A Deep Learning-Based Method for Predicting and Tracing Semiconductor Pollutants

Abstract

This invention discloses a deep learning-based method for predicting and tracing semiconductor contaminants, relating to the field of semiconductor pollution monitoring technology. The method includes simultaneously acquiring raw spectral data of gaseous contaminants from multiple monitoring devices in a semiconductor plant; generating standardized mass spectrometry time series data through time alignment and signal intensity normalization; comparing the data with a pollution source inventory database to identify contaminant types and preliminary concentrations; inputting the data into a pre-trained neural network model predicting spatiotemporal evolution to generate a predicted field for the spatiotemporal distribution of contaminant concentrations; calculating the mass migration flux of contaminants; and, combined with the plant's spatial structure, using a deconvolution diffusion model to simulate the diffusion process and deduce the location of potential release sources. This invention enables accurate identification, spatiotemporal prediction of concentrations, and precise source tracing of new gaseous contaminants in semiconductor plants, adapting to the contaminant control needs of complex semiconductor industrial plants.

Description

Technical Field

[0001] This invention belongs to the field of semiconductor pollution monitoring technology, specifically a method for predicting and tracing semiconductor pollutants based on deep learning. Background Technology

[0002] In current semiconductor plant gaseous pollutant control processes, gaseous pollutant spectral data are typically collected using individual monitoring devices. Pollutant types are identified through conventional comparison methods, and general simulation tools are used in conjunction with plant ventilation, temperature, and humidity parameters to extrapolate pollutant diffusion. The source location of pollutants is then determined manually based on concentration data from monitoring points. However, the composition of new gaseous pollutants in semiconductor plants is complex, the number of monitoring devices is large, and the collected data exhibits temporal and intensity variations. Conventional data processing methods do not provide standardized processing for multi-source spectral data, making pollutant identification results prone to inaccuracies.

[0003] Conventional pollutant analysis methods can only achieve forward concentration estimation and cannot establish a precise nonlinear mapping relationship between pollutant concentration and plant environmental parameters, resulting in insufficient accuracy in predicting the spatiotemporal distribution of pollutants. Existing source tracing methods cannot combine concentration gradients and airflow paths to quantify the pollutant migration process; they can only rely on simple models to perform forward diffusion extrapolation, making it difficult to reverse the pollutant diffusion path and accurately locate the pollutant release source.

[0004] The raw spectral data of gaseous pollutants collected by multiple monitoring devices have problems such as temporal misalignment and uneven signal intensity. They need to be standardized and accurately matched with a dedicated pollution source inventory database. It is also necessary to quantify the migration flux based on the spatiotemporal distribution results of pollutant concentrations and use a dedicated model to reverse the diffusion process to determine the location of pollutant release sources. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art;

[0006] To this end, this invention proposes a deep learning-based method for predicting and tracing semiconductor contaminants, comprising:

[0007] Multiple sets of raw spectral data of gaseous pollutants were simultaneously acquired from various monitoring devices deployed within the semiconductor industrial plant.

[0008] The original spectral data of multiple gaseous pollutants are time-aligned and signal intensity normalized to form standardized pollutant mass spectrum time series data. The standardized pollutant mass spectrum time series data is then compared with the characteristic mass spectra in a preset pollution source list database to identify the types and preliminary concentration information of new gaseous pollutants in semiconductor industrial plants.

[0009] The types and preliminary concentration information of the new gaseous pollutants identified in the semiconductor industrial plant are input into a pre-trained spatiotemporal evolution prediction neural network model. The pre-trained spatiotemporal evolution prediction neural network model has learned the nonlinear mapping relationship between the historical pollutant concentration distribution and the plant's ventilation layout, temperature and humidity field. The pre-trained spatiotemporal evolution prediction neural network model is used to calculate and generate a predicted spatiotemporal distribution field of new pollutant concentration in different areas in the future time period.

[0010] Based on the predicted spatiotemporal distribution field of the new pollutant concentration, the mass migration flux of the new pollutant driven by the concentration gradient and airflow path inside the plant is calculated. Combined with the spatial structure information inside the plant, the diffusion process of the pollutant is simulated in reverse using the deconvolution diffusion model, and the locations of one or more potential release sources of the new gaseous pollutant in the semiconductor plant are deduced.

[0011] Furthermore, the original spectral data of the gaseous pollutants from multiple sets are subjected to time alignment and signal intensity normalization to form standardized pollutant spectral time series data, including:

[0012] The raw spectral data of the gaseous pollutants includes the mass-to-charge ratio signal, signal intensity, and corresponding sampling timestamp;

[0013] A unified time reference is set, and the mass-to-charge ratio signals, signal strengths and sampling timestamps from different portable monitoring devices are interpolated and calibrated to ensure that the time coordinates of all data points are aligned.

[0014] For the signal strength of each of the mass-to-charge ratio signal channels, a baseline subtraction calculation is performed based on the historical background signal strength of the corresponding portable monitoring device;

[0015] For the signal strength data after baseline subtraction, the average response intensity of all monitoring devices in a clean air environment is used as the calibration reference, and intensity normalization scaling is performed to unify the signal strength of different devices to the same dimension scale.

[0016] The data of all mass-to-charge ratio signal channels, after being aligned in time and normalized in intensity, are integrated in chronological order to construct multidimensional standardized pollutant mass spectrum time series data. Each time point in the standardized pollutant mass spectrum time series data corresponds to a vector containing the signal intensities of multiple mass-to-charge ratio channels.

[0017] Furthermore, the standardized pollutant mass spectrum time series data is compared with the characteristic mass spectra in a pre-set pollution source inventory database to identify the types and preliminary concentration information of new gaseous pollutants in semiconductor industrial plants, including:

[0018] From the pollution source inventory database, standard characteristic mass spectra of various known gaseous pollutants from semiconductor processes are extracted. Each standard characteristic mass spectrum contains a specific mass-to-charge ratio and a corresponding standard relative intensity distribution.

[0019] In the standardized pollutant mass spectrum time series data, for each time slice vector, the similarity distance metric between it and each standard characteristic mass spectrum is calculated;

[0020] Select standard feature mass spectra with similarity distance metric below a preset recognition threshold, and determine the corresponding known semiconductor process gaseous contaminants as the contaminant types identified in the time slice;

[0021] Based on the ratio of the signal intensity of the standard characteristic mass spectrum corresponding to the identified pollutant type to the measured vector in the characteristic mass-to-charge ratio, and combined with the preset calibration curve corresponding to the identified pollutant type, preliminary concentration information is calculated.

[0022] By integrating the pollutant types identified continuously over time with the preliminary concentration information, a time series of identification results for new gaseous pollutants in the semiconductor plant is formed.

[0023] Furthermore, the types and preliminary concentration information of the newly identified gaseous pollutants in the semiconductor industrial plant are input into a pre-trained neural network model for predicting spatiotemporal evolution, including:

[0024] A deep recurrent neural network containing encoder and decoder structures is constructed as the basic architecture of the neural network model for predicting spatiotemporal evolution laws.

[0025] The types and preliminary concentrations of new gaseous pollutants in the semiconductor industrial plant at multiple historical moments are combined with temperature data, humidity data, and ventilation system wind speed and direction data collected at multiple points inside the plant at the same historical moment to form the model input feature vector sequence.

[0026] Use the actual pollutant concentration distribution field at the next moment in the historical time series as the target output label for model training;

[0027] The deep recurrent neural network is iteratively trained using massive amounts of historical working condition data until the error between its predicted output and the target output label is less than a preset threshold, thereby obtaining the pre-trained spatiotemporal evolution law prediction neural network model.

[0028] In the actual prediction stage, the latest identification result time series and real-time factory environment data are input into the pre-trained spatiotemporal evolution law prediction neural network model, and the model directly outputs the spatiotemporal distribution prediction field of new pollutant concentration in different regions within the future time period.

[0029] Furthermore, based on the predicted spatiotemporal distribution field of the new pollutant concentration, the mass migration flux of the new pollutant within the plant, driven by the concentration gradient and airflow path, is calculated, including:

[0030] From the spatiotemporal distribution prediction field of the new pollutant concentration, extract three-dimensional spatial concentration grid data for several consecutive future time periods;

[0031] The gradient of the three-dimensional spatial concentration grid data is calculated in three orthogonal directions of each spatial grid cell to obtain the concentration gradient field;

[0032] Obtain the steady-state airflow path field inside the factory, which is pre-generated by the factory ventilation system layout and computational fluid dynamics simulation;

[0033] The concentration gradient field and the steady-state airflow path field are multiplied in each spatial grid cell, and then multiplied by the molecular diffusion coefficient of the pollutant to calculate the preliminary mass migration flux vector of each spatial grid cell at each future time.

[0034] The preliminary mass migration flux vector is integrated over a specified time interval, and the reverse path is traced along the steady-state airflow path field to deduce the main mass transport path of pollutants in space.

[0035] Furthermore, by combining the spatial structure information inside the factory building and using a deconvolution diffusion model to simulate the diffusion process of pollutants, the locations of one or more potential release sources of new gaseous pollutants in the semiconductor factory building are deduced, including:

[0036] A convolutional diffusion forward model is established to describe the diffusion process of gaseous pollutants in the enclosed space of a factory. The kernel function of the convolutional diffusion forward model is determined by the spatial structure, ventilation conditions, and turbulence intensity parameters.

[0037] The current pollutant concentration distribution is used as the output of the convolutional diffusion forward model, and the model is run in reverse, with the historical pollutant source release distribution as the input to be solved.

[0038] Using an iterative inversion algorithm, the spatial distribution of historical release sources that minimizes the difference between the output of the convolutional diffusion forward model and the predicted field of the spatiotemporal distribution of the new pollutant concentration is obtained. The region in the spatial distribution of historical release sources with a release intensity greater than the background value is the location of the potential release source.

[0039] For each of the potential release source locations, calculate the historical time variation curve of its release intensity;

[0040] The spatial coordinates of the potential release source location and the time variation curve of the release intensity are compared with the geographical location of the equipment and its periodic operating conditions recorded in the equipment operation process parameter database.

[0041] Further, gradient calculations are performed on the three-dimensional spatial concentration grid data in three orthogonal directions of each spatial grid cell to obtain the concentration gradient field, including:

[0042] The concentration values ​​of each spatial grid cell in the three orthogonal directions (x, y, z) are obtained from the three-dimensional spatial concentration grid data.

[0043] The central difference method is used to calculate the concentration gradient component in the x-direction of each spatial grid cell. The value is the concentration value of the spatial grid cell in the positive x-direction adjacent grid cell minus the concentration value of the spatial grid cell in the negative x-direction, and then divided by twice the step size of the spatial grid in the x-direction.

[0044] Using the same central difference method, the concentration gradient components in the y and z directions of each spatial grid cell were calculated respectively;

[0045] The concentration gradient components in the three orthogonal directions calculated for each spatial grid cell are combined into a three-dimensional gradient vector, and the three-dimensional gradient vectors of all spatial grid cells together constitute the concentration gradient field.

[0046] The calculated concentration gradient field is smoothed using a Gaussian filter to suppress high-frequency fluctuations in the gradient field caused by spatial grid discretization and computational noise.

[0047] Furthermore, the method also includes:

[0048] The potential release source locations are correlated and matched with the equipment operation process parameter database to identify pollution source equipment that meets the emission characteristics of a specific process, and an early warning report is generated that includes pollution source equipment identification, pollutant type, predicted concentration evolution and source tracing information.

[0049] The step of associating and matching the potential release source locations with the equipment operating process parameter database to identify pollution source equipment that meets the emission characteristics of a specific process includes:

[0050] From the equipment operation process parameter database, obtain the spatial coordinates, equipment type, process formula parameters, and periodic operation logs of all process equipment emitting gaseous pollutants in the plant.

[0051] The spatial proximity analysis is performed between the calculated spatial coordinates of the potential release source location and the spatial coordinates of all process equipment to filter out a set of candidate equipment located within a set distance threshold of the potential release source location.

[0052] From the pollution source inventory database, query the typical emission processes and equipment types corresponding to the types of new gaseous pollutants in the semiconductor industrial plant, and compare them with the equipment types in the candidate equipment set to filter out the equipment subset with matching equipment types;

[0053] Further analysis of the process formula parameters and periodic operation logs of each device in the aforementioned equipment subset was conducted to extract operation events that showed a high correlation with the time variation curve of pollutant release intensity in terms of both period and phase.

[0054] The process equipment corresponding to the operational event is ultimately identified as the pollution source equipment that meets the emission characteristics of the specific process.

[0055] Furthermore, the method also includes a step of optimizing the parameters of the dust removal and purification equipment based on the prediction and source tracing results:

[0056] Based on the pollutant types and predicted concentration evolution information contained in the warning report, query the purification performance database of micro electrostatic dust removal device and series nano-mineral-based pollutant purification filter cartridge to obtain the single-pass purification efficiency parameters of the corresponding pollutants.

[0057] Based on the mass migration flux of the pollutants inside the plant driven by concentration gradient and airflow path, the concentration history of the pollutants diffused from the pollution source equipment to the inlet of the dust removal and purification equipment was simulated.

[0058] By combining the single-channel purification efficiency parameters with the design air volume parameters of the dust removal and purification equipment, the theoretical removal efficiency of the dust removal and purification equipment for pollutants is calculated.

[0059] The theoretical removal efficiency is compared with the target removal efficiency. If the theoretical removal efficiency is lower than the target value, an adjustment instruction is generated. The adjustment instruction is used to optimize the electrostatic voltage and electrode spacing parameters of the micro electrostatic dust removal device, as well as the catalyst loading and filter structure parameters of the series nano-mineral-based pollutant purification filter.

[0060] Furthermore, combining the single-channel purification efficiency parameter with the design airflow parameter of the dust removal and purification equipment, the theoretical removal efficiency of the dust removal and purification equipment for pollutants is calculated, including:

[0061] The primary single-pass purification efficiency of the micro electrostatic dust removal device for pollutants and the secondary single-pass purification efficiency of the series-connected nano-mineral-based pollutant purification filter cartridges are obtained.

[0062] Based on the series purification model, the primary single-channel purification efficiency and the secondary single-channel purification efficiency are combined to calculate the comprehensive single-channel purification efficiency of the dust removal and purification equipment under rated operating conditions.

[0063] The actual airflow rate passing through the dust removal and purification equipment is obtained from the factory ventilation system;

[0064] The concentration history of the pollutant diffused from the pollution source equipment to the inlet of the dust removal and purification equipment is multiplied by the actual airflow rate to obtain the mass of pollutants entering the dust removal and purification equipment per unit time.

[0065] By using the comprehensive single-channel purification efficiency and the mass of pollutants entering the dust removal and purification equipment per unit time, the mass of pollutants captured and removed by the dust removal and purification equipment per unit time is calculated, and thus the theoretical removal efficiency is obtained.

[0066] Compared with the prior art, the beneficial effects of the present invention are:

[0067] Multiple sets of raw gaseous pollutant spectral data acquired simultaneously by various monitoring devices within a semiconductor factory are time-aligned and signal intensity normalized to form standardized pollutant mass spectrum time series data. This standardized data is then directly compared with characteristic mass spectra in a pre-set pollution source inventory database. This eliminates the problem of asynchronous data acquisition from multiple devices, unifies the signal output intensity of different monitoring devices, and ensures consistency in the characteristic dimensions of the spectral data. When matched with characteristic mass spectra in the pollution source inventory database, the types of new gaseous pollutants can be accurately identified, providing accurate preliminary concentration information and reducing identification bias caused by non-standardized data.

[0068] Based on the spatiotemporal distribution prediction field of new pollutant concentration, the mass migration flux of new pollutants within the plant, driven by concentration gradient and airflow path, is calculated. Combined with the spatial structure information inside the plant, the diffusion process of pollutants is simulated in reverse using a deconvolution diffusion model. This model can fully reflect the migration driving logic of pollutants within the plant, closely match the actual spatial layout and airflow trajectory of the plant, and restore the source of pollutant propagation by reversing the diffusion path. This overcomes the limitation of conventional models that can only simulate forward and cannot accurately trace back to the source, and accurately determines the potential release source location of gaseous new pollutants within the semiconductor plant. Attached Figure Description

[0069] Figure 1 This is a flowchart illustrating the steps of a deep learning-based semiconductor contaminant prediction and tracing method described in this invention.

[0070] Figure 2 A flowchart for time alignment and signal strength normalization processing;

[0071] Figure 3 A flowchart for training and predicting neural network models for spatiotemporal evolution. Detailed Implementation

[0072] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0073] See Figure 1 This invention provides a method for predicting and tracing semiconductor contaminants based on deep learning, the specific implementation of which is as follows:

[0074] This method deploys multiple portable monitoring devices within a semiconductor industrial plant. These devices operate synchronously and collect raw spectral data of gaseous pollutants. The collected raw spectral data are time-aligned and signal intensity normalized to form a standardized time-series data set of pollutant mass spectra. This standardized data is compared with characteristic mass spectra from a pre-set pollution source inventory database to identify the specific types and preliminary concentrations of new gaseous pollutants within the semiconductor plant. The identified pollutant types and preliminary concentrations are then input into a pre-trained spatiotemporal evolution prediction neural network model. This model has learned from historical data and mastered the complex mapping relationship between pollutant distribution and the plant's ventilation layout, temperature, and humidity fields. Through calculations using this model, a predicted spatiotemporal distribution field of new pollutant concentrations in different areas of the plant over a future period is generated. Based on this predicted field, the mass migration flux of pollutants within the plant due to concentration gradients and predetermined airflow paths is calculated. Combined with the spatial structure information within the plant, a deconvolution diffusion model is used to simulate the pollutant diffusion process, ultimately deducing the locations of one or more potential pollutant release sources.

[0075] In one embodiment of the present invention, during the data preprocessing stage, the raw spectral data of gaseous pollutants acquired from each monitoring device includes the mass-to-charge ratio signal, signal intensity, and corresponding sampling timestamp, see reference. Figure 2A unified time reference is established, and interpolation calibration is performed on the mass-to-charge ratio signals, signal strengths, and sampling timestamps from different portable monitoring devices to align the time coordinates of all data points. For the signal strength of each mass-to-charge ratio signal channel, baseline subtraction is performed based on the historical background signal strength of each device. For the signal strength data after baseline subtraction, the average response intensity measured by all monitoring devices in a clean air environment is used as a calibration reference, and intensity normalization scaling is performed to unify the signal strengths of different devices to the same dimensional scale. All mass-to-charge ratio signal channel data after time alignment and intensity normalization are integrated in chronological order to construct multidimensional standardized pollutant mass spectrum time series data. Each time point in this time series corresponds to a vector containing the signal strengths of multiple mass-to-charge ratio channels. In the pollutant identification stage, standard characteristic mass spectra of various known gaseous pollutants from semiconductor processes are extracted from the pollution source inventory database. Each spectrum contains a specific mass-to-charge ratio and a corresponding standard relative intensity distribution. In the standardized pollutant mass spectrum time series data, for each vector in each time slice, the similarity distance metric between it and each standard characteristic mass spectra is calculated. Standard characteristic mass spectra with similarity distance metrics lower than a preset identification threshold are selected, and the known pollutants corresponding to them are identified as pollutant types in that time slice. Based on the signal intensity ratio of the standard characteristic mass spectra of the identified pollutant type and the measured vector in the characteristic mass-to-charge ratio, combined with the preset calibration curve of the pollutant type, preliminary concentration information is calculated. The pollutant types continuously identified in the time dimension are integrated with the preliminary concentration information to form a time series of identification results for new gaseous pollutants in semiconductor plants.

[0076] In practical implementation, multiple portable monitoring devices, such as online mass spectrometers, deployed within the semiconductor plant simultaneously collect raw spectral data of gaseous pollutants. Each set of raw spectral data for gaseous pollutants includes a complete mass-to-charge ratio signal sequence, the signal intensity corresponding to each mass-to-charge ratio signal, and a precise sampling timestamp. Due to slight differences in sampling start time and sampling frequency, the timestamps of raw spectral data from different portable monitoring devices are not perfectly aligned. In some embodiments, a unified time reference is established based on the time of the plant's central control computer. Linear interpolation is used to calibrate the mass-to-charge ratio signals, signal intensities, and sampling timestamps uploaded by all portable monitoring devices, ensuring a one-to-one correspondence between the data points of all portable monitoring devices on a unified time coordinate axis, thus achieving time alignment. Optionally, the time interval after alignment can be set to 1 second or 5 seconds as required. For the signal intensity of each mass-to-charge ratio signal channel, baseline subtraction is performed based on the average background signal intensity obtained from long-term statistical analysis of the corresponding portable monitoring device under pollution-free operating conditions. In practice, the signal strength data after baseline subtraction may have different dimensions due to variations in sensor sensitivity among different portable monitoring devices. The average response intensity recorded by all portable monitoring devices operating in a clean air environment for a period of time is used as a unified calibration reference. Intensity normalization scaling is then performed on the signal strength data of each mass-to-charge ratio signal channel for each portable monitoring device. This normalization scaling maps the signal strength of different portable monitoring devices to a unified, dimensionless relative intensity scale. Finally, the time-aligned and intensity-normalized data from all mass-to-charge ratio signal channels are integrated in chronological order to construct a multi-dimensional, standardized pollutant mass spectrum time series. Each time point in the standardized pollutant mass spectrum time series corresponds to a vector containing a fixed number of mass-to-charge ratio channel signal intensities.

[0077] In practical implementation, a pre-defined pollution source inventory database stores standard characteristic mass spectra of various known gaseous pollutants from semiconductor processes. Each standard characteristic mass spectra contains a set of characteristic mass-to-charge ratio sequences unique to that pollutant, as well as the standard relative intensity distribution corresponding to each characteristic mass-to-charge ratio. In the standardized pollutant mass spectrum time series data, for the signal intensity vector in each time slice, it is necessary to calculate the similarity distance metric between it and each standard characteristic mass spectra in the pollution source inventory database. In some embodiments, the similarity distance metric can be implemented by calculating the Euclidean distance between two vectors. Optionally, cosine similarity or other similarity metric algorithms can also be used. When the calculated similarity distance metric is lower than a preset identification threshold, the corresponding known gaseous pollutant from semiconductor processes is identified as the pollutant type identified in the current time slice. It is understood that the setting of the preset identification threshold needs to be balanced based on the actual monitoring accuracy and false alarm rate requirements. For identified pollutant species, preliminary concentration information is calculated based on the ratio between the standard relative intensity of the characteristic mass-to-charge ratio in its standard characteristic mass spectrum and the measured signal intensity of the same mass-to-charge ratio in the standardized pollutant mass spectrum time series data, combined with the concentration-intensity calibration curve pre-determined through calibration experiments for that pollutant species. In practice, the formula used to calculate the preliminary concentration information is:

[0078]

[0079] in: This indicates preliminary concentration information. The measured signal intensity representing the characteristic mass-to-charge ratio in standardized pollutant spectral time-series data. The function represents the standard relative intensity of the characteristic mass-to-charge ratio corresponding to the standard characteristic mass spectrum in the pollution source inventory database. This represents the mapping relationship determined by the calibration curve. The pollutant types continuously identified across all time slices are integrated with the calculated preliminary concentration information to form a time series of identification results for new gaseous pollutants in the semiconductor plant.

[0080] In one embodiment of the present invention, see [reference] Figure 3A deep recurrent neural network containing encoder and decoder structures is constructed as the basic architecture of the spatiotemporal evolution prediction neural network model. The types and preliminary concentrations of new gaseous pollutants in the semiconductor plant at multiple historical moments are combined with temperature, humidity, and ventilation system wind speed and direction data collected at multiple points inside the plant at the same historical moment to form the input feature vector sequence of the model. The actual pollutant concentration distribution field at the next moment in the historical time series is used as the target output label for model training. The deep recurrent neural network is iteratively trained using a large amount of historical operating data until the error between its predicted output and the target output label is less than a preset threshold, thereby obtaining a pre-trained spatiotemporal evolution prediction neural network model. In the actual prediction stage, the latest identification result time series and real-time plant environmental monitoring data are input into this pre-trained model, and the model directly outputs the spatiotemporal distribution prediction field of new pollutant concentrations in different areas in the future time period.

[0081] In practical implementation, constructing a neural network model for predicting spatiotemporal evolution requires first defining its basic architecture. This architecture employs a deep recurrent neural network (RNN) containing an encoder and decoder. The encoder extracts the spatiotemporal features of the input sequence, while the decoder generates the pollutant concentration distribution field for future moments. In some embodiments, the RNN can use a Long Short-Term Memory (LSTM) network or a gated recurrent unit (GRU) as its basic unit. The input feature vector sequence is composed of the types and preliminary concentrations of new gaseous pollutants in the semiconductor plant at multiple historical moments, along with temperature, humidity, and ventilation system wind speed and direction data collected at multiple points within the plant at the same historical moment. Optionally, the time window length of the input feature vector sequence can be set to the data from the past 60 minutes, with a sampling interval of 1 minute. The actual pollutant concentration distribution field for the next moment in the historical time series serves as the target output label for model training. This actual pollutant concentration distribution field is measured using a high-precision sensor network. The RNN is iteratively trained using massive amounts of historical operating data. During training, the network parameters are adjusted using a backpropagation algorithm. The loss function used in model training measures the error between the predicted output and the target output label. The loss function is defined as follows:

[0082]

[0083] in: Indicates training loss, Indicates the number of training samples. The neural network model representing the spatiotemporal evolution law predicts the first... The predicted output for each sample, Indicates the first The target output label for each sample. The square of the Euclidean norm is represented. Iterative training continues until the error between the predicted output of the spatiotemporal evolution law prediction neural network model and the target output label is less than a preset threshold, thus obtaining the pre-trained spatiotemporal evolution law prediction neural network model. It can be understood that the preset threshold is set according to the prediction accuracy requirements of the application scenario.

[0084] In practical implementation, during the actual prediction phase, the latest identification result time series and real-time factory environmental data are jointly input into a pre-trained spatiotemporal evolution prediction neural network model. The latest identification result time series includes the types and preliminary concentration information of newly identified gaseous pollutants in the semiconductor factory. The real-time factory environmental data includes temperature data, humidity data, and ventilation system wind speed and direction data at multiple points inside the factory at the current moment. Based on the input data, the pre-trained spatiotemporal evolution prediction neural network model directly outputs a predicted field of spatiotemporal distribution of new pollutant concentrations in different areas within a future time period. In some embodiments, the future time period can be set to the next 30 minutes. The spatial distribution prediction field is presented in the form of a three-dimensional grid. Optionally, the model output can include concentration distribution sequences for multiple future times. The operation of the pre-trained spatiotemporal evolution prediction neural network model is based on a learned nonlinear mapping relationship. It can be understood that the nonlinear mapping relationship relates the complex patterns between historical pollutant concentration distribution and factory ventilation layout, temperature and humidity fields.

[0085] In one embodiment of the present invention, three-dimensional spatial concentration grid data for several consecutive future time periods are extracted from the predicted spatiotemporal distribution field of new pollutant concentration. For this three-dimensional spatial concentration grid data, gradient calculation is performed in three orthogonal directions of each spatial grid cell to obtain a concentration gradient field. The steady-state airflow path field inside the plant, which is pre-generated by the plant ventilation system layout and computational fluid dynamics simulation, is obtained. The calculated concentration gradient field and the steady-state airflow path field are multiplied by each spatial grid cell, and then multiplied by the molecular diffusion coefficient of the pollutant to calculate the preliminary mass migration flux vector of each spatial grid cell at each future time. This preliminary mass migration flux vector is integrated over a specified future time interval, and reverse path tracing is performed along the steady-state airflow path field to deduce the main mass transport path of pollutants in the plant space.

[0086] In practical implementation, three-dimensional spatial concentration grid data for several consecutive future time periods are extracted from the predicted spatiotemporal distribution field of new pollutant concentrations. This three-dimensional spatial concentration grid data divides the internal space of the plant into multiple discrete grid cells. Each grid cell contains its spatial coordinates and the predicted pollutant concentration value. Table 1 shows a simplified example of a three-dimensional spatial concentration grid data segment, listing the coordinates of different spatial grid cells and their corresponding predicted concentration values ​​at a future time. Based on the three-dimensional spatial concentration grid data, gradient calculations are performed in the three orthogonal directions of each spatial grid cell to obtain the concentration gradient field. The calculation process involves obtaining the concentration values ​​of adjacent grid cells in the x, y, and z directions for each spatial grid cell and applying the central difference formula. It can be understood that the concentration gradient field reflects the rate and direction of change of pollutant concentration in space. The steady-state airflow path field inside the factory building, pre-generated by the factory ventilation system layout and computational fluid dynamics simulation, is obtained. This steady-state airflow path field describes the stable airflow velocity and direction formed inside the factory building under the action of the ventilation system in the form of a vector field. Its spatial resolution is consistent with the three-dimensional spatial concentration grid data. The calculated concentration gradient field and the steady-state airflow path field are multiplied in each spatial grid cell, and then multiplied by the molecular diffusion coefficient of the pollutants to calculate the preliminary mass migration flux vector for each spatial grid cell at each future time. The formula for calculating the preliminary mass migration flux vector is as follows:

[0087]

[0088] in: This represents the initial mass transfer flux vector. Indicates the molecular diffusion coefficient of pollutants. This represents the concentration gradient vector of the spatial grid cell. The steady-state airflow velocity vector of this spatial grid cell is shown in Table 1.

[0089] Table 1: Three-dimensional spatial concentration grid data table X-coordinate of grid cell (meters) Y-coordinate of grid cell (meters) Z-coordinate of grid cell (meters) Predicted concentration (mg / m³) 10.1 5.2 2.0 0.15 10.2 5.2 2.0 0.18 10.3 5.2 2.0 0.22 10.1 5.3 2.0 0.14 10.2 5.3 2.0 0.17

[0090] In specific implementation, the preliminary mass migration flux vector is integrated over a specified future time interval. The integration operation accumulates the flux change of each spatial grid cell along the time dimension. In some embodiments, the specified future time interval can be 30 minutes from the current moment. The integration operation obtains the cumulative mass migration flux of each spatial grid cell within the time interval, and reverse path tracing is performed along the steady-state airflow path field. The reverse path tracing starts from the current high concentration area and traces the path of cumulative mass migration flux transfer along the opposite direction to the steady-state airflow path field. Optionally, the path tracing algorithm can adopt the particle inverse trajectory method to deduce the main mass transport path of pollutants in the plant space. The main mass transport path indicates the macroscopic movement trajectory of pollutants spreading from potential release sources to monitoring points within the plant. It can be understood that the calculation of mass migration flux and reverse path tracing provide directional constraints for subsequent pollution source inversion.

[0091] In one embodiment of the present invention, a convolutional diffusion forward model describing the diffusion process of gaseous pollutants in a closed space of a factory is established. The kernel function of the model is determined by the factory space structure, ventilation conditions, and turbulence intensity parameters. The current pollutant concentration distribution is used as the output of this convolutional diffusion forward model. The model is then run in reverse, using the historical pollutant source release distribution as the input to be solved. An iterative inversion algorithm is used to solve for the historical release source spatial distribution that minimizes the difference between the output of the convolutional diffusion forward model and the predicted field of the new pollutant concentration spatiotemporal distribution. The region in the historical release source spatial distribution where the release intensity is significantly greater than the background value is the potential release source location. For each identified potential release source location, the historical time variation curve of its release intensity is calculated. In calculating the concentration gradient field, the concentration values ​​of adjacent grid cells in the three orthogonal directions (x, y, z) of each spatial grid cell are obtained from the three-dimensional spatial concentration grid data. The central difference method is used to calculate the concentration gradient component in the x direction of each spatial grid cell. The value is the concentration value of the adjacent grid cells in the positive x direction minus the concentration value of the adjacent grid cells in the negative x direction, and then divided by twice the step size of the spatial grid in the x direction. The same central difference method is used to calculate the concentration gradient components in the y and z directions of each spatial grid cell. The concentration gradient components in the three orthogonal directions calculated for each spatial grid cell are combined into a three-dimensional gradient vector. The three-dimensional gradient vectors of all spatial grid cells together constitute the concentration gradient field. The calculated concentration gradient field is smoothed by applying a Gaussian filter to suppress high-frequency fluctuations in the gradient field caused by spatial grid discretization and computational noise.

[0092] In practical implementation, a convolutional diffusion forward model is established to describe the diffusion process of gaseous pollutants in the enclosed space of the factory. The kernel function of the convolutional diffusion forward model is determined by the spatial structure, ventilation conditions, and turbulence intensity parameters. The mathematical expression of the convolutional diffusion forward model describes the evolution of the pollutant concentration field with time step. The current pollutant concentration distribution is used as the output of the convolutional diffusion forward model, and the model is run in reverse, using the historical pollutant source release distribution as the input to be solved. An iterative inversion algorithm is used to solve for the historical release source spatial distribution that minimizes the difference between the output of the convolutional diffusion forward model and the predicted spatiotemporal distribution of the new pollutant concentration. The objective function minimized in the solution process can be expressed as:

[0093]

[0094] in: This represents the value of the objective function that needs to be minimized. The kernel function represents the forward convolutional diffusion model. This represents the spatial distribution of the historical release sources to be solved. This represents the convolution operation. This represents a slice of the predicted spatiotemporal distribution field of new pollutant concentrations at the current moment. Denotes the square of the L2 norm. Represents the regularization parameter. This represents a regularization constraint term on the spatial distribution of historical release sources. The iterative inversion algorithm continuously updates the estimated value of the spatial distribution of historical release sources until the objective function converges. Regions in the spatial distribution of historical release sources with release intensities greater than the background value are identified as potential release source locations. For each identified potential release source location, the historical time-varying curve of its release intensity is calculated. The spatial distribution sequence of historical release sources, obtained by the iterative inversion algorithm, which varies with the time step, is calculated. The spatial coordinates of the potential release source locations and the time-varying curve of release intensity are compared with the equipment's geographical location and periodic operating conditions recorded in the equipment operation process parameter database. The correlation between the time-varying curve and the equipment operation log sequence is calculated.

[0095] In some embodiments, the concentration values ​​of adjacent grid cells in the three orthogonal directions (x, y, z) of each spatial grid cell are obtained from the three-dimensional spatial concentration grid data. Table 2 shows an example of the concentration values ​​of a spatial grid cell and its adjacent cells. The central difference method is used to calculate the concentration gradient component in the x-direction of each spatial grid cell. The value of the concentration gradient component in the x-direction is the concentration value of the adjacent grid cells in the positive x-direction minus the concentration value of the adjacent grid cells in the negative x-direction, divided by twice the step size of the spatial grid in the x-direction. The same central difference method is used to calculate the concentration gradient components in the y-direction and z-direction of each spatial grid cell. The concentration gradient components calculated for each spatial grid cell in the three orthogonal directions are combined into a three-dimensional gradient vector. The three-dimensional gradient vectors of all spatial grid cells together constitute the concentration gradient field. The calculated concentration gradient field is smoothed using a Gaussian filter to suppress high-frequency fluctuations in the gradient field caused by spatial grid discretization and computational noise. It can be understood that the standard deviation parameter of the Gaussian filter is set according to the size of the spatial grid. Calculating the concentration gradient field is a preliminary step in obtaining the initial mass migration flux, see Table 2.

[0096] Table 2: Concentration values ​​of spatial grid cells and their adjacent cells direction Relative coordinate offset of adjacent grid cells Concentration value (mg / m³) negative direction (i-1,j,k) 0.18 Central Unit (i,j,k) 0.22 positive x direction (i+1,j,k) 0.25 negative y direction (i,j-1,k) 0.20 y positive direction (i,j+1,k) 0.24 negative direction (i,j,k-1) 0.21 z positive direction (i,j,k+1) 0.23

[0097] In practical implementation, the iterative inversion algorithm can employ either the gradient descent method or the conjugate gradient method. The kernel function of the convolutional diffusion forward model is pre-calibrated through computational fluid dynamics simulation, using known information on the location and intensity of the release source. The matching degree analysis between the spatial coordinates of the potential release source location and the equipment operating process parameter database can be achieved by calculating spatial distance and temporal correlation; optionally, the Pearson correlation coefficient can be used for temporal correlation analysis. The calculated historical time variation curve of the release intensity reflects the emission intensity estimate of the potential release source at different historical moments. Applying a Gaussian filter to smooth the concentration gradient field is a common denoising method; it is understood that smoothing helps to obtain a more stable and physically reasonable gradient field for subsequent flux calculations.

[0098] In one embodiment of the present invention, the spatial coordinates, equipment type, process formula parameters, and periodic operation logs of all process equipment emitting gaseous pollutants in the plant are obtained from the equipment operation process parameter database. The spatial proximity analysis of the calculated potential release source locations is performed with the spatial coordinates of all process equipment to screen out a set of candidate equipment located within a set distance threshold of the potential release source locations. The typical emission processes and equipment types corresponding to the identified new gaseous pollutants in the semiconductor plant are queried from the pollution source list database and compared with the equipment types in the candidate equipment set to screen out a subset of equipment with matching equipment types. The process formula parameters and periodic operation logs of each equipment in this subset are further analyzed to extract operation events that are highly correlated with the historical time change curve of pollutant release intensity in terms of period and phase. The process equipment corresponding to such operation events is identified as pollution source equipment that meets the emission characteristics of a specific process. Based on this, an early warning report containing pollution source equipment identification, pollutant type, predicted concentration evolution, and source tracing path information is generated. Based on the pollutant types and predicted concentration evolution information in the early warning report, the purification performance database of the micro electrostatic precipitator and the series nano-mineral-based pollutant purification filter is queried to obtain the single-pass purification efficiency parameters of the corresponding pollutants. Based on the mass migration flux of pollutants driven by concentration gradient and airflow path inside the plant, the concentration history of pollutants diffusing from the locked pollution source equipment to the inlet of the dust removal and purification equipment is simulated. Combining the single-pass purification efficiency parameters and the design airflow parameters of the dust removal and purification equipment, the theoretical removal efficiency of the dust removal and purification equipment for pollutants is calculated. The calculated theoretical removal efficiency is compared with the set target removal efficiency. If the theoretical removal efficiency is lower than the target value, an adjustment command is generated. This adjustment command is used to optimize the electrostatic voltage and electrode spacing parameters of the micro electrostatic precipitator, as well as the catalyst loading and filter structure parameters of the series nano-mineral-based pollutant purification filter. When calculating the theoretical removal efficiency, the primary single-pass purification efficiency of the micro-electrostatic dust collector and the secondary single-pass purification efficiency of the series-connected nano-mineral-based pollutant purification filter are obtained. Based on the series purification model, the primary and secondary single-pass purification efficiencies are combined to calculate the comprehensive single-pass purification efficiency of the dust removal and purification equipment under rated operating conditions. The actual airflow rate through the dust removal and purification equipment is obtained from the plant ventilation system. The concentration history of pollutants diffusing from the pollution source equipment to the inlet of the dust removal and purification equipment is multiplied by the actual airflow rate to obtain the mass of pollutants entering the dust removal and purification equipment per unit time. Using the comprehensive single-pass purification efficiency and the mass of pollutants entering the dust removal and purification equipment per unit time, the mass of pollutants captured and removed by the dust removal and purification equipment per unit time is calculated, thus obtaining the theoretical removal efficiency.

[0099] In practice, the spatial coordinates, equipment type, process formula parameters, and periodic operation logs of all process equipment emitting gaseous pollutants within the plant are obtained from the equipment operation process parameter database. Spatial proximity analysis is performed between the calculated spatial coordinates of potential release source locations and the spatial coordinates of all process equipment. This analysis calculates the Euclidean distance between each potential release source location and the coordinates of each process piece of equipment, filtering out a set of candidate equipment located within a set distance threshold for potential release source locations. The set distance threshold is determined based on the plant layout and monitoring accuracy. The typical emission processes and equipment types corresponding to new gaseous pollutants in the semiconductor plant are queried from the pollution source inventory database and compared with the equipment types in the candidate equipment set, filtering out a subset of equipment with matching equipment types. Further analysis of the process formula parameters and periodic operation logs of each piece of equipment in the subset extracts operational events that are highly correlated with the pollutant release intensity time-varying curves in terms of period and phase. These operational events include the start-up of process equipment, the execution of specific process steps, and shutdown. The process equipment corresponding to these operational events is ultimately identified as pollution source equipment that meets the emission characteristics of a specific process. Based on the matching and judgment results described above, an early warning report is generated. The report includes the pollution source equipment identification, pollutant type, predicted concentration evolution, and source tracing information. In some embodiments, the equipment operating process parameter database is updated in real time through the factory manufacturing execution system, and the distance threshold for spatial proximity analysis can be set to 5 meters. Equipment type comparison is based on the equipment classification standards published by the Semiconductor Equipment and Materials International (SEMI), and correlation analysis uses the method of calculating the cross-correlation function of time-series data. Optionally, the early warning report is presented in the form of structured text and visual graphs.

[0100] In practice, based on the pollutant types and predicted concentration evolution information contained in the early warning report, the purification performance database of the micro-electrostatic precipitator and the series-connected nano-mineral-based pollutant purification filter is queried to obtain the single-pass purification efficiency parameters for the corresponding pollutants. The concentration history of pollutants diffusing from the pollution source equipment to the inlet of the dust removal and purification equipment is simulated based on the mass migration flux driven by concentration gradient and airflow path within the plant. The concentration history is a curve showing the change in pollutant concentration over time. The theoretical removal efficiency of the dust removal and purification equipment is calculated by combining the single-pass purification efficiency parameters with the design airflow parameters of the dust removal and purification equipment, and then compared with the target removal efficiency. If the theoretical removal efficiency is lower than the target value, an adjustment command is generated. This command optimizes the electrostatic voltage and electrode spacing parameters of the micro-electrostatic precipitator, as well as the catalyst loading and filter structure parameters of the series-connected nano-mineral-based pollutant purification filter. It can be understood that the target removal efficiency is pre-set by environmental emission standards or internal air quality requirements of the plant. The purification performance database is established through laboratory testing and on-site calibration, recording the purification efficiency of different pollutants under different operating parameters. The concentration history is simulated using computational fluid dynamics simulation or a simplified transport model based on mass migration flux. In some embodiments, if the theoretical removal efficiency is lower than the target value, the adjustment instructions specifically suggest increasing the electrostatic voltage of the micro-electrostatic precipitator, reducing its electrode spacing, increasing the catalyst loading of the series-connected nano-mineral-based pollutant purification filter, and optimizing its pore structure parameters. Optionally, the design airflow parameters are obtained from the product specifications of the dust removal and purification equipment or actual operation monitoring data.

[0101] In practical implementation, calculating the theoretical removal efficiency requires obtaining the primary single-pass purification efficiency of the micro-electrostatic dust collector and the secondary single-pass purification efficiency of the series-connected nano-mineral-based pollutant purification filter. Based on the series purification model, the primary and secondary single-pass purification efficiencies are combined to calculate the comprehensive single-pass purification efficiency of the dust removal and purification equipment under rated operating conditions. The formula for calculating the comprehensive single-pass purification efficiency is as follows:

[0102]

[0103] in: This indicates the overall single-channel purification efficiency. This indicates the primary single-pass purification efficiency of the micro electrostatic dust removal device. This indicates the secondary single-pass purification efficiency of the series-connected nano-mineral-based pollutant purification filter cartridges. The actual airflow rate through the dust removal and purification equipment is obtained from the factory ventilation system. The concentration history of pollutants diffusing from the pollution source equipment to the inlet of the dust removal and purification equipment is multiplied by the actual airflow rate to obtain the mass of pollutants entering the dust removal and purification equipment per unit time. The formula for calculating the mass of pollutants entering the dust removal and purification equipment per unit time is:

[0104]

[0105] in: Indicates the time interval The quality of pollutants entering the dust removal and purification equipment. This represents the concentration of pollutants at the inlet of the dust removal and purification equipment at time t. This represents the actual airflow rate through the dust removal and purification equipment at time t. The theoretical removal efficiency is obtained by calculating the mass of pollutants captured and removed by the equipment per unit time using the comprehensive single-pass purification efficiency and the mass of pollutants entering the equipment per unit time. The theoretical removal efficiency is the ratio of the mass of pollutants removed per unit time to the mass of pollutants entering per unit time.

[0106] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for predicting and tracing semiconductor contaminants based on deep learning, characterized in that, The method includes: Multiple sets of raw spectral data of gaseous pollutants were simultaneously acquired from various monitoring devices deployed within the semiconductor industrial plant. The original spectral data of multiple gaseous pollutants are time-aligned and signal intensity normalized to form standardized pollutant mass spectrum time series data. The standardized pollutant mass spectrum time series data is then compared with the characteristic mass spectra in a preset pollution source list database to identify the types and preliminary concentration information of new gaseous pollutants in semiconductor industrial plants. The types and preliminary concentration information of the new gaseous pollutants identified in the semiconductor industrial plant are input into a pre-trained spatiotemporal evolution prediction neural network model. The pre-trained spatiotemporal evolution prediction neural network model has learned the nonlinear mapping relationship between the historical pollutant concentration distribution and the plant's ventilation layout, temperature and humidity field. The pre-trained spatiotemporal evolution prediction neural network model is used to calculate and generate a predicted spatiotemporal distribution field of new pollutant concentration in different areas in the future time period. Based on the predicted spatiotemporal distribution field of the new pollutant concentration, the mass migration flux of the new pollutant within the plant, driven by concentration gradient and airflow path, is calculated. Combined with the spatial structure information within the plant, a deconvolution diffusion model is used to simulate the pollutant diffusion process in reverse, deducing one or more potential release source locations for the gaseous new pollutants in the semiconductor industrial plant, including: From the spatiotemporal distribution prediction field of the new pollutant concentration, extract three-dimensional spatial concentration grid data for several consecutive future time periods; The gradient of the three-dimensional spatial concentration grid data is calculated in three orthogonal directions of each spatial grid cell to obtain the concentration gradient field; Obtain the steady-state airflow path field inside the factory, which is pre-generated by the factory ventilation system layout and computational fluid dynamics simulation; The concentration gradient field and the steady-state airflow path field are multiplied in each spatial grid cell, and then multiplied by the molecular diffusion coefficient of the pollutant to calculate the preliminary mass migration flux vector of each spatial grid cell at each future time. The preliminary mass migration flux vector is integrated over a specified time interval, and the reverse path is traced along the steady-state airflow path field to deduce the main mass transport path of pollutants in space. A convolutional diffusion forward model is established to describe the diffusion process of gaseous pollutants in the enclosed space of a factory. The kernel function of the convolutional diffusion forward model is determined by the spatial structure, ventilation conditions, and turbulence intensity parameters. The current pollutant concentration distribution is used as the output of the convolutional diffusion forward model, and the model is run in reverse, with the historical pollutant source release distribution as the input to be solved. Using an iterative inversion algorithm, the spatial distribution of historical release sources that minimizes the difference between the output of the convolutional diffusion forward model and the predicted field of the spatiotemporal distribution of the new pollutant concentration is obtained. The region in the spatial distribution of historical release sources with a release intensity greater than the background value is the location of the potential release source. For each of the potential release source locations, calculate the historical time variation curve of its release intensity; The spatial coordinates of the potential release source location and the time variation curve of the release intensity are compared with the geographical location of the equipment and its periodic operating conditions recorded in the equipment operation process parameter database.

2. The semiconductor contaminant prediction and source tracing method based on deep learning according to claim 1, characterized in that, Multiple sets of raw spectral data of the aforementioned gaseous pollutants were subjected to time alignment and signal intensity normalization to form standardized pollutant spectral time series data, including: The raw spectral data of the gaseous pollutants includes the mass-to-charge ratio signal, signal intensity, and corresponding sampling timestamp; A unified time reference is set, and the mass-to-charge ratio signals, signal strengths and sampling timestamps from different portable monitoring devices are interpolated and calibrated to ensure that the time coordinates of all data points are aligned. For the signal strength of each of the mass-to-charge ratio signal channels, a baseline subtraction calculation is performed based on the historical background signal strength of the corresponding portable monitoring device; For the signal strength data after baseline subtraction, the average response intensity of all monitoring devices in a clean air environment is used as the calibration reference, and intensity normalization scaling is performed to unify the signal strength of different devices to the same dimension scale. The data of all mass-to-charge ratio signal channels, after being aligned in time and normalized in intensity, are integrated in chronological order to construct multidimensional standardized pollutant mass spectrum time series data. Each time point in the standardized pollutant mass spectrum time series data corresponds to a vector containing the signal intensities of multiple mass-to-charge ratio channels.

3. The semiconductor contaminant prediction and source tracing method based on deep learning according to claim 2, characterized in that, The standardized pollutant mass spectrum time series data is compared with the characteristic mass spectra in a pre-set pollution source inventory database to identify the types and preliminary concentration information of new gaseous pollutants in semiconductor industrial plants, including: From the pollution source inventory database, standard characteristic mass spectra of various known gaseous pollutants from semiconductor processes are extracted. Each standard characteristic mass spectrum contains a specific mass-to-charge ratio and a corresponding standard relative intensity distribution. In the standardized pollutant mass spectrum time series data, for each time slice vector, the similarity distance metric between it and each standard characteristic mass spectrum is calculated; Select standard feature mass spectra with similarity distance metric below a preset recognition threshold, and determine the corresponding known semiconductor process gaseous contaminants as the contaminant types identified in the time slice; Based on the ratio of the signal intensity of the standard characteristic mass spectrum corresponding to the identified pollutant type to the measured vector in the characteristic mass-to-charge ratio, and combined with the preset calibration curve corresponding to the identified pollutant type, preliminary concentration information is calculated. By integrating the pollutant types identified continuously over time with the preliminary concentration information, a time series of identification results for new gaseous pollutants in the semiconductor industrial plant is formed.

4. The semiconductor contaminant prediction and source tracing method based on deep learning according to claim 3, characterized in that, The identified types and preliminary concentration information of new gaseous pollutants in the semiconductor industrial plant are input into a pre-trained neural network model for predicting spatiotemporal evolution, including: A deep recurrent neural network containing encoder and decoder structures is constructed as the basic architecture of the neural network model for predicting spatiotemporal evolution laws. The types and preliminary concentrations of new gaseous pollutants in the semiconductor industrial plant at multiple historical moments are combined with temperature data, humidity data, and ventilation system wind speed and direction data collected at multiple points inside the plant at the same historical moment to form the model input feature vector sequence. Use the actual pollutant concentration distribution field at the next moment in the historical time series as the target output label for model training; The deep recurrent neural network is iteratively trained using massive amounts of historical working condition data until the error between its predicted output and the target output label is less than a preset threshold, thereby obtaining the pre-trained spatiotemporal evolution law prediction neural network model. In the actual prediction stage, the latest identification result time series and real-time factory environment data are input into the pre-trained spatiotemporal evolution law prediction neural network model, and the model directly outputs the spatiotemporal distribution prediction field of new pollutant concentration in different regions within the future time period.

5. The semiconductor contaminant prediction and source tracing method based on deep learning according to claim 4, characterized in that, The gradient of the three-dimensional spatial concentration grid data is calculated in three orthogonal directions of each spatial grid cell to obtain the concentration gradient field, including: The concentration values ​​of each spatial grid cell in the three orthogonal directions (x, y, z) are obtained from the three-dimensional spatial concentration grid data. The central difference method is used to calculate the concentration gradient component in the x-direction of each spatial grid cell. The value is the concentration value of the spatial grid cell in the positive x-direction adjacent grid cell minus the concentration value of the spatial grid cell in the negative x-direction, and then divided by twice the step size of the spatial grid in the x-direction. Using the same central difference method, the concentration gradient components in the y and z directions of each spatial grid cell were calculated respectively; The concentration gradient components in the three orthogonal directions calculated for each spatial grid cell are combined into a three-dimensional gradient vector, and the three-dimensional gradient vectors of all spatial grid cells together constitute the concentration gradient field. The calculated concentration gradient field is smoothed using a Gaussian filter to suppress high-frequency fluctuations in the gradient field caused by spatial grid discretization and computational noise.

6. The semiconductor contaminant prediction and source tracing method based on deep learning according to claim 1, characterized in that, The method further includes: The potential release source locations are correlated and matched with the equipment operation process parameter database to identify pollution source equipment that meets the emission characteristics of a specific process, and an early warning report is generated that includes pollution source equipment identification, pollutant type, predicted concentration evolution and source tracing information. The step of associating and matching the potential release source locations with the equipment operating process parameter database to identify pollution source equipment that meets the emission characteristics of a specific process includes: From the equipment operation process parameter database, obtain the spatial coordinates, equipment type, process formula parameters, and periodic operation logs of all process equipment emitting gaseous pollutants in the plant. The spatial proximity analysis is performed between the calculated spatial coordinates of the potential release source location and the spatial coordinates of all process equipment to filter out a set of candidate equipment located within a set distance threshold of the potential release source location. From the pollution source inventory database, query the typical emission processes and equipment types corresponding to the types of new gaseous pollutants in the semiconductor industrial plant, and compare them with the equipment types in the candidate equipment set to filter out the equipment subset with matching equipment types; Further analysis of the process formula parameters and periodic operation logs of each device in the aforementioned equipment subset was conducted to extract operation events that showed a high correlation with the time variation curve of pollutant release intensity in terms of both period and phase. The process equipment corresponding to the operational event is ultimately identified as the pollution source equipment that meets the emission characteristics of the specific process.

7. The semiconductor contaminant prediction and source tracing method based on deep learning according to claim 6, characterized in that, The method also includes a step of optimizing the parameters of the dust removal and purification equipment based on the predicted source tracing results: Based on the pollutant types and predicted concentration evolution information contained in the warning report, query the purification performance database of micro electrostatic dust removal device and series nano-mineral-based pollutant purification filter cartridge to obtain the single-pass purification efficiency parameters of the corresponding pollutants. Based on the mass migration flux of the pollutants inside the plant driven by concentration gradient and airflow path, the concentration history of the pollutants diffused from the pollution source equipment to the inlet of the dust removal and purification equipment was simulated. By combining the single-channel purification efficiency parameters with the design air volume parameters of the dust removal and purification equipment, the theoretical removal efficiency of the dust removal and purification equipment for pollutants is calculated. The theoretical removal efficiency is compared with the target removal efficiency. If the theoretical removal efficiency is lower than the target value, an adjustment instruction is generated. The adjustment instruction is used to optimize the electrostatic voltage and electrode spacing parameters of the micro electrostatic dust removal device, as well as the catalyst loading and filter structure parameters of the series nano-mineral-based pollutant purification filter.

8. The semiconductor contaminant prediction and source tracing method based on deep learning according to claim 7, characterized in that, Combining the single-channel purification efficiency parameters with the design airflow parameters of the dust removal and purification equipment, the theoretical removal efficiency of the dust removal and purification equipment for pollutants is calculated, including: The primary single-pass purification efficiency of the micro electrostatic dust removal device for pollutants and the secondary single-pass purification efficiency of the series-connected nano-mineral-based pollutant purification filter cartridges are obtained. Based on the series purification model, the primary single-channel purification efficiency and the secondary single-channel purification efficiency are combined to calculate the comprehensive single-channel purification efficiency of the dust removal and purification equipment under rated operating conditions. The actual airflow rate passing through the dust removal and purification equipment is obtained from the factory ventilation system; The concentration history of the pollutant diffused from the pollution source equipment to the inlet of the dust removal and purification equipment is multiplied by the actual airflow rate to obtain the mass of pollutants entering the dust removal and purification equipment per unit time. By using the comprehensive single-channel purification efficiency and the mass of pollutants entering the dust removal and purification equipment per unit time, the mass of pollutants captured and removed by the dust removal and purification equipment per unit time is calculated, and thus the theoretical removal efficiency is obtained.

Images