Design and Measurements of an Electrothermal Filter Using CMOS Technology ^†

Abstract

Electronic circuits and systems often require continuous monitoring of their temperature. For most sensors, voltage is the temperature-sensitive parameter; however, electrothermal filters are one of a few exceptions, for which signal frequency or phase is the measure of temperature. Such filters are an essential part of temperature sensors, based on the measurement of material thermal diffusivity, in which the input signal of the filter is a square wave. However, the phase shift introduced by the filter depends on the signal frequency. Thus, the authors decided to explore this dependence in more detail by measuring filter response to sinusoidal input signals. The investigations presented in this paper were carried out for an electrothermal filter designed and manufactured in an ASIC using 3 µm CMOS technology. The obtained measurement results confirmed the hypothesis that both the gain and the phase shift in the filter strongly depend on the input signal frequency. Accurate data on the thermal impedance of filters is crucial for the optimization of their performance.

1. Introduction

Integrated circuits and microprocessors manufactured using the latest technologies are currently the core of most electronic systems [1,2,3]. Considering that temperature is one of the most important factors influencing electronic system operation and reliability [4], thermal simulations are now an indispensable stage of the design process [5,6]. Moreover, in many applications, temperature monitoring should be carried out also in real time in order to protect systems from malfunction or destruction [7]. Different kinds of sensors can be used to measure circuit temperature, including simple thermistors [8], thermocouples [9], p-n diodes [10], bipolar transistors [11] or more sophisticated sensors such as Proportional-To-Absolute-Temperature (PTAT) sensors [12,13,14]. In all these sensors, voltage is used as the temperature-sensitive parameter.

In contrast, in the electrothermal filter (ETF) considered in this paper, the signal frequency or its phase is the measure of temperature. ETFs can be manufactured using the standard CMOS technologies, which are usually cheaper because they do not require any modifications of technological processes [15,16,17]. An electrothermal filter consists of a heater and a temperature sensor. Usually, the heater is a simple diffusion resistor, whereas the temperature sensor can consist of a thermopile made up of p^+-diffusion/aluminum thermocouples, which do not require any biasing. Increasing the number of thermocouples placed at the same distance from the heater increases the sensor signal magnitude and does not affect the phase shift [18]. Typically, ETFs are used in temperature sensors based on the thermal diffusivity measurement. Then, they are placed in frequency-locked loops of temperature-to-frequency converters. The square wave signal is provided to the ETF heater by a voltage-controlled oscillator, and its temperature-dependent output signal from the thermopile is digitized by a sigma-delta Analog-to-Digital Converter (ADC) [19].

Electrothermal filters and their applications have been developed for over a decade by the research group led by Makinwa from the Delft University of Technology in the Netherlands. Possibly the first practical realization of an ETF manufactured using 0.7 μm CMOS technology was presented in [20]. Then, the original design was optimized for rectangular [19] and circular geometries [21]. The objective of the optimization was to determine the best locations of thermopile hot and cold junctions for the signal frequency of 100 kHz. The circuit layout was further optimized in order to minimize the influence of lithographic process inaccuracies [22]. An overview of existing ETF practical realizations in different technologies is provided in [23]. The authors of [24] presented the final compact low-power version of the ETF-based temperature sensor developed for System-on-Chip thermal monitoring purposes manufactured using 40 nm CMOS technology. Another version of the sensor, discussed in [25], compares the output signals of two ETFs measuring the diffusivity of silicon and silicon oxide, which has lower temperature sensitivity. Owing to this solution, it was possible to eliminate the need for an external frequency reference.

Considering that the ETF heaters in such sensors are driven by a square wave containing multiple spectral components and that the phase shift introduced by the filter depends on the signal frequency, the authors decided to explore this dependence in more detail by measuring the ETF response to sinusoidal input signals. The investigations presented in this paper were carried out for an electrothermal filter designed and manufactured in an ASIC using 3 µm CMOS technology. The following section of this paper introduces the ETF principles of operation and its design. Next, the results of measurements and tests of the manufactured device are presented. Finally, the major conclusions are provided, and the possible future tasks are outlined. This paper follows on from earlier conference papers [26,27] and the journal article presenting the PTAT sensor manufactured with the same ASIC [14].

2. Electrothermal Filter

2.1. Principles of Operation

Electrothermal filters, also known as temperature-to-frequency converters, could be an interesting alternative to other temperature sensors because they are fairly insensitive to technological process parameter spread [23]. The principle of their operation is relatively simple. As already mentioned, they consist of a diffusion resistor heater and a temperature sensor in the form of a thermopile with its respective hot and cold junctions that are equally distant from the heater [21]. A sinusoidal signal of frequency f applied to the heater diffuses through the substrate and reaches the sensor with a certain delay. The phase shift between the heater and sensor signals φ is proportional to the distance from the resistor to the thermopile d. Then, for a substrate with the thermal diffusivity α, one can write the following relation [25]:

$$\phi ~ d \sqrt{\pi f/\alpha }$$

(1)

Considering that the dependence of silicon substrate thermal diffusivity α on the absolute temperature T is known [28], temperature values can be determined using the ETF, since for a given signal frequency, the phase shift depends on the circuit temperature. The heater signal is delayed and attenuated as it diffuses through the substrate, so the ETF can be regarded as a filter with temperature-dependent characteristics, the phase shift in which depends on temperature [25].

2.2. Practical Realization

The ETF investigated in this paper is part of a larger Application-Specific Integrated Circuit (ASIC) designed in full-custom mode and containing several other designs. This integrated circuit was manufactured using apparently obsolete 3 μm CMOS 5 V technology with one polysilicon and two metal layers [29]. Such cheap maskless technologies use direct wafer writing, and they are still maintained and developed because they allow fast prototyping and verification of various design concepts by academic centers or small private companies [30,31,32].

The layout of the entire ASIC is presented in Figure 1a. The dimensions of the silicon die are 2080 μm × 2160 μm. The ASIC was mounted in the ceramic DIL 40 package, shown in Figure 1b, whose lid could be open, rendering possible infrared measurements of circuit temperature. The microscopic photograph of the bonded silicon die taken with the lid removed is visible in Figure 1c. The location of the electrothermal filter, which is the main interest in this paper, is indicated in the figures with yellow boxes.

The designed ETF, the schematics of which are presented in Figure 2a, consists of a diffusion resistor heater, visible in the top left corner of the figure, with a nominal resistance of around 1386 Ω, and a sensor thermopile composed of 20 p^+-diffusion/aluminum thermocouples connected in series. The resistance of each 20 μm × 38 μm resistor in the thermocouples is 48 Ω; hence, the expected total nominal resistance of the thermopile is 960 Ω. These resistor values should render possible the effective heating of the ETF and, at the same time, reasonably high sensitivity of the thermopile. Similar values of the heater and the thermopile resistances were also used in the original designs, e.g., those presented in [18,20], which were previously discussed in the Introduction.

The ETF is in the bottom right corner in the ASIC layout, and it is placed within the 560 μm × 135 μm n-well pictured in Figure 2b. The resistor heater has dimensions of 464 μm × 18 μm, and the sensor is located along both sides of the heater. In order to ensure the same heat diffusion time for each thermocouple, their hot and cold junctions are equally distant from the heater edges, by 18 μm and 42 μm, respectively. Both the heater and the thermopile are directly connected to package pins, so that they can be individually powered and read-out.

2.3. Preliminary Tests

Initially, the resistor heating capabilities were confirmed by infrared measurements. After removing the package lid, the thermographic images of silicon die surface temperature were recorded using the Inframetrics SC 1000 infrared camera equipped with a microscopic lens. The magnified image of the region around the heater recorded for the heating resistor current of 3.6 mA is shown in Figure 2c. The resistor is visible as the horizontal strip of pixels, indicated by the frame, which corresponds to locations of noticeably higher temperature. The precise measurement of resistor temperature was not possible here because the width of the heater was comparable with the lens resolution and the infrared radiation wavelength.

3. Measurement Results

Firstly, the circuit was calibrated in a Binder MKF 115 temperature chamber [33], visible in Figure 3a, within the temperature range of −40 °C ÷ +150 °C with a step of 5 °C. First, the ASIC was placed in the chamber, cooled to −40 °C, and then maintained at this temperature for 20 min. Next, the temperature was ramped up to +150 °C, and temperature measurements were taken during the heating process. Owing to the slow rate of temperature increase of less than 1 °C/min, it was possible to ensure minimal differences in temperature between the surrounding ambient environment and the measured ASIC. Moreover, in order to further minimize this difference, measurements were taken with the package lid removed and the silicon die exposed directly to the environment, as shown in Figure 3b. Removing the lid accelerated the attainment of a thermal steady state after each change in the preset temperature and significantly reduced the measurement time without affecting the measurement results.

The designed integrated circuit was manufactured in a batch consisting of 10 different specimens, but one of them was damaged mechanically. Thus, the above-described calibration procedure was repeated only for the nine remaining samples. The calibration results obtained for these ASICs are presented in Figure 4a using different colors and markers. The heating resistor is denoted in the figure by R_heat, whereas the sensing thermopile of the ETF is denoted by R_sens. The differences in the resistor values among the ASICs result from the technological parameter spread. As can be seen, at a temperature of 20 °C, the heater resistance varies from 1350 Ω to 1420 Ω, and the sensor resistance falls in the range between 940 Ω and 980 Ω; thus, the accuracy of the manufactured resistor values due to the technological parameter spread is around ±3%. Additionally, the measurement results were fitted with the first-order polynomials, demonstrating that both resistance values increase linearly with temperature, and the relative changes are around 12% per 100 K, with the coefficients of determination R² values higher than 0.995. Thus, both devices can be effectively used for sensing the ASIC temperature during its operation. Since the spread of technological parameters is fairly low, the following analyses will be presented for the averaged sensor characteristics.

The next step in the experiments was to practically assess the heating resistor’s ability to increase the ASIC temperature. During these measurements, the voltage across the resistor was gradually increased until the power supply value was reached, and both heater and sensor resistance values were measured. The variation in resistance as a function of dissipated power is presented in Figure 4b. As can be seen, when the power value exceeds around 20 μW, the increase in resistance due to the change in temperature and the self-heating effect is observed. Thus, the resistor can be effectively used as a heating element in the electrothermal filter. However, when this resistor is used as a sensor, its biasing current should be kept sufficiently low.

Then, the dynamic behavior of the ETF was investigated. These measurements were carried out using the Tektronix DPO5054 oscilloscope [34]. First, as shown in Figure 5a, the response to the voltage step excitation was measured by applying the power supply voltage to the resistor heater (yellow curve) and recording the thermopile sensor response (magenta curve). Considering that the amplitude of the thermopile output signal is significantly attenuated, different scales were used on the vertical axis for the measurement channels, i.e., 500 mV/div for the heater and 10 mV/div for the sensor. As can be seen, the thermopile voltage initially tracks the excitation and then exponentially decays over approximately 600 ns, indicating the high-pass behavior of the ETF.

Considering that in different applications the electrothermal filter should operate with sinusoidal signals, the frequency characteristics of the ETF were investigated in detail. During these tests, the input signal, which consisted of the DC component equal to half of the power supply voltage superimposed with a sinusoidal signal of variable frequency, was provided to the heating resistor, and then the voltage response of the sensing thermopile was measured. Figure 5b presents an example of the sensor response (magenta line) compared with the input signal (green line) at a frequency of 1 MHz. The measurement channels have different sensitivity values, identical to those in Figure 5a, so that both sinusoidal curves have comparable height. Moreover, the phase shift between the measured signals is indicated in the figure. As can be seen, the output signal is significantly attenuated and visibly delayed.

This issue was investigated in more detail by measuring the amplitude and frequency characteristics of the ETF. Indeed, the amplitude characteristic of the thermopile sensor, shown in Figure 6a, is frequency dependent, and for low frequencies, the gain is very small. Then, at a frequency of around 30 kHz, the amplitude of the response increases rapidly and eventually saturates at around 20 MHz just over the value of 0.1 V/V, which confirms the earlier observed high-pass behavior of the device.

The phase shift dependence on temperature between the input heating resistor signal and the sensor thermopile output was measured for different frequency values, focusing on the gain transition region, between 500 kHz and 5 MHz. The measurement results obtained for five frequency values are plotted using different markers in Figure 6b. Additionally, these results were fitted with a first-order polynomial, represented by the lines. The linear equations resulting from this fitting and the coefficients of determination R² are also presented in the figure. Moreover, selected data concerning the dependence of gain and sensitivity are provided in

Table 1. Data concerning the dependencies of measured gain and sensitivity on signal frequency.

Frequency [MHz]	Gain [mV/V]	Sensitivity [deg/K]
0.562	5.987	0.0313
1.000	10.692	0.0312
1.778	17.453	0.0309
3.163	26.242	0.0230
5.623	33.210	0.0138
10.000	36.655	0.0025

.

As can be seen, the gain increases with frequency, but then the sensitivity of the phase shift to temperature decreases. The sensitivity characteristics are fairly linear, with a determination coefficient value over 0.9 in the gain transition region between 1 MHz and 3 MHz. The use of higher frequency signals would render the phase shift insensitive to temperature. In contrast, the low gain values at low frequencies might adversely affect the measurement accuracy because of the limited dynamic input range of ADCs used in the measurement equipment.

4. Conclusions

This paper presents an electrothermal filter designed and manufactured entirely using CMOS technology. Such circuits might constitute an interesting alternative to other types of temperature sensors, such as thermocouples or ring oscillators, because they are inherently independent from the range of manufacturing process parameters. The presented measurement results prove that the manufactured devices operate properly, and, after calibration, they could be employed for circuit temperature monitoring purposes. They can also be used as phase references for thermal diffusivity sensors.

The circuit operation has been tested for various input signal frequencies in a wide range of temperature values. The main finding of the paper is that both the filter gain and the phase shift depend on temperature. Thus, in order to optimize ETF performance, the thermal impedance of the filter should be accurately determined; however, its value is difficult calculate analytically. The ETF layout optimization method presented in [21] and [35] was carried out using simplified analytical models based on thermal impedance theory, assuming that the heat generated by the heater diffuses in a semi-infinite medium at the source frequency of 100 kHz.

Further research should focus on determining the actual thermal impedance of the ETF. According to the principles of the Network Identification by Deconvolution (NID) method [36], this could be realized by recording ETF thermal responses to power unit step functions. Then, the recorded curves can be further processed by computing the spectral density of thermal resistances and the Bode plots of thermal impedances. The main goal of these analyses would be to determine the optimal input signal frequency, ensuring the maximal temperature sensitivity of the phase shift.