A machine learning method for detecting at least one anomaly in a plasma system.

Abstract

The present invention relates to a method for machine learning detection of at least one anomaly in a plasma system, in particular in an RF powered plasma processing system, comprising: providing at least one input signal (210) having at least one anomaly characteristic indicative of an anomaly in the plasma system, said input signal (210) being related to an analog signal of a power supply system (1) of the plasma system and / or another characteristic of the power supply system (1) and / or the plasma system; The method includes performing a machine learning procedure (310) and processing at least one input signal (210) having at least one anomaly feature by a programmable circuit (10) to train detection of anomalies in a plasma system.

Description

Technical Field

[0001] The present invention relates to a method for machine learning to detect at least one abnormality in a plasma system. The present invention also relates to a system and a data carrier signal.

Background Art

[0002] Digital signal processing, abbreviated as "DSP", can be used in radio frequency systems for various applications, such as filtering and / or detecting specific situations of a system. To improve the processing speed and / or performance speed, DSP can be executed at least partially within hardware, for example, by an integrated circuit, abbreviated as "IC". However, the parameterization of integrated circuits can be time-consuming and technically complex. The complexity of the parameterization process can affect digital signal processing in many undesirable ways. One of them is a partial reduction in the DSP efficiency of the integrated circuit.

[0003] Plasma power supply systems can measure radio frequency, or RF, signals. Typical RF measurements include forward and reverse power, as well as RF current and voltage at various points within the plasma power supply system. Measured signals generally vary in frequency, amplitude, noise or spurious components, and pulse length. Some variations, such as frequency, may be adequately controlled by the power supply system, while others may not be controlled, especially when supplying RF power to a plasma processing chamber. These variations can introduce various types of interference into the measured RF signal. For example, one form of interference may be the generation of harmonics due to the nonlinear nature of the plasma itself. Another form of interference may result from different frequencies used to generate the RF signal. In advanced plasma systems, several RF generators with different frequencies may be used for plasma excitation. However, in addition to the desired generation of the intended frequency set, mixed frequencies, such as combinations of the fundamental frequency and its harmonics (addition or subtraction), may also be generated.

[0004] Many interferences found in measured RF signals can be suppressed, or at least reduced, by using appropriate filters before digitizing the measured RF signal. Nevertheless, unwanted sidebands close to the base frequency may occur, requiring further filtering by a DSP, or imposing additional filtering requirements. Actual input signals are never ideal, as they always contain unavoidable distortions and other artifacts caused by the signal source, transmission path, measuring equipment, etc. Optimizing filters or filter parameters is crucial for reconstructing a distortion-free, ideal input signal to stably tune the network. However, each filter has its advantages and disadvantages. For example, if filtering increases signal delay too much, the control may not be able to respond in a timely manner to deviations from the intended power level. Furthermore, filtering can remove relevant parts of the signal that may be necessary to evaluate the quality of the plasma process, such as the presence of anomalies like secondary plasma (which may exist outside the intended area within the plasma chamber), arcs, and other plasma instabilities. Therefore, careful selection of not only the type of filter but also the parameter configuration is necessary to balance these characteristics and optimize performance.

[0005] In general, insufficient interference reduction can lead to fluctuations in the measurement of the output power of an RF generator. These fluctuations can be understood as "measurement artifacts." Since measurements may be used in various applications, such as controlling output power levels, these applications may depend on compromised measurement signals. In the control example, the output power level may incorrectly follow measurement artifacts rather than the correct measurement.

[0006] If plasma instability is suppressed by filtering, abnormalities such as secondary plasma or arcs may not be detected, and as a result, countermeasures may not be initiated. In semiconductor plasma processing systems, uncontrolled plasma instability, such as arc discharges, can damage the semiconductor wafers being processed.

[0007] US20060262889A1 discloses a power supply capable of generating a signal when driven by a drive signal from a controller. At radio frequencies, this may be either synchronous undersampling or synchronous oversampling.

[0008] In DE102015212242A1 / US20180122625A1, the sampling frequency is changed to eliminate interference. Both documents describe specific methods for correcting interference frequencies, including sampling frequency adaptation, but they do not describe machine learning, particularly the optimization of filter parameters using machine learning procedures.

[0009] The common use of programmable circuits, particularly field-programmable gate arrays (FPGAs) for accelerating simulations, may be known from prior art such as WO2007095574A1 or DE102011103861A1.

[0010] Furthermore, prior art has shown that intelligent systems exist for detecting specific conditions during plasma processing. However, this detection often lacks the reliability and processing speed to initiate corrective measures before the event is completed in time.

[0011] The use of specialized solutions for detecting specific conditions in plasma processing has been reported in prior art, e.g., US8989888B2.

[0012] In particular, the use of specialized solutions for controlling arc generation in plasma processing has been reported in several prior art documents.

[0013] EP3254295B1 specifically describes an arc treatment apparatus and method for suppressing arcs by shutting off RF power for a specific period of time.

[0014] US9316675B2 describes a method for detecting secondary plasma by monitoring changes in RF voltage and current. This method can distinguish between low-level and severe anomalies.

[0015] In US8890537B2, arcs can be detected through appropriate filtering and harmonic analysis.

[0016] The use of dynamic limits on RF-related signals such as reflected power, voltage, and current to identify arcs, as well as various methods for suppressing arcs, is described in particular in US7761247B2.

[0017] US10209294B2 demonstrates how arc detection can be improved by combining several observation signals. [Overview of the Initiative] [Problems that the invention aims to solve]

[0018] These prior art documents describe hardware or software solutions tuned to operate under specific process conditions, such as those of a sputtering system, or under specific anomalies (e.g., secondary plasma or arc only). In particular, all solutions require that an anomaly actually exists before the system detects it. In the case of arcs, the described arc control methods are effective, in particular, only in suppressing further arcs after at least one arc event has already occurred.

[0019] In contrast, the present invention described below may have broader applicability and greater versatility due to the fact that it can use a statistical approach to detect anomalies before they fully progress, and in particular, prevent their occurrence.

[0020] Therefore, an object of the present invention is to overcome at least partially the aforementioned drawbacks. In particular, an object of the present invention is to improve digital signal processing in RF systems.

[0021] The above-mentioned objectives are addressed by the method having the features of claim 1, the system having the features of claim 16, and the data carrier signal having the features of claim 17. Further features and details of the present invention are disclosed in the respective dependent claims, specification, and drawings. Features and details described in relation to the method according to the present invention also apply to the system and the data carrier signal according to the present invention, and vice versa. The individual embodiments of the present invention can always be referenced to one another. [Means for solving the problem]

[0022] In contrast to prior art literature, the methods, systems, and carrier signals of the present invention enable the optimization of digital signal processing, particularly in radio frequency systems. For this purpose, it is a particular idea of ​​the present invention that the processing procedure can be used, in particular, for training the system of the present invention. Training may be performed using reference data in a training mode to determine the optimized parameterization of a programmable circuit and / or algorithm, and the system of the present invention may then be switched to an application mode in which the optimized parameterization is used in real time for processing the relevant input signal. The programmable circuit may be a programmable signal processing circuit, such as an FPGA, for processing the input signal by executing an algorithm.

[0023] Further advantages of the method, system, and carrier signal of the present invention over the prior art may be achieved by executing the training operation mode and the real-time operation mode in the same system. Furthermore, further training using a new training data set, i.e., machine learning of new situations, can be advantageously performed at a later point in time and the existing parameterization can be further optimized (hereinafter, this is specifically referred to as "switching of the process from the real-time mode to the training mode"). Furthermore, for example, by using a user interface, it is also possible to transfer the optimized parameterization to other systems of the same type.

[0024] Advantageously, by using the same hardware components (FPGA, control, memory) in both the training mode and the application mode, the transfer of parameterization between systems becomes easier and future upgrades of the plasma processing system for yet unknown new situations become possible. The training can be performed, for example, in a test bench system or a laboratory plasma processing system using existing input signals recorded in the system used in the field. The variability of these signals can be numerically increased by scaling, filtering, superposition, etc. Subsequently, the optimized parameterization can be easily transferred from the training system to the plasma processing system in the field.

[0025] Furthermore, by using existing input signals from the memory, for example, in the factory acceptance test of an RF system, the functions of the RF system in various situations can be tested. In this case, the input signals are supplied by the memory instead of or in combination with ADC data, and the performance of the adjustment loop, filtering of RF signals, pattern recognition, and response operations, such as system operations such as warnings or interlocks, can be verified in a fully operating system.

[0026] This object is solved in particular by a method of machine learning for the detection of at least one anomaly in a plasma system, in particular a radio frequency (RF) power plasma system, preferably an RF power plasma processing system. The method may include the following (executed successively or in any order). · Providing at least one (or more) input signals each having at least one anomaly feature related to an analog signal of the power supply system of the plasma system and / or another characteristic of the power supply system and / or the plasma system, indicating an anomaly in the plasma system, · Executing a machine learning procedure, where at least one input signal having at least one anomaly feature is processed by a (preferably programmable) circuit to train for the detection of anomalies in the plasma system, particularly where this processing is at least partially executed by a detection procedure.

[0027] This has the advantage of being able to provide effective and fast machine learning. In particular, the use of a (preferably programmable) circuit for the machine learning procedure can further accelerate this process. Furthermore, the detection of anomalies can provide a safer and more reliable operation in the plasma system. According to another aspect of the present invention, a machine learning procedure can be configured to be at least part of a processing procedure. When executing the processing procedure and / or the machine learning procedure, at least one input signal can be iteratively processed by a circuit, preferably a programmable circuit, preferably a detection procedure, using at least one configurable parameter of the processing (i.e., processing performed particularly within the processing procedure and / or the machine learning procedure, preferably in the form of digital signal processing). The at least one configurable parameter may also be at least one parameter of the detection procedure. The configuration of the at least one parameter may be changed and particularly modified for each processing of at least one input signal within the processing procedure and / or the machine learning procedure to obtain each processing result. The configurable parameter may be a parameter of the detection procedure, particularly a weight or algorithm parameter of a neural network. The detection procedure may include digital signal processing in a radio frequency system, particularly a radio frequency power supply system. Each processing result may be a result or output of the detection procedure. The configuration can be modified to improve the detection procedure and, therefore, the digital signal processing.

[0028] At least one input signal may relate to a radio frequency signal of a radio frequency system, and / or another characteristic of the power supply system and / or the plasma system, such as a signal from the plasma chamber (e.g., an optical signal), preferably a signal measured within the plasma chamber, and / or a signal from an external source.

[0029] The method according to the present invention may further include the following: In particular, as a result of machine learning training, to optimize the configuration based on the processing results, determine at least one parameter result, including a selection of (one) modified configurations, preferably a set of weights and / or values. • To improve digital signal processing, and especially for the optimization of circuits using parameter results and, in particular, selection, it provides at least one determined parameter result, in particular a determined configuration.

[0030] The provided and / or determined parameter results may preferably include a set of weights for a neural network and may be used for parameterizing the circuit. The parameterized circuit may perform a detection procedure, which may include an application of a neural network using the parameter results, in particular the set of weights given by the parameter results. Thus, the circuit may implement a neural network, or in other words, it may implement a parameterized neural network using the parameter results, in particular the set of weights.

[0031] The parameterization of the circuit may include parameterization of the detection procedure based on the parameter results, i.e., preferably weights of a neural network according to the parameter results. The performance of the parameterized detection procedure may be aimed at detecting at least one anomaly in the normal (in-situ) operation of the plasma system.

[0032] Optionally, the programmable circuit may be configured as a programmable integrated circuit, preferably a digital signal processor, a composite programmable logic device, or a field-programmable gate array.

[0033] Advantageously, within the scope of the present invention, it may be provided that a programmable circuit, in particular, performs at least partially a detection procedure, which includes a pattern recognition or pattern matching and / or algorithm, particularly a neural application, preferably a neural network, to identify at least one anomalous feature. The weights and / or activation functions and / or neurons of the neural network may be provided by the detection procedure implemented, in particular by the circuit. Some of the activation functions and / or weights may be manually predefined.

[0034] The machine learning procedure may include the following: Using a detection procedure and at least one configurable parameter of the detection procedure, particularly the weights of a neural network or the parameters of an algorithm, the circuit performs processing of at least one input signal, preferably repeatedly, and for each processing of the input signal to obtain each processing result, the configuration of at least one parameter is changed, particularly modified. • As a result of machine learning training based on the processing results, determine at least one parameter result, including the selection of a modified configuration.

[0035] The parameter results and / or training results are sometimes referred to as the classification of the neural network.

[0036] Furthermore, the parameter results may be distributed for use in other circuits implementing the detection procedure. A data carrier signal, preferably a data carrier having a data carrier signal, may include the parameter results for this distribution. Therefore, a data carrier signal or data carrier including parameter results is also subject to the present invention.

[0037] Within the scope of the present invention, a machine learning procedure, particularly the processing of at least one input signal, includes repeated processing steps, wherein the same at least one input signal is processed, in particular by a programmable circuit, but using modified configurations with different parameters to obtain each processing result assigned to a configuration used, preferably, the evaluation of the modified configuration is performed by comparing each of the obtained processing results with a baseline result. The evaluation may be part of a machine learning procedure that enables training in the sense of optimizing the configuration (e.g., by using a conventional optimization algorithm). Thus, machine learning may, in addition or alternatively, include the automatic resolution of optimization problems by optimization algorithms, or simply the iterative search for the best parameters of an algorithm to identify at least one anomaly feature, as understood not only in the scope of artificial intelligence. The evaluation may be performed by another device outside the control circuit or programmable circuit.

[0038] Furthermore, optionally, the evaluation of the modified configuration may depend on the matching of the processing result with the baseline result, and at least one configuration with the highest evaluation may be selected from the modified configurations as at least one determined parameter result.

[0039] Furthermore, the baseline result is optionally provided to be a predetermined indicator of at least one anomaly. Therefore, it depends on how close the processing result is to the ideal baseline result.

[0040] An advantageous further development of the present invention can provide that each of at least one input signal relates to the following: • Radio frequency signals used to power the plasma system, and / or • Another characteristic of the power supply system and / or plasma system, Preferably, the anomaly is specific to the arc generated in the plasma processing system or to the probability of arc generation. The anomaly may be a specific pattern indicating the occurrence of a particularly expected arc or another undesirable anomaly.

[0041] Arcs can be understood as electrical flashovers or arc discharges, particularly during plasma processing. In plasma-supported coating processes, this anomaly can disrupt the plasma coating. Therefore, it is advantageous if the detection procedure can detect the arc in the shortest possible time.

[0042] Optionally, machine-learned detection of at least one anomaly may be used for arc detection and / or arc prevention and / or arc management. A warning message may be output as a result of the detection. Further actions, such as shutoff, enable arc management. Arc prevention becomes possible if an expected arc is detected.

[0043] Preferably, the machine learning procedure may be provided to provide iterative determination of the configuration of at least one parameter of a detection procedure for detection, in particular the configuration of at least one weight of a neural network or at least one configuration of an algorithm. In particular, the neural network or algorithm is at least partially implemented by a programmable circuit. The determined configuration is then used in a detection procedure in the field operation of a power supply system, which may include: • When an anomaly is detected by the detection procedure, a warning message is output.

[0044] Furthermore, providing at least one input signal is advantageous if it includes the following: • Record and convert analog signals from a power supply system in the form of radio frequency signals and / or other characteristics of the power supply system and / or plasma system to obtain a digital representation of the analog signals and / or other characteristics. • Provides a digital representation of an analog signal and / or other characteristics as at least one input signal to a particularly programmable circuit.

[0045] A further advantage may be provided, particularly in programmable circuits, that a detection procedure for detecting at least one anomaly can be implemented by combining functional blocks of circuitry for performing calculations in real time. This also enables high-speed execution of neural networks. Furthermore, within the scope of the present invention, the implemented detection procedure may be repeatedly used in real-time computation during machine learning procedures for processing at least one input signal. This makes it possible to accelerate machine learning.

[0046] Furthermore, the machine learning procedure is configured, in particular, as a training procedure for the parameterization of programmable circuits, so that the exact same circuit or another circuit with the same parameterization can be used for detecting at least one anomaly during field operation of a power supply system, in particular for arc detection and / or arc prevention and / or arc management.

[0047] Furthermore, it may be provided that, particularly in field operations, if at least one abnormality is detected, at least one of the following operations is initiated, preferably individually or in combination. • The shutdown of at least part or all of the power supply system, in particular the shutdown of only a part of the power supply system, or the shutdown of all parts of the power supply system. • Stopping the power supply system without restarting the power supply system. Temporary shutdown of the power supply system, followed by restart of the power supply system. • Temporary decrease in the output power of the power supply system. • Temporary modification of the output frequency of the power supply system, particularly the output frequency of the RF power signal and / or analog signal of the power supply system. • Temporary modification of at least one adjustable element of the impedance matching network.

[0048] This enables effective arc control and / or prevention. Generally, a transient change in at least one tunable parameter of a power supply system, such as a change in the RF power signal frequency or a change in the value of a tunable component of an impedance matching circuit, may be initiated, preferably individually or in combination with the operations described above, upon detection of at least one anomaly. The exact parameter used for this operation may be predefined depending on the specifications and exact configuration of the power supply system. The most suitable parameter may be found experimentally or empirically using a particular power supply system.

[0049] Within the scope of the present invention, it is even more advantageous if the anomaly detection includes probabilistic detection for detecting an arc before it occurs in the plasma processing system, in other words, for detecting an expected arc, preferably for determining the probability of arc occurrence, and / or for arc prevention. The anomaly detection may also include probabilistic detection for detecting another anomaly before it occurs in the plasma processing system.

[0050] Another aspect of the present invention is a system for detecting at least one anomaly in a plasma system, more particularly an RF power plasma system, preferably an RF power plasma processing system, which includes the following: • At least one sensor, particularly a directional coupler and / or a voltage-current sensor, for detecting at least one radio frequency signal of the plasma supply system and / or the power supply system and / or other characteristics of the plasma system of the plasma system. • At least one converter, in particular an analog-to-digital converter, connected to the sensor to convert at least one detected radio frequency signal and / or other characteristic into at least one input signal. A particularly programmable circuit for digital signal processing of at least one input signal, trained to detect at least one anomaly according to the method of the present invention. Therefore, the system according to the present invention provides the same advantages as those described in detail with reference to the method according to the present invention. The method according to the present invention has the advantage that multiple processing results specific to each configuration of parameters used are available, and the best configuration can be selected from the modified configurations based on the processing results. In other words, the parameters can be optimized by changing the configuration (also known as "changing the parameters"). Therefore, the selection of parameter results may include at least one of the modified configurations that yield the best processing result within the processing procedure. The use of parameter results for circuit parameterization can achieve improvements in digital signal processing (DSP).

[0051] The processing procedure may include iterative processing by a circuit, particularly digital signal processing, but may also include further steps such as evaluation and / or modification of the configuration, as will be described in more detail below. Furthermore, the processing procedure may be considered at least in part as a machine learning procedure. In other words, the processing procedure may include a machine learning procedure. The machine learning procedure may include (at least) evaluation, and / or storage of the determined parameter results into memory, and / or determination and / or provision of the parameter results, and / or learning and / or optimization of the configuration, preferably parameter values ​​and / or at least one weight of a neural network.

[0052] Within the scope of the disclosure of this invention, indefinite and definite articles, or numerical representations such as "1" and "2," should be understood as expressions of "at least," unless otherwise explicitly stated. Even when there is an explicit expression of "at least" or similar, it should not be concluded that a limitation in the sense of, for example, "exactly 1," is implied by another simple use of the article or numerical representation without this explicit expression. Furthermore, numerical representations, as well as representations of process parameters and / or apparatus parameters, should likewise be understood in a technical sense, i.e., with normal tolerances.

[0053] The use of the term "best possible outcome" implies the following: • Optimal results in digital signal processing, such as maximum suppression of unwanted frequencies, or the maximum number of specific abnormal signals detected from among a large number of abnormal signals. • Optimal results for measurement parameters related to RF power supply systems, e.g., RF power output with the least variation. • Optimal detection of a specific pattern in one of the measured signals. • Optimal results for measurement parameters related to plasma processing, such as DC self-bias or ion concentration. • Optimal results for measurement parameters related to the properties of articles processed in a plasma chamber, such as the thickness of the deposited layer.

[0054] To provide at least one input signal, the execution of the processing procedure may be initiated and controlled by a control circuit (hereinafter also referred to as "control") electrically connected to a programmable circuit. Furthermore, the determination and provision may be performed by the control circuit.

[0055] The term "digital signal processing" is generally understood to mean the processing of digital signals. Specifically, the digital signal processing according to the present invention may include processing of an RF signal after it has been detected by at least one sensor and subsequently converted into a digital signal in real time by an analog-to-digital converter, or ADC, of ​​the RF system. In contrast to digital signal processing, a "processing procedure" may be considered a training procedure that does not use an input signal that is actually detected in real time (instead, it may be pre-recorded or simulated on a computer). The algorithms used in processing procedures and processing in digital signal processing may be the same. However, the configuration of the algorithm is first trained in processing procedures and then applied in digital signal processing.

[0056] The term "RF system" refers to a system that outputs at least one electrical RF signal. An RF system may also include an RF generator, or be part of an RF generator. Therefore, the RF signal may be the output of an RF generator. The term "radio frequency power supply system" may refer to a specific configuration of an RF system, and specifically to the power supply of a plasma system, particularly a plasma generator system. This system may further include transmission lines and / or impedance matching networks. A plasma system may be suitable for generating plasma used in plasma processing, preferably plasma etching or plasma deposition. As is well known to those skilled in the art, plasma processing applications typically require a power signal of 100 W or more (amplified) to properly ionize a gas and generate plasma within a plasma processing chamber, and often signals exceeding 1 kW. However, the control signals involved in controlling these high-power signals may include much lower power signals and may be processed according to low-power signal processing techniques. Detecting the characteristics of high-power signals, such as RF signals, may require various types of sensors, and it may be necessary to pick up characteristic quantities of the signal, including only a small portion of the amplitude of the high-power signal.

[0057] The “at least one input signal” may preferably be a digital signal related to a radio frequency signal (also abbreviated as an RF signal or RF power signal) of a radio frequency system. In particular, this means that at least one input signal may be specific to an RF signal and / or may be obtained by sensing, and in particular by measuring, an RF signal. Thus, the RF signal may be a physical signal in the radio frequency range that is the actual electrical output of or within an RF system. The RF range may be a frequency range of 20 kHz to 300 GHz, preferably 300 kHz to 300 MHz (mid-frequency band, high-frequency band, and very high-frequency band). The input signal may also be a DC signal obtained from the RF signal and varying at low frequencies, for example, a measured DC bias voltage present in a plasma chamber in a pulsed application, measured forward and / or reflected power, reflection coefficient, phase between RF voltage and current, or a scalar signal from an external sensor.

[0058] After a machine learning procedure has been executed, the trained anomaly detection can be used to control the signal generation unit. In other words, in field operation, a parameterized programmable circuit may be used to detect anomalies using a detection procedure, and control of the signal generation unit may be initiated when at least one anomaly is detected. At least one anomaly can be detected before it actually occurs. That is, the parameterized programmable circuit can predict the likelihood of an anomaly occurring and initiate actions such as controlling the signal generation unit to prevent the anomaly from occurring. Thus, anomaly "detection" may also include anomaly "prediction". The trained detection may be part of digital signal processing.

[0059] The results of digital signal processing, particularly detection procedures, may be used to control a signal generation unit, such as a DDS core, or further circuitry in the RF system. The signal generation unit may also be used to adjust the amplitude, frequency, and phase of the output signal, which is converted into an RF power signal by a digital-to-analog converter (DAC) and a power amplification stage. As a result, the RF power signal from the output of the RF generator can be provided to the plasma processing system via an impedance matching network. In this way, for example, the results of filtering the input signal from the directional coupler at the output of the RF generator can be directly used to adapt the RF power signal, for example, to stabilize the RF power supplied to the plasma and minimize reflected power from the plasma.

[0060] In particular, the RF range may be 300 kHz to 500 kHz, or 1 MHz to 3 MHz, or 10 MHz to 16 MHz, or 24 MHz to 30 MHz, or 37 MHz to 43 MHz, or 57 MHz to 63 MHz, or 77 MHz to 83 MHz, or 159 to 165 MHz. For example, in the RF range of 10 MHz to 16 MHz, a specific frequency of the radio frequency signal may be 13.56 MHz. The aforementioned frequency ranges are used particularly in plasma applications and / or RF generating equipment.

[0061] Within the processing procedure, at least one input signal is repeatedly processed by the circuit. The circuit may be suitable for executing an algorithm for processing at least one input signal, depending on the configuration of at least one parameter. Thus, the parameter may be a parameter of an algorithm. In the case of filtering, the algorithm may be a filtering algorithm, and the parameter may be a filter parameter and / or filter coefficient. In the case of arc detection, the algorithm may be a pattern recognition algorithm, and the parameter may be the probability of a pattern matching a particular reference pattern. In other words, the probability may represent the similarity between a pattern and a particular reference pattern. The circuit may also be suitable for applying a neural network. In this case, the parameter may be at least one weight of the neural network. The circuit may be programmable in the sense that the configuration of at least one parameter can be changed.

[0062] "Iterative processing" means repeatedly processing the same input signal using the same algorithm, but it may be understood as using different configurations. Therefore, the configuration of at least one parameter is changed for each processing of at least one input signal to obtain each processing result. Each processing result, in other words, may be the respective result of the algorithm when run with each configuration.

[0063] Advantageously, at least one parameter may be a processing parameter that affects the processing result. If the processing involves filtering at least one input signal, at least one parameter may include at least one filter parameter. For example, at least one parameter may include an adjustable cutoff frequency and / or threshold.

[0064] "Determining at least one parameter result" can be understood as selecting at least one, or exactly one, of the modified configurations that yield the best processing result (and therefore, based on the processing result). This allows for optimization of the parameter configuration. "Providing at least one determined parameter result for improved digital signal processing" may be understood as parameterizing the circuit using the selected configuration.

[0065] The parameter results may include a specific configuration selected from the modified configurations, and this selection may be made depending on the processing result. For example, a criterion such as a metric or baseline result can be provided that is used to evaluate each of the processing results. Each processing result corresponds to the configuration used to obtain that result. As a result, each processing result can be used to evaluate, in particular, the configuration. Thus, the exact configuration corresponding to the best evaluated or best rated processing result can be selected. In other words, the configuration is optimized so that the difference between the corresponding processing result and the baseline is minimized. "Configuration optimization" is also called "optimization of at least one parameter," and at least one optimized parameter (i.e., the selected configuration) may then be used to parameterize (or another) programmable circuit for the "field" operation, "real-time" operation, or "application mode" of the radio frequency system. These terms are used synonymously below.

[0066] "Evaluation" may be understood as a comparison between the processing result and a standard, such as an ideal processing result. The better the processing result matches the standard, the higher the evaluation of the corresponding configuration. Evaluation may be performed by a control circuit or by a dedicated part of a programmable circuit. A variety of state-of-the-art methods can be used for evaluation, such as statistical methods for signal evaluation, multivariate statistical methods, and empirical comparative time series analysis.

[0067] In one embodiment, at least one parameter result includes at least two different parameter results for different applications of the DSP. Alternatively, or additionally, the DSP and digital signal processing may include filtering that is parameterized by configuration. The choice of filtering configuration has a significant impact on reducing interference in the RF signal. However, one particular configuration may not work for all combinations of RF power levels in multi-frequency applications, nor for all processing performed within the plasma chamber of a plasma system. Therefore, it may be necessary to determine a configuration optimized for different plasma processing conditions. This can be achieved by using multiple parameter results for different applications.

[0068] Similarly, in arc detection, one particular configuration may be optimized only for detecting a specific anomaly, but may not work for other types of anomalies. Therefore, detecting several different anomalies may be achieved by sequential and / or parallel processing of the input signals, applying different configurations to each process. Thus, a machine learning procedure may be run multiple times for different types of anomalies, such that the anomaly features are specific to different types of anomalies.

[0069] For each application and / or type of anomaly, a different set of input signals can be provided during the training process, which can be used individually in the processing procedure, thus yielding different parameter results. Depending on the sequence of different processing conditions in the steps of a plasma processing recipe, for example, if the processing conditions change during real-time operation, fast switching between different configurations may be advantageous. Therefore, it is possible to store different parameter results in memory for fast switching.

[0070] In digital signal processing and / or processing procedures, at least one input signal may be predefined by providing an already digitized signal that serves, for example, as a training input signal or a reference input signal. The already digitized signal can be an existing signal that is previously recorded or, for example, a simulated signal. Therefore, an ADC is not necessarily required for processing. However, in real-time operation, an ADC may convert an analogously sensed RF signal into a digital representation of the signal, which will be referred to below as the real-time input signal. In training mode, at least one predefined input signal may be predefined with desirable characteristics and features for the processing procedure and therefore may only "represent" an actual radio frequency signal. However, in real-time mode, a programmable circuit may process a real-time input signal from a sensed RF signal instead of a predefined input signal.

[0071] The RF system may include at least one sensor for performing detection, in particular measurement, of the RF signal used to supply RF power to the plasma system. Measurements may include measuring the RF signal in the form of forward power and reverse power and / or RF current and / or RF voltage, particularly at various points within the RF system. It is also intended to adjust the output power of the RF generator, for this purpose the RF signal is measured by the sensor in the form of an RF waveform, particularly at the output of the generator. A directional coupler may be used to provide the sensor with forward and reverse waveforms. Thus, the RF signal may be an analog signal, which is then converted by an ADC with a specific sample rate to provide a digital real-time input signal. Similarly, other sensors, e.g., in an impedance matching network, such as phase amplitude sensors, voltage-current sensors, DC bias sensors, or sensors directly installed in the plasma chamber, such as a Langmuir probe or optical sensor, can measure various parameters, such as RF voltage and current, the DC self-bias of the plasma, or the intensity of specific spectral lines in the plasma spectrum. These measurements are used to control adjustable elements to maximize power transfer to the plasma or to characterize the plasma properties. The measured parameters may be provided as real-time input signals or represented as training input signals.

[0072] The programmable circuit may be configured as a programmable logic device, such as a digital signal processor (DSP), a composite programmable logic device (CPLD), or an FPGA. According to the present invention, a programmable circuit may be provided, at least in part, by performing a detection procedure that includes the application of a (artificial) neural network, preferably pattern recognition or pattern matching using a neural network, to identify at least one anomaly feature. The detection procedure may provide evaluation of measurement data obtained by measuring an input signal, in particular an analog signal.

[0073] In algorithms for pattern recognition or pattern matching, or preferably for identifying at least one anomaly feature in an input signal, particularly in arc detection, a variety of algorithms can be used, as described below. A statistical pattern recognition algorithm can process measured data to obtain a set of characteristics, such as the mean and standard deviation, correlation coefficient, noise level, amplitude and direction of the deviation, and especially slow-changing background patterns and fast-changing anomalies. To adapt such an algorithm to a specific input signal for optimal detection and classification of anomalies, parameters such as deviation thresholds and correlation coefficients can be used. Furthermore, the accuracy of anomaly detection and classification may be improved by using a combination of several features, each of which may have an adjustable weight. The template matching algorithm divides the input signal into intervals containing a specific structure, which can then be compared to a stored anomaly template. The parameters for dividing the input signal into meaningful intervals, e.g., periods of steady-state operation and periods with large deviations from the mean and / or noise level, may be optimized for the best detection of features having similarity to the stored template. Furthermore, the scaling coefficient or scaling coefficient range used to normalize the input signal so that it can be compared to the template is also a parameter that can be modified and optimized. In addition, variations in the correlation coefficient threshold can be used for further algorithm optimization.

[0074] As an alternative to conventional deterministic algorithms, neural networks can be used to automatically extract features from input signals and classify them. In particular, the following types of networks can be used for time series analysis and anomaly detection, such as arcs: • A feedforward artificial neural network comprising an input layer, one or more hidden layers, and an output layer, such as a multilayer perceptron (MLP) network with a nonlinear activation function. • In particular, convolutional neural networks (CNNs) that include convolutional layers to utilize the local correlation of time-series input signals. Recurrent neural networks, particularly long-term short-term memory (LSTM) networks, are well-suited for identifying and predicting individual anomalies in input signals that may be inconspicuous over long periods of time.

[0075] When neural networks are used for pattern recognition of input signals, parameters that control neuron weights, activation function thresholds, or the shape of nonlinear activation functions may be parameters that can be modified to optimize the neural network for optimal detection and classification of anomaly features. Backpropagation using gradient descent may be used to train the neural network for optimizing these parameters.

[0076] In training an algorithm or neural network that can be used to process input signals, i.e., in a machine learning procedure, a training dataset, i.e., a number of existing input signals (as at least one input signal according to the present invention), can be used and can be generated and provided by using at least one of the following elements: Each of the various anomalies detected may need to be represented by a sufficient number of recorded input signals in the training dataset, e.g., 10% or 20%, so that the anomaly training dataset contains both anomaly and non-anomaly input signals. • Certain anomalies may be artificially induced by appropriate parameters of the plasma process or RF power supply system, and may occur more frequently or with greater intensity compared to normal operation. In this way, a large number of input signals with specific anomalies can be recorded in a short period of time. For example, an input signal with specific anomalies, evaluated and recorded by an engineer, may be used to automatically generate a set of more similar input signals by scaling, shifting, expanding, compressing, and / or offsetting features, by adding noise with specific characteristics, by adding a slowly changing baseline, or by filtering operations. In this way, batches of input signals with similar anomalies can be generated and included in the training dataset. Furthermore, it is possible to generate specific anomalies detectable over longer time scales, such as characteristic trends indicating the degradation of specific components, from the simulation and convert them into input signals that are added to the training dataset. Each input signal stored in the training dataset can also be complemented by its correct classification and / or expected processing result. In training mode, parameters can be modified for all signals in the training set to minimize the deviation between the processed result and the expected result.

[0077] In particular, in training mode, at least one input signal can include multiple predefined batches or multiple existing input signals (i.e., previously recorded or, for example, simulated). Each input signal can be associated with a radio frequency (RF) signal, i.e., it can be specific to or represent an actual RF signal or a simulated RF signal. It can also be easily derived from a measured RF signal, for example, by generating a digital representation from the measured RF signal. Furthermore, each input signal may contain a known pattern, and the corresponding reference result may exhibit this known pattern, which will be discussed in more detail below. Existing input signals can be considered as references or reference signals because they contain specific features and / or can provide a corresponding reference result. Configuration optimization, i.e., learning the optimal parameterization of a programmable circuit, can be achieved by comparing the reference result with the processed result. The reference result can include a target waveform, a target frequency, or a desired pattern detection result. The comparison can be achieved by using an appropriate metric that indicates the degree of discrepancy between the reference result and the processed result.

[0078] After providing at least one input signal in the form of a batch of input signals, i.e., (at least) a first input signal and a second input signal, in particular at least 10, 100, or 1000 additional input signals, the processing procedure, also known as the training procedure, may include the following processing flow: In the first processing step, the first input signal is processed using the first configuration, the second input signal is processed using the first configuration, and further input signals are also processed using the first configuration, thereby obtaining the first processing result for the first configuration. In the second processing step, the first input signal is processed using the second configuration, the second input signal is processed using the second configuration, and further input signals are also processed using the second configuration, thereby obtaining the second processing result for the first configuration. • In the further processing steps, the first input signal is processed using the further configuration, the second input signal is processed using the further configuration, and the further input signal is also processed using the further configuration, thereby obtaining the corresponding processing result for each of the further configurations.

[0079] Therefore, the process is repeated with different parameter configurations, and the configuration changes with each process. In the training procedure, the configuration change (and thus the parameter change) can be performed by established methods such as gradient descent or backpropagation techniques.

[0080] Within the scope of the present invention, it may be advantageous for the programmable circuit to be configured as a programmable integrated circuit, preferably a digital signal processor (DSP), a composite programmable logic device (CPLD), or a field-programmable gate array (FPGA). Alternatively, the circuit may be configured as a fixed circuit, such as an application-specific integrated circuit (ASIC) or another integrated circuit. In general, any computer or controller (e.g., PC, microcontroller, DSP, CLPD, FPGA, etc.) can be used for machine learning, particularly with machine learning procedures. FPGAs, in particular, are well-suited for high-speed real-time digital signal processing and can accelerate machine learning procedures. Furthermore, the use of programmable circuits, especially FPGAs, enables flexible and automated training. Many applications of radio frequency systems, such as plasma applications, are evolving in various ways, and FPGA solutions have the advantage of being able to handle future applications requiring adaptive processing, such as yet-to-be-known filter designs, with the same hardware architecture. In contrast, fixed circuits such as ASICs need to be modified or completely redesigned, which increases both the workload and cost. The method according to the present invention preferably allows for easy addition of new configurations via software upgrades, which may include more advanced filter configurations. Users can select one of these configurations for new applications.

[0081] Furthermore, within the scope of the present invention, the present invention may optionally further include the following: • Parameterize the same or another programmable circuit using at least one determined parameter result. • Detects radio frequency signals from a radio frequency system, • Perform digital signal processing by using processing of detected radio frequency signals with at least partially parameterized circuits. The monitoring of radio frequency signals by repeatedly performing detection and digital signal processing, preferably the processing procedure can be a training procedure used to perform parameterization of a programmable circuit in the training mode of the processing using an existing signal as a training input signal, and / or can be configured using parameterizations that may have been previously determined in the training mode and obtained from memory, on the other hand, the monitoring in particular is performed in the real-time mode of the processing in which the radio frequency signal is detected as a real-time input signal and processed. This means that the radio frequency system can actually be used in real time, and that processing can be switched from training mode to real time mode.

[0082] At a later point, the radio frequency system can be switched back from real-time mode to training mode, allowing for further optimization of parameterization by iteratively processing additional existing signals, such as previously recorded or simulated signals.

[0083] In particular, a key advantage of the present invention is that the programmable circuit enables both real-time processing of the input signal and training of the system by iteratively processing existing training signals to determine the optimal configuration of the real-time processing mode. The existing signal may be a previously recorded signal or a computer-simulated signal, among others.

[0084] These two features make the RF system highly versatile, allowing it to be trained to handle new plasma processing conditions or detect new situations without changing the hardware or algorithms. Only the parameterization of the programmable circuitry changes. Instead of using an external computer to analyze measured signals and determine an improved configuration, such an optimized configuration can be determined within the actual RF system itself by switching it into training mode. In this way, the programmable circuitry of the RF system possesses the characteristics of a machine learning-enabled system or on-device inference framework, both of which can adapt to new applications.

[0085] A further advantage of the present invention is that the programmable circuits employed, particularly FPGAs, perform digital signal processing much faster, for example, 10,000 to 100,000 times faster, compared to standard model simulations using standard computers and tools such as Matlab / Simulink.

[0086] The parameter result may include an optimized parameter set, i.e., the configuration of each of at least one parameter optimized using a processing procedure, so that the programmable circuit may be parameterized using this parameter set of the parameter result. As a result, the method according to the present invention achieves improved digital signal processing by using highly accelerated optimization of at least one parameter in the first step. In the second step of real-time operation, the processing procedure and the programmable circuit are used, making the selection of the parameter set for digital signal processing faster and smarter. For digital signal processing of a radio frequency signal detected during real-time operation, the programmable circuit can receive at least one real-time input signal converted from the detected radio frequency signal by an analog-to-digital converter.

[0087] Furthermore, the circuitry for optimizing at least one parameter can be the same as that used in the final application for real-time operation of the radio frequency system.

[0088] Furthermore, within the scope of the present invention, at least one input signal processed in the processing step is configured as at least one training input signal, and the method may further include the following: The parameterization of the programmable circuit is completed by configuring at least one parameter using previously determined parameter results, and preferably, the configuration is then fixed. - Switching the processing from training mode to real-time mode, in particular in real-time mode, the real-time input signal is processed in real time by a programmable circuit using a fixed configuration, particularly for filtering the real-time signal, preferably the real-time input signal is configured as an actually measured signal received from at least one analog-to-digital converter that converts radio frequency signals detected by at least one sensor into at least one real-time signal. - Switching the processing from real-time mode to training mode, preferably in training mode, by re-executing the processing procedure with at least one additional training input signal as the processed training input signal, thereby further optimizing the existing previously fixed configuration, and performing re-parameterization of the programmable circuit by configuring at least one parameter using the newly determined parameter result in response to a comparison of the newly determined parameter result with the previously determined parameter result. This allows for further optimization of the existing parameter set. The user can switch the system back to training mode to decide whether to load additional reference input signals and reference results for the existing configuration. Loading or uploading can be done via an electronic interface or a user interface. Alternatively or additionally, additional reference input signals and reference results may be recorded by the RF system itself and loaded directly into memory in real time for later use in training mode. The old and new sets of parameter configurations are then evaluated, and the better parameter set is subsequently stored in memory. Furthermore, switching from real time mode to training mode allows for training of new situations at a later point in time; that is, switching to training mode, loading new reference input signals and reference results into memory, and determining and saving the optimized parameter set to memory.

[0089] Training results using new reference input data can be used to optimize existing parameter configurations, i.e., existing parameter configurations can be overwritten with optimized ones, or the optimized parameter configuration can be used for new applications, stored in memory as a separate parameter configuration, and, especially when used for new applications, can be instructed by the user via an electronic interface to the RF system.

[0090] According to another embodiment of the method of the present invention, the iterative processing of at least one input signal is performed by an iterative processing step, in which the same at least one input signal is processed by the same programmable circuit using a modified configuration, and each processing result is specific to and / or assigned to the modified configuration, in particular, at least one input signal is provided as one or more predetermined reference input signals, for each reference input signal a reference result is provided that represents the desired result of processing the corresponding input signal, and at least one parameter result for configuration optimization based on the processing results is determined, which includes: The modified configuration is evaluated by comparing each of the processing results obtained for at least one reference input signal with the corresponding reference result, preferably the evaluation depending on the agreement between the processing result and the reference result, and in particular, identifying the deviation between the processing result and the reference result. In other words, the parameters can be optimized by changing the configuration. The same iterative processing of at least one input signal may be used, and different configurations in the sense of different parameter values ​​may be applied. As a result, the processing results will likely be slightly different. Each processing result can be used to evaluate the performance of the particular configuration used to obtain the said processing result.

[0091] In particular, the reference, reference input signal, and / or reference result referred to within the scope of the present invention are also called reference data.

[0092] Furthermore, by using evaluation, it is possible to provide configuration optimization in the sense that the parameter results can include a set of parameters that yield a processing result that minimizes the deviation from the reference result for all reference input signals. Configuration optimization can also be called parameter optimization. Therefore, the processing procedure may be called a training procedure and / or an optimization procedure. At least one given reference input signal and its corresponding reference result may be called training data. The determined at least one parameter result can then be verified by performing the processing and evaluation of the processing result again on a second set of reference data (including a new input signal and its corresponding reference result). In this way, the problem of overfitting the parameter configuration to the training data is avoided.

[0093] Determining at least one parameter result for optimizing the configuration based on the processing results can be provided within the scope of the present invention, and includes the following: • Select at least one configuration with the highest rating as a result of at least one determined parameter. In particular, providing at least one determined parameter result for improved digital signal processing is important. • Parameterize the same or different programmable circuit using at least one selected configuration. Here, preferably, the at least one predetermined reference input signal and the corresponding reference result includes a plurality of predetermined reference input signals and the corresponding reference results for different applications of digital signal processing, wherein determining the at least one parameter result for optimizing the configuration based on the processing result is performed individually for each application, and preferably, providing the at least one determined parameter result includes storing the selected configuration in memory, and parameterizing the same or a different programmable circuit includes selecting the configuration to be used from the stored configurations depending on the application. In other words, different training datasets can be provided for different digital signal processing applications. This allows for switching between parameter sets for different applications. In further possibilities, at least one input signal is composed of multiple existing input signals (also known as a "batch of signals"), and preferably the processing procedure includes a repeating processing step, each processing step including: • Processing the same input signal one after another, The control circuit sets the configuration of at least one parameter used in this processing step. This method may also include the following: After each processing step, the control circuit evaluates the deviation of the processing result from a predetermined reference result. After each evaluation, the configuration of at least one parameter used for the next processing step is modified based on the evaluation. Preferably, the control circuit is electrically connected to the programmable circuit. It is configured separately from the programmable circuit, In particular, the configuration of at least one parameter used is changed for each different processing step, and the improved digital signal processing configuration is iteratively optimized. The control circuit or control unit may be configured as an additional control processor separate from the programmable circuit, having access to a memory containing training data, such as at least one input signal and / or multiple batches of existing input signals and / or corresponding reference results. The existing input signal may be a previously recorded signal or a simulated signal, etc. The control circuit can also be implemented on top of the programmable circuit. The programmable circuit may be an FPGA. Providing at least one input signal, in particular selecting training data for a processing procedure (e.g., based on the application to be trained), and / or changing the configuration within the processing procedure does not have to be performed directly by the control circuit rather than by the programmable circuit itself. For example, with each iteration of the process, the control circuit may supply the input signal and / or the modified configuration to the programmable circuit. The control circuit can then perform an evaluation and, depending on that evaluation, decide on the next change to the configuration. To accelerate the training procedure, the evaluation may also be implemented at least partially by the programmable circuit.

[0094] Alternatively or additionally, the programmable circuit can be directly connected to shared memory, or at least a portion of the memory may be implemented by the programmable circuit. In this case, the control circuit can instruct the programmable circuit from which address range of memory to load the training data. The programmable circuit can then directly access the memory, load and process the training data sequentially, and evaluate the processing results. This method has the advantage of significantly speeding up reads from memory.

[0095] The processing procedure may be configured as a training procedure for iteratively optimizing a configuration for different specific situations in a radio frequency system application, and the resulting parameter results determined for different situations may be configured as different optimized parameter sets for those different situations, which may be further provided within the scope of the invention. The different situations may be different digital signal processing applications. The different parameter sets may include different configurations adapted to these different situations. This allows for flexible switching of parameter sets depending on a given situation. Thus, the training procedure enables training for different situations, which can result in parameter sets optimized for different situations.

[0096] In another advantageous embodiment, at least one input signal includes a temporal signal course, preferably instantaneous and historical values, in particular a measurement of a radio frequency signal or a measurement obtained from that signal, and the processing result is obtained by processing the signal course, indicating the probability of detecting a particular pattern in the signal course. Thus, not only instantaneous values ​​of the input signal are considered, but a series of values ​​from a specific time interval prior to the most recent point in time are also processed, and the probability of detection, i.e., the expected result, can be calculated. In training mode, instantaneous and historical values ​​may be predefined and / or previously recorded. However, in real-time mode, the real-time input signal may also include a temporal signal course, preferably instantaneous and historical values, in particular a measurement of a radio frequency signal or a measurement of a signal derived from a radio frequency signal. In other words, in real-time mode, the processing considers not only instantaneous measurements of the radio frequency signal, but also a series of values ​​from a specific time interval prior to the most recent measurement point. This allows the processing to be used in different applications, such as probabilistic evaluation.

[0097] At least one input signal, i.e., a training input signal and / or a real-time input signal, may be configured as at least one digital signal. The digital signal may be provided by an analog-to-digital converter that converts the sensed analog RF signal into a digital signal. The digital signal can be a digitized representation of the sensed RF signal and can therefore be treated as a one-dimensional array of data points describing the instantaneous amplitude of the measured RF signal. For time-sequential measurements, the array can be further extended so that the most recent data point represents the most recent measurement. The temporal measurement rate of the measurement, and therefore the temporal measurement rate of the array extension, may depend on the sampling rate of the digital-to-analog converter. By selecting a time frame of a specific length that always ends with the most recent data point, a series of one-dimensional arrays updated at the measurement rate or sampling rate of the digital-to-analog converter can be obtained by shifting older data points from one end of the array and adding new data points to the other end of the array.

[0098] In particular, real-time measurements require identifying critical situations and addressing them before they escalate into undesirable events. Such methods are especially useful in handling overvoltage or overcurrent conditions in electrical equipment, or in managing arc events during plasma processing. Consequently, it may be necessary to continuously analyze streams of measurement data regarding the presence of specific patterns.

[0099] Advantageously, the method according to the present invention further includes the detection of at least one pattern in the detected radio frequency signal and / or at least one input signal. To improve the reliability of detecting a particular situation, radio frequency signals measured by several sensors can be used in combination such that at least one pattern is present in each signal at a particular time. For example, successful detection of an event may require that a particular pattern be detected in two RF signals at the same time or with a certain time delay in between. Therefore, by evaluating at least two signals in combination, it is possible to avoid false detections due to measurement artifacts or interference that may occur when evaluating only one signal.

[0100] In particular, “performing digital signal processing by at least partially using processing of a radio frequency signal detected by a parameterized circuit” and / or “performing a processing procedure in which at least one input signal is repeatedly processed by a programmable circuit” each include the detection of at least one pattern in the detected radio frequency signal and / or at least one input signal. The detected radio frequency signal may be configured as a real-time input signal and therefore as a digital signal representation as described above. Therefore, at least one input signal is also used in the processing procedure, and it may be configured as a training input signal and therefore as a digital signal representation as described above. In either case, the processing may be designed to provide the detection of at least one pattern, preferably in the form of pattern recognition. Therefore, the configuration may be a configuration for pattern detection, in particular a configuration of a pattern matching algorithm or a pattern recognition algorithm, and / or a configuration of weights for individual neurons of a neural network. Pattern matching algorithms or pattern recognition algorithms have the advantage of being less technically complex to use than neural networks. However, in some situations, neural networks can more reliably detect patterns in noisy or fluctuating signals than simple pattern matching algorithms. It is also possible to combine both. Therefore, the configurable parameters may be configured as algorithms or neuron parameters of a neural network. The processing procedure may also be configured as a machine learning procedure for machine learning and / or training a neural network, or may include such a procedure, or may be configured as part of a machine learning procedure, and digital signal processing may be relevant to the application of the trained neural network. At least one parameter result for configuration optimization may be further configured as a training result in the form of a set of neuron weights. The occurrence of events may be detected and / or predicted based on the detection of at least one pattern.

[0101] Furthermore, it may be provided that an action is initiated based on detection or prediction. In particular, the result of pattern detection is the probability of the pattern occurring. In addition, or alternatively, detection and / or prediction based on detection involves determining the probability of an event occurring. The initiation of an action may depend on comparing this probability with a threshold. This makes it possible to initiate an action, in particular, before the event is completed. The action may also be a corrective action to counteract the occurrence of an event, such as shutting down the RF system.

[0102] In more advanced embodiments, the probabilistic detection or prediction of an event can be based on several conditions. This can be done by monitoring the progression of the probability of an event occurring over a specific period. Rather than simply detecting an event by comparing an instantaneous probability value to a threshold, additional conditions for more reliable detection may include, for example: • The probability value that consistently exceeds a specific threshold over a specific period of time. • The average probability value that exceeds a specific threshold over a specific period. • A probability value that increases after exceeding a certain threshold. • The derivative of the probability value within a specific period that exceeds a threshold. • Correlation of probability values ​​determined for several monitoring signals. • Any combination of one or more of the above conditions and any further conditions.

[0103] With the help of more complex conditions, the reliability of event detection can be increased, and reliance on instantaneous fluctuations in probability values ​​can be reduced. Advantageously, by selecting a sufficiently short monitoring period for probability values, events can be suppressed before they are completed by initiating action. Therefore, system training can include parameterization for detecting events as early as possible and / or classifying events as accurately as possible. As a result, higher weighting can be applied to the parameterization.

[0104] It is possible to provide the ability to detect a specific pattern in at least one digital signal, particularly in the form of a line array as described above. Preferably, pattern detection is trained using a training input signal as a digital signal and / or performed in real time using a real-time input signal as a digital signal. The latter makes it possible to identify events in operation of the RF system based on pattern detection. Events may occur in the RF system or plasma processing system and may relate to problems requiring action, such as the shutdown of the RF system. Furthermore, events may indicate the deterioration of a particular component of the RF system or plasma processing system. Events can be reported by notifying the user, for example, via a specific analog or digital signal that can be provided via an electronic interface or user interface, allowing the user to respond to the situation with additional corrective action or to schedule repair and / or preventive maintenance at a later time. Certain patterns may include interferences, and / or anomalies and / or changes in characteristics, as described above. These patterns may be related to specific phenomena in plasma processing, plasma processing systems, or RF power supply systems. When an RF signal is repeatedly detected by continuous measurements, the digital signal represents a continuously detected signal. By continuously processing the digital signal, particularly the resulting sequence of one-dimensional arrays, it may be possible to monitor the temporal changes of specific patterns.

[0105] In the case of periodic pulse modulation of RF power signals, abrupt changes in RF power, such as switching between two levels in pulse-mode operation, may be recorded as spurious arc events by pattern matching algorithms and / or neural networks. Such intentional power changes can be flagged as arc events, especially when evaluating at least two signals in combination. Similarly, when RF power is turned on and the plasma is first ignited, an arc event may also be flagged.

[0106] To avoid such false positives, anomaly detection by pattern matching algorithms and / or neural networks can be briefly switched off ("dead time") whenever the plasma processing recipe or pulse pattern requires a switch in RF power level. In this way, only actual arc events occurring within a certain RF power period can be detected.

[0107] Furthermore, by using artificial neural networks, the detection of specific patterns in their entirety can be learned by optimizing the weights of individual neurons, and using the optimized set of weights, the network can recognize the occurrence of specific patterns in real time. To detect events before they are fully completed, an appropriate output function of the neural network can be selected to generate the probability of a specific pattern in a real-time input signal or data stream.

[0108] The same result can also be achieved by calculating the cross-correlation between a reference pattern and a stream of actually measured signal data. The reference pattern is continuously shifted in time with respect to the signal waveform.

[0109] If an event progresses to completion, the pattern will be recognized with a probability close to 1. However, prior to that point, the calculated probability has already gradually increased from 0 or close to 0 towards 1. As soon as the probability exceeds a user-adjustable threshold, or a combination of some of the above conditions, the event can be reported before it actually completes. In this way, anomalies in the stream of measurement data that correlate with undesirable or catastrophic events can be flagged before the event occurs. In this case, measures can be initiated to avoid the event progressing to completion and to reduce and / or suppress serious consequences. Probability thresholds and further conditions as described above can be set individually for each type of event. For example, catastrophic events can be handled with a low threshold, while minor anomaly events can be handled with a higher threshold. The advantage of using a low threshold for serious anomalies is that there is more time to perform corrective actions before the event progresses to completion. On the other hand, a higher threshold allows for classifying events with higher accuracy.

[0110] For example, an arc on the surface of a semiconductor wafer in a plasma processing system can cause significant defects and loss of one or more chips on the wafer. If the arc is strong enough, the plasma processing sequence can be severely disrupted, potentially resulting in the discarding of the entire wafer. Therefore, in the case of such catastrophic arc events, a low probability threshold is advantageous so that the event can be suppressed early. Examples of using this probabilistic method to suppress arcs before they fully progress are discussed further below. In another example, a higher probability threshold may be more appropriate for less critical events. In plasma chambers used for film deposition, not only the semiconductor wafer but also parts of the plasma chamber are coated with the deposited layer. After processing a large number of wafers, these chamber components are covered with increasingly thick layers of deposited material, from which particles can fall. During plasma processing, such particles can cause plasma anomalies that are to a certain extent acceptable. Nevertheless, it is important to accurately classify and measure all particles associated with an event so that preventive maintenance, in this case cleaning of the chamber, can be initiated if the number of particles associated with the event exceeds a threshold. In this case, a high probability threshold allows particle-related events to be isolated from other anomalies, thus avoiding premature cleaning of the chamber caused by the measurement of false particle events.

[0111] After detection, classification of the detected anomalies into different anomaly classes may be performed. The classification may be trained in the same way as the detection, i.e., the classification can be trained by performing a machine learning procedure, and the detected anomalies can be trained to be classified into different anomaly classes by processing the detected anomalies with a programmable circuit. The detection of anomalies in the input signal and the classification of these anomalies into different anomaly classes may relate to a particular aspect of deviation from normal operation, or to a malfunction of an RF power supply system, plasma system, or plasma process, and may include various signal preprocessing techniques within the detection algorithm or neural network, thereby including, for example, at least one of the following: • Subtraction of slowly changing baseline fluctuations and offset (which may be observed even in normal input signals without abnormalities). • Scaling operation (temporal and / or amplitude-wise) before comparing the input signal to a normalized anomaly pattern. • Especially when evaluating a combination of multiple input signals, time-domain shift operations such as adding or subtracting a time offset are used to adjust for anomalies. • Calculation of correlation between at least two input signals • Removal of background noise through filtering operations, etc. • Removal of at least one specific frequency range from the input signal. After preprocessing the input signal using one or more of the techniques described above, an actual pattern detection algorithm, including probability calculations, can be applied to the resulting normalized input signal.

[0112] Implementing an optimized neural network on a programmable logic device, such as an FPGA, is straightforward because all parts of the neural network can be transferred to their corresponding logic components. Similarly, pattern matching algorithms can be easily implemented on an FPGA using shift registers and simple logic operations. Changing the configuration or size of the neural network, or modifying the pattern matching algorithm, can be done by loading the new configuration onto the FPGA without changing the hardware.

[0113] If an application requires the detection of multiple pattern classes, multiple parallel neural networks or pattern matching algorithms optimized for each specific class can be configured. Alternatively, multiple pattern classes can be detected using a single common neural network or pattern matching algorithm, for example, by sequentially evaluating the input signal in different, distinct steps, specifically one per class. In this case, instead of a single probability value as the network output, an array of probabilities, one for each class, is generated. Further pattern class detection can be added later. Neural network parameterization or pattern matching algorithms can be optimized through additional teaching as needed.

[0114] This allows for saving different configurations for different applications and activating them as needed. In particular, it enables rapid switching between different configurations, for example, by activating them one after another for different sequential processing steps. Subsequently, by activating teaching mode on the installed system, it may be possible to further optimize existing configurations or learn new situations with additional reference data. Thus, probabilistic pattern recognition in measured signals is also possible by applying parallel (simultaneous) classification to several pattern types, even before the pattern is fully unfolded.

[0115] Furthermore, an electronic interface may be provided for uploading at least one input signal or an existing configuration to the system of the present invention, particularly the programmable circuit, or the control unit, or to memory. The method of the present invention then further parameterizes the programmable circuit using the uploaded existing configuration. If the input signal is uploaded to the control unit, the control unit may load the signal into memory before supplying it to the programmable circuit. At least one input signal can be configured as multiple input signals, which are stored in memory after uploading.

[0116] The electronic interface may be configured as a user interface, allowing the user to upload signals using at least one input signal, and / or user input signals, and / or user configurations, and / or control signals for switching the system from training mode to real-time mode, and / or specific configurations stored in memory for parameterizing programmable circuits. Uploaded input signals are generally formulated training data, which may be supplied to programmable circuits for digital signal processing directly or via a control unit, or stored in memory directly or via a control unit. User input signals mean at least one input signal from the user. User configurations may be existing configurations that the user loads into the interface from, for example, cloud or on-site storage, or a removable data carrier medium that can be connected to the user interface. User configurations may also be configurations created by using the method of the present invention on user input signals. Generally, training data may include at least one input signal and a reference result corresponding to each input signal. Training data may be configured as existing (previously recorded or simulated) data, in particular for a training procedure, which can be initially loaded into memory and processed one by one (in various iterations) until an optimized parameter set is found. Uploaded existing configurations can be stored in the same memory and / or different memory, and / or uploaded existing configurations can override configurations already implemented according to any of the methods of the present invention described above. When uploaded existing configurations are stored in the same memory, they can function as supplemental / additional parameterization of programmable circuits. The advantage of storing uploaded configurations in a separate memory or memory space is that a backup configuration is always available. Restricting access to configurations other than those created by the user is also an advantage for manufacturers. In this case, loading parts of the configuration into a secure and reliable memory space is also optional.The advantages of overwriting the configuration of a programmable circuit by uploading a different configuration may include new parameterization of the programmable circuit and a reduction in the required memory space.

[0117] It is possible to provide an electronic interface to access parameterized and / or fixed configurations, and / or at least one input signal and / or data from memory, / or programmable circuits, and / or control units, particularly for reading and / or writing and / or overwriting. However, it is also possible to restrict access via the electronic interface so, for example, that reading and / or overwriting and / or exchanging specific data and / or algorithms or neural networks is prohibited. Thus, some of the data or information stored in memory and / or programmable circuits and / or control units can be locked, for example, by using encryption methods. In other words, it may be possible to use the electronic interface to perform the uploading of configurations stored in memory for use in a particular application. These configurations may be the result of training procedures performed on similar circuits using appropriate training data. In particular, manufacturers can allow users to upload configurations, but at the same time protect the basic factory configuration and the algorithms or neural networks themselves from user access. This has the advantage that if the user is not satisfied or the system stops working with the new configuration, it can be reverted to the factory configuration.

[0118] By using only programmable circuits, user interfaces, memory, and controls as a training bench, it is also possible to perform training for circuit parameterization and determine the optimal parameter set. Similar to the training mode of a complete radio frequency system, already digitized training data is read from memory and provided by the controls as input signals for digital signal processing. Parameter changes and evaluation of processing results are also performed in the same way as with a complete RF system. This training bench system, with a common commercially available FPGA, can determine optimized parameter results in a much shorter time, for example, 10,000 to 100,000 times faster, compared to standard model simulations using a standard computer and tools such as Matlab / Simulink.

[0119] Optimizing the parameterization of typical filtering or pattern matching algorithms used in radio frequency systems can take hours or even days using Matlab / Simulink simulations, whereas it can be determined in minutes using the training bench system described.

[0120] For example, even with a large neural network containing thousands of neurons arranged in 500 to 1000 layers, a training bench system can determine the optimal set of neuron parameters within a few hours. For even more complex algorithms or neural networks, and / or very large sets of training input data, determining the optimal configuration of the programmable circuit can be performed on a dedicated high-performance supercomputer.

[0121] Independent of the above possibilities for determining the optimal configuration using the RF system itself, a training bench system, a standard computer with simulation tools, or a high-performance supercomputer, the determined configuration can be uploaded to the RF system's memory via a user interface by control, and then used to parameterize programmable circuits that can process the RF system's input signals in real time for specific applications.

[0122] Similarly, a configuration determined in one RF system can be exported via the user interface and uploaded to a similar RF system for parameterization of programmable circuits. In this way, implementing an existing configuration from the first RF system into a second system of the same type avoids the need for separate training of the second system. Therefore, since a copy of an already optimized configuration can be used in multiple systems, there is the advantage of saving time when setting up a new system.

[0123] Due to unavoidable small variations between systems, it may be necessary to further optimize the parameterization of a new RF system. Nevertheless, this parameter optimization usually converges faster when starting from a known good configuration than when starting from the default parameter configuration, which is the case for entirely new applications.

[0124] At least one input signal can be configured as a predefined reference input signal, for which a desired processing result is provided as a corresponding predetermined reference result. The training data may also include the predefined reference input signals and their corresponding predefined reference results.

[0125] To provide processing for at least one input signal, the circuit may be capable of executing an algorithm parameterized by at least one configurable parameter of the circuit. An interface, in particular a user interface, can be used to select an appropriate parameter configuration before processing begins. Alternatively, or additionally, the interface may be used to switch from an application operating mode or a real-time operating mode to a training mode. The processing procedure may be configured as a learning cycle to optimize the configuration. It may also be possible for the user to control the interface to load a reference input signal and / or a reference result and / or training data into memory, in particular the memory of a radio frequency generator, and / or execute a learning cycle and / or execute another learning cycle and / or store the new parameter configuration in memory for later use in the application mode. In training mode, the memory can be used to simulate a continuous data stream from the analog-to-digital converter from the application mode. Thus, the memory is connected to a control circuit, which can read at least one input signal from the memory and provide it to the programmable circuit in training mode. To accelerate processing, it is also possible for the programmable circuit to read at least one input signal directly from the memory and perform digital signal processing. On the other hand, in application mode or real-time operation mode, the programmable circuit can be directly connected to the analog-to-digital converter to receive the input signal as a real-time input.

[0126] A further option is that at least one input signal is configured as at least 1,000 or at least 1 million different input signals, each of which represents at least one waveform related to a radio frequency signal, and the processing is aimed at evaluating the waveform according to the parameter configuration, thereby each processing result constitutes an evaluation result. This can further improve the performance of the processing. Furthermore, automated training controlled by a control circuit, for example, can accelerate the optimization of at least one parameter.

[0127] In another advantageous embodiment, the programmable circuit may include a logic element for decomposing an input signal into amplitude and phase using an algorithm, and the algorithm can be parameterized by a modified configuration. Thus, algorithms such as IQ demodulation, Goertzel's algorithm, and the Discrete Fourier Transform may depend on many parameters. For example, in the case of IQ demodulation, the parameters can be the coefficients of a digital filter. Thus, the configuration of the parameters can be specific values ​​selected for these parameters.

[0128] Furthermore, the evaluation of a modified configuration can be performed by comparing each of the processing results associated with the modified configuration with a predetermined baseline result, which includes a predetermined ideal result of an algorithm executed by a programmable circuit. Thus, the predetermined ideal result serves as a benchmark for the quality of the currently set configuration. Changing the parameter configuration can generate different processing results that can be compared with the baseline result. The deviation between the processing result and the baseline result indicates the performance of the configuration used, which can be understood as an evaluation of the configuration.

[0129] Furthermore, another aspect of the present invention is an RF measurement system for a system, in particular a radio frequency system, preferably a radio frequency power supply system or a radio frequency generator or radio frequency plasma system, which includes the following: • At least one sensor for detecting radio frequency signals of a radio frequency system, in particular a directional coupler or a voltage-current ("V / I") sensor, • At least one analog-to-digital converter for converting at least one detected radio frequency signal into at least one input signal. • A circuit for digital signal processing of at least one input signal, particularly a programmable circuit. The circuit can be provided to be parameterized by at least one parameter result determined according to the method of the present invention. Accordingly, the system according to the present invention provides the same advantages as those described in detail with reference to the method according to the present invention. The system may also include an analog anti-aliasing filter and / or a transmission line and / or a matching circuit, in particular an impedance matching circuit. The radio frequency system may be configured in particular as a radio frequency power supply system. The radio frequency power supply system may further be configured as a power source for a plasma system, in particular a plasma power supply system. An analog-to-digital converter may be provided for use in the power supply and for converting detected radio frequency signals detected by sensors in the radio frequency system.

[0130] The system may further include an electronic interface for uploading existing configurations to programmable circuits, or to a control unit, or to memory, for the parameterization of the circuit. Uploading and / or writing and / or reading using the interface can be performed using data transfer. The electronic interface may be electronically connected to a control unit, memory, or programmable circuit to perform uploads using (particularly direct) data transfer. The electronic interface may be configured as a data interface, such as a serial interface, which can connect to another electronic device, such as a computer. The electronic interface may also run a computer program to enable data transfer. The computer program may also provide more advanced interface technologies, such as a local area network, wireless local area network, or Bluetooth. The electronic interface may be accessed via a control unit, memory, or programmable circuit.

[0131] The system according to the present invention features any of the above-described features of the method according to the present invention and can further demonstrate the advantageous features described.

[0132] Another aspect of the present invention is a data carrier signal carrying at least one parameter result determined according to the method of the present invention. Thus, the data carrier signal according to the present invention provides the same advantages as the corresponding features described in detail for the method and system according to the present invention. The data carrier signal can be a signal on a data carrier such as a hard disk or flash memory. Furthermore, the data carrier signal can be implemented as a download, etc. A further possible aspect of the present invention is a computer program containing the parameter results determined according to the method of the present invention, and / or a computer-readable data carrier storing the computer program and / or the parameter results determined according to the method. The data carrier signal may be provided for the user to perform a software upgrade, i.e., to parameterize a programmable circuit using the determined parameter results of the data carrier signal. Alternatively, the data carrier signal may be provided to perform parameterization / readout of a programmable circuit, so that the configuration is stored in the memory of the RF system or exported, i.e., stored, on a removable data carrier medium connected to the user interface of the RF system. The removable data carrier medium can then be connected to a second RF system or training bench system to upload the exported configuration, which can then be used to parameterize a programmable circuit in the second RF system.

[0133] Further advantages, features, and details of the present invention will become apparent from the following description, in which embodiments of the present invention will be described in detail with reference to the drawings. In this regard, the features described in the claims and specification may each be essential to the present invention individually or in any combination. [Brief explanation of the drawing]

[0134] [Figure 1a]This document shows one embodiment of the RF measurement system 1 and the data carrier signal of the present invention. [Figure 1b] This document illustrates one embodiment of a plasma power system using the RF measurement system 1 of the present invention. [Figure 2] This document describes one embodiment of the present invention for improving the DSP in a radio frequency system. [Figure 3a] This shows an example of a complete arc event unfolding without corrective action. [Figure 3b] A similar example to Figure 3a is shown, where arc prevention is initiated when the probability of arcing exceeds a threshold. [Figure 3c] This example demonstrates how to identify an actual arc event during pulse mode operation. [Figure 4] This document shows one embodiment of the RF measurement system of the present invention in a training mode configuration. [Figure 5] This shows one embodiment of the RF measurement system of the present invention in application (real-time operation) mode. [Modes for carrying out the invention]

[0135] In the following figures, the same technical features are referred to by the same reference numerals, even in examples of different embodiments.

[0136] Figure 1 shows an embodiment of the RF system 1 of the present invention, which is a radio frequency generator, preferably a radio frequency power supply system 1, or a radio frequency generator or radio frequency plasma system. The system includes at least one sensor 20, in particular a directional coupler, a voltage-current sensor, or a plasma sensor 32 for detecting radio frequency signals of the radio frequency system 1. A directional coupler 20 is a type of sensor commonly used to probe radio frequency signals, for example, in a power supply circuit. However, other sensors can also be implemented, and Figure 1a further shows a radio frequency RF power signal transmission line 60 and two signal branches, where the upper branch may represent forward power and the lower branch may represent reflected power of the RF power signal picked up by the sensor 20. The structure of the system 1 of the present invention described herein can be suitable for either or both of the measured forward power or reflected power. The structure of System 1 of the present invention as described herein can also be adapted to any combination of RF phase having an RF voltage, RF current, reflection coefficient (Γ or "gamma value"), or measured DC self-bias, or RF current, RF voltage, measured forward power or reflected power or both, DC self-bias, or reflection coefficient.

[0137] System 1 further includes at least one analog-to-digital converter, ADC30, for converting at least one detected radio frequency signal into at least one input signal 230. The ADC converts time-continuous and continuous values, i.e., analog signals, into time-discrete and discrete values, i.e., digital signals. The analog signals may be signals picked up by a directional coupler 20, preferably a transmission line 60. The number of samples per second is given by the sampling rate, and the number of different values ​​is given by the number of bits in the ADC30. System 1 may further include an analog anti-aliasing filter 50 and matching circuits 40, in particular an impedance matching circuit 40, between the sensor and the ADC30.

[0138] System 1 preferably includes a programmable circuit 10 for digital signal processing. Circuit 10 is suitable for digital signal processing of at least one input signal 230. The input signal 230 may be a signal measured during the operation of an RF generator, i.e., a real-time signal 230. The circuit is further suitable for digital signal processing of at least one input signal 210 stored in memory 70. Either the input signal 230 or the stored input signal 210 may be a signal related to the radio frequency signal of a power supply system. Thus, these can be directly measured signals or simulated signals. The stored input signal 210 may be further processed by a controller or control circuit 80 before entering circuit 10. Circuit 10 may be further parameterized by at least one parameter result determined according to the method of the present invention shown in Figure 2. Parameterization may occur before the input signal 210 or 230 reaches circuit 10. The parameterized circuit 10 then outputs a processing result 220 according to the input signal (either from memory 210 or real-time mode 230). Subsequently, the processing result 220 can pass through the control circuit 80 and be stored in the memory 70 as the parameter result 240.

[0139] Figure 1a also shows a data carrier signal 240 that carries at least one parameter result 240 determined from the processing result 220. The data carrier signal 240 can be a data package downloadable by an authorized person or any third party. The data carrier signal can then be used to parameterize the programmable circuit 10, in particular the FPGA 10. The processing result 220 used for the parameter result 240 can be determined from the method of the present invention by at least one of the methods described herein. The system may further include an electronic interface 90 for uploading an existing configuration to the programmable circuit 10, the control unit 80, or the memory 70 for parameterization of the circuit 10. The electronic interface 90 can be electronically connected to the control unit 80, the memory 70, or the programmable circuit 10 to perform the upload using direct data transfer. Furthermore, the electronic interface 90 can also be used to perform a download of the configuration, for example, to transfer it to a different RF system for a similar application. The electronic interface 90 may be a user interface.

[0140] Figure 1b shows an embodiment of a plasma power system 5 using the RF measurement system 1 of the present invention. In particular, the plasma power system 5 includes an RF generator 2, a matching network 4, and a plasma process system 3, for example, a plasma chamber 31 equipped with electrodes supplied with RF power.

[0141] The RF generation device 2 includes an RF measurement system 1 which includes a sensor 20, an analog-to-digital converter 30, a programmable circuit 10, a control unit 80, a memory 70, and an electronic interface 90, also called a user interface 90. Furthermore, the RF generation device 2 includes a signal generation unit 16, for example, a direct digital synthesis (DDS) circuit and a power amplification stage 23 powered by a DC power supply 21, and the signal generation unit 16 can be implemented on the programmable circuit 10. Within the scope of the present invention, it should be noted that the signal generation unit 16 may be a separate circuit or chip outside the programmable circuit 10.

[0142] The matching network 4 includes input sensors 41, e.g., phase and amplitude detectors, an impedance matching circuit 40, typically including variable inductors and capacitors, a matching control unit 42, and optionally an output sensor 43, e.g., a voltage-current sensor or a DC bias sensor. The function of the matching network 4 is to match the output impedance (typically 50 ohms) of the RF generator 2 to the variable impedance of the plasma process system 3, so that the RF power coupled to the plasma is maximized and the power reflected from the plasma process system 3 is minimized.

[0143] The plasma process system 3 includes a plasma chamber 31 supplied with RF power from a matching network 4. Within the plasma chamber 31, a gas or gas mixture is excited and partially ionized by the supplied RF power. Electrical coupling of the RF power to the plasma can be achieved by connecting the RF power to electrodes, for example, a metal plate on the opposite side of a similar plate connected to ground. In this way, the plasma is capacitively coupled to the two electrodes. Alternatively, the RF power can be connected to an inductor placed around the plasma chamber 31 so that the plasma is inductively coupled.

[0144] Typically, a semiconductor wafer is loaded into a plasma chamber 31 and exposed to RF plasma for a specific time to etch material from the top layer of the wafer or deposit material on the top layer of the wafer.

[0145] The harmonized network 4 and the plasma process system 3 may include sensors 41, 43, and 32 that measure RF signals or signals characteristic of the plasma process. Such sensor signals can be connected to a dedicated sensor support of the RF generator 2 and converted into a digital representation by an analog-to-digital converter (ADC) 30, for example, a single-channel and / or multi-channel ADC, and used as input signals for digital signal processing performed by the programmable circuit 10.

[0146] Figure 2 shows one embodiment of the method of the present invention for improving digital signal processing in a radio frequency system, in particular a radio frequency power supply system 1. The method includes providing 101 at least one input signal 210 related to radio frequency signals of the radio frequency system 1, and performing 102 a processing procedure 310 in which the at least one input signal 210 is iteratively processed by a programmable circuit 10 using at least one configurable parameter of processing. The configuration of the at least one parameter therein may be changed or particularly modified for each processing of the at least one input signal 210 to obtain a respective processing result 220. The method further includes determining 103 at least one parameter result for optimization of the configuration based on the processing result 220 (e.g., in particular from a modified parameter to an optimized parameter), and providing 104 at least one determined parameter result for improved digital signal processing (in particular for parameterization of the circuit). In the method of the present invention, the programmable circuit 10 can be configured as a programmable integrated circuit 10, preferably an FPGA.

[0147] The method of the present invention may further include parameterization 105 of the same or another programmable circuit 10 using at least one determined parameter result, the parameterization 105 including detection 106 of a radio frequency signal of a radio frequency system 1 and execution of digital signal processing 107 by using at least partial processing of the detected radio frequency signal by the parameterized circuit 10. The method of the present invention may further include monitoring of a radio frequency signal by execution of repeated detection 106 and digital signal processing 102. Regardless of the details of the parameterization 105 of the circuit 10, the processing procedure 310 of the programmable circuit 10 may be configured as a training procedure for performing the parameterization 105 of the programmable circuit 10 in a training mode of processing, using an input signal 210 as an existing (e.g., previously recorded or simulated) training input signal 210, while monitoring is performed in a real-time mode of processing, in which the radio frequency signal is detected and processed as a real-time input 230. At least one input signal 210 processed in processing step 310 can be configured as at least one training input signal 210. The method of the present invention may further include finalizing the parameterization 105 of the programmable circuit 10 by configuring at least one parameter using the previously determined parameter results, and the configuration is then fixed. The method of the present invention may further include switching processing from training mode to real-time mode, in which real-time input signals 230 are processed in real time by the programmable circuit 10 using a fixed configuration, particularly for pattern detection of the real-time input signals 230. Preferably, the real-time input signals are configured as measured signals received from at least one analog-to-digital converter 30 that convert radio frequency signals detected by at least one sensor 20 into at least one real-time input signal. The method of the present invention may further include switching processing from real-time mode to training mode. In training mode, the existing previously fixed configuration can be further optimized by running processing step 310 again with at least one additional training input signal 210 as processed training input signals 210 to newly determine at least one parameter result. In response to a comparison between the newly determined parameter results and the previously determined parameter results, the programmable circuit 10 is reparameterized by configuring at least one parameter using the newly determined parameter results.

[0148] The iterative processing of at least one input signal 210 can be performed by iterative processing steps, in which the same at least one input signal 210 is processed by the same programmable circuit 10 using a modified configuration of at least one parameter, so that each processing result 220 is specific to and / or assigned to the modified configuration. At least one input signal 210 can be configured as one or more predetermined reference input signals 257, for each of which a reference result is provided that represents the desired result of processing the corresponding input signal 210. The determination of at least one parameter result 103 for configuration optimization based on the processing results 220 may further include evaluation of the modified configuration by comparing each of the processing results 220 obtained for at least one reference input signal 257 with the corresponding reference result. The evaluation depends on the agreement between the processing result 220 and the reference result, or the deviation between the processing result 220 and the reference result. The determination of at least one parameter result 103 for optimizing the configuration based on the processing result 220 may further include selecting at least one configuration having the highest evaluation as at least one determined parameter result. Criteria results can be provided by the user, such as by uploading them using the user interface 90. Thus, the user can define at any time what criteria are used for evaluation. Input signals, and by extension training data, can also be defined by the user and uploaded via the user interface 90 for, for example, reasons. The provision of at least one determined parameter result 104 for improved digital signal processing may further include parameterization 105 of the same or a different programmable circuit 10 using at least one selected configuration.Preferably, the at least one predetermined reference input signal 257 and the corresponding reference result include a plurality of predetermined reference input signals 257 and the corresponding reference results for different applications of digital signal processing, so that the determination of the at least one parameter result for optimizing the configuration based on the processing result 220 is performed separately for each application. Providing the at least one determined parameter result 104 may further include storing the selected configuration in memory, and parameterization of the same or another programmable circuit 10 includes selecting the configuration to be used from the stored configurations, depending on the application. At least one input signal 210 can be configured as a plurality of existing input signals 210. The existing signals can be previously recorded or, for example, simulated. The processing procedure 310 may include repeated processing steps, each of which may include processing the same input signal 210 successively and setting the configuration of at least one parameter used in this processing step by the control circuit 80. The method of the present invention may further include, after each processing step, the control circuit 80 evaluating the deviation of the processing result 220 from a predetermined reference result, and after each evaluation, modifying the configuration of one parameter used for the next processing step based on the evaluation. Thus, the control circuit 80 may be electrically connected to the programmable circuit 10 and configured separately from the programmable circuit.

[0149] The configuration of at least one parameter used can be changed depending on the processing step to iteratively optimize the configuration for improved digital signal processing.

[0150] The processing procedure 310 can be configured as a training procedure for iteratively optimizing the configuration for different specific situations in the application of the radio frequency system 1, and as a result, the parameter results determined for different situations are configured as different optimized parameter sets for these different situations.

[0151] At least one input signal 210 may preferably include instantaneous and historical values, in particular, the temporal signal progression from a measurement of a radio frequency signal. The processing result 220 is obtained by processing the signal path and can indicate the probability of detecting a particular pattern in the signal path.

[0152] An electronic interface 90 can be provided to perform the uploading of at least one input signal 210 to the programmable circuit 10. This allows at least one input signal 210 to be configured as multiple input signals 210, which are then stored in memory after uploading.

[0153] At least one input signal 210 can be configured as at least 1,000 or at least 1,000,000 different input signals 210, each of which represents at least one waveform associated with a radio frequency signal. The processing here may be aimed at evaluating the waveforms depending on the parameter configuration. Each processing result 220 may include the results of the evaluation.

[0154] Existing configurations can also be uploaded via the electronic interface 90, in particular to the programmable circuit 10, the control unit 80, or the memory 70. The parameterization of the programmable circuit 10 then uses the existing configuration. The uploaded existing configurations can be stored in the same memory 80 and / or in different memories. The uploaded existing configurations can override any configurations calculated using the method of the present invention, or can be stored in memory in addition to other configurations calculated using the method of the present invention or previously uploaded and stored in memory.

[0155] The evaluation of the modified configuration can be performed by comparing each of the processing results 220 associated with the modified configuration with a predetermined baseline result. The predetermined baseline result may include a predetermined ideal result of the algorithm executed by the programmable circuit 10.

[0156] Figure 3a shows an example of an arc event observed with two different sensor signals, a first sensor signal 402 and a second sensor signal 403. The bottom curve represents RF power 404. The x-axis represents time 405. In this case, the arc only needs to be monitored, no precautions are initiated, and the arc may proceed fully until the discharge is complete. After training a pattern matching algorithm or neural network for this particular case of arc, the algorithm may provide a probability value p(t)406 (shown in the upper curve of Figure 3a representing probability 401) that increases from zero at time tstart to 1 at time tend. The first signs of the arc event are prominent at time tstart, and the event is completed at time tend. After the arc event, the probability value p(t)406 may again rapidly decrease to 0. The first sensor signal 402 may be configured as a current sensor signal, and / or the second sensor signal 403 may be configured as a voltage sensor signal.

[0157] Figure 3b shows a similar situation. Similar to Figure 3a, the top curve represents probability 401, the two curves below it represent the first sensor signal 402 and the second sensor signal 403, and the bottom curve represents RF power 404. The x-axis represents time 405. In contrast to Figure 3a, the probability value p(t)406 is given by the threshold p threshold It can be compared to 407, and the probability value p(t)406 is this threshold p threshold If it exceeds 407, corrective measures may be initiated. As shown in Figure 3a, the probability value p(t)406 of arc event p(t)406 may gradually increase starting from time tstart. When the probability value p(t)406 reaches the threshold p at time tstop... thresholdWhen the value reaches 407, a temporary reduction in RF power 404 may be triggered. This action prevents arc events and allows the plasma processing to return to normal operation. In particular, the spike in the first sensor signal 402 in Figure 3a is significantly reduced, which may indicate that the arc is being suppressed.

[0158] Figure 3c shows an example of a method for efficiently suppressing false positive detection. The x-axis represents time 405. In this example, the RF power output 404 can be modulated with a pulse pattern, which is a regular switching between two RF power levels. By disabling arc detection at time intervals t1a...t2a, t1b...t2b, t3a...t4a, t3b...t4b, t5a...t6a, t5b...t6b, t7a...t8a, t7b...t8b, spikes in the transition between low and high RF power levels are ignored, and only two arc events unrelated to RF power switching can be detected. In this example, the two arc events can be classified into two different classes: arc event class A408 and arc event class B409 (which should be evident from their different shapes).

[0159] In the case of periodic pulse modulation of the RF power signal 404, sudden changes in RF power 404, such as when switching between two levels in pulse mode operation, may be registered as spurious arc events by pattern matching algorithms or neural networks. In particular, these intentional power changes may be flagged as arc events even when evaluating at least two signals in combination.

[0160] To avoid such false positives, anomaly detection via pattern matching algorithms or neural networks can be switched off for a short period ("dead time") whenever the pulse pattern requires a switch in the RF power level. In this way, only actual arc events occurring within a period of constant RF power 404 can be detected. An example of the use of such a dead time is shown in Figure 3c.

[0161] Figure 4 shows an embodiment of the RF measurement system 1 of the present invention configured to operate in training mode, such as a training bench system. The operating mode 253 of the system 1 can be set to training mode via the user interface 90.

[0162] In the following, the term “signal processing circuit” as referred to by reference numeral 17 refers to a portion of the programmable circuit 10 that is specialized for performing signal processing, particularly digital signal processing. The programmable circuit 10 may include further circuits, such as an evaluation circuit 18 used in training mode to determine the deviation between the processing result 220 and a reference result. Further circuits used in real-time mode may be implemented within the programmable circuit 10, but may be deactivated in training mode.

[0163] In training mode, connections to the power amplification stage 23, impedance matching network 4, and plasma process system 3 of system 1 are not required. Instead of real-time input signals provided by the RF generator 2, matching network 4, or sensors within the plasma process system 3, system 1 may process an existing reference input signal 257 that is stored in memory 70 or uploaded via memory 70. Furthermore, reference results 256 can also be stored in memory 70 or provided via the user interface 90. These reference results indicate the expected outcome of signal processing of the reference input signal 257. In addition, parameter configurations 254 that can be used to parameterize the signal processing circuit 17 can also be stored in memory 70 or provided via the user interface 90 as a user parameter set 251. Furthermore, a digital signal processing algorithm, or, for example, a neural network circuit layout 255, which can be implemented in the programmable circuit 10, can also be stored in the memory 70. Typically, uploading the algorithm and circuit layout 255 via the user interface and implementing them in the programmable circuit 10 is restricted to the manufacturer and is not accessible to general users of System 1.

[0164] In training mode, the controller 80 can be configured to read the user parameter set 251, user reference result 252a, and user reference input signal 252b from the user interface 90 and store them in the memory 70. Furthermore, the controller 80 can read the stored parameter set 254 from the memory 70 and configure the programmable circuit 10 to parameterize it accordingly.

[0165] Furthermore, the controller 80 can be configured to read reference input signals 257 from memory 70 or user interface 90 and provide them to the signal processing circuit 17 as input signals 210. Alternatively, the signal processing circuit 17 can directly access memory 70 and directly read the parameter configuration 254 and training data, i.e., the reference input signals 257 and reference results 256, thereby accelerating the reading and processing of training data.

[0166] Furthermore, the evaluation circuit 18 shown in Figure 4 can be implemented within the programmable circuit 10, or it can be implemented as a software algorithm within the controller unit 80. The evaluation results, i.e., the optimized parameter set 254, can be stored in memory 70 by retrieving it from the programmable circuit 10 and writing it to memory 70 via the controller 80, or by direct memory access from the programmable circuit 10.

[0167] Furthermore, it should be noted that, as an alternative to the configuration in Figure 4, at least a portion of the controller unit 80 and / or at least a portion of the memory 70 can be implemented in the programmable circuit 10. This has the advantage of significantly accelerating the reading of the reference input signal 257 and reference result 256 from the memory 70, the subsequent signal processing of the reference input signal 257, and the evaluation of the processing result 220 for the reference result 256.

[0168] Figure 5 shows an embodiment of the RF measurement system 1 of the present invention in application (real-time operation) mode. In application mode, the signal processing circuit 17 is parameterized using an existing parameter set 254 stored in memory 70 during previous training or uploaded to memory 70 via the user interface 90. Depending on a process recipe 258 that may be provided via the user interface 90, a dedicated parameter configuration 254 may be loaded into the signal processing circuit 17 from memory 70 or via the controller 80 from the user interface 90 for each recipe step. The thus parameterized signal processing circuit 17 processes input signals in real time, which are continuously supplied from an RF system, such as an RF generator 2, an impedance matching network 4, or sensors in a plasma process system 3, and digitized by the ADC 30.

[0169] Depending on the algorithm or neural network implemented within the programmable circuit 10, the results of signal processing may be used to control a signal generation unit 16, for example, a DDS core, or further circuitry 19. The signal generation unit 16 is used, for example, to adjust the amplitude, frequency, and phase of the signal that is converted into an RF power signal by a digital-to-analog converter (DAC) 22 and a power amplification stage 23. The RF power signal obtained at the output of the RF generator 2 can then be supplied to the plasma process system 3 via an impedance matching network 4. In this way, for example, the results of filtering the input signal 210 from the directional coupler 20 at the output of the RF generator can be directly used to adapt the RF power signal so that, for example, the RF power supplied to the plasma is stabilized and the reflected power from the plasma is minimized.

[0170] Furthermore, the signal processing circuit 17 may also be configured to monitor the input signal 210 for specific events or undesirable situations, such as high-voltage discharge or secondary plasma in the plasma processing system 3, or degradation of hardware components related to plasma processing, such as particle accumulation in the plasma system, or patterns indicating plasma impedance drift that indicate cleaning or other preventive maintenance of the plasma chamber 31 is required. Such events can be directly acted upon by further circuits 19 within the programmable circuit 10, which can either flag 259 to the user via the user interface 90 or trigger corrective action 250. For example, in the case of an arc event or high-voltage discharge, or if signal processing determines that such an event is likely to occur, the programmable circuit 10 may initiate a reduction or complete shutdown of RF power for a certain period. This can be achieved, for example, by reducing the amplitude of the RF signal generated in the signal generation unit 16, or by disabling the DAC 22, or by disconnecting the connection between the DAC 22 and the power amplification stage 23 for a certain period.

[0171] The circuit layout of the programmable circuit 10, or the algorithm used by the programmable circuit 10, can also be stored in the memory 70, and different circuit layouts 255 can be implemented in the programmable circuit 10 depending on the application.

[0172] As mentioned earlier regarding training mode, even in application mode, at least a portion of the controller 80 and / or a portion of the memory 70 can be implemented within the programmable circuit 10. This has the advantage of significantly speeding up the switching between different configurations of the signal processing circuit 17.

[0173] The above description of embodiments explains the present invention based solely on examples. Of course, within a technically reasonable scope, the features of each embodiment can be freely combined without departing from the spirit of the present invention. [Explanation of Symbols]

[0174] 1 System 2 RF generator 3. Plasma Processing System 4. Unified Network 5. Plasma Power System 10 circuits 16 Signal Generation Unit 17 Signal Processing Circuits 18 Evaluation Circuit 19 Further circuits 20 sensors 21 DC power supply 22. Digital-to-Analog Converter (DAC) 23 Power Amplifier Stage (PA) 30 Analog-to-Digital Converters (ADCs) 31 Plasma Chamber 32 Plasma Sensor 40 Impedance Matching Circuit 41 Input Sensors 42 Match Control Unit 43 Output sensors, V / I sensors 50 Anti-aliasing filters 60 Transmission lines 70 memory 80 Control circuits 90 Electronic Interfaces, User Interfaces 101 Providing at least one input signal, first method step 102 Execution, Method Step 2 103 Decision, Third Method Step 104 Providing at least one determined parameter result, fourth method step 105 Parameterization, Method Step 5 106 Detection, 6th Method Step 107 Execution, Method Step 7 210 (Stored / Existing) Input Signal 220 Processing results 230 (Real-time) Input Signals 240 parameter results 250 Corrective measures 251 User Parameter Set 252a User-defined results 252b User reference input signal 253 Operating Modes 254 parameter set 255 Circuit Layouts 256 Standard results 257 Recorded measurement dataset, reference input signal 258 Process Recipes 259 Event Flags 310 Processing steps, machine learning steps 401 Probability 402 First sensor signal 403 Second sensor signal 404 RF power 405 Time 406 p(t) 407 p threshold 408 Arc event Class A 409 Arc event Class B

Claims

1. A machine learning method for detecting at least one anomaly in a plasma system, particularly an RF power plasma processing system, wherein the method is: The present invention provides an analog signal of the power supply system (1) of the plasma system, and / or an input signal (210) having at least one abnormal feature indicating the abnormality in the plasma system, relating to the power supply system (1) and / or another characteristic of the plasma system. This includes performing a machine learning procedure (310) such that the at least one input signal (210) having the at least one anomaly feature is processed by a programmable circuit (10) to train the detection of the anomaly in the plasma system, The programmable circuit (10) at least partially performs a detection procedure including, in particular, the application of a neural network, preferably pattern recognition or pattern matching using the neural network, or an algorithm, in order to identify the at least one anomaly feature. The aforementioned machine learning procedure (310) is, The programmable circuit (10) performs the processing of the at least one input signal (210) using the detection procedure and at least one configurable parameter of the detection procedure, in particular the weights of the neural network or the parameters of the algorithm, wherein the configuration of the at least one parameter is changed, in particular, modified for each processing of the input signal (210) to obtain the respective processing result (220). The training results of the machine learning based on the processing results (220) include determining at least one parameter result, in particular, including the selection of the modified configuration. The detection procedure for detecting the at least one anomaly is implemented in the programmable circuit (10) by combining functional blocks of the programmable circuit (10) for performing calculations in real time. The implemented detection procedure is repeatedly used in real-time computations during the machine learning procedure (310) for the processing of the at least one input signal (210), and is also used to train the plasma system by repeatedly processing an existing training signal, thereby determining the optimal configuration of the real-time processing mode.

2. The method according to claim 1, characterized in that the programmable circuit (10) is configured as a programmable integrated circuit (10), preferably a digital signal processor (DSP), a decoderable programmable logic device (CPLD), or a field-programmable gate array (FPGA).

3. The method according to claim 1, wherein the machine learning procedure includes iterative processing steps, in which the same at least one input signal (210) is processed by the programmable circuit (10), but using the modified configuration with different parameters, the respective processing results (220) assigned to the configuration used are obtained, and the evaluation of the modified configuration is performed by comparing each of the obtained processing results (220) with a reference result.

4. The method according to claim 3, characterized in that the evaluation of the modified configuration depends on the matching of the processing result (220) with the reference result, and the at least one configuration having the highest evaluation is selected from the modified configuration as the at least one parameter result determined.

5. The method according to claim 3, characterized in that the aforementioned criterion result is a predetermined indicator of the at least one abnormality.

6. The method according to any one of claims 1 to 5, wherein each of the at least one input signal (210) is a radio frequency signal used to supply power to the plasma system, and / or relates to another characteristic of the power supply system (1) and / or the plasma system, and the anomaly is specific to an arc occurring in the plasma system or specific to the probability of the arc occurring.

7. The method according to any one of claims 1 to 5, characterized in that the machine-learned detection of the at least one anomaly is used for arc detection and / or arc prevention and / or arc control.

8. The machine learning procedure (310) provides an iterative determination of the configuration of at least one parameter of the detection procedure for the detection, in particular the configuration of at least one weight of a neural network or at least one parameter of an algorithm, in particular the configuration of the neural network or the algorithm which is at least partially implemented by the programmable circuit (10), the determined configuration which is later used in the detection procedure in the field operation of the power supply system (1). The method according to any one of claims 1 to 5, characterized in that when the abnormality is detected by the detection procedure, warning information is output.

9. Providing the aforementioned at least one input signal (210) is, The method according to any one of claims 1 to 5, characterized in that the analog signal of the power supply system (1) is recorded and converted in the form of a radio frequency signal and / or another characteristic of the power supply system and / or the plasma system, and the digital representation of the analog signal and / or the other characteristic is provided to the programmable circuit (10) as the at least one input signal (210).

10. The method according to any one of claims 1 to 5, characterized in that the machine learning (310) is configured as a training procedure for parameterization of the programmable circuit (10), and that the exact same programmable circuit (10) or another circuit (10) having the same parameterization can be used for the detection of the at least one anomaly during field operation of the power supply system (1), in particular for arc detection and / or arc prevention and / or arc control.

11. In the aforementioned field operation, if at least one abnormality is detected, the following actions are taken: The shutdown of at least part or all of the aforementioned power supply system (1), Without restarting the power supply system (1), the power supply system (1) is stopped. Temporary shutdown of the power supply system (1), and subsequent restart of the power supply system (1), A temporary decrease in the output power of the power supply system (1), Temporary correction of the output frequency of the power supply system (1), Temporary modification of at least one legally permissible element in an impedance matching network (4), The method according to claim 10, characterized in that it is initiated by at least one of the above individually or in combination.

12. The method according to any one of claims 1 to 5, wherein the detection of the abnormality includes, in particular, probabilistic detection for detecting the arc before it occurs in a plasma processing system, preferably for determining the probability of the arc occurring and / or for arc prevention.

13. A digital signal processing method for at least one input signal implemented in a programmable circuit of a plasma system, A digital signal processing method in which the programmable circuit is trained according to the method of any one of claims 1 to 5 so that the input signal is processed according to an optimal configuration for a real-time processing mode.

14. A system for detecting at least one anomaly in a plasma system, particularly an RF power plasma processing system, At least one sensor (230), particularly a directional coupler and / or a voltage-current sensor, for detecting at least one radio frequency signal of the power supply system (1) of the plasma system and / or other characteristics of the power supply system (1) and / or the plasma system, Connected to the sensor (230) to convert the at least one detected radio frequency signal and / or the other characteristic into at least one input signal (210), at least one converter (30), in particular an analog-to-digital converter (30), The programmable circuit (10) for digital signal processing of the at least one input signal (210) includes, The system is characterized in that the programmable circuit (10) is trained for the detection of the at least one anomaly according to the method described in any one of claims 1 to 5.

Images