Neural network based prediction of semiconductor device response.
A neural network-based method predicts semiconductor device performance by analyzing wafers, addressing defects and improving yield by identifying suitable regions for fabrication, enhancing device performance prediction accuracy.
Patent Information
- Application Number
- JP2024529631
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2021-11-19
- Filing Date
- 2022-11-04
- Publication Date
- 2024-11-19
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Semiconductor fabrication processes introduce defects due to particles and process steps, affecting device performance and yield, which current evaluation methods struggle to predict effectively.
A neural network-based method that analyzes semiconductor wafers using image and electrical data to predict the electrical response of devices before fabrication, identifying suitable regions for device fabrication and improving yield.
Enhances yield and reduces waste by accurately predicting device performance and suitability across a wafer, utilizing convolutional neural networks to learn both known and unknown features affecting device performance.
Smart Images

Figure 2024543097000001_ABST
Abstract
Description
[Technical field]
[0001] cross reference This application claims the benefit of priority to U.S. Provisional Patent Application No. 63 / 281,206, filed November 19, 2021.
[0002] Federally funded research and development The U.S. Government owns an ownership interest in this invention. For licensing inquiries, please contact the U.S. Naval Research Laboratory, Office of Technology Transfer, Code 1004, Washington, DC 20375, USA, +1.202.767.7230, techtran@nrl.navy.mil, refer to Navy Case No. 210892.
[0003] Technical Field The present invention relates to the evaluation of semiconductor devices, and in particular to a method for predicting the electrical response of a semiconductor device prior to its fabrication based on characterization or measurement of the semiconductor wafer on which the semiconductor device is fabricated, in combination with a properly designed machine learning model trained on characterization and measurement data of similar wafers and devices fabricated on the wafer. [Background technology]
[0004] Modern semiconductor fabrication involves dozens of steps, including boule growth, cutting, polishing, epitaxy, implants, and lithography. However, these process steps can introduce defects that adversely affect device performance and reduce device yields from wafers that contain such defects.
[0005] One common defect in semiconductor wafers results from the presence of particles that can accumulate on the wafer surface, even in the cleanroom facilities used in modern semiconductor manufacturing. Similarly, process steps such as etching and deposition can generate unique particles in the wafer material that alter the electrical response of the finished device. The presence of such particles can be identified by visual inspection or by predefined computerized image analysis using optical or electron micrographs or related imaging techniques.
[0006] Other process steps, such as epitaxy and dopant implantation, can introduce defects into the wafer that can be easily identified by photoluminescence, electroluminescence, cathodoluminescence, Auger spectroscopy, or similar techniques that examine the radiative or non-radiative transitions of the defect. Again, these types of defects are easily identified by those skilled in the art by analysis of metrology images, or similarly by analysis of predefined shapes found by computer algorithms based on features that have themselves been classified by experts.
[0007] Semiconductor manufacturing employs numerous data science algorithms for process analysis, process control, process optimization, and overall yield improvement. Rich data sets are generated by the process parameters of the fabrication tools used and from in-situ measurements by the tools themselves.
[0008] A set of data science analytical tools has long been applied to find correlations between process and test data to enable process optimization and general yield analysis. For example, it is known from statistical analysis that the utilization of certain workflows (e.g., specific selection of wafer locations in various deposition systems coupled with specific on-wafer die locations during lithography and etching steps) has a predictable correlation with the production of devices with improved metrics (e.g., maximum operating frequency). In addition, typically at later stages of fabrication, metallized test structures are usually incorporated to provide in-process or post-process electrical test data. It is common to use the electrical response of the final device or test structures on the wafer as one parameter in the statistical analysis of the entire fabrication process. An example is a correlation analysis that relates process parameters or metrology metrics at various stages of processing to the electrical response of the final device.
[0009] Evaluating wafers before device fabrication is complete is essential to maximize yield and reduce the incidence of defective devices. (See U.S. Patent No. 5,339,633.) Using a general framework for such evaluation, it is possible to identify, within statistical likelihood, features that lead to yield loss at each process step or throughout the entire process flow. Optical metrology is one evaluation method that can be performed to inspect the semiconductor surface before or after a particular process step. (See U.S. Patent No. 5,339,633.) After device processing on a wafer is complete, such metrology techniques can be used to trace back to the specific features or defects that created the failing devices at their physical locations on the wafer.
[0010] Recently, machine learning approaches have been introduced to enhance statistical analysis and optimization approaches. Moriya et al. used a machine learning approach to enhance the optimization of process conditions for PECVD processes. See Non-Patent Document 1. Traditionally, in etch and lithography, it is common to utilize a combination of visual inspection and rule-based approaches to correct the process steps or the mask itself to correct prototype features to a target feature set. Hooker et al. used a fully connected neural network to enhance the lithography target correction process. See Non-Patent Document 2.
[0011] They discussed the application of machine learning in yield prediction, process deviation detection, and process flow simplification. See Non-Patent Document 3. In general, the introduction of machine learning techniques has been used to replace or enhance established data analysis techniques already applied to process optimization, process control, and yield evaluation in semiconductor fabrication processes. Such statistical or data analysis per se replaces the analysis performed by a person skilled in the art or finds relationships that a semiconductor field expert would intuitively expect. [Prior art documents] [Patent documents]
[0012] [Patent Document 1] U.S. Patent No. 11,415,518 [Patent Document 2] U.S. Patent Application Publication No. 2021 / 0389126 [Non-patent literature]
[0013] [Non-Patent Document 1] T. Moriya, "Machine Learning Approaches Optimizing Semiconductor Manufacturing Processes," 2021 5th IEEE Electron Devices Technology & Manufacturing Conference (EDTM), 2021, pp. 1-3, doi: 10. 1109 / EDTM50988.2021.9420955 [Non-Patent Document 2] K. Hooker et al., "Using machine learning etch models in OPC and ILT correction," Proc. SPIE 11614, Design Process-Technology Co-optimizationXV, 116140B (13 April 2021); doi: 10.1117 / 12.2587225 [Non-Patent Document 3] C. He et al., “Applications for Machine Learning in Semiconductor Manufacturing and Test”, 2021 Electron Devices Technology and Manufacturing Conference (EDTM) [Non-Patent Document 4] J. C. Gallagher, “High Voltage Vertical GaN Diodes for Shipboard Power Conversion,” NRL Technical Report, DTIC Accession No. ADI 145251 (2021) [Non-Patent Document 5] M. Kuball, “Raman spectroscopy of GaN, AlGaN and AIN for process and growth monitoring / control,” Surf. Interface Anal. 2001; 31: 987-999 [Non-Patent Document 6] J. Gallagher, et al., “Predicting Vertical GaN Diode Quality using Long Range Optical tests on Substrates,” in 2020 International Conference on Compound Semiconductor Manufacturing Technology, 2020, pp. 207-210 [Non-Patent Document 7] JC Gallagher et al., “Long range, non-destructive characterization of GaN substrates for power devices,” Cryst. Growth, vol. 506, pp. 178-184, Jan. 2019 Summary of the Invention
[0014] This Summary is intended to introduce a selection of concepts in a simplified form that are further described in the Detailed Description. This Summary is not intended to identify key features or essential features of the claimed subject matter, nor is it intended to be used as an aid in determining the scope of the claimed subject matter. Instead, this Summary is merely presented as a brief overview of the subject matter described and claimed herein.
[0015] The present invention provides a computer-implemented method for evaluating semiconductor wafers, whereby, using a suitably designed neural network, a computer can take image data for the wafer in question, as well as image and electrical data for previous wafers and devices fabricated on those previous wafers, to find relationships to and between known and previously unidentified structural features that may degrade the performance of devices fabricated on the wafer or reduce the device yield of the wafer.
[0016] In some embodiments, neural networks can be used to analyze semiconductor wafers and predict electrical performance, such as the current-voltage (IV) response and capacitance-voltage (CV) response of devices under various bias conditions, prior to critical processing or testing of the semiconductor wafers.
[0017] In some embodiments, response predictions obtained using the neural network analysis techniques of the present invention can be used as inputs into a pass / fail yield metric for devices.
[0018] In some embodiments, the response predictions obtained using the neural network analysis techniques of the present invention can be used to develop models designed to enable neural networks to learn visual information. The neural network models are based on one or more layers of a convolutional neural network (CNN) that train a series of filters that scan the input image or the output of a previous CNN layer. This type of neural network is particularly effective for images, since individual image pixels are not important; instead, patterns of blocks of pixels are more informative. By appropriately training the weights of the multi-layer CNN, a model is created that essentially identifies features of the input image that affect the output response. Such input image features identified by the trained filters include features known or expected by those skilled in the art, as well as features that are unknown or non-obvious to those skilled in the art.
[0019] In some embodiments, the response predictions can be used to identify areas on a wafer that may be more or less suitable for fabricating certain types of devices, allowing fabrication of devices that are more suited to a particular wafer, thereby improving yields and reducing waste.
[0020] This approach can be easily applied to entire wafers to predict the performance of multiple types of electronic and optoelectronic devices that can be fabricated on the wafer, such as transistors, diodes, integrated circuits, light emitters, capacitors, inductors, etc. Furthermore, properly designed models can predict circuits, subcomponents of larger circuits, and circuit traces. [Brief description of the drawings]
[0021] [Figure 1] 1A and 1B are Raman maps of an exemplary semiconductor wafer evaluated in accordance with the prior art, where FIG. 1A is a map of Raman A1 vibrational peak locations and FIG. 1B is a map of Raman E2 vibrational peak locations. [Diagram 2] 2A-2D are optical profilometry images of the same wafer area used in FIG. 1B that provide additional information about defects in semiconductor wafers. [Diagram 3] FIG. 1 is a block diagram illustrating an exemplary embodiment of a neural network-based assessment of semiconductor wafers according to the present invention, in which optical images of a GaN wafer are input into a multi-layer convolutional neural network (CNN) model to predict the performance of semiconductor devices fabricated on the wafer. [Figure 4] 4A and 4B are graphs showing the measured DC on-resistance performance (FIG. 1A) and maximum DC forward current (FIG. 4B) of a semiconductor device compared to the response predicted by a machine learning model in accordance with the present invention. [Diagram 5] 1 is a plot of exemplary current-voltage measurements taken from −10 V (reverse bias) to +10 V (forward bias) of three exemplary PIN vertical diodes, showing the current-voltage performance of a properly functioning (good) device, a device with low (forward) on-current, and a device with high leakage under reverse bias (i.e., negative voltage). [Figure 6]1 is a plot comparing the current-voltage response of an as-fabricated PIN diode with the response predicted by a machine learning model using only optical images of the wafer surface as input in accordance with the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0022] The aspects and features of the present invention summarized above can be implemented in various forms. The following description shows, by way of example, combinations and configurations in which the aspects and features can be implemented. It should be understood that the described aspects, features, and / or embodiments are merely examples, and that those skilled in the art may utilize other aspects, features, and / or embodiments and make structural and functional changes without departing from the scope of the present disclosure.
[0023] As described in more detail below, the present invention provides a computer-implemented method for evaluating semiconductor wafers, whereby, using a suitably designed neural network, a computer can take image data for the wafer in question, as well as image and electrical data for previous wafers and devices fabricated on those previous wafers, to find relationships to and between known and previously unidentified structural features that may degrade the performance of devices fabricated on the wafer or reduce the device yield of the wafer.
[0024] In some embodiments, neural networks can be used to analyze semiconductor wafers and predict electrical performance, such as the current-voltage (IV) response and capacitance-voltage (CV) response of devices under various bias conditions, prior to critical processing or testing of the semiconductor wafers.
[0025] In some embodiments, response predictions obtained using the neural network analysis techniques of the present invention can be used as inputs into a pass / fail yield metric for devices.
[0026] In some embodiments, the response predictions obtained using the neural network analysis techniques of the present invention can be used to develop a model that predicts the response of a diode or transistor. The model is trained with input and output data of similar wafers. One of the input data for GaN-based devices is an optical image. The neural network is designed to effectively learn visual information. A suitable design is a multi-layer convolutional neural network, where each layer learns a set of filters that are convolved with the input image or the output of the previous layer in the network. A suitable design of the neural network allows for a training procedure where the learned weights of the network can accurately predict relevant outputs such as the current-voltage response or capacitance-voltage response of the device and related metrics.
[0027] In some embodiments, the response prediction values can be used to identify areas on a wafer that may be more or less suitable for fabricating certain types of devices, allowing fabrication of devices that are suitable for a particular wafer, thereby improving yields, reducing costs, and improving the efficient utilization of tools used in the fabrication process.
[0028] The approach can be easily applied to entire wafers to predict the performance of multiple types of electronic and optoelectronic devices that can be fabricated on the wafer, including transistors, diodes, integrated circuits, light emitters, capacitors, inductors, and more.
[0029] A method for evaluating semiconductor wafers according to the present invention includes the use of a neural network trained with input data for the wafer being evaluated, output data from evaluations of previous wafers, and the input and output data of previously processed wafers, which takes the data for the previous evaluations and applies it to the data for the wafer being evaluated to obtain information about devices that may be processed from the wafer. Such information may include a prediction of the electrical response of one or more types of devices that may be processed in a particular region of the wafer, a map of the wafer identifying locations on the wafer that may be more or less suitable for producing one or more types of data, an identification of types of devices that may be more or less suitable for fabrication on the wafer, or a comprehensive evaluation of a large region containing many devices or the entire wafer.
[0030] The input data set for the wafer being evaluated may include structural, optical, and / or luminescence images and / or metrics processed from the image set. In an exemplary embodiment, the input may include the results of any one or more optical imaging techniques known in the art, such as optical profilometry, Nomarski, fluorescence, and ultraviolet images. In other embodiments, the input data may include one or more of photoluminescence, electroluminescence, cathodoluminescence, and Auger spectroscopy images of the wafer being evaluated or previous wafers. In yet other embodiments, the input data may include scanning electron micrographs, atomic force microscopy images, X-ray diffraction (XRD) mapping images, or Raman mapping images of the wafer being evaluated or previous wafers. In still other embodiments, the input data may include or be based solely on metrics derived from input data for previous wafers already evaluated.
[0031] The output data from the previous evaluation used in the neural network method for evaluating the target wafer may include data regarding the results of optical imaging of the previous wafer, data regarding the type, size, and arrangement of one or more devices fabricated on the previous wafer, and data regarding the electrical performance of such devices (e.g., data regarding the current-voltage and capacitance-voltage performance of the devices).
[0032] According to the present invention, the neural network receives these inputs and calculates the output by a series of mathematical operations. The neural network has a series of layers consisting of matrix operations with learnable weights. Typically, activation functions are applied to the output of each layer to increase the complexity of the relationships the network can learn. The multi-dimensional input data is transformed into multi-dimensional output data by the matrix and non-linear operations of the layers. The output data becomes the input to the next layer of the neural network. Each layer in the network performs operations in turn from input to output, and is typically designed so that the output of the last layer has the dimensions of the target data.
[0033] From matrix algebra, it is known that a series (or layers) of linear matrix operations are equivalent to one (i.e., one layer). Activation functions increase the complexity of the mapping from input to output of a layer in the network. Typically, activation functions are nonlinear functions such as rectified linear unit (ReLu), hyperbolic tangent (TanH), sigmoid, swish, softmax, etc. Typically, only one type of activation function is applied to a particular layer, but different activation functions can be applied to the outputs of different layers.
[0034] It is known that the weights of a neural network can be efficiently learned by applying the backpropagation algorithm to a particular model of the training data set from the input to the output.
[0035] The neural network may include any neural network known in the art, such as, for example, a convolutional neural network (CNN), a fully-connected neural network, and a recurrent neural network, including a gated recurrent unit (GRU) network or a long short term memory (LSTM) network.
[0036] Using such neural networks, the computer can process such input data and provide output data indicative of the electrical performance of devices that may be fabricated on the wafer (e.g., data regarding possible current-voltage (IV) or capacitance-voltage (CV) responses, etc.). The data may relate to the entire wafer or only to certain regions of the wafer, providing information regarding which regions of the wafer are more or less suitable for fabricating particular devices or devices having particular sizes. In some embodiments, the computer can generate "maps" or wafers indicative of the performance of devices fabricated at various locations on the wafer, or indicative of the optimal geometric configuration or maximum or minimum size of such devices.
[0037] The first step in the neural network-based semiconductor wafer evaluation method of the present invention is to obtain a set of input / output data for the wafer of interest, and similar previous wafers must also be obtained. Often the input data is obtained from Raman spectroscopy and / or optical profilometry data for the wafer in question, and the output data consists of current-voltage (IV) data for PIN vertical diodes fabricated on previous samples.
[0038] Raman spectroscopy and optical profilometry are particularly useful for detecting the presence of defects in semiconductor wafers.
[0039] Using Raman spectroscopy, the A in the wafer was measured as shown in Figures 1A and 1B. l By evaluating the positions of the E1 and E2 peaks, changes in crystal stress and electron carrier concentration can be detected.
[0040] For the wafer shown in Figure 1A and Figure 1B, the Raman spectrum of the wafer was excited using a 532 nm laser and acquired over a relatively large sample area. The sample was a type II-a wafer with a regular crystalline defect pattern. The black box in the figure represents the area covered by the anode metal of the vertical device.
[0041] Two peaks in the Raman spectrum are used to identify areas of the wafer that contain defects that may affect device performance.
[0042] The A1(LO) peak is often used to detect changes in carrier concentration. As shown in Figure 1A, this A l (LO) The position of the peaks reveals that the sample has a regular pattern of circular regions of relatively high carrier concentration.
[0043] In the E2 peak map shown in Figure IB, points of high crystallographic stress are often observable as distinct spots in the plot, but the shifts can be subtle and the spots are not always present in the map. When present in the plot, they indicate areas where there is a change in crystallographic stress, and A l The increased carrier concentration as indicated by the (LO) peak can degrade the performance of devices fabricated on the wafer. (See Non-Patent Document 4.) A detailed description of such peaks and how to calculate carrier concentrations can be found, for example, in Non-Patent Document 5.
[0044] Optical profilometry (e.g., ZYGO optical profilometry or other suitable methods) can be used to detect abnormal features (such as bumps and pits) on the wafer surface and is another method for identifying potential defects in the wafer. Such surface anomalies are often caused by defects below the wafer surface and can severely degrade the performance of devices fabricated on such defective areas of the wafer. Example optical profilometry images are shown in Figures 2A and 2B, which show the presence of surface anomalies on the imaged wafer, as indicated by the circled areas in the figures.
[0045] One example of a prior art approach to evaluating such semiconductor wafers in a non-machine learning environment divides a full wafer into a grid to estimate the size of devices that can be fabricated on the wafer, and then uses long-range optical techniques to estimate yield using generalized ESD testing to detect outliers in combination with surface RMS measurements. (See Non-Patent Document 6; see also Patent Document 2, supra.) This method is accurate enough for screening wafers, but it has not yet been demonstrated whether certain types of surface features can be used to predict device failure.
[0046] Several devices were fabricated on the imaged wafer, at locations indicated by squares in the figures. These devices either had bumps or pits greater than 1.5 μm deep in the edge termination region of the device. Thus, the circles in the optical profilometry image shown in Figure 2A highlight the locations of structural defects on the wafer, and the two circles in Figure 2B highlight the locations of surface pits on the wafer. As shown in Figures 2C and 2D, devices fabricated in areas with significant surface defects exhibited large current losses, as indicated by the white squares in Figures 2C and 2D, while devices fabricated in other areas of the wafer exhibited much less loss. However, it should be noted that even devices fabricated on areas exhibiting smooth optical profilometry images exhibited degraded performance, as indicated by the black squares in Figures 2C and 2D.
[0047] The neural network evaluation method of the present invention uses this input data to provide predictions of IV (current-voltage), CV (capacitance-voltage), or both IV and CV. The IV and CV predictions can be used to extract specific metrics including ideality factor, (on) current, on voltage, and on resistance at forward voltage, (off) leakage current at reverse voltage.
[0048] Similar network designs can be trained to directly output and predict specific metrics including ideality factor, (on) current, on voltage, and on resistance at forward voltage, (off) leakage current at reverse voltage, where the output of the network is a vector with dimensions equal to the dimensions of the output metric.
[0049] For designs where a particular metric is extracted from predicted IV and CV data or where the metric is calculated directly, a pass / fail threshold can be defined for one or more metrics. The number of devices that pass based on the pass / fail criteria can be used to estimate the local yield at a particular spatial location. Similarly, the number of devices that pass can be summed over a larger area or an entire wafer to estimate the overall yield.
[0050] Machine Learning Algorithms
[0051] According to the present invention, by using optical profilometry as input data and IV curves of diodes as output data, machine learning algorithms for evaluating semiconductor wafers and their performance can be trained to accurately predict the IV and / or CV response of devices fabricated on the wafer being evaluated, as well as other metrics (e.g., turn-on voltage, breakdown voltage, etc.). Similarly, metrics such as rise time to an on state and fall time to an off state in response to a change in input can be made a function of time. Additionally, models can be trained to predict responses such as output power and gain as a function of frequency. Backpropagation algorithms work by iteratively adjusting the weights of the mathematical operations of a network to fit the output of the network to a known target output.
[0052] The neural network-based method for evaluating semiconductor wafers and predicting the performance of semiconductor devices fabricated thereon is described below in the context of a convolutional neural network (CNN) as known in the art, although one of ordinary skill in the art will readily appreciate that other suitable neural networks and / or machine learning approaches may also be used, and the use of all such approaches is considered to be within the scope of the present invention. In some embodiments of the neural network-based method for predicting the performance of semiconductor devices according to the present invention, the model may directly predict the IV data of the device, while in other embodiments, device metrics may be derived from the predicted IV and / or CV performance obtained by the neural network model.
[0053] Thus, FIG. 3 is a block diagram illustrating aspects of an exemplary neural network-based assessment of a semiconductor wafer and prediction of the performance of devices fabricated thereon in accordance with the present invention. As shown in FIG. 3, a non-destructive optical image of the wafer, having undergone pre-processing (e.g., ZYGO mapping), is input to a neural network model consisting of three convolutional layers 302a / 302b / 302c and two fully connected layers 304, 305 that output a prediction of the device output (e.g., current-voltage (IV) response). Each layer of the convolutional neural network (CNN) consists of a series of filters that are repeatedly scanned across the entire image or the data output of the previous CNN layer, with the dimension of the final output being equal to the dimension of the current-voltage target data. Since the image contains only surface height information, the color image is simply a representation of a one-dimensional [N×N] plane, where N is the height and width of a given pixel in the square image.
[0054] An image 301 [of dimensions N×N×l] is the input data of the first convolution layer 302a. In the first convolution layer, the blocks and arrows indicate a dot product calculation on a [5×5×1] filter weight matrix and pixels of a [5×5×1] portion of the image. This dot product is repeated over the entire input image to create an [N×N×l] plane. The filter operation is repeated for a given C number of filters (e.g., C=16) to create an [N×N×C] output volume represented by 302b. The output of the first convolution layer, represented by such an [N×N×C] data block 302b, is the input data of the next convolution layer. The mathematical process of the second convolution is similar to that of the first convolution, except that the data block of 302b is 16 channels deep. In the second convolution layer of 302b, the blocks and arrows indicate a dot product calculation on a [5×5×16] filter weight matrix and data points of a [5×5×16] portion of the data. This dot product is repeated over the data block of 302b to create an [N×N×l] plane. The filter operation is repeated for a given C2 number of filters (e.g., C2=32) to create an [N×N×C2] output volume represented as 302c. The data block of 302c is the output of the second convolution and the input to the third convolution layer. The mathematical process of the third convolution is similar to the output of the third convolution represented as data block 303. In each convolution step 302a, 302b, 302c, a max pooling function is applied to the output of the convolution, followed by a rectified linear activation unit (ReLu).
[0055] The 3D data block in 303 is converted to a 1D matrix by a flattening operation. This 1D data is the input to a fully connected layer, whose output is represented by the 1D block 304. The first fully connected layer contains ReLu activation and a dropout function to improve the robustness of the prediction, but is not shown in the image.
[0056] The output data of the first fully connected layer, represented as a one-dimensional block at 304, is the input data of the second fully connected layer.
[0057] The output of the second fully connected layer is represented as a one-dimensional block 305 having the same shape as the experimental output data (e.g., the current response over a particular voltage range).
[0058] Parameters such as the number of layers, the number of filters, the filter dimension, the filter stride length, the number of types of activation functions, and the application of pooling layers can be adjusted according to the input data and output data (including the dimensions of the input data and output data) and the characteristic dimensions of the input data features.
[0059] The previous discussion of Figure 3 showed that the last layer of the network is designed to generate an overall output whose dimensions match those of the experimental data (e.g., current-voltage or capacitance-voltage experimental data). A similar network can directly predict metrics of a device. Here, the dimension of the output from the last layer is equal to the dimension of the vector containing the target metric. For example, a neural network can be designed to predict the Ideality Factor, Max Forward Current, and fwc )(A / cm 2 ), 200V reverse leakage current (200 VI rev )(A / cm 2 ), on-resistance (R on )(Ω cm 2 ), and turn-on voltage (V on ) to generate a vector of five values representing five specific metric values.
[0060] To test the accuracy of the prediction, the neural network model was trained on a selection of 80% of the device data and tested on the remaining 20% of the device data. The test data was not used to fit the parameters of the model. It is known that separation of training data and test data is necessary to evaluate the accuracy and robustness of the model. Table 1 below shows exemplary prediction results of ideality factor, maximum forward current, 200V reverse leakage current, on-resistance, and turn-on voltage of the wafers evaluated using the neural network-based method of the present invention.
[0061] [Table 1]
[0062] In tests, all of these properties were predicted within 6% of the experimental results.
[0063] The graphs in Figures 4A and 4B show the accuracy of the on-resistance and maximum forward current points tested by the machine learning model in accordance with the present invention, with Figure 4A showing the measured DC on-resistance on a test data set compared to the on-resistance response predicted by the machine learning model, and Figure 4B showing the measured maximum DC forward current measured on a test data set compared to the response predicted by the machine learning model. The 1:1 ratio line in both figures is included for visual clarity.
[0064] From these plots, we can see that 91% of the tested points are within 5% of the predicted values, indicating that the algorithm is effective at predicting forward bias conditions. This is expected, since large surface roughness, dust particles, and hillocks are likely to result in large contact resistance or epitaxial layer resistance, which can be clearly detected by this technique (see Non-Patent Document 7).
[0065] In some embodiments, the pass / fail yield of devices may be predicted as an additional step in the methods described above. For example, predetermined thresholds may be assigned as pass / fail points for one or more metrics. For example, a pass device may have an absolute current level of 2·10 at a reverse voltage of −200V. -7 A / cm 2 Less than or equal to the on-resistance (R on ) is 2mΩ / cm 2 It is necessary that the value is less than 100%.
[0066] 5 and 6 further illustrate the results of device evaluation using the neural network-based method according to the present invention.
[0067] FIG. 5 is a plot of exemplary current-voltage measurements taken from −10 V (reverse bias) to +10 V (forward bias) of three exemplary PIN vertical diodes, showing the current-voltage performance of a properly functioning (good) device (501), a device with low (forward) on-current (502), and a device with high leakage under reverse bias (i.e., negative voltage) (503).
[0068] FIG. 6 is a plot comparing the current-voltage response of an as-fabricated PIN diode with the response predicted by a machine learning model using only optical images of the wafer surface as input in accordance with the present invention. There is good agreement between the experimental current-voltage data and the response predicted by the model. The sudden increase in source current is characteristic of the turn-on of the pn junction, and the model accurately represents this turn-on mechanism. The agreement between the experimental and predicted current-voltage responses demonstrates the validity of this neural network-based approach.
[0069] Benefits and New Features
[0070] The method of the present invention allows the electrical response of devices to be fabricated on a target semiconductor wafer to be determined prior to or at an early stage of fabrication, and allows the method of the present invention to determine the location, number, size, and shape of devices that can be fabricated on the wafer, reducing waste and increasing the efficiency of device fabrication.
[0071] What is unique about the method is the use of a neural network capable of predicting electrical output of a vertical semiconductor diode, including current-voltage data and capacitance-voltage data, based on optical images of a wafer, and the use of a neural network capable of predicting electrical output parameters of a semiconductor device based on optical images of a wafer.
[0072] The present invention provides a method for predicting the electrical output of a semiconductor device without test electrical data or post-final lithography imaging.
[0073] The present invention also provides a method for predicting the pass / fail yield of die on a semiconductor wafer, either by directly predicting the pass / fail of each die, or by first predicting the electrical response of devices subsequently fabricated on similar wafers.
[0074] The present invention provides a method for analyzing the electrical performance of devices fabricated on a wafer by neural networks based on non-destructive structural characterization of the wafer at an early stage of fabrication.
[0075] The present invention provides a method for evaluating a single device or multiple devices across an entire wafer.
[0076] The processed devices can be any lateral electronic device such as a diode, MOSFET, JFET, HEMT, HBT, or any vertical electronic device such as a diode, MOSFET, JFET, thyristor, HBT, or the like.
[0077] In other cases, the processed device can be any optoelectronic device such as an LED or micro LED, a laser diode, a solar cell, a thin film transistor, a photovoltaic cell, etc.
[0078] Semiconductor materials that can be evaluated using the method of the present invention include, for example, SiC, BN, GaN / AlGaN / InGaN / InAlGaN, ZnO / MgZnO, ZnSe / ZnMgSe, CdTe / CdMnTe, ZnS, Ga2O3 / (Al x Ga 1-x )2O3, or diamond.
[0079] Semiconductor materials that can be evaluated using the methods of the present invention also include any mid or narrow bandgap semiconductor, such as Si, Ge, SiGe, GaAs, AlGaAs, InGaAs, InP, GaP, CuO2, CuO, CuS, CuInGaSe2, CuZnSnS2, and similar alloys.
[0080] Materials that can be evaluated using the methods of the present invention further include organic semiconductor materials such as pentacene, poly(3-hexylthiophene) (P3HT), poly(3-alkylthiophene) (P3AT), poly(3-octylthiophene) (P3OT), phenyl-C61-butyric acid methyl ester (PCBM), and poly(3,4-ethylenedioxythiophene) polystyrene sulfonate (PEDOT:PSS).
[0081] The present invention provides an approach to predicting the response of semiconductor devices based on optical, structural, and electrical wafer maps and images taken or extracted prior to final processing including contact metallization. Properly designed neural networks inherently learn and respond not only to features that are easily identified by experts, but also to features or relationships between such features that are currently unknown or nearly imperceptible to experts.
[0082] While particular embodiments, aspects, and features have been described, one of ordinary skill in the art will readily appreciate that the invention described herein is not limited to only such embodiments, aspects, and features, but also contemplates all modifications and alternative embodiments that are within the spirit and scope of the underlying invention described and claimed herein. The present application contemplates all modifications that are within the spirit and scope of the underlying invention described and claimed herein, and all such modifications and alternative embodiments are deemed to be within the scope and spirit of the present disclosure.
Claims
1. 1. A neural network based semiconductor wafer evaluation method, comprising: performing an initial evaluation of the semiconductor wafer; inputting data of the initial evaluation into a neural network, the neural network performing iterative evaluations of the wafer based on the initial evaluation and subsequent evaluations performed by the neural network; outputting data relating to at least one performance metric of electronic devices fabricable on said wafer; A semiconductor wafer evaluation method comprising:
2. 2. The method of claim 1, wherein the data input to the neural network includes data from a previous evaluation of a wafer to the neural network.
3. 2. The method of claim 1, wherein the data input to the neural network includes optical microscope images, optical profilometry images, Nomarski images, fluorescent images, or ultraviolet images of the wafer being evaluated or a previous wafer.
4. 2. The method of claim 1, wherein the data input to the neural network includes photoluminescence images, electroluminescence images, cathodoluminescence images, and / or Auger spectroscopy images of the wafer being evaluated or a previous wafer.
5. 2. The method of claim 1, wherein the data input to the neural network includes scanning electron micrographs, atomic force microscope images, X-ray diffraction (XRD) mapping images, or Raman mapping images of the wafer being evaluated or a previous wafer.
6. 2. The semiconductor wafer evaluation method of claim 1, wherein the neural network is a multi-layer convolutional neural network (CNN).
7. 2. The method of claim 1, wherein the neural network is a gated recurrent unit (GRU) network, a long short-term memory (LSTM) network, or a similar recurrent neural network.
8. 2. The semiconductor wafer evaluation method according to claim 1, wherein the performance index is a current-voltage (IV) response or a capacitance-voltage (CV) response of an electronic device fabricated on at least one region of the wafer.
9. The semiconductor wafer evaluation method according to claim 1 , wherein the performance index is an on-resistance performance, a stable operating voltage, or a maximum operating frequency of the electronic device.
10. 2. The method of claim 1, wherein the performance index is a turn-on time or rate, or a turn-off time or rate, relative to an input change as a function of time.
11. 2. The method of claim 1, wherein the performance index is output power or gain as a function of frequency.
12. 2. The semiconductor wafer evaluation method of claim 1, wherein the output data includes data indicative of performance of devices fabricated at multiple locations on the wafer.
13. 2. The method of claim 1, wherein the output data includes data indicative of an optimum geometric configuration or maximum or minimum size of a device to be fabricated at a given location on the wafer.
14. 2. The method of claim 1, wherein the output data comprises data indicative of a Boolean pass / fail metric for devices fabricated at given locations on the wafer.
15. 2. The method of claim 1, wherein the output data includes data indicative of a probability that a device fabricated at a given location on the wafer will exceed a predetermined functionality threshold.
16. The semiconductor wafer evaluation method of claim 1 , wherein the output data includes data on at least one performance index of an optoelectronic device fabricable on the wafer.
17. The semiconductor wafer evaluation method according to claim 1 , further comprising the step of outputting data relating to at least one performance index of integrated circuits that can be fabricated on the wafer.
Citation Information
Patent Citations
Method and apparatus for data analysis
JP2009283947A
Captioning region of image
JP2019008778A
Evaluating quality of a product such as a semiconductor substrate
US20190362221A1
Method for predicting electric characteristics of semiconductor element
WO2020234685A1
US11,415,518
