A rapid reliability assessment method for semiconductor refrigeration devices

Abstract

This invention is based on the failure mechanism of semiconductor cooling devices. It uses the abnormal changes in the package interface resistance and thermal resistance as the characterization factors of the reliability degradation of semiconductor cooling chips. It generates machine learning training data by randomly generating numerical simulation models to form a reliability assessment model. It uses the time-domain distribution of temperature in thermal infrared images as input data and outputs the reliability status of the device. This enables non-contact rapid assessment of the reliability of semiconductor cooling devices during manufacturing and service, and can identify and trace the location of failed thermoelectric arms.

Description

Technical Field

[0001] The invention belongs to the field of intelligent detection of thermoelectric devices, and in particular relates to a method for rapid reliability evaluation of semiconductor refrigeration devices. Background Art

[0002] Semiconductor refrigeration devices are refrigeration devices made of semiconductor materials. They have the advantages of high refrigeration efficiency, small size, no noise, and high reliability. They are widely used in thermal management of various optoelectronic devices and precision components.

[0003] The main reasons for the reliability degradation of semiconductor refrigeration devices are: first, the solder and thermoelectric arm have poor welding factors such as cold soldering during the packaging process; second, the interface interconnection structure is subjected to thermal stress during the service of the device, resulting in defects such as interface delamination and cracks; third, under the action of electromigration, defects such as hillocks or holes are formed in the metal interconnection structure due to the aggregation or consumption of atoms, thereby accelerating the failure extension of the device. The interface failure extension of semiconductor devices will lead to increased interface resistance and thermal resistance, affecting the thermoelectric transport performance of the device. In order to improve the service reliability of semiconductor refrigeration devices, it is necessary to evaluate and predict the reliability status of devices leaving the factory and in service. The current main method is to measure the device resistance and determine the reliability index of the device according to a certain threshold. The disadvantage is that it is difficult to achieve non-contact rapid evaluation. Therefore, it is of great significance to study a non-contact method for rapid reliability evaluation of semiconductor refrigeration devices. Summary of the invention

[0004] In view of this, the purpose of the present invention is to provide a method for rapid reliability evaluation of semiconductor refrigeration devices, construct random interface resistance thermal resistance conditions, obtain virtual model physical parameters of a large number of samples through simulation calculations, use the simulation results as machine learning training data to obtain an evaluation model, and input infrared thermal imaging time domain data to obtain reliability status evaluation information.

[0005] In order to achieve the above object, the present invention adopts the following scheme:

[0006] A method for rapid reliability assessment of a semiconductor refrigeration device, characterized in that it comprises the following steps:

[0007] S1: Establish a finite element model of mechanical, electrical and thermal coupling of semiconductor refrigeration chips;

[0008] S2: Obtain the time-varying temperature distribution of the semiconductor refrigeration chip under different settings;

[0009] S3: training the semiconductor refrigeration chip temporal temperature distribution anomaly detection model;

[0010] S4: Detect abnormal temperature conditions of semiconductor refrigeration chips.

[0011] Furthermore, S1 is specifically:

[0012] The finite element simulation model includes a thermally conductive but insulating ceramic plate, a conductive copper electrode and a thermoelectric arm. Parameters such as the number of thermoelectric arms, the size of the thermoelectric arms, the spacing between the thermoelectric arms, the thickness of the copper electrode, the distance between the thermoelectric arm and the edge of the copper electrode, the thickness of the ceramic plate, the width of the ceramic plate, and the length of the ceramic plate are set. A virtual temperature sensor is set at the center of each thermoelectric arm, and a value range table of the interface resistance between each thermoelectric arm and the copper electrode and the thermal resistance are set.

[0013] Furthermore, S2 is specifically:

[0014] Interface condition samples are randomly selected from the S1 value range table, and finite element simulation boundary conditions are set according to the service conditions of the semiconductor refrigerator. Time-domain finite element calculations are performed to obtain the distribution results of the temperature of the virtual sensors at each position of the semiconductor refrigerator sample over time.

[0015] Furthermore, S3 is specifically:

[0016] Let X t represents the finite element simulation temperature distribution matrix at time t, S = [X1, X2, ..., X n ] represents the temperature distribution time series, where n represents the current moment. A time series autoencoder anomaly detection model is constructed using a convolutional long short-term memory neural network (ConvLSTM). Two ConvLSTM networks with the same structure are constructed as the encoder network and the decoder network. First, S is input into the encoder network, and the encoder network outputs the implicit representation vector v of S. S , v S As the decoder network input, the sequence S is reconstructed to generate the reconstructed sequence S'. The network is trained until convergence with minimizing the reconstruction error l = ||S-S'|| as the training goal of autoencoder anomaly detection.

[0017] Furthermore, S4 is specifically:

[0018] The temperature distribution of the semiconductor refrigeration device after power-on is collected by infrared thermal imager, and the temperature distribution time series to be detected is obtained. It is input into the encoder end of the abnormality detection network, and the decoder end outputs the corresponding reconstruction sequence to obtain the value of the reconstruction error l. If l is greater than the set threshold α, it is determined that the current refrigeration piece is abnormal, and the difference between the current reconstruction error and the threshold is used as its abnormality measurement index. The larger this index is, the lower the reliability of the current refrigeration piece is.

[0019] Furthermore, the value ranges of the interface resistance and thermal resistance in S1 are obtained based on actual measurements of a large number of devices when the interface is not failed.

[0020] Furthermore, the number of samples in S2 is not less than 100.

[0021] Furthermore, the frame rate of the infrared thermal imager in S4 is above 30.

[0022] Compared with the prior art, the present invention has the following beneficial effects:

[0023] The present invention combines numerical simulation, artificial intelligence data-driven methods and non-contact thermal infrared image detection, takes abnormal changes in electrical resistance and thermal resistance as the core characterization factors of reliability attenuation of semiconductor refrigeration plates, generates machine learning training data through virtual simulation, forms a reliability assessment model, uses thermal infrared images as input data, outputs the device reliability status, and thus realizes non-contact rapid assessment of the reliability of semiconductor refrigeration devices before leaving the factory and during service, and can identify and trace the location of failed thermoelectric arms. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 Flow chart of the rapid reliability assessment method for semiconductor refrigeration devices. DETAILED DESCRIPTION

[0025] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0026] A method for quickly evaluating the reliability of a semiconductor refrigeration device comprises the following steps:

[0027] S1: Establishing a finite element model of mechanical, electrical and thermal coupling of semiconductor refrigeration chips

[0028] The finite element simulation model includes a thermally conductive but insulating ceramic plate, a conductive copper electrode and a thermoelectric arm, which contains 7 pairs of thermoelectric arms with a size of 0.8×0.8×0.6mm. 3 , the distance between the thermoelectric arms is 0.8mm, the thickness of the copper electrode is 0.2mm, the distance between the thermoelectric arm and the edge of the copper electrode is 0.2mm, the thickness of the ceramic plate is 0.76mm, the width of the ceramic plate is 10mm, the length of the ceramic plate is 10mm, and a virtual temperature sensor is set at the center of each thermoelectric arm; the interface resistance between each thermoelectric arm and the copper electrode ranges from 0.9e -11 -1.1e -11 Ω·m 2 , the thermal resistance setting range is 0.9e -6 -1.1e -6 m 2 ·K / W.

[0029] S2: Obtain the time-varying temperature distribution of the semiconductor refrigeration chip under different settings

[0030] 100 samples of interface conditions were randomly selected from the S1 value range table, and finite element simulation boundary conditions were set according to the service conditions of the semiconductor refrigeration chip. Time domain finite element calculations were performed to obtain the distribution results of the virtual sensor temperature at each position of the semiconductor refrigeration chip sample over time.

[0031] S3: Training the semiconductor refrigeration chip time series temperature distribution anomaly detection model

[0032] Let X t The finite element simulation temperature distribution matrix at time 0 is represented. The temperature distribution time series S = [X1, X2, ..., X 10 ], repeat the above simulation process 100,000 times, and form the obtained sequence into a training data set. Construct a ConvLSTM neural network time series autoencoder anomaly detection model, and train the network with the above training data until convergence.

[0033] S4: Detect abnormal temperature conditions of semiconductor refrigeration chips

[0034] The temperature distribution of the semiconductor refrigeration device after power-on is collected with an infrared thermal imager with a frame rate of 30, and the temperature distribution time series to be detected is obtained. It is input into the encoder end of the anomaly detection network, and the decoder end outputs the corresponding reconstruction sequence, and the reconstruction error l is obtained to be 0.85. Compared with the anomaly detection threshold α=0.5, the reconstruction error is too large, so it is determined that the current refrigeration plate has an abnormality.

[0035] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for rapid reliability assessment of semiconductor refrigeration devices, characterized in that: The steps include: S1: Establish a finite element model of mechanical, electrical and thermal coupling of semiconductor refrigeration chips; S2: Obtain the time-varying temperature distribution of the semiconductor refrigeration chip under different settings; S3: training the semiconductor refrigeration chip temporal temperature distribution anomaly detection model; S4: Detecting abnormal temperature conditions of semiconductor refrigeration chips; S1 is specifically: The finite element simulation model includes a thermally conductive but insulating ceramic plate, a conductive copper electrode and a thermoelectric arm, and sets the number of thermoelectric arms, the size of the thermoelectric arms, the spacing between the thermoelectric arms, the thickness of the copper electrode, the distance between the thermoelectric arm and the edge of the copper electrode, the thickness of the ceramic plate, the width of the ceramic plate, and the length of the ceramic plate. A virtual temperature sensor is set at the center of each thermoelectric arm, and a value range table of the interface resistance and thermal resistance between each thermoelectric arm and the copper electrode is set. S2 is specifically: Randomly select interface condition samples from the S1 value range table, set finite element simulation boundary conditions according to the service conditions of the semiconductor refrigeration piece, perform time domain finite element calculations respectively, and obtain the distribution results of the virtual sensor temperature at each position of the semiconductor refrigeration piece sample over time; S3 is specifically: Let X t represents the finite element simulation temperature distribution matrix at time t, S = [X1, X2, ..., X n ] represents the temperature distribution time series, where n represents the current moment. A time series autoencoding anomaly detection model is constructed using a convolutional long short-term memory neural network (ConvLSTM). Two ConvLSTM networks with the same structure are constructed as the encoder network and the decoder network. First, S is input into the encoder network, and the encoder network outputs the implicit representation vector v of S. s , v s As the decoder network input, the sequence S is reconstructed to generate the reconstructed sequence S′, so as to minimize the reconstruction error l = ||SS′|| as the training target of autoencoder anomaly detection and train the network until convergence; S4 is specifically: An infrared thermal imager is used to collect the temperature distribution of the semiconductor refrigeration device over time after power is turned on, and the temperature distribution time series to be detected is obtained. The temperature distribution time series is input into the encoder end of the anomaly detection network, and the decoder end outputs the corresponding reconstruction sequence to obtain the value of the reconstruction error l. If l is greater than the set threshold α, it is determined that the current refrigeration piece is abnormal, and the difference between the current reconstruction error and the threshold is used as the abnormality measurement index. The larger this index is, the lower the reliability of the current refrigeration piece is.

2. A method for rapid reliability assessment of a semiconductor refrigeration device according to claim 1, characterized in that: The value ranges of the interface resistance and thermal resistance in S1 are obtained based on actual measurements of a large number of devices when the interface is not failed.

3. A method for rapid reliability assessment of a semiconductor refrigeration device according to claim 2, characterized in that: The number of samples in S2 is not less than 100.

4. A method for rapid reliability assessment of a semiconductor refrigeration device according to claim 3, characterized in that: The frame rate of the infrared thermal imager in the S4 is above 30.

Images