1. Introduction

Recently, the pursuit of high contrast, high pixel resolution, and fast response speed displays has shifted from LCDs (liquid-crystal displays) to micro LED (light-emitting diode) displays [1–4]. The self-luminous LED sub-pixels can be transferred and mounted on various substrates [5,6], making them an ideal technology for widespan applications, including virtual reality (VR) [7,4], augmented reality (AR) [8,9], and commercial billboards [10]. Micro-LED displays are now shifting from research [11–14] to commercial demonstration [15]. However, manufacturing micro-LEDs for displays encounters several obstacles. For example, luminescence efficiency degradation at micron-scale chip size will seriously impede power consumption [16–19]. Mass transfer of many chips from wafers to templates inevitably incurs LED die loss or misplacement that leads to defects of the panel [20]. The electrical and optical properties of the mounted LEDs are a critical issue but are rarely mentioned. The micron-scale chips are too small and time-consuming for direct probing. Going forward, a chip size smaller than 4 × 4 µm² for AR applications is even more challenging [21,22]. On top of extracting light output from probing, each chip's radiation field must be examined to meet the optics design on the display substrate to achieve uniform light output [23].

The general strategy of micro-LED display manufacturers to avoid defects in the panel is to improve the LED wafer yield as high as possible and to have redundant LEDs for each sub-pixel [24]. Considering an LED wafer production yield shy of 100%, redundant LEDs have to be unavoidably used with the drawback of increasing the cost. Furthermore, before mounting the chips to the display panel, techniques such as photoluminescence (PL) and electroluminescence (EL) are employed to pre-screen the quality of dies. PL has the benefit of being non-contact, which protects LED chips from being damaged during detection. However, the electrical properties of each LED are not recorded, and the actual EL light output isn't properly monitored.

In order to address the issue of inspecting micro-LED performance, various mass detection methods have been proposed [25–27]. LiLi et al. used a two-dimensional detecting system with an integrated camera to correlate microscopic LED surface luminescence with the measured light output intensity [25]. The contact electrodes of a single current-injected LED are all fabricated on the wafer. In addition, to prevent probing damage, capacitive current injection (C2I), which is the non-contact EL method, was proposed to be integrated into the camera detection system [26]. The detection capacitor is close to the p-electrode, and the displacement current was generated. An integrated camera captures the LED luminescence. The quality of the LED is judged visually from the photograph. Furthermore, using deep learning algorithms for image recognition has gained significant attention in various applications. Yufend. et al. [27] presented an inspection method that combined LED chip radiation fields with a convolutional neural network (CNN) model to detect surface defects in LED appearance.

In this study, we propose a mass detection method to examine the luminescence and radiation profile of the microscale LED array. Luminous profiles are extracted from the 13 × 27 LED array sharing the p-type contact from an external ITO (indium-tin oxide) glass and on-wafer n-type ground pads. Light output intensity from each LED is determined following the multi-variable regression analysis to rule out the influence of transmission resistance between the probe and device electrode. To justify the proper radiation profile of each LED subpixel, we developed 2-D convolutional neural network (CNN) models. The proposed methods provide a practical methodology to massively detect micron-scale LEDs’ electrical and optical properties.

2. Materials, fabrication, measurement, and deep learning models

2.1 Device structure and fabrication of the micro-LED array

The micro-LED arrays were fabricated on a 4-inch wafer. They were diced into 1 cm x 1 cm in sample sizes for the mass detection experiment in this work. The epi-structure was grown on the patterned sapphire substrate (PSS) by metal-organic chemical vapor deposition (MOCVD). The epi-layers consist of a 5 µm undoped GaN layer, a 2 µm-thick Si-doped n-type GaN layer, 10 periods of InGaN/GaN multiple quantum wells (MQW), and a 250-nm-thick Mg-doped p-GaN layer. The device fabrication starts from a 1 µm-deep isolation etching in the periphery of the LED array using inductively coupled plasma-reactive ion etching (ICP-RIE), followed by a mesa definition of each device. N-type contact pads, Ti/Al/Ni/Au (25/125/50/125 nm), were then evaporated and alloyed at 900 °C for 30 s by rapid thermal annealing (RTA). Next, a thin Ni/Au (5/5 nm) metal stack was deposited on the p-GaN surface and alloyed at 550 °C for 15 mins to form p-type ohmic contact. In order to protect the device and suppress leakage current from sidewall defects, we then deposited a 100-nm-thick Si₃N₄ dielectric layer by plasma-enhanced chemical vapor deposition (PECVD). Finally, Ti/Au (30 /800 nm) probe pads were evaporated after VIA holes were opened by reactive ion etching (RIE). The microscopic image of the LEDs is shown in Fig. 1(a), along with a schematic device structure in Fig. 1(b). The micro-LED array is arranged with 13 × 27 devices and is 38 µm and 52 µm away from each other at y-and x- direction, respectively. The mesa size of each micro-LED is 13 × 20 µm². The LED array was prepared by connecting n-type contact pads of all micro-LEDs.

 

Fig. 1. Micro-LED characteristic and measurement setup. (a) Top view microscopic image of the whole micro-LED array. (b) The cross-sectional schematic structure. (c)The mass detection measurement setup from ITO-glass attachment to the whole measurement setup. (d) Individual measurement setup.

Download Full Size | PDF

2.2 Measurement setup

Figure 1(c) and Fig. 1(d) illustrate the measurement setup of mass detection and individual LED testing. The sample is first mounted on the glass substrate. As shown in Fig. 1(c), an indium tin oxide (ITO) coated glass substrate with the dimension of 1.5 × 1.5 cm² is flipped and attached to the LED array with proper pressure. During the measurement, a Cu pad connected with a wire is attached to the ITO glass near the 20^th column of the LED array. And the n-type contact pad, the rectangular yellow pad in Fig. 1(c) and (d) not covered by the ITO glass, is attached by a probe. An electrical current source (Agilent 4155C semiconductor parameter analyzer) is used to inject current. The luminescent image was recorded by a Silicon CMOS high resolution camera (SP932U Beam Profiling Camera). The camera parameters and photography are controlled by the BeamMic software.

As for individual device testing, a probe is attached to the p-metal contact of each device (see Fig. 1(d)). Agilent 4155C semiconductor parameter analyzer is used to provide the bias current. A silicon CMOS high-resolution camera is used to extract light output.

2.3 2-D CNN-based training models of mass detection

The fundamental component of a CNN is a multi-tier network that comprises convolution, pooling, and fully connected layers. These layers work together to extract features such as lines and edges. In this work, we utilize three models: VGG16 [28], Deep Residual Network ResNet-18, and ResNet-50 [29], with two class outputs for image recognition. The learning model architectures are shown in Fig. 2. The accuracy, precision, and recall of three models based on the input of micro-LED’s luminescent images will be computed. Since there is a significant disparity between the good and destructive examples in the training data, the F-score calculation is used for more accurate results.

 

Fig. 2. The learning model architectures of (a) VGG16, (b) Resnet-18, and (c) Resnet-50.

Download Full Size | PDF

3. Methods and results

3.1 Methodology of mass detection for micro-LED arrays

Since each micro-LED pixel cannot be tested due to the probe size constraint, we propose a mass detection strategy to connect all the devices under test with the bias voltage applied on the p-type and n-type connectors. Such an approach inevitably incurs serial transmission resistance from the bond pad to the LED at various locations. Hence, the methodology of mass detection was developed to calibrate the luminescence by considering the resistance from the bond pads to the individual device. Figure 3 shows the flow of mass detection of each LED in an array. It is divided into two steps to examine the luminescence quality and radiation profile. The micrographic luminescence of the LED array was taken by applying the bias to the ITO and n-contact. Light exposure parameters of the camera are fixed among different photography runs to avoid variations in luminescence imaging. Because the contact resistance between the probe pad and the LED is location-dependent, a calibration method was developed to compensate for the resistance variations incurred from the detection. The calibration strategy starts by developing a circuit model based on the computer-aided design (CAD) tool, LTspice, to estimate the current and luminescence behaviors of the individual device in the LED array. Once the simulated light output from each LED, after considering the transmission resistance, is determined, multi-variable regression analysis is employed to establish the relationship between the actual luminescence of the individual device (dependent variable) and the location (independent variables of the x- and y-axis location). Through the multi-variable regression analysis, the actual luminescence can be reconstructed.

 

Fig. 3. Methodology of mass detection measurement. It consists of steps to determine micro-LED’s luminescence and quality of the radiation profile.

Download Full Size | PDF

As for examining the radiation profile of the LEDs, we proposed two recognizing methods. The first separates LEDs into functioning and defeated categories, and a network model was trained for recognition. The second divides an LED luminescent image into four quarters. The LED is considered as functioning only when at least three-quarters of the LED image passes the test. Three different learning models were used for training, namely VGG16, Resnet-18, and Resnet-50. The evaluation is based on accuracy, precision, and recall. Results from the luminescence and radiation profiles were used to determine the quality of each micro-LED for displays.

3.2 Micro-LED array image of mass detection

The LED luminescent image taken from the mass detection technique is shown in Fig. 4(a). The micro-LED array is biased at 5.5 V with a total injection current of 45.63 mA. The luminescent intensity profile can be extracted and numerically recorded. For example, luminescence along the 1^st row and 1^st column is demonstrated in Fig. 4(b) and 4(c), respectively. Though the luminescent image can be recorded for 351 devices at a time, the results in Fig. 4 could not reflect the actual light output because of the serial resistance from the probe (bond) pad to the individual device. A multi-variable regression is then employed to de-embed the serial resistance next.

 

Fig. 4. (a) The 13 × 27 micro-LED array image taken from a camera (we intentionally selected an image with several defective LEDs). The color bar in the right indicates relative luminescence (b) The luminescence distribution along the first row of the image in (a). The unit of x-axis is expressed by camera pixel. (c) The luminescence distribution along the first column in (a)

Download Full Size | PDF

3.3 Electrical and optical modeling of mass detection based on CAD simulation

In order to obtain the luminescence distribution of the micro-LED array under the test condition of location-dependent resistance variations, we use LTspice to simulate the luminescent distribution in the LED array. The parameters for modeling I-V (current-voltage) and L-I (luminescence-current) curves of the individual device are first extracted by fitting to the measurement results.

The electrical characteristic of an LED in the CAD model is defined following the equation,

R_s, I_sat, and n are the diode's series resistance, saturation current, and ideality factor. v is the voltage applied across the diode in volts, e is the charge of the electron (1.6 × 10⁻¹⁹ coulombs), k is the Boltzmann’s constant (1.38 × 10⁻²³), and T is the junction temperature in Kelvins. Curve fitting to the measured diode I-V relationship is shown in Fig. 5(a), with the parameters R_s, I_sat, and n extracted to be 214.1 Ω, 1.386 × 10-8 A, and 24.176, respectively. The parameters are then plugged into the LTspice model.

 

Fig. 5. Simulation of current and light output of the LED array using the LTspice. (a) Measured I-V curve (solid line) and the corresponding fitting (dashed line) of a typical single LED following Eq. (1). (b) Measured L-I curve and the corresponding fitting of a typical single LED following Eq. (2). (c) Equivalent circuit of the micro-LED array for LTspice simulation. The array consists of 13 × 27 micro-LED elements. (d) (Dots) Simulated luminescence distribution profile in the first row of the array and (red line) the corresponding curve fitting following Eq. (3). (e) (Dots) Simulated luminescence distribution profile in the first column of the array and (red line) the corresponding curve fitting following Eq. (4).

Download Full Size | PDF

Similarly, the luminescence-injection current (L-I) correlation in the LTspice was obtained by fitting the measured L-J curve, which is shown in Fig. 5(b) by the following empirical equation,

Once the individual chip's I-V and L-I curves are modeled, the ideal light output of each LED in the array is simulated. The nomenclature D_{i_j} is used to indicate the device in the i^th column and j^th row, assuming the top-left most LED is D_{1_1}. The circuit in Fig. 5(c) consists of a current source and an LED array with a resolution of 13 rows x 27 columns.

The resistance between LEDs is determined by the ITO and n-type metal resistivity. Considering a device-to-device spacing of 52 µm (x-axis) and 38 µm (y-axis) and a mesa size of 13 × 20 µm², the ITO resistance is 38.4 mΩ for the device-to-device connections in the x-axis direction and 27.2 mΩ in the y-direction. The n-contact resistance is 640 mΩ in the y-axis. The n-contact resistance in the x-axis is negligible because each device is placed next to the n-type transmission line in the y-axis. Also, since a Cu metal pad is attached to the ITO glass near the 20^th column of the LED array for p-type wire connection, in the circuit model, the power source is connected to the p-type contact of D_{20_1}. One can thus imagine that the current flow maximum will occur at the device near the 20^th column.

The current distribution profile is simulated by applying a total injection current of 45.63 mA to the circuit in Fig. 5(c). And by plugging in Eq. (2) to the CAD model, we can the obtain light output of the LEDs at different locations. Figure 5(d) and 5(e) demonstrate the simulated luminescence distribution profiles of D_{1_1}, D_{2_1}… D_{27_1} and D_{1_1}, D_{1_2}…D_{1_13} in the x- and y- direction, respectively. In Fig. 5(d), the device current is maximum at the 20^th column because it is adjacent to the p-type probe contact pad. The y-direction current distribution is nearly exponentially increasing toward the n-type probe contact because the high resistivity n-metal is the limiting factor of voltage drop in the column. The location-dependent EL of the LEDs suggests that the transmission resistances are critical to the current injection to each LED. It is one of the key factors for calibrating mass detection optical output.

In the x-direction, we employed a Gaussian function to fit the luminescence obtained from the CAD model:

3.4 Multi-variable regression analysis based on CAD simulation

The LED array's simulated light output distribution profile is considered the ideal case when the ITO glass is mounted. For the actual luminescent image, the light output of the LED varies to each other because of the fabrication uniformity and the voltage drop of the mass detection. As a result, each device's real light output intensity has to de-embed the resistance effect. We next attempt to obtain the actual optical power of LEDs based on the CAD simulation results by ruling out the resistance effect following the multi-variable regress analysis.

First, the measured optical output, luminescence taken from a CMOS (complementary metal oxide semiconductor) camera by mass detection, is now correlated to the LTspice simulation results. For example, we show, in Fig. 6(a), the luminescence fitting along the x-axis of Fig. 4(a) in the 3^rd row. The error of curve fitting is 0.023. And Fig. 6(b) is the fitting along the y-axis in the 4^th column, with an error of curve fitting 0.012. In order to calibrate the luminescence of each LED in the array from mass detection, we employed multi-variable regression analysis to establish the location-dependent calibration function. Before building the regression model, calibration functions along x- and y-axis are first derived from simulation results. From the image in Fig. 4(a), the average expected luminescence of a single micro-LED is 3600 a.u. (arbitrary unit). By using the back-calculation method, the calibration function of the x-axis is derived as

 

Fig. 6. The calibrated luminescence based on the simulation using multi-variable regression analysis. (a). The luminescence of the first row. The black dots represent results from the LTspice simulation, while the red dots are from the actual measurements. (b). The luminescence of the first column. The black dots represent results from the LTspice simulation, while the red dots are from the actual measurements. (c) Correlation between the luminescence based on the mass detection and calibration and actual luminescence. Ideally, the data will align with the blue line.

Download Full Size | PDF

On the other hand, the calibration function of the y-axis is expressed as

After two independent variables, K(i) and N(j), and the dependent variable (the calibration factor by dividing the actual luminescence by the luminescence from mass detection) are determined, the multi-variable regression model can be built. The coefficient of determination (R2), a parameter suggesting the total variance, is 0.8. Based on K(i) and N(j), we obtain the regression equation:

The regression equation can be used to calculate the actual luminescence by:

To verify, we measured the actual light output power from 100 randomly selected LEDs. As shown in Fig. 6(c), the actual light output of the selected single LED is correlated to extracted output power following the above calibration method. An average error of 6.89% is obtained from the devices.

3.5 Training of micro-LED radiation images based on three different learning models

We first perform training to identify radiation patterns of micro-LEDs using Dataset 1. Table 1 shows the training results from three different learning models with two-class classification outputs for identifying good and destructive LEDs. It suggests the Resnet-50 architecture yields the highest accuracy and precision (96.07 and 96.29%, respectively), compared to the other two. The results also imply that the more stack convolution layers in the Resnet architecture, the better a model is in learning features. While the recall of three learning models shares similar results, the probability of being predicted as good using Resnet-50 in the original good micro-LED is the highest. Because precision is a metric that measures how often a machine learning model correctly predicts the positive class, it is the most critical parameter in the training models, as the ultimate purpose of performing mass detection for micro-LED displays is to identify good LEDs precisely.

Table 1. Comparison of average loss, accuracy, precision, and recall in the training models. Dataset 1 is employed

View Table | View all tables in this article

An essential perspective of image classification is identifying defective details of the LED radiation profile because they may seriously affect display quality. We next employ Dataset 2 in the Resnet-50 learning model. Table 2 shows the performance evaluation of different identification regulations. ‘Original’ represents the images from individual micro-LEDs, which are the same as Dataset 1. ‘Split 3’ represents the identification of three out of four single micro-LED images as good after splitting the micro-LED images into four pieces; otherwise, they will be classified as destructive LEDs. The results show that the non-split micro-LED images have the highest accuracy, precision, and recall. Unfortunately, Split 3 and 4 cannot deliver better results. The split of the LED luminescent profile introduces more variations that lead to misjudgments by the model.

Table 2. Comparison of average loss, accuracy, precision, and recall using Dataset 2 for Resnet-50

View Table | View all tables in this article

The above analysis performs the data processing using an Intel Core i5-13600 K CPU (central processing unit) and a GeForce RTX-4070 GPU (graphic processing unit). Table 3 shows that VGG16 and Resnet-18 architectures have similar running times, while Resnet-50 architecture is the slowest. It’s mainly due to the number of layers in the architecture that affect the duration of training and testing. Furthermore, since the mechanism of Resnet-18 utilizes a residual structure that results in a relatively shallower model and faster execution speed, the time consumption is slightly shorter than VGG16, which possesses a deeper architecture with more convolutional layers and parameters.

Table 3. Comparison of running time between VGG16, Resnet-18, and Resnet-50 for each epoch based on i5-13600 K of CPU and RTX-4070 of GPU

View Table | View all tables in this article

For all three learning models, Dataset 1 and 2 convergence typically occurs in around 10 epochs. The rapid convergence is attributed to the minimal variation in the training data, the micro-LEDs luminescent images. Furthermore, it may also be attributed to a relatively small training dataset (< 50000 images) with only two-class classifications. Based on the results above, Resnet-50 provides the highest accuracy but at the cost of longer running time. While the results of Resnet-18 sacrifice some accuracy, its inference speed is more than twice as fast as Resnet-50. Therefore, precision and time consumption tradeoffs must be considered when deciding on a learning model for micro-LED’s mass detection.

4. Discussions

The methodology of mass detection proposed in this work consists of two parts, predicting light output intensity by correlating simulation and measurement, and determining good LEDs based on the radiation profiles using convolutional neural networks.

For a 13 × 27 micro-LED array demonstrated in this work, light output intensity is extracted from mass detection with an average deviation from actual light output of 6.89%. The variation is mainly attributed to the fitting error in Fig. 5(e), in which the current and luminescence are not ideal Gaussian distribution. A wire-connected Cu pad is to mimic the measurement setup in the future if the LED array in the whole wafer is employed for mass detection. Only a certain contact pads on the ITO glass can be wire connected during the chip probing test. In the current setup, because of the current spreading issue in the p-type probe pad, the deviation is more severe when the row is close to the pad. For example, Fig. 7(a), 7(b) and 7(c) show the CAD simulated luminescence distribution of the 1^st, 6^th, and 13^th rows and the corresponding Gaussian fitting curves. The R2 values of these fitting results are 0.99, 0.997, and 0.999, respectively. It is clearly seen that the light output prediction on LEDs in the D_{1_20} is less accurate. Furthermore, the current injection to each LED will be identical if the transmission resistance is ideally zero, which implies that CAD modeling and regression analysis are not required. In practical measurement, the deviation between the CAD model and the actual light output profile will be minimized with decreased p- and n- metal pad resistance.

 

Fig. 7. The simulated luminescence distribution and the corresponding Gaussian fitting curves at (a) the 1^st row, (b) the 6^th row, and (c) the 13^th row.

Download Full Size | PDF

Tradeoffs are made among different approaches for determining the quality of radiation profiles using CNN. Resnet-50 performs best with 96.07% accuracy, 96.29% precision, and 99.45% recall. However, more time is required to recognize the patterns. It is up to the manufacturers to decide the tradeoff. However, considering the cost of rework and repair in a micro-LED display, adopting the approach with the highest precision may be worthwhile.

5. Conclusions

We demonstrated a mass detection method to examine LED arrays’ luminescence and radiation profiles for micro-LED displays. Mass detection consists of two stages. In the first stage, luminous profiles are extracted from the 13 × 27 LED array sharing the same p-type contact on ITO glass and n-type ground pads on the n-type GaN epi-layer. A non-uniform luminescence image in the LED array is observed because of the fabrication non-uniformity and the transmission resistance on the p-type (ITO) and n-type contact pads. We thus developed a calibration method based on the LTspice to rule out the LEDs’ injection and output power variations incurred from transmission resistance. The multi-variable regression analysis results in an average variation as low as 6.89% for devices predicted from luminescent image capturing and actual optical power measurement. In the second stage, we determine a micro-LED's defective or non-uniform illumination profile by constructing and comparing different 2-D CNN models. The model we developed can recognize the radiation patterns with 96.07% accuracy, 96.29% precision, and 99.45% recall. The proposed methods provide a practical methodology to detect micron-scale LEDs’ electrical and optical properties massively.