Infrared focal plane process diagnosis and optimization method based on machine learning

By extracting the spectral feature vectors of the infrared focal plane array and combining them with machine learning and physical simulation, the problem of the disconnect between spectral features and physical mechanisms was solved, achieving high-precision response rate prediction and process diagnosis, and improving the imaging quality and detection performance of the infrared system.

CN121236033AActive Publication Date: 2025-12-30SHANGHAI INSTITUTE OF TECHNICAL PHYSICS CHINESE ACADEMY OF SCIENCES

Patent Information

Application Number
CN202511397753.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-09-28
Publication Date
2025-12-30
Estimated Expiration
2045-09-28

AI Technical Summary

Technical Problem

Existing technologies struggle to establish a direct link between spectral characteristics and the physical mechanisms of infrared focal plane arrays, resulting in low performance prediction accuracy and an inability to trace process defects, thus affecting the detection range, temperature resolution, and imaging quality of infrared systems.

Method used

By collecting photocurrent spectrum data from a focal plane array, preprocessing it, extracting spectral feature vectors, combining it with a machine learning model to predict the response rate, and using semiconductor physics simulation to establish a quantitative correlation between key spectral features and manufacturing process parameters, process diagnosis and optimization can be achieved.

Benefits of technology

It achieved high-precision prediction of the response rate of each pixel in the infrared focal plane array, revealed the process root cause of performance non-uniformity, guided process optimization, and improved imaging quality and detection performance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121236033A_ABST
    Figure CN121236033A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of semiconductor photoelectric device test and diagnosis, in particular to an infrared focal plane process diagnosis method based on machine learning, which comprises the following steps of: firstly, collecting photocurrent spectrum data of a plurality of pixel units on a focal plane array, and extracting a spectrum feature vector after preprocessing; secondly, a machine learning model is trained to obtain a performance prediction model, and key spectral features are obtained through feature importance evaluation; and finally, in combination with semiconductor physical simulation, establishing quantitative correlation between the key spectral characteristics and manufacturing process parameters, and completing diagnosis of production process links. The invention also comprises an optimization method. According to the method, spectral deep form information neglected by a conventional method is deeply mined and quantified, high-precision prediction of the response rate of each pixel of the whole focal plane array is realized by constructing characteristics closely related to a device microscopic physical mechanism, and correlation between spectral characteristics and process structure parameters is established, so that the accuracy of the device is improved. And the performance optimization and the process rapid iteration of the infrared detector chip are guided.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of semiconductor optoelectronic device testing and diagnostic technology, and in particular to a machine learning-based method for infrared focal plane array process diagnostics and optimization. Background Technology

[0002] Infrared focal plane array detectors are core components of infrared thermal imaging systems, with wide applications in military, aerospace, industrial, and civilian fields. The responsivity uniformity of the detector array is one of the key indicators for evaluating its imaging quality. However, due to limitations in complex material growth and device fabrication processes, responsivity non-uniformity is a common problem among pixel units in large-scale focal plane arrays. This non-uniformity severely affects the detection range, temperature resolution, and imaging quality of infrared systems. Therefore, the ability to quickly and accurately predict the responsivity of all pixels on a focal plane array and identify the process-related causes of persistent performance non-uniformity is crucial for guiding process optimization and iteration, and improving device performance.

[0003] Currently, existing methods attempt to fit empirical or semi-empirical functions to the entire photocurrent spectrum, using optimization algorithms to extract the parameters of the fitted function formula as features describing the spectrum. These mathematically significant features are then used to build a machine learning model with performance parameters. However, the drawback of this method is that the extracted spectral features lack clear physical meaning and cannot be directly linked to the physical characteristics of the device and the manufacturing process, thus making it difficult to provide effective guidance for performance evaluation and process improvement.

[0004] Meanwhile, some existing technologies also apply machine learning to improve the performance of infrared focal plane arrays, but their technical approaches do not utilize the physical information of the device. For example, some methods focus on post-processing in the image domain, using neural networks to filter or compensate the gray values ​​of already imaged pixels to correct blind pixels or eliminate fixed-pattern noise. These methods extract local statistical features of pixel gray values, essentially belonging to image inpainting, and cannot reveal the device's physical mechanisms behind performance non-uniformity, let alone provide guidance for front-end manufacturing process optimization. Other methods remain at the level of using general deep learning models (such as generative adversarial networks) to predict and repair the output of defective pixels. Their features mainly come from the gray-level distribution patterns around the pixels, similarly lacking a connection with physical processes such as device band structure and carrier transport. Because these methods fail to deeply utilize the physical information of the device, they cannot fundamentally solve the performance uniformity problem.

[0005] Some studies also use machine learning to predict device performance based on theoretical design parameters (such as the thickness, composition, and doping concentration of each layer). While these methods are valuable in device design and simulation, their precise internal physical parameters are not directly known for already fabricated chips, making them unsuitable for performance prediction and process diagnostics in actual products.

[0006] Meanwhile, physical enhancement methods that combine physical knowledge with machine learning have achieved significant results in research in other fields. Therefore, there is an urgent need in this field for a spectral feature extraction method with physical implications, which can be linked to performance data and combined with machine learning models to further reveal the physical mechanisms and technological roots behind performance fluctuations. Summary of the Invention

[0007] The purpose of this invention is to provide an infrared focal plane process diagnosis and optimization method based on machine learning, which mainly solves the problems of the existing technology, such as the disconnect between spectral features and device physical mechanisms, low performance prediction accuracy, and inability to trace process defects.

[0008] To achieve the above objectives, the technical solution adopted by this invention is to provide a machine learning-based infrared focal plane array process diagnostic method. This method utilizes the spectral data of a focal plane array to predict and diagnose the manufacturing process of the focal plane array. The method is characterized by comprising the following steps: Step S100: Collect photocurrent spectrum data of multiple pixel units on the focal plane array.

[0009] Step S200: Preprocess the photocurrent spectrum data to obtain preprocessed spectral data.

[0010] Step S300: Based on the preprocessed spectral data, extract spectral feature vectors. Each spectral feature vector corresponds one-to-one with a pixel unit.

[0011] Step S400: Using the spectral feature vector as input and the measured response rate of the corresponding pixel unit as output, train a machine learning model to obtain a performance prediction model for predicting the response rate of any pixel unit within the focal plane array.

[0012] Step S500: Perform feature importance assessment on the performance prediction model and obtain the key spectral features that have the greatest impact on the response rate from the spectral feature vector.

[0013] Step S600: Combine semiconductor physics simulation to establish a quantitative correlation between the key spectral features and manufacturing process parameters, which is used to diagnose the production process.

[0014] Furthermore, after step S600, the process also includes a verification of the key spectral features, comprising the following steps: Step S701: Based on the reference manufacturing process parameters, the first simulated absorption spectrum of each pixel unit is obtained using semiconductor physical simulation.

[0015] Step S702: Preprocess the first simulated absorption spectrum to obtain the first preprocessed simulated spectrum data.

[0016] Step S703: Based on the first preprocessed simulated spectral data, extract the first simulated spectral feature vector. The first simulated spectral feature vector corresponds one-to-one with the pixel unit.

[0017] Step S704: Fine-tune the manufacturing process parameters and use semiconductor physics simulation to obtain the second simulated absorption spectrum of each pixel unit.

[0018] Step S705: Preprocess the second simulated absorption spectrum to obtain the second preprocessed simulated spectrum data.

[0019] Step S706: Based on the second preprocessed simulated spectral data, extract the second simulated spectral feature vector. The second simulated spectral feature vector corresponds one-to-one with the pixel unit.

[0020] Step S707: Compare the first simulated spectral feature vector and the second simulated spectral feature vector, and find the spectral feature with the greatest difference as the key spectral feature for simulation.

[0021] Step S708: Compare the simulated key spectral features with the key spectral features. If they belong to the same spectral feature, the key spectral feature verification is passed; otherwise, the key spectral feature is determined not to meet the physical process requirements.

[0022] Further, in step S200, step S702, or step S705, the preprocessing first normalizes the photocurrent spectrum data, the first simulated absorption spectrum, or the second simulated absorption spectrum of each pixel unit, and then performs first-order differential processing on the normalized data to obtain the differential spectrum, which serves as the preprocessed spectral data, the first preprocessed simulated spectral data, and the second preprocessed simulated spectral data.

[0023] Further, in step S300, step S703, or step S706, the spectral feature vector, the first simulated spectral feature vector, and the second simulated spectral feature vector are composed of the spectral characteristics of multiple peaks in the same pixel unit, obtained from the differential spectrum in the preprocessed spectral data, the first preprocessed simulated spectral data, or the second preprocessed simulated spectral data. The spectral characteristics of each peak include peak position, full width at half maximum (FWHM), positive peak value, negative peak value, and symmetry.

[0024] Furthermore, the spectral feature vector, the first simulated spectral feature vector, and the second simulated spectral feature vector all originate from the main peak, the first-order peak, the second-order peak, and the third-order peak within the same pixel unit.

[0025] Furthermore, the peak position is determined by the zero-crossing position of the peak pair formed by the left positive peak and the right negative peak in the differential spectrum. The full width at half maximum (FWHM) is represented by the wavelength position difference between the left positive peak and the right negative peak. The peak value of the positive peak is represented by the peak value of the left positive peak, and the peak value of the negative peak is represented by the peak value of the right negative peak. The symmetry is represented by the sum of the peak values ​​of the left positive peak and the right negative peak.

[0026] Further, in step S400, training the machine learning model includes the following steps: Step S401: Associate the spectral feature vector and the measured response rate corresponding to the same pixel unit, so that each pixel unit corresponds to one sample. The samples of all the pixel units form a sample set.

[0027] Step S402: Divide the sample set into a training set and a test set. The number of samples in the training set is greater than the number of samples in the test set.

[0028] Step S403: The training set is fed into the machine learning model to generate the performance prediction model.

[0029] Step S404: Verify the effectiveness of the performance prediction model using the test set. If the test fails, proceed to step S403 to continue training. If the test passes, save the performance prediction model as output.

[0030] Furthermore, the machine learning model is a random forest model.

[0031] Furthermore, in step S600, the manufacturing process parameters include the offset of grating overlay in the X direction, the offset of grating overlay in the Y direction, the offset of etching depth, and the deflection of overlay angle.

[0032] This invention also discloses an optimization method for the above-mentioned machine learning-based infrared focal plane array process diagnostic method, characterized by comprising the following steps: Step S801: Obtain the performance prediction model obtained in step S400.

[0033] Step S802: Use the performance prediction model to perform reverse optimization to find the optimal combination of spectral features corresponding to the maximum response rate.

[0034] Step S803: Obtain the quantitative correlation between the key spectral features obtained in step S600 and the manufacturing process parameters.

[0035] Step S804: Using the quantization correlation, the manufacturing process parameters are deduced from the optimal spectral feature combination to obtain the optimal manufacturing process parameters.

[0036] In view of the above technical features, this invention utilizes photocurrent response spectroscopy to accurately predict the responsivity of each pixel unit in an infrared quantum well focal plane array detector and diagnoses process fluctuations that cause performance inhomogeneity. Compared with existing technologies, it has the following significant advantages: 1. This invention, by normalizing and then differentiating the original photocurrent spectrum, highlights the minute spectral deviation signal caused by process fluctuations that determines performance differences, and further sharpens the details of the deviation signal.

[0037] 2. This invention employs a physics-enhanced machine learning method to extract spectral features closely related to the physical processes of the device. In particular, it reveals the key indicative role of secondary absorption peak features in performance, enabling in-depth mining and quantification of deep spectral morphological information that is ignored by conventional methods, and achieving high-precision prediction of the response rate of each pixel in the entire focal plane array.

[0038] 3. After verifying the high-precision model obtained from the training, this invention uses the high-precision model to identify the key spectral features that play a decisive role in the response rate, and establishes the correlation between spectral features and process structure parameters by combining semiconductor physical simulation. By extracting physically meaningful spectral features, it provides a basis for tracing process problems, thus forming a closed loop from performance prediction to process diagnosis, which is used to guide the performance optimization and process diagnosis improvement of infrared detector chips. Attached Figure Description

[0039] Figure 1 This is a flowchart of a preferred embodiment of the infrared focal plane process diagnostic method based on machine learning of the present invention. Figure 2 This is a diagram illustrating the preprocessed photocurrent spectrum data in a preferred embodiment of the infrared focal plane process diagnostic method based on machine learning of the present invention. Figure 3 This is a schematic diagram illustrating the extraction of spectral feature vectors in a preferred embodiment of the infrared focal plane process diagnostic method based on machine learning of the present invention. Figure 4 This is a scatter plot of the predicted response rate and the actual response rate after the machine learning model is trained in a preferred embodiment of the infrared focal plane process diagnostic method based on machine learning of the present invention. Figure 5 This is a preferred embodiment of the infrared focal plane array process diagnostic method based on machine learning of the present invention, in which an importance analysis is performed to obtain a ranking diagram of key spectral features; Figure 6 This is a heatmap of simulated spectral feature changes when verifying key spectral features in a preferred embodiment of the infrared focal plane process diagnostic method based on machine learning of the present invention. Figure 7 This is a flowchart of a preferred embodiment of the optimization method of the infrared focal plane process diagnostic method based on machine learning of the present invention. Detailed Implementation

[0040] The present invention will be further described below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0041] Please see Figure 1 This invention discloses a machine learning-based method for diagnosing infrared focal plane array processes. It collects spectral data of the focal plane array, combines it with a machine learning model to identify the spectral features that most significantly affect the performance of the infrared focal plane array, and then uses this information to further predict and diagnose the manufacturing process of the focal plane array. As shown in the figure, a preferred embodiment includes the following steps: Step S10: Collect raw spectral data.

[0042] For a finished infrared focal plane array, the photocurrent spectrum data of each pixel unit is read to obtain the responsivity data for each pixel. In this embodiment, a quantum well infrared focal plane detector chip is used as the research object, and its infrared focal plane array has a size of 320×256, that is, it contains a total of 81920 pixel units. In this step, a total of 81920 sets of data need to be collected, corresponding to the photocurrent spectrum and responsivity data of each pixel unit.

[0043] Step S20: Preprocess the photocurrent spectrum data.

[0044] In the same infrared focal plane array, the light intensity received by different pixel units will vary depending on their position. In order for the subsequent machine learning model to effectively process the information of each pixel unit, all photocurrent spectrum data are preprocessed to eliminate the differences introduced by position, and the preprocessed spectral data is used by the machine learning model.

[0045] The preprocessing involves two steps. First, the raw spectral data is normalized to obtain the normalized photocurrent spectrum. Then, the derivative of the photocurrent spectrum is calculated based on the normalized photocurrent spectrum to obtain the differential spectrum. The differential spectrum serves as the final preprocessed spectral data and is used in subsequent stages. Figure 2 This embodiment shows some of the preprocessed photocurrent spectrum data.

[0046] Step S30: Extract the spectral feature vector.

[0047] In the preprocessed spectral data, specified spectral feature vectors are extracted to form the input of the machine learning model. Each pixel unit has its own preprocessed spectral data, so the extracted spectral feature vectors also correspond one-to-one with each pixel unit.

[0048] Please see Figure 3 This demonstrates the details of extracting spectral feature vectors using differential spectroscopy. Each spectrum contains multiple peaks. For each peak, five features are extracted: peak position, full width at half maximum (FWHM), positive peak value, negative peak value, and symmetry, which are used as the spectral characteristics of that peak. In this embodiment, five features are extracted from each of the four peaks in each spectrum; that is, for each pixel unit, its main peak, first-order peak, second-order peak, and third-order peak are extracted, totaling twenty features, which are used as the spectral feature vector of that pixel unit. Specifically, these features are retrieved and extracted using differential spectroscopy. In a spectrum, a peak point corresponds to a point where its first derivative equals zero. By identifying peak pairs consisting of a left positive peak and a right negative peak in the differential spectrum, the peak position of the spectrum can be accurately located, and then the FWHM, the peak value of the positive peak, the peak value of the negative peak, and the symmetry can be calculated. First, the peak position is determined by the zero-crossing position of the peak pair consisting of the left positive peak and the right negative peak in the differential spectrum, thereby locating the positions of the main peak, the first-order peak, the second-order peak, and the third-order peak. Then, for each peak, the half-maximum width (WHM) is obtained from the wavelength difference between the left positive peak and the right negative peak; the positive peak value is obtained from the peak value of the left positive peak; the negative peak value is obtained from the peak value of the right negative peak; and the sum of the peak values ​​of the left positive peak and the right negative peak represents the symmetry. Deviations in symmetry reflect the asymmetry of the spectral shape and are related to physical mechanisms such as non-uniform broadening.

[0049] The position and width of the main peak primarily reflect the most basic band structure design of the device and the general properties of the material. Secondary peaks originate from more complex physical processes, such as transitions to higher energy levels, coupling between energy levels, interface scattering effects, or coupling of diffraction modes introduced by microstructures such as gratings with quantum wells. The morphology of secondary peaks is highly sensitive to minute fluctuations in the device's microstructure (such as layer thickness uniformity, interface roughness, and process deviations such as photolithography overlay precision).

[0050] Step S40: Train the machine learning model.

[0051] Using the spectral feature vector of a pixel as input and the measured response rate of the pixel as output, a machine learning model is trained to obtain a performance prediction model. This performance prediction model is capable of predicting the response rate of any pixel within a focal plane array based on its spectral feature vector.

[0052] Step S41: Form a sample set.

[0053] The spectral feature vector and measured response rate are classified on a pixel-by-pixel basis. Each pixel should have a corresponding spectral feature vector and a sample of its measured response rate. All 81,920 pixels correspond to 81,920 samples, which together constitute the sample set.

[0054] Step S42: Divide the dataset into training and test sets.

[0055] The sample set is divided into a training set and a test set. The training set is used to train the machine learning model, while the test set is used to evaluate the training results. The number of samples in the training set must be greater than the number of samples in the test set. In this embodiment, 81,920 samples are randomly divided into training and test sets. In this embodiment, 10% (approximately 8,192 pixel samples) is used as the test set, and the remaining 90% is used as the training set.

[0056] Step S43: Train the machine learning model.

[0057] In this embodiment, the Random Forest algorithm, which has high stability and does not require complex parameter tuning, is selected as the performance prediction model, and training data is read for training.

[0058] Step S44: Verify the training results.

[0059] After the performance prediction model is trained, the test set is fed into the model to obtain the predicted response rate for each pixel. The predicted response rate is then compared with the measured response rates in the test set to evaluate the effectiveness of the performance prediction model. For example, the predicted and measured response rates can be plotted as a scatter plot, and the coefficient of determination R² can be calculated. (See also...) Figure 4In the scatter plot of a trained performance prediction model in this embodiment, it can be seen that the data points are closely clustered near the diagonal of y=x, and the corresponding coefficient of determination R² is 0.92. This indicates that the established performance prediction model has extremely high prediction accuracy.

[0060] If the test fails, proceed to step S43 to continue training, or provide more original samples for retraining. Once the test passes, save the performance prediction model as output.

[0061] Step S50: Obtain key spectral features.

[0062] Using a performance prediction model, feature importance analysis is performed to evaluate the performance. The spectral features that have the greatest impact on the response rate are identified from the spectral feature vector and used as key spectral features.

[0063] In this embodiment, the feature importance evaluation function built into the random forest model was used to rank the relative roles of twenty spectral features in the input spectral feature vector in response rate prediction. The results are as follows: Figure 5 As shown, the spectral features related to the first and third order peaks are ranked first. Therefore, we can conclude that the decisive factor for the final response rate is not the spectral features related to the main peak, but rather the spectral features related to the first and third order peaks.

[0064] Step S60: Diagnose the process steps.

[0065] By combining semiconductor physics simulations, a quantitative correlation is established between key spectral features and manufacturing process parameters. Using these conclusions, the production process can be diagnosed. Manufacturing process parameters include grating overlay offset in the X-direction, grating overlay offset in the Y-direction, etching depth offset, and overlay angle deflection.

[0066] In this embodiment, the conclusion is that the spectral characteristics of the first and third secondary peaks play a decisive role. According to semiconductor physics simulations, the spectra of the first and third secondary peaks are related to microstructural deviations in finer physical processes within the quantum well structure (such as inter-level coupling, interface scattering, and layer thickness uniformity). Therefore, process engineers should not be limited to adjusting the parameters related to the main peak when improving the process; instead, they can use simulations to study which specific process parameter fluctuations are most sensitive to the spectral shape deviations of the first and third secondary peaks, thereby enabling targeted process adjustments and improving the uniformity of device performance from the source.

[0067] Step S70: Verify key spectral features.

[0068] After obtaining the specific process parameters involved, semiconductor physics simulation is used again to simulate the deviation of the process parameters and evaluate whether the resulting changes in spectral characteristics are consistent with the key spectral characteristics predicted by the machine learning model.

[0069] Step S71: Obtain the reference simulated absorption spectrum.

[0070] Based on baseline manufacturing process parameters, a baseline simulated absorption spectrum for each pixel unit is obtained using semiconductor physics simulation, denoted as the first simulated absorption spectrum. For the infrared focal plane array in this embodiment, which has a size of 320×256 and contains a total of 81,920 pixel units, 81,920 sets of data were obtained after simulation, corresponding to the simulated absorption spectrum of each pixel unit.

[0071] Step S72: Preprocess the first simulated absorption spectrum.

[0072] The preprocessing of the simulated absorption spectrum is consistent with the preprocessing of the actual photocurrent spectrum acquired from the device. That is, the preprocessing of the first simulated absorption spectrum also involves two steps: first, the first spectral data is normalized to obtain the normalized spectrum; then, the derivative of the spectrum is calculated based on the normalized spectrum to obtain the differential spectrum. The differential spectrum, as the final preprocessed simulated spectral data, is delivered to subsequent stages.

[0073] Step S73: Extract the first simulated spectral feature vector.

[0074] Similar to the actual photocurrent spectrum, the first simulated spectral feature vector is extracted from the first preprocessed simulated spectral data derived from the simulation data in the same manner. Each simulated spectral feature vector corresponds one-to-one with a pixel unit. In this embodiment, for the simulated spectrum derived from the baseline process fabrication parameters, five characteristics are extracted from each of the four peaks. That is, for each pixel unit, its main peak, first-order peak, second-order peak, and third-order peak are extracted, totaling twenty features, which serve as the baseline spectral feature vector for that pixel unit, denoted as the first simulated spectral feature vector.

[0075] Step S74: Obtain the deviation simulation absorption spectrum.

[0076] The manufacturing process parameters are fine-tuned, and semiconductor physics simulation is used to obtain the reference simulated absorption spectrum for each pixel unit again, denoted as the second simulated absorption spectrum. The fine-tuned manufacturing process parameters include the offset of the grating overlay in the X direction, the offset of the grating overlay in the Y direction, the offset of the etching depth, and the deflection of the overlay angle, etc. For the infrared focal plane array in this embodiment, its size is 320×256, containing a total of 81920 pixel units. After completing the simulation, 81920 sets of data are obtained, corresponding to the simulated absorption spectrum of each pixel unit.

[0077] Step S75: Preprocess the second simulated absorption spectrum.

[0078] The preprocessing of the simulated absorption spectrum is consistent with the preprocessing of the actual photocurrent spectrum acquired from the device. That is, the preprocessing of the second simulated absorption spectrum also involves two steps: first, the second spectral data is normalized to obtain the normalized spectrum; then, the derivative of the spectrum is calculated based on the normalized spectrum to obtain the differential spectrum. The differential spectrum, as the final second preprocessed simulated spectral data, is delivered for subsequent use.

[0079] Step S76: Extract the second simulated spectral feature vector.

[0080] Similar to the actual photocurrent spectrum, the second preprocessed simulated spectral data from the simulation data is also processed in the same way to extract the second simulated spectral feature vector. Each second simulated spectral feature vector corresponds one-to-one with a pixel unit. In this embodiment, for the simulated spectrum from the baseline process fabrication parameters, five characteristics are extracted from each of the four peaks. That is, for each pixel unit, its main peak, first-order peak, second-order peak, and third-order peak are extracted, totaling twenty features, which serve as the baseline spectral feature vector for that pixel unit, denoted as the second simulated spectral feature vector.

[0081] Step S77: Obtain the key spectral features of the simulation.

[0082] The first simulated spectral feature vector reflects the effect after production with baseline process parameters, while the second simulated spectral feature vector reflects the effect after production with deviation (fine-tuning) process parameters. By comparing the two, and comparing twenty features one by one, the spectral feature with the greatest difference is identified as the key spectral feature for simulation. In this embodiment, a heatmap of the changes in simulated spectral features after process parameter adjustment is used to find the spectral feature with the greatest difference. Figure 6 A thermogram showing the changes in simulated spectral characteristics related to etching depth is presented.

[0083] Step S78: Compare key spectral features.

[0084] The key spectral features obtained from the semiconductor physics simulation reflect the correlation between changes in process parameters and changes in spectral features during the physics simulation. The key spectral features themselves come from the inference results of the machine learning model, reflecting the correlation between changes in process parameters and changes in spectral features given by the machine learning model. Comparing the two, if the key spectral features obtained by the two methods are consistent, then the conclusion given by the machine learning model is reliable and the verification is passed; otherwise, it is determined that the key spectral features do not satisfy the physical process and relearning is required.

[0085] The following explains the principle and basis of the infrared focal plane array process diagnostic method based on machine learning in this invention. The photocurrent spectrum of a quantum well infrared detector directly reflects its band structure, electronic states, and carrier transport characteristics. Every detail in the spectrum is closely related to the physical state and manufacturing process of the device. For example, the position of the main absorption peak with the highest intensity in the spectrum directly corresponds to the energy of the ground-state electron transition to the first excited state in the quantum well. Its peak width includes uniform broadening caused by phonon scattering, electron-electron scattering, etc., and also includes non-uniform broadening caused by process defects such as interface roughness, material composition inhomogeneity, and small fluctuations in quantum well layer thickness. However, the results of this embodiment show that the secondary absorption peaks, in addition to the main absorption peak, are more revealing of the root causes of performance differences between pixels. These secondary absorption peaks reflect more complex physical processes, and any small change in their morphology contains valuable information about specific process fluctuations. Therefore, quantifying and extracting features with clear physical meaning from the photocurrent spectrum, especially the secondary absorption peaks, can provide a direct window for understanding the physical root causes of device performance differences. By combining the aforementioned physical knowledge with machine learning methods, a complete analytical chain can be constructed, from spectral microscopic features to macroscopic performance and then to front-end process parameters, thereby revealing the physical mechanisms behind performance fluctuations and the root causes of process defects.

[0086] To demonstrate the technical advantages of this invention, we compared it with a conventional method in the art. This conventional method extracts empirical parameters such as the peak intensity, peak position, and full width at half maximum (FWHM) of the main absorption peak from the spectrum to model the data. Under the same test conditions, its prediction model has a coefficient of determination (R²) of only 0.82. In contrast, this invention, by extracting spectral physical features closely related to the device's physical processes, achieves a model R² value as high as 0.92. Figure 4 The results show that this invention reduces the model's prediction error by more than 55%, achieving an improvement in prediction accuracy. This improvement is not a simple parameter aggregation; the key lies in the fact that the physical feature extraction of this invention can deeply mine and quantify the deep spectral morphological information that conventional methods ignore.

[0087] Please see Figure 7 The present invention also discloses an optimization method for infrared focal plane array process diagnostics based on machine learning. As shown in the figure, a preferred embodiment includes the following steps: Step S81: Obtain the performance prediction model.

[0088] An infrared focal plane array process diagnostic method based on machine learning is used to obtain a performance prediction model after machine learning. For example, in this embodiment, a performance prediction model trained using a random forest model is used.

[0089] Step S82, reverse optimization.

[0090] By employing a performance prediction model for backpropagation optimization, the optimal spectral feature combination corresponding to the maximum response rate is found. In this embodiment, the random forest model has a backpropagation interface. By combining it with a global optimization algorithm, the predicted maximum response rate is used as the optimization objective, and the optimal feature combination that achieves this objective is searched within the multidimensional spectral feature space. This optimal feature combination describes the key spectral morphology that an ideal device should possess.

[0091] Step S83: Obtain the quantitative correlation.

[0092] By using a machine learning-based infrared focal plane array process diagnostic method, a quantitative correlation between key spectral features and manufacturing process parameters is obtained.

[0093] Step S84 yields the optimal manufacturing process parameters.

[0094] Quantitative correlation reveals the relationship between key spectral features and manufacturing process parameters. By using the optimal combination of spectral features as input, quantitative correlation can be used to deduce the manufacturing process parameters that achieve the optimal combination of spectral features. These optimal manufacturing process parameters can not only diagnose deviations in existing processes but also proactively guide the design and process parameter setting of next-generation devices, fundamentally achieving a leapfrog improvement in performance.

[0095] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A machine learning based infrared focal plane process diagnostic method, utilizing spectral data of a focal plane array, to predict and diagnose production processes of the focal plane array, characterized by, The method comprises the following steps: In step S100, photoelectric current spectrum data of a plurality of pixel units on the focal plane array is collected. In step S200, the photoelectric current spectrum data is preprocessed to obtain preprocessed spectrum data. In step S300, a spectrum feature vector is extracted based on the preprocessed spectrum data. The spectrum feature vector corresponds to the pixel unit one by one. In step S400, a machine learning model is trained by taking the spectrum feature vector as input and the measured response rate of the pixel unit as output, to obtain a performance prediction model for predicting the response rate of any pixel unit in the focal plane array. In step S500, the performance prediction model is evaluated for feature importance, and the key spectrum feature that has the greatest impact on the response rate is obtained from the spectrum feature vector. In step S600, a quantitative correlation between the key spectrum feature and the manufacturing process parameters is established by combining semiconductor physical simulation, for diagnosing the process of production.

2. The machine learning based infrared focal plane process diagnostic method of claim 1, wherein, After step S600, the key spectrum feature is reviewed, comprising the following steps: In step S701, a first simulation absorption spectrum of each pixel unit is obtained by semiconductor physical simulation based on the reference manufacturing process parameters. In step S702, the first simulation absorption spectrum is preprocessed to obtain first preprocessed simulation spectrum data. In step S703, a first simulation spectrum feature vector is extracted based on the first preprocessed simulation spectrum data. The first simulation spectrum feature vector corresponds to the pixel unit one by one. In step S704, a second simulation absorption spectrum of each pixel unit is obtained by semiconductor physical simulation after fine-tuning the manufacturing process parameters. In step S705, the second simulation absorption spectrum is preprocessed to obtain second preprocessed simulation spectrum data. In step S706, a second simulation spectrum feature vector is extracted based on the second preprocessed simulation spectrum data. The second simulation spectrum feature vector corresponds to the pixel unit one by one. In step S707, the first simulation spectrum feature vector and the second simulation spectrum feature vector are compared to find the spectrum feature with the greatest difference as the simulation key spectrum feature. In step S708, the simulation key spectrum feature and the key spectrum feature are compared. If they belong to the same spectrum feature, the review of the key spectrum feature is passed, otherwise it is determined that the key spectrum feature does not meet the physical process.

3. The machine learning based infrared focal plane process diagnostic method according to any of claims 1 or 2, characterized in that, In step S200, step S702 or step S705, the preprocessing is first normalizing the photoelectric current spectrum data, the first simulation absorption spectrum or the second simulation absorption spectrum of each pixel unit, and then performing first-order differential processing on the normalized data to obtain a differential spectrum as the preprocessed spectrum data, the first preprocessed simulation spectrum data and the second preprocessed simulation spectrum data.

4. The machine learning based infrared focal plane process diagnostic method according to any of claims 1 or 2, wherein, In step S300, step S703 or step S706, the spectral feature vector, the first simulated spectral feature vector and the second simulated spectral feature vector are composed of spectral characteristics of multiple peaks in the same pixel unit, which are obtained from the differential spectrum in the preprocessed spectrum data, the first preprocessed simulated spectrum data or the second preprocessed simulated spectrum data. The spectral characteristics of each peak include peak position, half-peak width, positive peak value, negative peak value and symmetry.

5. The machine learning based infrared focal plane process diagnostic method of claim 4, wherein, The spectral feature vector, the first simulated spectral feature vector and the second simulated spectral feature vector are from the main peak, the first secondary peak, the second secondary peak and the third secondary peak in the same pixel unit.

6. The machine learning based infrared focal plane process diagnostic method of claim 4, wherein, The peak position is determined by the zero-crossing position of the peak pair composed of the left positive peak and the right negative peak in the differential spectrum. The half-peak width is represented by the wavelength position difference between the left positive peak and the right negative peak. The positive peak value is represented by the peak value of the left positive peak, and the negative peak value is represented by the peak value of the right negative peak. The symmetry is represented by the sum of the peak values of the left positive peak and the right negative peak.

7. The machine learning based infrared focal plane process diagnostic method of claim 1, wherein, In step S400, the machine learning model is trained, including steps: Step S401, associate the spectral feature vector corresponding to the same pixel unit with the measured response rate, then each pixel unit corresponds to a sample. The samples of all pixel units form a sample set; Step S402, divide the sample set into a training set and a test set. The number of samples in the training set is greater than the number of samples in the test set; Step S403, input the training set into the machine learning model to generate the performance prediction model; Step S404, verify the effectiveness of the performance prediction model using the test set. If the test fails, go to step S403 to continue training. If the test passes, save the performance prediction model as output.

8. The machine learning based infrared focal plane process diagnostic method of any of claims 1 or 7, wherein, The machine learning model is a random forest model.

9. The machine learning based infrared focal plane process diagnostic method of claim 1, wherein, In step S600, the manufacturing process parameters include grating overlay offset in X direction, grating overlay offset in Y direction, etching depth offset and overlay angle deflection.

10. An optimization method using the machine learning-based infrared focal plane process diagnostic method of claim 1, characterized by, Including steps: Step S801, obtain the performance prediction model obtained in step S400; Step S802, use the performance prediction model to perform inverse optimization to find the optimal spectral feature combination corresponding to the maximum response rate; Step S803, obtain the quantitative association between the key spectral features and the manufacturing process parameters obtained in step S600; Step S804, use the quantitative association to back-propagate the manufacturing process parameters from the optimal spectral feature combination to obtain the optimal manufacturing process parameters.

Citation Information

Patent Citations

  • Multispectral ranging / imaging sensor arrays and systems

    CN112805584A

  • Method for simulating electrical property of wafer chip and semiconductor process method

    CN114823398A

  • Substrate process endpoint detection using machine learning

    CN117441177A

  • Real-time vision and characteristic spectrum laser welding interconnection monitoring method and device

    CN118371857A

  • Ancient ceramic breaking method based on infrared spectrum

    CN118447944A

Cited By

  • Optical measurement model acquisition method, optical measurement method and device, and storage medium

    CN121682779A