 4.09E-07 3.61 E-07 3.96E-07 3.92E-07 3.77E-07 3.59E-07 3.74E-07 3.48E-07 3.25E-07 4.43E-07 4.18E-07 3.81 E-07 3.97E-07 4.09E-07 3.83E-07 4.36E-07 4.47E-07 4.56E-07 4.84E-07 8.84E-07 8.57E-07 5.28E-07 5.74E-07 4.41 E-07 3.65E-07 5.01 E-07 4.60E-07 4.56E-07 4.30E-07 4.01 E-07 4.02E-07 6.76E-07 7.1OE-07 4.58E-07 0.072 0.143 0.072 0.143 0.072 0.143 0.072 0.14 0.07 0.14 0.07 0.136 0.068 0.136 0.068 0.14 0.07 0.214 0.107 0.143 0.072 0.143 0.071 0.143 0.072 0.147 0.074 0.147 40 0.074 20.7 40 20.7 20.7 40 0.143 0.072 0.143 0.072 0.143 0.072 0.143 0.072 0.147 20.7 20.7 20.7 20.7 40 40 40 20.7 40 40 40 21.9 21.9 40 40 20.7 40 20.7 40 50 0.074 0.143 0.072 4.09E-07 Lot Wafer 104.1 104.1 16 16 13 13 13 13 13 13 13 13 20 20 20 20 20 20 20 20 20 20 20 20 20 20 9 9 99.1 99.1 99.1 99.1 99.1 99.1 99.1 99.1 99.1 99.1 99.1 99.1 99.1 99.1 99.1 99.1 99.1 99.1 99.1 99.1 99.1 99.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 99.1 99.1 99.1 99.1 99.1 99.1 99.1 Die Meas R. F2L F2L F2 F2 F2 F2 F2 F2 F2 F2 F2 F2 F2 F2 19.6 19.6 10.5 10.5 10.5 10.5 F2 9 9 9 9 17 17 17 17 17 17 17 17 13 13 13 13 13 13 13 Part 10 F2 F2 F2 F2 F2 F2 F2 F2 F2 F2 F2 F2 F2 F2 F2 F2 F2 F2 F2 F2 F2 F2 F2 F2L F2L F2L F2L F2L F2L F2L 10.4 10.4 10.1 10.1 12.1 12.1 12.1 12.1 12.1 12.1 12 12 12 12 11.7 11.7 11.4 11.4 12.4 12.4 11.8 11.8 11.6 11.6 12.5 12.5 17.2 17.2 12.5 12.5 12.2 12.2 25.3 25.3 26.6 26.6 24.1 24.1 25.3 51 e 40 40 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 Vhigh b Vn/V 0.14 0.07 0.105 0.053 0.105 0.053 0.105 0.053 5.52E-08 2.96E-08 3.37E-07 1.79E-07 3.13E-07 1.73E-07 2.84E-08 1.62E-08 3.04E-08 1.66E-08 3.83E-07 1.89E-07 3.72E-07 1.84E-07 3.99E-07 1.81 E-07 3.70E-08 2.17E-08 3.23E-08 1.93E-08 3.23E-08 1.88E-08 3.21 E-08 1.67E-08 4.OOE-08 2.22E-08 3.84E-08 1.74E-08 2.78E-08 1.76E-08 5.05E-08 2.67E-08 9.99E-07 4.41 E-07 3.54E-08 1.91 E-08 3.39E-08 1.89E-08 4.58E-08 2.42E-08 5.16E-08 2.86E-08 4.04E-08 2.11 E-08 4.29E-08 3.81 E-07 3.73E-07 3.21 E-06 3.37E-06 2.98E-06 3.25E-06 2.57E-07 2.56E-07 2.76E-07 2.56E-07 3.44E-06 3.42E-06 3.35E-06 3.34E-06 0.104 0.052 0.111 0.055 0.111 0.055 0.111 0.055 0.11 0.055 0.111 0.055 0.109 0.054 0.109 0.055 0.113 0.056 0.115 0.058 0.108 0.054 0.113 0.056 0.116 0.058 0.114 0.057 0.112 0.056 0.162 0.081 0.163 0.082 0.157 0.079 0.162 3.60E-06 3.29E-06 3.27E-07 3.60E-07 2.76E-07 2.97E-07 2.85E-07 3.05E-07 2.89E-07 2.81 E-07 3.48E-07 3.76E-07 3.30E-07 2.84E-07 2.37E-07 2.55E-07 4.45E-07 4.69E-07 8.61 E-06 7.60E-06 2.92E-07 2.62E-07 2.90E-07 2.89E-07 2.77E-07 2.80E-07 3.14E-07 3.40E-07 2.53E-07 2.47E-07 2.60E-07 Lot 99.1 99.1 99.1 99.1 99.1 99.1 99.1 99.1 99.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 Wafer Die Da r Mfa . R . F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L 25.3 34.6 34.6 36.8 36.8 31.5 31.5 28.9 28.9 12.9 12.9 12 12 12.4 12.4 12.9 12.9 28.3 28.3 28 28 28 28 28 28 28 28 28 28 13.3 13.3 13.3 13.3 13.3 13.3 13.8 13.8 13.3 13.3 11.2 11.2 13.8 13.8 13.3 13.3 52 Dep. VhIgh b VnN 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 60 0.081 0.18 0.09 1.99E-08 7.51 E-08 3.92E-08 7.59E-08 3.85E-08 7.14E-08 3.72E-08 6.96E-08 4.17E-08 4.04E-08 2.40E-08 3.09E-08 1.71 E-08 2.88E-08 1.68E-08 3.43E-08 1.76E-08 8.79E-08 4.32E-08 5.02E-08 2.91 E-08 4.37E-08 2.20E-08 4.80E-08 2.63E-08 4.35E-08 2.60E-08 4.82E-08 2.71 E-08 4.82E-08 2.62E-08 3.29E-08 1.81E-08 3.42E-08 1.62E-08 4.91 E-08 2.55E-08 4.29E-08 1.94E-08 3.25E-08 1.69E-08 3.62E-08 1.75E-08 3.50E-08 1.92E-08 2.22E-07 Time 0.19 0.1 0.175 0.088 0.169 0.085 0.112 0.056 0.109 0.054 0.111 0.055 0.113 0.056 0.169 0.085 0.168 0.084 0.169 0.085 0.169 0.084 0.168 0.085 0.169 0.084 0.114 0.057 0.114 0.057 0.114 0.057 0.116 0.058 0.114 0.057 0.112 0.056 0.116 0.058 0.114 0.057 4.16E-07 4.33E-07 3.85E-07 3.28E-07 4.01 E-07 3.97E-07 4.02E-07 4.57E-07 3.51E-07 3.96E-07 2.69E-07 2.61 E-07 2.46E-07 2.55E-07 2.90E-07 2.58E-07 5.1OE-07 4.65E-07 2.82E-07 2.86E-07 2.47E-07 2.08E-07 2.82E-07 3.08E-07 2.43E-07 2.52E-07 2.79E-07 2.99E-07 4.21 E-07 4.54E-07 2.72E-07 2.51 E-07 2.93E-07 2.52E-07 4.11E-07 3.93E-07 3.69E-07 3.05E-07 2.84E-07 2.79E-07 3.09E-07 2.90E-07 3.01E-07 3.13E-07 Lot Wafer Die Part Meas R. Dep. Time Vhigh b Vn/V 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 17 17 17 17 17 17 17 17 17 17 17 17 17 17 6 6 7 7 7 7 9 9 9 9 10 10 11 11 F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L 13.3 13.3 34.3 34.3 34.3 34.3 33.9 33.9 33.9 33.9 29.6 29.6 10.6 10.6 60 60 60 60 60 60 60 60 60 60 60 60 60 60 0.114 0.057 0.184 0.092 0.184 0.092 0.183 0.092 0.183 0.092 0.175 0.088 0.109 0.054 4.30E-08 2.13E-08 6.73E-08 3.25E-08 7.72E-08 3.72E-08 1.57E-07 8.31 E-08 1.99E-07 8.27E-08 7.51E-08 3.54E-08 3.44E-08 1.98E-08 3.70E-07 3.43E-07 3.64E-07 3.46E-07 4.18E-07 3.98E-07 8.54E-07 8.83E-07 1.08E-06 8.84E-07 4.21E-07 3.69E-07 3.01E-07 3.15E-07 104.1 17 12 F2L 12.8 60 0.113 4.94E-08 4.31E-07 60 60 60 60 60 60 60 60 60 0.056 0.116 0.058 0.116 0.058 0.111 0.055 0.107 0.054 2.55E-08 3.43E-08 1.75E-08 4.24E-08 2.32E-08 3.46E-08 2.14E-08 7.54E-08 3.1OE-08 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 104.1 17 17 17 17 17 17 17 17 17 12 13 13 14 14 15 15 16 16 F2L F2L F2L F2L F2L F2L F2L F2L F2L 12.8 13.8 13.8 13.8 13.8 12.4 12.4 11.6 11.6 53 4.31E-07 2.76E-07 2.14E-07 3.52E-07 3.47E-07 3.1OE-07 3.82E-07 7.03E-07 5.63E-07 54 Appendix B Annealing Experiment Measurements Before Baking Lot 54 54 54 54 54 54 54 54 54 54 54 54 54 54 88 88 88 88 88 88 88 88 88 88 88 Wafer 10 10 10 10 10 10 10 10 10 10 10 10 10 10 21 21 21 21 21 21 21 21 21 21 21 Die 5 5 7 7 8 8 10 10 10 10 13 13 15 15 1 1 5 5 7 7 8 8 10 10 13 Part F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L R 41.4 41.4 37.2 37.2 41.4 41.4 46.6 46.6 41.4 41.4 41.4 41.4 41.4 41.4 41.4 41.4 41.4 41.4 41.4 41.4 41.4 41.4 41.4 41.4 41.4 Vhigh 0.202 0.101 0.192 0.096 0.202 0.102 0.214 0.107 0.202 0.102 0.202 0.101 0.202 0.101 0.202 0.101 0.202 0.101 0.202 0.101 0.202 0.102 0.202 0.101 0.202 55 b 2.35E-07 1.22E-07 2.51E-07 1.22E-07 2.17E-07 1.09E-07 2.27E-07 1.13E-07 2.18E-07 1.05E-07 3.22E-07 1.55E-07 2.26E-07 1.11E-07 2.26E-07 1.11E-07 2.15E-07 1.1OE-07 2.11E-07 1.08E-07 2.32E-07 1.16E-07 2.22E-07 1.09E-07 2.22E-07 VnN 1.16E-06 1.20E-06 1.31E-06 1.26E-06 1.07E-06 1.05E-06 1.06E-06 1.05E-06 1.08E-06 1.03E-06 1.59E-06 1.53E-06 1.12E-06 1.1OE-06 1.12E-06 1.09E-06 1.05E-06 1.06E-06 1.04E-06 1.06E-06 1.15E-06 1.13E-06 1.1OE-06 1.07E-06 1.1OE-06 Lot 88 88 88 106 106 106 106 106 106 106 106 106 106 106 106 106 106 106 106 106 106 106 106 Wafer 21 21 21 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 11 Die 13 15 15 1 1 1 1 5 5 5 5 7 7 8 8 9 9 10 10 13 13 15 15 Part F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L R 41.4 46.6 46.6 28.7 28.7 28.7 28.7 28.7 28.7 28.7 28.7 28.7 28.7 26.6 26.6 28.7 28.7 26.6 26.6 26.6 26.6 28.7 28.7 Vhigh 0.101 0.215 0.107 0.169 0.085 0.169 0.085 0.169 0.084 0.169 0.085 0.168 0.084 0.162 0.081 0.169 0.085 0.162 0.081 0.162 0.081 0.168 0.084 b 1.1OE-07 2.26E-07 1.04E-07 1.63E-07 8.03E-08 1.76E-07 8.39E-08 1.48E-07 7.13E-08 1.81E-07 9.06E-08 3.67E-07 1.54E-07 1.73E-07 8.91E-08 1.85E-07 8.68E-08 1.59E-07 7.87E-08 1.60E-07 8.36E-08 1.70E-07 8.93E-08 Vn/V 1.07E-06 1.05E-06 9.63E-07 9.61E-07 9.31E-07 1.04E-06 9.74E-07 8.71E-07 8.31E-07 1.07E-06 1.05E-06 2.18E-06 1.83E-06 1.07E-06 1.1OE-06 1.09E-06 1.01E-06 9.80E-07 9.65E-07 9.86E-07 1.01E-06 1.01E-06 1.05E-06 1.45E-07 6.78E-08 1.36E-07 6.94E-08 1.35E-07 7.02E-08 1.37E-07 6.89E-08 1.48E-07 7.30E-08 1.43E-07 6.98E-08 1.33E-07 6.57E-08 1.56E-07 8.02E-08 1.53E-07 7.77E-08 1.49E-07 7.1OE-08 1.54E-07 7.65E-08 9.86E-07 9.14E-07 9.19E-07 9.26E-07 9.45E-07 9.69E-07 9.31E-07 9.27E-07 1.OOE-06 9.70E-07 9.37E-07 9.02E-07 9.OOE-07 8.90E-07 9.63E-07 9.89E-07 9.69E-07 9.72E-07 9.47E-07 8.94E-07 9.77E-07 9.61 E-07 Once Baked 54 54 54 54 54 54 54 54 54 54 54 54 54 54 88 88 88 88 88 88 88 88 10 10 10 10 10 10 10 10 10 10 10 10 10 10 21 21 21 21 21 21 21 21 1 1 5 5 7 7 8 8 9 9 10 10 13 13 1 1 5 5 7 7 8 8 F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L 21.9 21.9 21.9 21.9 20.7 20.7 21.9 21.9 21.9 21.9 23.3 23.3 21.9 21.9 26.6 26.6 24.8 24.8 24.8 24.8 24.8 24.8 0.147 0.074 0.148 0.074 0.143 0.072 0.147 0.074 0.147 0.074 0.152 0.076 0.147 0.073 0.162 0.081 0.157 0.079 0.157 0.079 0.157 0.079 56 Lot 88 88 88 88 88 88 106 106 106 106 106 106 106 106 Wafer 21 21 21 21 21 21 11 11 11 11 11 11 11 Die 9 9 10 10 15 15 1 1 5 5 8 8 9 9 10 10 106 106 11 11 11 106 11 106 11 54 54 54 54 10 10 10 10 9 9 10 54 10 10 54 10 54 54 54 10 10 13 13 5 5 7 7 8 8 15 15 1 1 13 13 13 13 Part F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L R 24.8 24.8 24.8 24.8 26.6 26.6 19.6 19.6 20.7 20.7 19.6 19.6 20.7 20.7 Vhigh 0.157 0.079 0.157 0.079 0.162 0.081 0.139 0.07 0.143 0.072 0.139 0.07 0.143 18.6 0.071 0.136 18.6 19.6 0.068 0.139 19.6 0.07 b 1.53E-07 7.47E-08 1.54E-07 8.02E-08 1.54E-07 8.16E-08 1.30E-07 6.53E-08 1.37E-07 6.83E-08 1.32E-07 6.41E-08 1.37E-07 6.81 E-08 1.23E-07 6.28E-08 1.32E-07 6.18E-08 VnN 9.72E-07 9.36E-07 9.81 E-07 1.01 E-06 9.47E-07 9.99E-07 9.34E-07 9.29E-07 9.53E-07 9.42E-07 9.44E-07 9.01 E-07 9.58E-07 9.53E-07 9.02E-07 9.12E-07 9.46E-07 8.72E-07 1.41E-07 7.57E-08 1.40E-07 7.16E-08 8.35E-08 4.23E-08 1.38E-07 7.44E-08 1.38E-07 6.76E-08 1.58E-07 7.31 E-08 1.35E-07 6.79E-08 1.31 E-07 7.27E-08 1.22E-07 6.16E-08 1.24E-07 6.35E-08 1.33E-07 6.91 E-08 1.24E-07 6.74E-08 1.58E-07 7.60E-08 1.50E-07 9.53E-07 1.01E-06 9.42E-07 9.48E-07 8.47E-07 8.43E-07 9.27E-07 9.75E-07 9.25E-07 8.93E-07 1.06E-06 9.83E-07 9.09E-07 9.02E-07 8.82E-07 9.61 E-07 8.69E-07 8.76E-07 9.11E-07 9.23E-07 9.46E-07 9.59E-07 8.87E-07 9.59E-07 Twice Baked 54 54 54 54 54 54 54 106 106 106 106 106 106 106 106 88 88 88 10 10 10 10 10 10 10 10 10 11 11 11 11 11 11 11 11 21 21 21 15 15 9 9 9 9 8 F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L 21.9 21.9 21.9 21.9 9.56 9.56 21.9 21.9 21.9 21.9 21.9 21.9 21.9 21.9 21.9 21.9 19.6 19.6 18.6 18.6 19.6 19.6 19.6 19.6 26.6 26.6 24.8 0.148 0.074 0.148 0.074 0.098 0.049 0.148 0.074 0.148 0.074 0.148 0.074 0.148 0.074 0.148 0.074 0.14 0.07 0.136 0.068 0.14 0.07 0.14 0.07 0.162 0.082 0.157 57 9.71 E-07 9.16E-07 9.49E-07 Lot 88 88 88 88 88 88 88 88 88 88 88 88 88 Wafer 21 Die 8 Part F2L 21 21 21 21 7 7 F2L 21 21 21 5 5 13 13 15 21 21 15 10 21 21 21 10 1 1 F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L F2L R 24.8 24.8 24.8 24.8 24.8 24.8 24.8 24.8 24.8 23.3 23.3 24.8 24.8 Vhigh 0.078 0.157 0.078 0.157 0.078 0.157 0.079 0.157 0.078 0.152 0.076 0.157 0.079 58 b 8.05E-08 1.51 E-07 7.40E-08 1.50E-07 7.87E-08 1.43E-07 7.32E-08 1.47E-07 7.18E-08 1.43E-07 7.23E-08 1.44E-07 7.36E-08 VnN 1.02E-06 9.59E-07 9.30E-07 9.49E-07 9.94E-07 9.13E-07 9.23E-07 9.32E-07 9.03E-07 9.35E-07 9.42E-07 9.15E-07 9.18E-07 Bibliography [1] W. Lentz. Characterization of Noise in Uncooled IR Bolometer Arrays. Massachusetts Institute of Technology, Cambridge MA, 1998. [2] C. D. Motchenbacher and J. A. Connelly. Low-Noise Electronic System Design. J. Wiley and Sons, New York, 1993. [3] R. F. Pierret. Semiconductor Fundamentals. Addison-Wesley Modular Series on Solid State Devices. Addison-Wesley, Reading MA, 1988. [4] S. M. Sze. Physics of Semiconductor Devices, Second Edition. J. Wiley and Sons, New York, 1981. [5] W. L. Wolfe and G. J. Zissis, editors. The Infrared Handbook. Environmental Research Institute of Michigan, Ann Arbor, Michigan, 1985. 59