Method for ascertaining a luminance noise level of a light-emitting semiconductor diode

Abstract

A method for ascertaining a luminance noise level of a light-emitting semiconductor diode comprises steps for measuring an electrical voltage dropping across the semiconductor diode during operation of the semiconductor diode at a constant current intensity in order to obtain a voltage dataset, and for predicting the luminance noise level based on the voltage dataset using a trained model.

Description

[0001] Method for determining a brightness noise level of a light-emitting semiconductor diode

[0002] DESCRIPTION

[0003] The present invention relates to a method for determining a brightness noise level of a light-emitting semiconductor diode.

[0004] This patent application claims priority from German patent application 10 2023 132 576 . 0 , the disclosure of which is hereby incorporated by reference.

[0005] Light-emitting semiconductor diodes intended for use in measuring devices for measuring time-resolved signals must, in many applications, exhibit high temporal stability of the brightness of the emitted light. To identify suitable light-emitting semiconductor diodes, it is known in the art to determine the brightness noise level of a candidate semiconductor diode. For this purpose, the brightness of the light emitted by the light-emitting semiconductor diode is recorded and evaluated while the semiconductor diode is operating at a constant current over a prolonged period of time, for example, 25 seconds.

[0006] One object of the present invention is to provide a method for determining a brightness noise level of a light-emitting semiconductor diode. This object is achieved by a method having the features of claim 1. Various further developments are specified in the dependent claims.

[0007] A method for determining a brightness noise level of a light-emitting semiconductor diode comprises steps of detecting an electrical voltage dropped across the semiconductor diode during operation of the semiconductor diode at a constant current in order to obtain a voltage data set, and of predicting the brightness noise level on the basis of the voltage data set using a trained model. This method allows the brightness noise level to be determined on the basis of an electrical measurement instead of an optical measurement. The method is based on the finding that there is a relationship between the brightness noise level of a light-emitting semiconductor diode and noise of an electrical voltage dropped across the semiconductor diode, which relationship allows the brightness noise level to be determined on the basis of the noise of the electrical voltage.This makes the process particularly simple to implement and less susceptible to interference. A particular advantage of the process is that it can be performed in parallel on several light-emitting semiconductor diodes without the light emitted by individual semiconductor diodes distorting the measurement results for the other semiconductor diodes. This allows for parallelization, which ultimately makes the process faster and more cost-effective.

[0008] In one embodiment of the method, the voltage data set is acquired over a period of less than 5 seconds, in particular over a period of less than 2 seconds. Experiments have shown that even a voltage data set acquired over such a short period is sufficiently informative to determine the brightness noise level of the light-emitting semiconductor diode with high accuracy. This advantageously allows the method to be carried out particularly quickly and thus also cost-effectively.

[0009] In one embodiment of the method, predicting the brightness noise level comprises a step of transforming the voltage data set into the frequency domain to obtain a spectral data set. The brightness noise level is predicted based on the spectral data set. Advantageously, particularly meaningful features can be extracted from the spectral data set transformed into the frequency domain, from which conclusions can be drawn about the brightness noise level of the semiconductor diode.

[0010] In one embodiment of the method, the voltage data set is transformed using short-time Fourier transformation. This advantageously allows a voltage data set acquired over only a short period of time to be transformed into the spectral data set.

[0011] In one embodiment of the method, predicting the brightness noise level comprises extracting features from the spectral data set using a dimension-reducing transformation. The brightness noise level is predicted based on the extracted features. Advantageously, the dimension-reducing transformation enables a reduction of the spectral data set to particularly characteristic features. This simplifies the prediction of the brightness noise level using the trained model.

[0012] In one embodiment of the method, the model comprises a regression model. Such models are advantageously easy to handle and have proven useful for predicting the brightness noise level.

[0013] In one embodiment of the method, the semiconductor diode is a light-emitting diode. For light-emitting diodes, determining a brightness noise level can be particularly important.

[0014] In one embodiment, the method is carried out while the semiconductor diode is in a wafer assembly. This advantageously allows the method to be carried out at an early stage during the manufacture of the light-emitting semiconductor diode. In this case, if a semiconductor diode proves to be unsuitable, for example, further processing can be dispensed with, which can enable a significant reduction in costs. Determining the brightness noise level at an early stage of the process can also enable early detection and correction of any interference during wafer production and processing. This can potentially reduce the amount of waste, which can also result in a reduction in costs.

[0015] One embodiment of the method is carried out in parallel on several semiconductor diodes of the wafer assembly. Advantageously, the light emitted by the several semiconductor diodes does not mutually influence the measurement results. The parallelizability of the method, together with the short duration of the method, can make it possible to carry out the method for a large number or even all semiconductor diodes of a wafer assembly.

[0016] In one embodiment of the method, this includes prior training of the model.The training comprises steps of providing a plurality of training semiconductor diodes, of recording a training brightness noise level and an associated training voltage data set for each training semiconductor diode, wherein the training brightness noise level of each training semiconductor diode is recorded by measuring a brightness curve of a light emitted by the respective training semiconductor diode during operation of the respective training semiconductor diode with a constant current, wherein an electrical voltage dropping across the respective training semiconductor diode during operation of the respective training semiconductor diode with a constant current is recorded in order to obtain the respective training voltage data set, and of training the model to predict the training brightness noise levels on the basis of the associated training voltage data sets.An advantage of this method is that the training data sets required to train the model can be obtained in a simple and reliable manner. The training data sets used each comprise a training brightness noise level and a corresponding training voltage data set. The training brightness noise level can be recorded in a known manner by optically measuring the brightness of the light emitted by the respective training semiconductor diode.

[0017] In one embodiment of the method, training the model comprises a step of transforming the training voltage data sets into the frequency domain to obtain a plurality of training spectral data sets. The model is trained to predict the training brightness noise levels based on the associated training spectral data sets. From the training spectral data sets transformed into the frequency domain, meaningful features for predicting the training brightness noise levels can advantageously be extracted particularly easily.

[0018] In one embodiment of the method, training the model involves analyzing the training spectral data sets using an analysis method to determine the dimension-reducing transformation. The analysis method identifies particularly meaningful features and feature combinations of the training spectral data sets.

[0019] In one embodiment of the method, the analysis method is a principal component analysis or a non-negative matrix factorization. These analysis methods have proven advantageously suitable for determining a dimension-reducing transformation that enables the extraction of meaningful features from the training spectral data sets as well as from the spectral data sets determined in the subsequent application of the method.

[0020] In one embodiment of the method, training the model comprises extracting training features from the training spectral data sets using the dimension-reducing transformation. The model is trained to predict the training brightness noise levels based on the corresponding extracted training features. This advantageously enables the model to subsequently be used to predict the brightness noise level of a light-emitting semiconductor diode under investigation based on the features extracted according to the method.

[0021] In one embodiment of the method, the training semiconductor diodes originate from different wafers. This advantageously makes it possible to subsequently use the model to predict the brightness noise level of different light-emitting semiconductor diodes.

[0022] The above-described properties, features, and advantages of this invention, as well as the manner in which they are achieved, will become clearer and more readily understood in connection with the following description of the embodiments, which are explained in more detail in connection with the drawings. In each case, a schematic representation shows:

[0023] Figure 1 shows a wafer assembly with several light-emitting semiconductor diodes;

[0024] Figure 2 shows a prediction method for determining a brightness noise level of a light-emitting semiconductor diode based on a voltage data set;

[0025] Figure 3 shows a training procedure for training a model to predict a brightness noise level; and

[0026] Figure 4 shows a method for analyzing spectral data sets.

[0027] Figure 1 shows a schematic representation of a plurality of light-emitting semiconductor diodes 10. The light-emitting semiconductor diodes 10 are arranged in a wafer assembly 15 and are not yet separated. The light-emitting semiconductor diodes 10 are designed to emit electromagnetic radiation, for example, visible light. The light-emitting semiconductor diodes 10 can be, for example, light-emitting diodes.

[0028] For various applications of the light-emitting semiconductor diodes 10, it may be necessary to know a brightness noise level of the light-emitting semiconductor diodes 10. This may be the case, for example, when the light-emitting semiconductor diodes 10 are used to detect time-resolved signals. The brightness noise level represents a measure that characterizes a temporal variation in the brightness of the light emitted by the respective light-emitting semiconductor diode 10.

[0029] It is known in the art to determine the brightness noise level of a light-emitting semiconductor diode 10 by operating the light-emitting semiconductor diode 10 at a constant current and detecting the brightness of the light emitted by the light-emitting semiconductor diode 10 over a period of, for example, 25 seconds, for example, using a photodiode. The data thus acquired are bandpass filtered, and then the brightness noise level is calculated as the effective value of the noise signal normalized to the average brightness.

[0030] In the following, a prediction method 100 is described with reference to Figure 2, which makes it possible to determine a brightness noise level 140 of a light-emitting semiconductor diode 10 in a shorter time and without an optical light measurement.

[0031] The prediction method 100 begins with a measuring process 105 for detecting an electrical voltage drop across the semiconductor diode 10 during operation of the semiconductor diode 10 at a constant current in order to obtain a voltage data set 110. The voltage data set 110 can be detected over a period of time that is shorter than 25 seconds. The period can even be shorter than 5 seconds, shorter than 2 seconds or shorter than one second. The detection of the voltage data set 110 can begin just a few seconds after the start of current supply to the light-emitting semiconductor diode 10, for example after a time of 3 seconds or 5 seconds. The voltage data set 110 can be detected, for example, at a data rate in the range of kHz, so that the voltage data set 110 can comprise, for example, several thousand measurement points.

[0032] Experiments have shown that the brightness noise level 140 of the light-emitting semiconductor diode 10 can be predicted from the voltage data set 110 thus obtained, for example, using a trained model 135. This can, for example, include the method steps described below.

[0033] In this example, predicting the brightness noise level 140 first involves transforming 115 the voltage data set 110 into the frequency domain to obtain a spectral data set 120. The transformation 115 can be performed, for example, using the short-time Fourier transformation method. The brightness noise level is then predicted based on the spectral data set 120.

[0034] For this purpose, features 130 can first be extracted from the spectral data set 120 using a dimension-reducing transformation 125. The dimension-reducing transformation 125 can, for example, be determined in a training process in a manner described in more detail below. The trained model 135 predicts the brightness noise level 140 of the light-emitting semiconductor diode 10 based on the extracted features 130.

[0035] The trained model 135 can, for example, comprise a known regression model, such as AdaBoost or ExtraTrees. The described prediction method 100 can be performed while the semiconductor diode 10 under investigation is still located in the wafer assembly 15. The prediction method 100 can even be performed in parallel for multiple semiconductor diodes 10 of the wafer assembly 15.

[0036] A training method 200 for training the model 135 is described below with reference to Figure 3. The training method 200 is performed once before applying the prediction method 100 described above.

[0037] The training method 200 begins with providing a plurality of training semiconductor diodes 20. The training semiconductor diodes 20 are configured like the light-emitting semiconductor diodes 10 later examined by means of the prediction method 100. It is expedient if the training semiconductor diodes 20 originate from different wafers and thus have different properties.

[0038] Subsequently, in a measuring process 205, a training brightness noise level 240 and an associated training voltage data set 210 are recorded for each training semiconductor diode 20. The training brightness noise level 240 of each training semiconductor diode 20 is recorded in the conventional manner described above by measuring a brightness curve of a light emitted by the respective training semiconductor diode 20 during operation of the respective training semiconductor diode 20 at a constant current. The training voltage data set 210 is recorded for each training semiconductor diode 20 by recording an electrical voltage drop across the respective training semiconductor diode 20 during operation of the respective training semiconductor diode 20 at a constant current.The acquisition of the training brightness noise level 240 and the associated training voltage data set 210 can occur simultaneously for each training semiconductor diode 20. The measurement process 205 thus provides a training data set 250 for each training semiconductor diode 20, which includes the training brightness noise level 240 and the associated training voltage data set 210. In a subsequent training process 255, the model 135 is trained to predict the training brightness noise level 240 for each training data set 250 based on the associated training voltage data set 210.

[0039] If the prediction method 100 is configured as described above with reference to Figure 2, the training process 255 comprises transforming 115 the training voltage data sets 210 of the training data sets 250 into the frequency domain in order to obtain a training spectral data set 220 for each training data set 250. The model 135 is then trained to predict the training brightness noise level 240 based on the associated training spectral data sets 220.

[0040] For this purpose, the training method 200 further comprises, for each training data set 250, extracting training features 230 from the respective training spectral data set 220 using the dimension-reducing transformation 125. The model is trained to predict the training brightness noise level 240 for each training data set 250 based on the associated extracted training features 230.

[0041] The training of the model 135 further comprises an analysis method 300, shown schematically in Figure 4, in which the training spectral data sets 220 of the training data sets 250 are analyzed using an analysis method 310 in order to determine the dimension-reducing transformation 125. The analysis method 310 can, for example, be a principal component analysis or a non-negative matrix factorization. The dimension-reducing transformation 125 thus determined is suitable for extracting the characteristic features 130 even from the spectral data sets 120 obtained in the prediction method 100 subsequently performed.

[0042] The invention has been illustrated and described in more detail using preferred embodiments. However, the invention is not limited to the disclosed examples. Other variations may be devised by those skilled in the art.

[0043] LIST OF REFERENCE SYMBOLS light-emitting semiconductor diode wafer composite training semiconductor diode prediction method measurement process voltage data set transformation spectral data set dimension-reducing transformation features model brightness noise level training method measurement process training voltage data set training spectral data set extracted training features training brightness noise level training data set training process method for analysis analysis method

Claims

PATENT CLAIMS 1. A method for determining a brightness noise level (140) of a light-emitting semiconductor diode (10) comprising the following steps: - detecting an electrical voltage drop across the semiconductor diode (10) during operation of the semiconductor diode (10) with a constant current in order to obtain a voltage data set (110); - predicting the brightness noise level (140) based on the voltage data set (110) using a trained model (135).

2. The method according to claim 1, wherein the voltage data set (110) is acquired over a period of less than 5 seconds, in particular over a period of less than 2 seconds.

3. A method according to any one of the preceding claims, wherein predicting the brightness noise level (140) comprises the following step: - Transforming the voltage data set (110) into the frequency domain to obtain a spectral data set (120); wherein the brightness noise level (140) is predicted based on the spectral data set (120).

4. The method according to claim 3, wherein the transformation of the voltage data set (110) is carried out using short-time Fourier transformation (115).

5. The method according to one of claims 3 and 4, wherein predicting the brightness noise level (140) comprises extracting features (130) from the spectral data set (120) by means of a dimension-reducing transformation (125), wherein the brightness noise level (140) is predicted based on the extracted features (130).

6. The method according to any one of the preceding claims, wherein the model (135) comprises a regression model.

7. Method according to one of the preceding claims, wherein the semiconductor diode (10) is a light-emitting diode.

8. The method according to any one of the preceding claims, wherein the method is carried out while the semiconductor diode (10) is located in a wafer assembly (15).

9. The method according to claim 8, wherein the method is carried out in parallel for a plurality of semiconductor diodes (10) of the wafer assembly (15).

10. The method according to any one of the preceding claims, wherein the method comprises a prior training of the model (135), wherein the training comprises the following steps: - Providing a plurality of training semiconductor diodes (20); - detecting a training brightness noise level (240) and an associated training voltage data set (210) for each training semiconductor diode (20), wherein the training brightness noise level (240) of each training semiconductor diode (20) is detected by measuring a brightness profile of a light emitted by the respective training semiconductor diode (20) during operation of the respective training semiconductor diode (20) with a constant current, wherein an electrical voltage dropping across the respective training semiconductor diode (20) is detected during operation of the respective training semiconductor diode (20) with a constant current in order to obtain the respective training voltage data set (210); - Training the model (135) to predict the training brightness noise levels (240) based on the associated training voltage data sets (210).

11. The method according to claim 10, wherein training the model (135) comprises the following step: - Transforming the training voltage data sets (210) into the frequency domain to obtain a plurality of training spectral data sets (220); wherein the model (135) for predicting the training brightness noise levels (240) is trained based on the associated training spectral data sets (220).

12. The method according to claims 5 and 11, wherein training the model (135) comprises analyzing the training spectral data sets (220) with an analysis method (310) to determine the dimension-reducing transformation (125).

13. The method according to claim 12, wherein the analysis method (310) is a principal component analysis or a non-negative matrix factorization.

14. The method according to one of claims 12 and 13, wherein training the model (135) comprises extracting training features (230) from the training spectral data sets (220) by means of the dimension-reducing transformation (125), wherein the model (135) is trained to predict the training brightness noise levels (240) based on the associated extracted training features (230).

15. The method according to any one of claims 10 to 14, wherein the training semiconductor diodes (20) originate from different wafers.

Images