Simultaneous Submicron Temperature Mapping of Substrate and Channel in P-GaN/AlGaN/GaN HEMTs Using Raman Thermometry

Abstract

In this study, we introduce a high-resolution, high-speed thermal imaging technique using Raman spectroscopy to simultaneously measure the temperature of a substrate and a channel. By modifying the Raman spectrometer, we achieved a measurement speed faster than commercial spectrometers. This system demonstrated a sub-micron spatial resolution and the ability to measure the temperatures of the Si substrate and GaN channel simultaneously. During high-current operation, we observed significant self-heating in the GaN channel, with hotspots 100 °C higher than the surroundings, while the Si substrate showed an even temperature distribution. The ability to detect hotspots can help secure the reliability of devices through early failure analysis and can also be used for improvement research to reduce hotspots. These findings highlight the potential of this technique for early defect inspection and device improvement research. This study provides a novel and effective method for measuring the sub-micron resolution temperature distribution in devices, which can be applied to various semiconductor devices, including SiC-based power devices.

1. Introduction

Power devices are used widely in trains, aircraft, and power plants that require high power control. Additionally, their usability has recently increased with the commercialization and popularization of electric vehicles [1,2,3,4]. Until now, power devices have been dominated by Si-based devices such as IGBTs and MOSFETs [3,4]. However, existing Si-based power devices are showing limitations due to several issues [3,4,5,6]. Si cannot be free from the degradation of operational stability caused by hot electrons generated at high temperatures due to its low bandgap of 1.12 eV. Furthermore, limitations were encountered due to the relatively small bandgap for high voltage operation and insufficient electron mobility for high current operation. There have been efforts to bypass these limitations through research on the structure and operation method of Si power devices. Although some minor improvements have been made, it is not easy to overcome the material’s fundamental limitations. GaN/AlGaN High Electron Mobility Transistors (HEMTs) have been widely studied in recognition of their potential to overcome the fundamental limitations of Si with its larger bandgap and higher mobility [1,2,3,4,5,6].

GaN is a material with a wide band gap of 3.4 eV and has a much higher temperature and higher voltage stability than Si. Additionally, it has high electron mobility due to the two-dimensional electron gas (2DEG) formed at the interface between AlGaN and GaN, enabling high current operation. It has a high efficiency in high-frequency operation due to its high switching speed [1,2,3,4,5,6]. It has already been commercialized as a high frequency RF device.

GaN/AlGaN HEMTs are typically grown and manufactured on substrates such as sapphire, Si, GaN, and SiC [7,8,9,10,11,12]. SiC boasts an excellent thermal conductivity of 300–490 W/mK and superior operating characteristics, but its high substrate price is a significant drawback [7,10,12,13,14]. Sapphire substrates have been extensively studied due to their use in GaN-based LED production, and the growth technology of GaN on sapphire is well established. However, sapphire’s low thermal conductivity of ~30 W/mK hampers its ability to dissipate the heat generated by self-heating in HEMTs, leading to poor performance [10,11,12]. This is a critical disadvantage for power devices operating at high voltages and currents. GaN substrates offer advantages such as being stress-free and having a low defect density, making them ideal for GaN epitaxial materials. However, their high cost, low productivity, and lower thermal conductivity (~220 W/mK) compared with SiC limit their widespread use [15,16]. Despite its relatively low thermal conductivity (~130 W/mK) and poorer crystal quality than SiC substrates, Si has recently become the most widely used substrate for GaN/AlGaN HEMTs—particularly in power electronics—owing to its low cost and its higher thermal conductivity compared with sapphire [8,9,10,12].

Nevertheless, several issues remain critical to the efficiency, lifetime, and reliability required for commercial power devices. Performance degradation, reduced lifetime, and sudden failure due to self-heating are still commonly reported [17,18,19,20]. Therefore, research on minimizing the self-heating effect and improving heat dissipation in GaN/AlGaN HEMTs grown on Si substrates is essential for enhancing the performance, lifetime, and reliability of these devices.

To study the self-heating effect and heat dissipation in power devices, it is very important to measure the distribution of heat generated in the device. The channel length of power devices is gradually decreasing to achieve performance improvements and the high integration of the device, and the current commercialized GaN/AlGaN HEMT power devices have a channel length of several tens of μm. Due to the shortened channel length, a spatial resolution of less than 1 μm is required to measure the heat distribution. In addition, according to the results of several simulation studies, a very localized hotspot of 1~2 μm size, which is significantly higher in temperature than the surrounding area around the gate electrode in HEMTs, can be formed, and this can cause device failure [21,22,23,24,25,26,27]. Also, to find hotspots, a spatial resolution of less than 1 μm is also required.

The methods for measuring the heat generated from devices include infrared (IR) thermography, thermal reflectance measurement, gate resistance thermometry, and Raman measurement. Gate resistance thermometry provides an indirect estimation of temperature through variations in electrical characteristics. However, it suffers from several limitations, such as inaccurate absolute temperature measurements and the inability to resolve spatial temperature distributions. IR thermography is a method for measuring the temperature distribution using an infrared image of several tens of μm by utilizing the temperature-dependent characteristics of the wavelength and intensity of blackbody radiation generated from a material [28,29,30,31,32]. This method has the advantage of being fast because it images the temperature of a large area using a photodetector pixel array, but since it uses a wavelength of several tens of μm, it has a spatial resolution limit of 15 to 20 μm mainly due to the diffraction limit [33,34]. Therefore, it is not suitable for measuring the submicron heat distribution of an HEMT. Additionally, there are limitations in measuring the temperature of GaN because GaN is transparent to IR wavelengths [34].

Thermal reflectance measurements use the characteristic that reflectance changes with temperature [35,36]. However, since the change in reflectance due to temperature is very small (~10⁻⁴), it can only be measured through heat modulation, repeated measurements, and signal amplification using a lock-in amplifier [35,36,37,38]. This measurement method faces challenges in accurately determining absolute temperature because reflectance can be influenced by various physical factors such as electric fields, strain, and others, necessitating a rigorous calibration process [39,40]. Additionally, only metals can have a valid heat reflection coefficient, so measurements can only be made on electrodes.

The temperature measurement method using Raman spectroscopy is a method that measures temperature by changing the Raman spectrum according to temperature [41,42,43,44,45]. Since Raman spectroscopy usually uses a visible light laser, it basically has a spatial resolution of less than 1 μm. There are three representative methods: measuring the energy shift in a specific Raman peak, measuring the FWHM, and using the intensity ratio of the Stokes signal and the anti-Stokes signal [41,42,43,44,45,46,47,48]. The methods for measuring the energy shift and FWHM of the Raman peak use peak characteristics that change according to temperature, and they require a high spectral resolution. In order to achieve a high spectral resolution with Raman measurements, a grating with a high number of grooves must be used, which inevitably reduces the intensity of the Raman signal and therefore, increases the measurement time. In addition, peak energy and FWHM do not depend only on temperature, but on other factors such as electric field, strain, and composition fluctuation.

On the other hand, the method using the ratio of the intensity of the Stokes signal and the anti-Stokes signal uses the principle that the probability of anti-Stokes scattering is affected by temperature [42,43,46,47,48]. The Stokes signal only serves as a reference for measuring the intensity of the anti-Stokes signal. The intensity of the anti-Stokes signal depends only on temperature since it is directly proportional to the number of phonons, and therefore, it is a method of measuring temperature more directly than the previous methods. However, it is difficult to use an optical system, for example, a long-pass filter to block elastically scattered light, to measure the Stokes and anti-Stokes signals simultaneously. A Raman measurement system that can measure a relatively wide spectral range simultaneously is required. Fortunately, this method does not require an ultra-high spectral resolution because the effect of the peak shift due to temperature is relatively small. Since the anti-Stokes scattering is about eight times smaller than the already-small Stokes Raman signal at room temperature, it takes a long time to map a whole area. The long measuring time is a very big disadvantage for mapping. For example, if it takes about 1 s to measure the anti-Stokes signal of GaN using a commercial spectrometer, it takes about 11 h to map a 100 × 100 μm² area at 0.5 μm spatial resolution. This long measurement time can cause changes in the device characteristics during the measurement time, and this makes the method impractical, and significantly reduces the reliability of measurements as well.

Raman measurements typically employ lasers with photon energies lower than the bandgap of the material. When performing Raman spectroscopy on GaN in Si-based HEMTs, signals from both the GaN/AlGaN channel and the Si substrate can be simultaneously detected, as the GaN layer is transparent to the excitation source. Since each signal contains temperature information specific to its respective material, the temperatures of the channel and the substrate can be measured independently. However, when the GaN layer is thick, the channel temperature may be underestimated, as the measurement reflects a volumetric average over the entire GaN layer. However, the advantage of being able to measure the temperature of the channel and the substrate simultaneously is obvious, which is difficult to achieve with IR thermography or thermal reflectance measurements. Of course, this advantage is only valid for devices with a transparent top layer, and measurements cannot be performed when the top layer is obscured by opaque materials such as electrodes.

In this study, we simultaneously measured the submicron temperature distributions of the substrate and channel in GaN on Si-based GaN/AlGaN HEMTs during operation by measuring the temperature using the intensity ratio of the Stokes and anti-Stokes signals. To overcome the long measurement time, we constructed a highly sensitive spectral measurement system at the expense of slightly sacrificing the spectral resolution (to the extent that GaN and Si can be distinguished), and as a result, we achieved a speed of about 50 ms to obtain sufficient anti-Stokes signals of GaN, which is 20 times faster than the conventional method. Therefore, it took only about 35 min to map a 100 × 100 μm² area at 0.5 μm intervals. This method can be applied not only to GaN on Si-based power devices but also to GaN/AlGaN HEMTs fabricated on other substrates. It is expected that it can be applied to other semiconductor devices such as SiC-based power devices through optimization, and it presents an effective and novel method to measure the submicron resolution temperature distribution of the device.

2. Materials and Methods

2.1. Modified Raman Spectroscopy Setup

Raman spectroscopy is a method for measuring the Raman scattering spectrum, which involves the inelastic scattering occurring in a material. Raman scattering is a phenomenon in which an incident electromagnetic wave loses or gains energy while generating or absorbing phonons within a material [49]. Scattering that loses energy while generating phonons is called Stokes scattering, whereas scattering that gains energy while absorbing phonons is called anti-Stokes scattering. Since the energy of phonons is influenced by factors such as the type of material, bonding structure, stress, and temperature, the analysis of the Raman scattering spectrum provides valuable information about these characteristics [41,42,43,44,45,46,47,48,49,50,51,52,53,54,55].

Because Raman scattering requires interaction between electromagnetic waves and phonons, its probability of occurrence is very low—approximately 1 in 100,000 photons. This probability is significantly smaller compared with other elastic interactions such as Rayleigh scattering or reflection. As a result, elastically scattered light must be filtered out, and a highly sensitive measuring system is required.

Stokes scattering occurs when phonons are generated, making it largely independent of the number of phonons already present in the material. In contrast, anti-Stokes scattering occurs when existing phonons are absorbed, meaning that higher temperatures—where more phonons are already present—lead to an increased probability of occurrence. Consequently, the temperature of a material can be determined by measuring the intensity of anti-Stokes scattered light. However, a challenge arises in that the intensity fluctuates during the Raman measurement process. To address this issue, a widely known method involves calculating the temperature using the ratio of Stokes scattering intensity to anti-Stokes scattering intensity [42,43,46,47,48]. The relationship between this intensity ratio and temperature can be expressed as Equation (1) [42,43].

$${\displaystyle \frac{I_s}{I_a}}={\left({\displaystyle \frac{\mu +\nu }{\mu -\nu }}\right)}^4\left[\mathit{exp}\left({\displaystyle \frac{\hslash \omega }{k_{B}T}}\right)-1\right]$$

(1)

where I_s and I_a represent the intensities of Stokes scattering and anti-Stokes scattering, respectively; μ and ν represent the wavenumbers of the excitation light and Raman scattering, respectively. ℏ, ω, and k_B are Planck’s constant, Raman scattering wavenumber, and Boltzmann’s constant, respectively.

Generally, the calculated temperature is hardly affected by peak shift, because the difference in μ and ν is very small. Therefore, the spectroscopic system for anti-Stokes Raman thermography does not require a high spectral resolution. However, in order to measure Stokes scattering and anti-Stokes scattering simultaneously, a wide spectral range must be measured, and a very sensitive measuring system is required to measure anti-Stokes scattering, which is about eight times smaller than Stokes scattering at room temperature. This has a significant impact on the mapping time for Raman thermography.

A typical Raman spectrometer consists of optical fibers, a slit, a reflective grating, and spherical mirrors. While using a reflective grating and slit provides good resolution, it also results in significant signal loss. Additionally, commercial spectrometers receive signals through binning, meaning the vertical focus of the Raman signal on the CCD or CMOS detector is not a primary consideration. In fact, the vertical focus size of a commercial Raman spectrometer has been observed to exceed 300 μm. However, to enhance the speed of Raman mapping, it is crucial to focus the Raman signal onto the narrowest possible area. Doing so reduces the number of pixels required by the detector, improving the measurement speed and increasing the sensitivity, as the signal threshold of each pixel is more easily exceeded. To implement this, a measurement setup incorporating a transparent diffraction grating and a cylindrical lens was constructed, as illustrated in Figure 1.

To verify the effectiveness of this measurement system, the Raman spectra of Si and GaN on Si substrates were measured. Figure 2 presents the Raman spectrum of GaN on Si. In this spectrum, the Raman peaks of GaN and Si are clearly separated, enabling the simultaneous measurement of temperature distributions in both materials. Additionally, the time required to distinguishably measure the anti-Stokes signal of GaN was only 50 ms at 150 mW laser power. Consequently, the mapping speed for Raman thermography in GaN on Si-based HEMTs is remarkably fast, reaching up to 50 ms per point.

In order to verify the spatial resolution of the Raman thermography measuring system configured by the above method, a reference pattern sample for verifying the spatial resolution was produced. The sample was manufactured using p-doped Silicon on Insulator (SOI). As shown in Figure 3, it has a structure in which one ground electrode and electrodes are attached to the 1, 2, 3, and 4 μm wire patterns, respectively. The Si layer is 4 μm thick, and a bridge structure connecting equal-width lines was introduced to prevent pattern collapse. In addition, to suppress collapse caused by the undercutting of the Si patterns, a few hundred nanometers of Si were intentionally left unetched.

When a voltage of 150 V is applied to a 1 μm line width pattern using this sample, the temperature increases due to Joule heating, and the temperature distribution is confirmed using Raman thermography. Figure 4a shows the image of the SOI reference pattern viewed through an optical microscope and the result of the Raman thermography overlay. Interestingly, a temperature of about 200 °C is measured even for the 2–4 μm pattern without applied voltage. This is likely due to the effect of leakage current from the less-etched p-doped Si in the bridge structure designed to prevent pattern collapse, as shown in Figure 4b. The two images coincide in a local area, while the temperature difference is clear with and without the pattern, and even the 1 μm pattern can be identified. This demonstrates that the measurement system is capable of taking thermal images at a sub-micro spatial resolution.

2.2. Sample

The sample used in this study is an unpackaged p-GaN/AlGaN/GaN on Si-based HEMT bare chip. The simplified structure of the sample is shown in Figure 5. The GaN buffer that has a thickness of 4 μm and p-GaN/AlGaN/GaN heterostructures are grown on a Si substrate. A 2DEG is formed at the AlGaN and GaN interface due to the difference in band gap between the two materials [56,57,58,59,60]. However, since the 2DEG has high electron mobility even when no gate voltage is applied, the HEMT operates normally on. For normally off operation, heavily p-doped GaN is grown on top of AlGaN/GaN layers, which is etched away except for the gate region. As a result, the area under the gate electrode is depleted by the p-GaN, preventing the formation of the 2DEG unless a voltage is applied. When a forward voltage is applied to the gate, the 2DEG is formed, and the channel is connected. This structure and operation method are called e-MODE and are generally used in normally off GaN HEMT devices [61,62,63].

The field plate is connected to the source electrode. It provides a stable electric field to the channel, increases electrical stability, and has a heat dissipation effect [64,65,66]. The overall structure of the sample features a comb-shaped electrode design enabling high-current operation and power efficiency [67,68].

The sample is bonded with Al wire on the PCB substrate. The source and drain electrodes are connected to the PCB with 10 wire bonds each for high-voltage applications, while the gate electrode is connected with a single wire since it does not require a high voltage.

3. Results

In this study, the temperature distribution during the operation of an AlGaN/GaN HEMT was measured using the Raman thermography measurement system introduced above. Two measurements were performed to compare low-current and high-current conditions. The gate voltage was applied at the same 6 V, and the source-to-drain voltage was applied so that the drain current flowed at 50 mA and 1.2 A, respectively. The mapping speed was 50 ms/point, and the mapping interval was 0.5 μm to measure the area between the gate and drain. The time taken for measurements was about 40 s.

Figure 6a,b show the temperature maps of the Si substrate and GaN channel, respectively, measured under low-current conditions. For the Si substrate, the average temperature was measured to be about 80 °C, which is higher than the average temperature of the GaN channel at 18 °C. This is because the temperature of the Si was measured to be high due to laser absorption. This phenomenon occurs because Si absorbs a 532 nm laser at 1.12 eV, while GaN does not absorb a 532 nm laser at 3.4 eV. As can be seen from these results, there was almost no self-heating effect due to the current under low-current conditions.

The increase in Si temperature due to laser heating raises concerns about the accuracy of the temperature measurements. One method to mitigate this effect is to calibrate the temperature while accounting for laser-induced heating; however, this approach is cumbersome, as it requires material-specific calibration. Alternatively, reducing the laser power can minimize the heating effect. To determine the appropriate laser power level, we measured the temperature variation in Si and GaN as a function of laser power, as shown in Figure 7. The temperature of GaN remained relatively stable across the power range, whereas Si exhibited a linear increase in temperature with increasing laser power. A very low error was observed in the linear fitting. The fitting parameters and associated errors are provided in Figure 7, and repeated measurements yielded virtually identical results. To sufficiently suppress laser-induced heating, the laser power must be reduced below 5 mW. However, this significantly increases the measurement time by approximately 30 times compared with using a 150 mW laser. Furthermore, the anti-Stokes peak of GaN was not observed at power levels below 40 mW, indicating the need for further investigation to overcome these limitations.

Figure 8a,b show the temperature maps of the Si substrate and GaN channel, respectively, measured under high-current conditions. The average temperature of the GaN channel was measured to be 350 °C, which is higher than the Si substrate’s temperature of 250 °C. This indicates that the temperature increase is due to the self-heating effect caused by the current, rather than the heat generation from the laser. Additionally, while the Si substrate exhibited a relatively even temperature distribution, a hotspot measuring 1–3 μm in size and 100 °C higher than the surroundings was observed in the GaN channel. The presence of such hotspots has been reported in previous studies, and they are known to significantly impact the performance, lifetime, and reliability of the device [22,24,26]. The fact that these hotspots are not observed on the Si substrate indicates that the heat generated at the hotspots is not sufficiently dissipated through the Si substrate. This result is a new finding that has not been reported in previous studies and was discovered because the temperatures of the Si substrate and the GaN channel were measured simultaneously. These results suggest that the hotspots may have been caused by micro-sized cracks or voids between the Si substrate and the GaN channel. It is possible that the hotspots formed because the heat generated in the channel was unable to escape to the substrate through micro-cracks or voids with very low thermal conductivity. Further research is needed to clarify this. This hotspot detection capability can help secure the reliability of devices through early failure analysis and can also be used for improvement research to reduce hotspots. If the above measurement results are measurement artifacts, similar phenomena should be observed in Figure 6 measured at low-current conditions, but this phenomenon is not observed in Figure 6 because the temperature deviation is about 20 degrees. This is strong evidence that the measured hotspots are real and not artifacts.

4. Conclusions

In this study, we introduce a sub-micron high-resolution, high-speed thermal imaging technique capable of simultaneously measuring the temperature of a substrate and a channel. To simultaneously measure the temperature of the substrate and the channel and to achieve a high resolution, we used Raman spectroscopy using visible light. To overcome the slow speed of Raman spectroscopy, we designed a modified Raman spectrometer for fast measurement at the expense of spectral resolution. Instead of the commonly used beam splitter to minimize optical loss, a 42° tilted bandpass filter was used. Additionally, pin hole and aperture were not used. In addition, a cylindrical lens with a short focal length was used in front of the CMOS detector for fast speed and a high signal-to-noise ratio. Through this, we built a measurement system with a measurement speed approximately 20 times faster than that of a commercial Raman spectrometer.

Through measurements of SOI reference patterns, we confirmed that patterns with 1 μm spacing can be clearly distinguished and measured, confirming that this measuring system has a sub-micron spatial resolution. In addition, we confirmed that the Raman signals of the Si substrate and GaN channel were separated, confirming that the temperatures of the substrate and channel can be measured simultaneously.

Additionally, the temperature distribution during the operation of an AlGaN/GaN on Si HEMT was measured in low- and high-current conditions. As a result, it was confirmed that there was almost no self-heating due to current in the low current situation, and it was confirmed that the temperature increased by self-heating to 250 °C for the Si substrate and 350 °C for GaN at the high current. In addition, hotspots measuring 1–3 μm in size were found in the GaN channel, which were 100 °C higher than the surroundings, while these were not found in the Si substrate. The fact that hotspots were not observed on the substrate may suggest that hotspots were formed when the heat generated in the channel was prevented from escaping to the substrate by micro voids or cracks.

Hotspots are reported to be a major factor in the deterioration of the performance and lifetime of devices and in the deterioration of reliability due to sudden failures, and directly measuring them can be of great help in early defect inspection and device improvement research. In addition, the fact that these hotspots are not visible on the substrate is a new result that has not been reported before, and it is expected to play an important role in identifying the cause of hotspots and in improving and resolving them.