Prediction system and mass production method

Abstract

The objective of the present invention is to alleviate a work load at the time of start-up in a prediction system equipped with a reservoir. This prediction system includes: a calibration-verified AI device in which a plurality of physical reservoirs for processing time-series signals are connected to each other; and a trained prediction device that reads outputs from the plurality of physical reservoirs included in the calibration-verified AI device while adding reading weights and outputs a predicted value. The calibration-verified AI device has the plurality of physical reservoirs verified to be calibrated so that the physical properties of the physical reservoirs are the same by adding M-series noise to the time-series signals. The trained prediction device has read weights duplicated from another trained prediction device that reads outputs from a plurality of other physical reservoirs having the same physical properties as the plurality of physical reservoirs while adding reading weights and outputs a predicted value, and that has read weights updated by training.

Description

Prediction system and mass production method

[0001] The present disclosure relates to a forecasting system and a mass production method.

[0002] Generally, when AI devices such as nanomolecular reservoirs are mass-produced, there are individual differences between the mass-produced AI devices. Therefore, when multiple prediction systems each equipped with a mass-produced AI device are generated, each prediction system cannot use the learning results of other prediction systems and must be trained from scratch at startup.

[0003] Tanaka H., Akai-Kasaya M., Termeh AY, Hong L., Fu L., Tamukoh H., Tanaka D., Asai T., and Ogawa T., A molecular neuromorphic network device consisting of single-walled carbon nanotubes complexed with polyoxometalate Nature Communications, vol. 9, p. 2693 (2018)Atsushi Uchida, Ryan McAllister, and Rajarshi Roy, Consistency of Nonlinear System Response to Complex Drive Signals, PHYSICAL REVIEW LETTERS, vol.93, 244102 (2004)

[0004] The present disclosure reduces the workload at the time of startup in a prediction system equipped with an AI device.

[0005] A prediction system according to one aspect of the present disclosure has, for example, the following configuration: That is, a prediction system including: a calibration-verified AI device in which a plurality of physical reservoirs that process time-series signals are connected to each other; and a trained prediction device that reads out outputs from the plurality of physical reservoirs possessed by the calibration-verified AI device, adds weights to the readout, and outputs a predicted value, wherein the calibration-verified AI device has the plurality of physical reservoirs that have been verified to be calibrated to have equal physical properties by adding M-sequence noise to the time-series signals, and the trained prediction device is another trained prediction device that reads out outputs from other plurality of physical reservoirs that have equal physical properties to the plurality of physical reservoirs, adds weights to the readout, and outputs a predicted value, and has readout weights copied from the other trained prediction device whose readout weights have been updated by learning.

[0006] This reduces the workload at the start-up of a prediction system equipped with an AI device.

[0007] FIG. 1 is a diagram illustrating an application example of a trained first prediction system. FIG. 2 is a diagram illustrating an example of the system configuration of a trained first prediction system. FIG. 3 is a diagram illustrating an example of the detailed configuration of a calibrated and verified AI device and a trained prediction device in the trained first prediction system. FIG. 4 is a first diagram illustrating a flow from generation to startup of a first prediction system. FIG. 5 is a second diagram illustrating a flow from generation to startup of a first prediction system. FIG. 6 is a diagram illustrating a specific example of a calibration verification process for an AI device. FIG. 7 is a diagram illustrating an example of a processing unit of a calibrated and verified AI device. FIG. 8 is a diagram illustrating a specific example of a learning process for a prediction device in the first prediction system. FIG. 9 is a diagram illustrating a specific example of a read weight sharing process for a prediction device in the first prediction system. FIG. 10A is a diagram illustrating a specific example of a verification process for a trained first prediction system. FIG. 10B is a diagram illustrating a specific example of a re-learning process for a trained prediction device in the trained first prediction system. FIG. 11 is a diagram illustrating an example of the system configuration of a trained second prediction system. FIG. 12 is a diagram illustrating an example of the detailed configuration of an AI device, a trained correction device, and a trained prediction device in the trained second prediction system. FIG. 13 is a first diagram showing the flow from generation to startup of a second prediction system. FIG. 14 is a second diagram showing the flow from generation to startup of a second prediction system. FIG. 15 is a diagram showing a specific example of a learning process for a correction device. FIG. 16 is a diagram showing a specific example of a verification process for a learned correction device. FIG. 17 is a diagram showing a specific example of a learning process for a prediction device in the second prediction system. FIG. 18 is a diagram showing a specific example of a read weight sharing process for a prediction device in the second prediction system. FIG. 19A is a diagram showing a specific example of a verification process for a learned second prediction system. FIG. 19B is a diagram showing a specific example of a re-learning process for a learned prediction device in the learned second prediction system.

[0008] Hereinafter, each embodiment will be described with reference to the accompanying drawings. In this specification and drawings, components having substantially the same functional configurations are designated by the same reference numerals, and redundant description will be omitted.

[0009] [First embodiment] <Application example of trained first prediction system> First, an application example of a trained first prediction system according to a first embodiment will be described, which is generated by training a prediction system (hereinafter referred to as the first prediction system). Fig. 1 is a diagram illustrating an application example of the trained first prediction system.

[0010] In the example of Figure 1, the learned first prediction system is described as being applied to a substrate processing apparatus, but the application of the learned first prediction system is not limited to substrate processing apparatuses, and may also be to apparatuses that perform other manufacturing processes.

[0011] In addition, in Figure 1, in addition to the substrate processing apparatus ((b)) to which the trained first prediction system is applied, a substrate processing apparatus ((a)) to which the trained first prediction system is not applied is shown as a comparative example, and the differences between the two are explained by comparing them as appropriate.

[0012] 1A and 1B, the substrate processing apparatuses 110 and 120 include chambers 111 and 121 for processing substrates, sensors a 112a and 122a, and sensors b 112b and 122b. The substrate processing apparatuses 110 and 120 also include management devices 113 and 123, control devices 115 and 125, and actuators 117 and 127.

[0013] 1A, which is a comparative example, in a substrate processing apparatus 110, a sensor a 112a and a sensor b 112b measure physical quantities while a substrate is being processed in a chamber 111 and output the measured physical quantities as time-series sensor data a and sensor data b. In the case of prediction processing, the time-series sensor data a output from the sensor a 112a is processed in a state prediction and management unit 114 of a management device 113 to predict the process state, and the time-series sensor data b output from the sensor b 112b is input to the state prediction and management unit 114 of the management device 113 and used in the learning processing.

[0014] Furthermore, the time-series sensor data a output from the sensor a 112a is processed in the control unit 116 of the control device 115 to calculate a control amount. At this time, the control unit 116 may correct the control amount based on a predicted value. The control amount calculated by the control unit 116 is output to the actuator 117, and the actuator 117 notifies the chamber 111 of a control command based on the control amount.

[0015] 1A, graph 130 is an example of time-series sensor data a output from sensor a 112a, with the horizontal axis representing time and the vertical axis representing signal strength. Also, in FIG. 1A, graph 140 is time-series sensor data a' when processed by the state prediction and management unit 114 of the management device 113, with the horizontal axis representing time and the vertical axis representing signal strength.

[0016] In general, the processing period T(b) when the management device 113 predicts the process state is longer than the measurement period T(a) when the sensor a 112a measures the time-series sensor data a. Therefore, the state prediction and management unit 114 cannot capture short behaviors that appear in the time-series sensor data a as shown in graph 130 (see graph 140), and as a result, when predicting the process state, it is difficult to obtain sufficient prediction accuracy.

[0017] 1B, in the substrate processing apparatus 120, a sensor a 122a and a sensor b 122b measure physical quantities while a substrate is being processed in a chamber 121 and output the measured values ​​as time-series sensor data a and b. In the case of prediction processing, the time-series sensor data a output from the sensor a 122a is processed in a trained first prediction system 128 to predict a process state. Then, the predicted value predicted by the trained first prediction system 128 is output to the management unit 124 of the management device 123 and the control unit 126 of the control device 125. The time-series sensor data b output from the sensor b 122b is input to the management unit 124 of the management device 123 and is used as process state data (correct answer data) when re-learning (details will be described later) is performed on the trained first prediction system.

[0018] Furthermore, the time-series sensor data a output from the sensor 122a is processed in the control unit 126 of the control device 125 to calculate a control amount. At this time, the control unit 126 may correct the control amount based on a predicted value. The control amount calculated by the control unit 126 is output to the actuator 127, and the actuator 127 notifies the chamber 121 of a control command based on the control amount.

[0019] 1(b), graph 131 is an example of time-series sensor data a output from sensor a 122a, with the horizontal axis representing time and the vertical axis representing signal strength. Also, in FIG. 1(b), graph 141 is time-series sensor data a' when processed in the trained first prediction system 128, with the horizontal axis representing time and the vertical axis representing signal strength.

[0020] The trained first prediction system 128 processes the time-series sensor data a' using reservoir computing to predict the process state. Therefore, the processing cycle T(c) for predicting the process state is significantly shorter than the processing cycle T(b) for predicting the process state by the management device 113 in FIG. 1( a). As a result, the trained first prediction system 128 can capture short-term behaviors that appear in the time-series sensor data a shown in graph 131, thereby improving prediction accuracy.

[0021] <System Configuration of Trained First Prediction System> Next, a description will be given of the system configuration of the trained first prediction system 128. Fig. 2 is a diagram illustrating an example of the system configuration of the trained first prediction system.

[0022] As shown in FIG. 2, the trained first prediction system 128 includes an I / O control device 201, a voltage modulation device 202, a calibration-verified AI device 203, and a trained prediction device 204.

[0023] The I / O control device 201 controls the input and output of digital signals. Specifically, when time-series sensor data a output from the sensor a 122 a is input, the I / O control device 201 notifies the voltage modulation device 202 of sensor data a′.

[0024] The time-series sensor data a input to the I / O control device 201 may be one type of time-series sensor data or a sensor data set consisting of multiple types of time-series sensor data. For simplicity of explanation, the following description will be given assuming that one type of time-series sensor data is input.

[0025] Furthermore, the I / O control device 201 acquires the predicted value (the value resulting from the prediction of the process state) output by the trained prediction device 204 and transmits it to the management apparatus 123 or the control device 125. Furthermore, when process state data (correct answer data) to be used when re-learning (details will be described later) is performed in the trained first prediction system 128 is input from the management apparatus 123, the I / O control device 201 notifies the trained prediction device 204.

[0026] The voltage modulation device 202 is an example of a modulation device, and converts the time-series sensor data a' notified by the I / O control device 201 into voltage data to be input to the calibration-verified AI device 203. Specifically, the voltage modulation device 202 acquires and modulates the sensor data a' at a sampling frequency of 1 MHz or higher, thereby converting it into voltage data.

[0027] The voltage modulation device 202 may convert the time-series sensor data a′ notified by the I / O control device 201 into voltage data corresponding to the predicted value output by the learned prediction device 204, for example.

[0028] The calibration-verified AI device 203 is a device that outputs reservoir features and has been verified to be calibrated to eliminate individual differences between mass-produced AI devices. The AI ​​device here is, for example, a device that has a physical reservoir, receives a voltage signal based on voltage data, converts the output current signal from the physical reservoir into voltage data, and outputs reservoir features. The reservoir features are numerical values ​​that quantitatively represent the characteristics of the time-series sensor data a' output by the physical reservoir based on the values ​​from the past to the present when each value of the time-series sensor data a' is input to the physical reservoir.

[0029] The trained prediction device 204 is a prediction device that adds read weights to the reservoir features output from the calibration-verified AI device 203, reads them, and outputs a predicted value, and the read weights are optimized by a learning process. The trained prediction device 204 outputs the predicted value to the I / O control device 201 and the voltage modulation device 202.

[0030] In addition, when the learned prediction device 204 performs re-learning (details will be described later), it re-updates the read weights so that the predicted value is correlated with the process status data (correct data) input by the I / O control device 201.

[0031] <Details of the process state prediction system> Next, we will explain the detailed configuration of each device (here, the calibration-verified AI device 203 and the trained prediction device 204) in the trained first prediction system 128 shown in Fig. 2. Fig. 3 is a diagram showing an example of the detailed configuration of the calibration-verified AI device and the trained prediction device in the trained first prediction system.

[0032] (1) Detailed Configuration of Calibration-Verified AI Device First, we will explain the detailed configuration of the calibration-verified AI device 203. As shown in Fig. 3, the calibration-verified AI device 203 has a D / A conversion unit 301, a noise addition unit 302, a processing unit 303, and a reservoir feature output unit 304.

[0033] The D / A conversion unit 301 D / A converts the time-series voltage data input by the voltage modulation device 202 to generate a time-series analog voltage signal, and inputs it to the noise addition unit 302 .

[0034] The noise adding section 302 adds an M-sequence noise signal to the time-series analog voltage signal.

[0035] The processing unit 303 has a plurality of physical reservoirs connected to each other, which process a time-series analog voltage signal to which an M-sequence noise signal has been added.

[0036] Generally, a physical reservoir does not refer to a reservoir that exists on a computer, but rather to a reservoir that physically exists (or could exist) in the real world. A reservoir refers to a complex network of interconnected reservoir nodes, which retains reservoir features for voltage signal inputs from the past to the present. A reservoir node refers to a basic component of a reservoir, which receives one or more voltage signal inputs, performs linear or nonlinear transformations on the values ​​(or the values) and outputs a current signal. A reservoir node is not a static element, but a dynamic element (the state of the node at the next time is determined based on the current state of the node and the states of other connected nodes). The current signal output by a reservoir node follows the current voltage signal input while forgetting past voltage signal inputs.

[0037] In the first embodiment, the multiple physical reservoirs included in the processing unit 303 each serve as multiple reservoir nodes included in the reservoir. That is, in the first embodiment, the processing unit 303 is the reservoir, and the multiple physical reservoirs are multiple reservoir nodes. In addition, in the first embodiment, the physical reservoir includes at least one of a nanomolecule reservoir, a spin reservoir, an optical reservoir, a CMOS reservoir, and a memristor reservoir. As an example, multiple nanomolecule reservoirs are used as the multiple physical reservoirs. A nanomolecule reservoir is a physical reservoir composed of POM molecules and carbon nanotubes. A POM molecule is a polyacid molecule that accumulates electric charge and has the property of releasing the accumulated electric charge when the accumulated amount exceeds a certain threshold.

[0038] Furthermore, in the calibration-verified AI device 203, it has been verified that the multiple physical reservoirs possessed by the processing unit 303 are calibrated to have equal physical properties by adding an M-sequence noise signal. It has also been verified that the physical properties of the multiple physical reservoirs possessed by the processing unit 303 are calibrated to have equal physical properties as well as those of other mass-produced physical reservoirs by adding an M-sequence noise signal. Note that the other mass-produced physical reservoirs refer to physical reservoirs other than the physical reservoir mounted in the processing unit 303 of the calibration-verified AI device 203.

[0039] In addition, in the calibration-verified AI device 203, the multiple physical reservoirs possessed by the processing unit 303 are configured such that the output of the connected physical reservoir is averaged over a predetermined period of time, a weight is added between the connected physical reservoir and the output is then input to the connected physical reservoir.

[0040] The reservoir feature output unit 304 reads out a current signal from each of the plurality of physical reservoirs, converts it into voltage data, and outputs it to the trained prediction device 204 as a reservoir feature.

[0041] In this way, by configuring the device to use multiple physical reservoirs, the calibration-verified AI device 203 can capture short behaviors that appear in the sensor data a and output reservoir features.

[0042] (2) Detailed Configuration of Trained Prediction Device Next, we will explain the detailed configuration of the trained prediction device 204. As shown in Fig. 3, the trained prediction device 204 has a FORCE learning unit 311 that uses the recursive least squares method.

[0043] In the first embodiment, the FORCE learning unit 311 using the recursive least squares method is set with read weights calculated by performing FORCE learning processing using the recursive least squares method. Alternatively, in the first embodiment, the FORCE learning unit 311 using the recursive least squares method is set with read weights that are duplicated and calculated by another FORCE learning unit using the recursive least squares method by performing FORCE learning processing using the recursive least squares method. The FORCE learning unit 311 using the recursive least squares method is an example of a prediction unit, and reads out reservoir features that are output from the reservoir feature output unit 304, adds read weights to them, and outputs a predicted value.

[0044] The read weights set in the FORCE learning unit 311 using the recursive least squares method may be updated again by performing re-learning.

[0045] <Flow from generation to start-up of first prediction system> Next, the flow from generation to start-up of the first prediction system will be described with reference to Fig. 4 and Fig. 5. Fig. 4 and Fig. 5 are first and second diagrams showing the flow from generation to start-up of the first prediction system.

[0046] 1 to 3, the trained first prediction system 128 is applied to the substrate processing apparatus 120, etc., but the substrate processing apparatus to which the trained first prediction system is applied is generally mass-produced and a start-up process is performed at the delivery destination. For this reason, it is desirable that the trained first prediction system is also mass-produced by a method suitable for mass production and start-up.

[0047] On the other hand, mass-producing the trained first prediction system requires mass-producing physical reservoirs, but generally, when physical reservoirs are mass-produced, individual differences arise between the mass-produced physical reservoirs.

[0048] Therefore, when a first prediction system is formed using a mass-produced physical reservoir, individual differences arise in the first prediction system, and the substrate processing apparatus to which the first prediction system is applied cannot use the learning results of other mass-produced first prediction systems. As a result, the substrate processing apparatus to which the first prediction system is applied must perform learning from scratch at startup. In other words, the workload at startup of the substrate processing apparatus increases.

[0049] On the other hand, if the individual differences among mass-produced physical reservoirs can be eliminated, thereby eliminating the individual differences among the first prediction systems, there is no need to perform learning from scratch at the time of startup. For example, if learning is performed using one specific first prediction system and readout weights are calculated, trained first prediction systems can be generated for other first prediction systems simply by duplicating the calculated readout weights.

[0050] As a result, when starting up the substrate processing apparatus, instead of performing learning from scratch, it is sufficient to re-update the read weights by performing re-learning on the trained first prediction system according to the machine differences of the substrate processing apparatus, thereby reducing the workload at the time of start-up. Hereinafter, a flow from generation to start-up of the first prediction system, including the mass production method according to the first embodiment, will be described with reference to FIGS.

[0051] 4, in the first embodiment, first, a physical reservoir generation process 401 is executed. The physical reservoir generation process 401 generates a plurality of physical reservoirs. As described above, the plurality of physical reservoirs mass-produced in the physical reservoir generation process 401 have different physical properties and individual differences exist.

[0052] Next, the AI ​​device generation process 402 is executed. As described above, the AI ​​device includes the D / A conversion unit 301, the noise addition unit 302, the processing unit 303, and the reservoir feature output unit 304, and the processing unit 303 includes multiple physical reservoirs. Since there are individual differences among the multiple physical reservoirs included in the processing unit 303, there are also individual differences among the multiple AI devices mass-produced in the AI ​​device generation process 402. However, it is assumed that the multiple AI devices mass-produced in the AI ​​device generation process 402 are generated so that the number of physical reservoirs included in each processing unit 303 is equal and the connection relationships are equal.

[0053] Next, an AI device calibration verification process 403 is executed. In the AI ​​device calibration verification process 403, a calibration verification signal (including an M-sequence noise signal) is input to the processing unit of each of the multiple AI devices. As a result, in the AI ​​device calibration verification process 403, it is verified that calibration is performed so that the physical characteristics of the multiple physical reservoirs included in the processing units of the AI ​​devices are equal. In addition, in the AI ​​device calibration verification process 403, it is verified that calibration is performed so that the physical characteristics of the multiple physical reservoirs included in the processing unit 303 of one AI device and the physical characteristics of the multiple physical reservoirs included in the processing unit 303 of the other AI device are equal.

[0054] If it is verified in the AI ​​device calibration verification process 403 that the physical characteristics of the multiple physical reservoirs are calibrated to be equal, a first prediction system generation process 404 is executed. The first prediction system generated by executing the first prediction system generation process 404 includes an I / O control device 201, a voltage modulation device 202, a calibration-verified AI device 203, and a prediction device.

[0055] The prediction device shown in the first prediction system generation process 404 is a device equipped with a FORCE learning unit using the recursive least squares method before learning of the readout weights is performed.

[0056] In this way, in the first prediction system generation process 404, first prediction systems that have been verified to be calibrated so as to have no individual differences are mass-produced.

[0057] 5, a prediction device learning process 501 is executed. In the prediction device learning process 501, a learning signal is input to a specific one of the mass-produced first prediction systems, thereby learning the specific one first prediction system. As a result, the read weight of the specific one first prediction system is optimized, and a learned first prediction system is generated.

[0058] Next, a readout weight sharing process 502 is executed. The readout weight sharing process 502 copies the readout weights of the trained first prediction system to a plurality of first prediction systems other than the specific one first prediction system. As a result, trained first prediction systems are mass-produced for the plurality of first prediction systems other than the specific one first prediction system.

[0059] Next, a trained first prediction system verification process 503 is executed. The trained first prediction system verification process 503 inputs the same system verification signal to a plurality of mass-produced trained first prediction systems to verify that the systems output the same predicted values ​​(or within a predetermined error range).

[0060] Next, a substrate processing apparatus mounting process 504 is executed. In the substrate processing apparatus mounting process 504, a plurality of trained first prediction systems that have been determined to output the same predicted value (or within a predetermined error range) as a result of verification by the trained first prediction system verification process 503 are mounted on mass-produced substrate processing apparatuses. The substrate processing apparatuses on which the trained first prediction systems are mounted are delivered to their respective destinations, where installation work is carried out.

[0061] After the installation work is completed, the substrate processing apparatus undergoes a start-up process. At this point, the prediction system installed in the substrate processing apparatus has already been trained, so the start-up process involves executing a trained prediction device re-learning process 505 to re-learn the prediction system according to differences between the substrate processing apparatuses.

[0062] Specifically, the learned prediction device re-learning process 505 performs re-learning by inputting a re-learning signal to the learned first prediction system installed in the substrate processing apparatus, and re-updates the read weights.

[0063] In this way, according to the flow from generation to start-up of the first prediction system, including the mass production method according to the first embodiment, it is possible to reduce the workload at the time of start-up.

[0064] <Details of Each Process> Next, details of each process from generation to startup of the first prediction system will be described. Here, details of the AI ​​device calibration verification process 403, the first prediction system generation process 404, the prediction device learning process 501, the read weight sharing process 502, the trained first prediction system verification process 503, and the trained prediction device relearning process 505 will be described.

[0065] (1) Details of AI Device Calibration Verification Processing First, details of the AI ​​device calibration verification processing 403 will be described. Fig. 6 is a diagram showing a specific example of the AI ​​device calibration verification processing. As shown in Fig. 6, in the AI ​​device calibration verification processing 403, calibration verification signals are input to multiple physical reservoirs included in the processing unit 303.

[0066] Specifically, a signal in which an M-sequence noise signal is superimposed on a time-series analog voltage signal is input to multiple physical reservoirs as a calibration verification signal. This is to utilize the phenomenon that occurs when the same random noise is input to multiple physical reservoirs, resulting in the multiple physical reservoirs having the same outputs after a transient state (a synchronization phenomenon, also called consistency). In Fig. 6, reference numeral 601 is an example of a time-series analog voltage signal, and reference numeral 602 is an example of an M-sequence noise signal. Furthermore, reference numeral 603 is an example of a calibration verification signal in which an M-sequence noise signal is superimposed on a time-series analog voltage signal.

[0067] 6, a calibration verification signal (reference numeral 603) is input to each of the plurality of physical reservoirs included in the processing unit 303, and a current signal is output from each of the plurality of physical reservoirs. The current signals output from each of the plurality of physical reservoirs are subjected to an averaging process at predetermined time intervals. As a result, the M-sequence noise signal components are removed from the current signals output from each of the plurality of physical reservoirs, and outputs indicated by reference numerals 611 to 614 are obtained.

[0068] As described above, if the input of the calibration verification signal continues, a synchronization phenomenon occurs after the transient state, and the outputs indicated by reference numerals 611 to 614 become equal to each other. This verifies that the multiple physical reservoirs included in the processing unit 303 are calibrated to have equal physical properties, and the AI ​​device calibration verification process 403 is terminated.

[0069] In the example of Figure 6, one processing unit 303 is shown, but by performing similar processing in parallel on multiple processing units other than processing unit 303, it can be verified that the physical properties of all physical reservoirs are calibrated to be equal.

[0070] (2) Details of the First Prediction System Generation Process Next, details of the first prediction system generation process 404 will be described. Fig. 7 is a diagram showing a specific example of the first prediction system generation process. The first prediction system generation process 404 generates a processing unit 303 that has: a plurality of physical reservoirs that have been verified in the AI ​​device calibration verification process 403 to be calibrated to have equal physical properties, and each physical reservoir is configured to average the output of the source physical reservoir over a predetermined time period, add a predetermined weight to the destination physical reservoir, and then input the averaged output to the destination physical reservoir.

[0071] Note that the example in Figure 7 shows only one processing unit 303 generated in the first prediction system generation process 404, but in the first prediction system generation process 404, it is assumed that the processing unit 303 shown in Figure 7 and multiple processing units having physical reservoirs that are verified to be equal in terms of: the number of physical reservoirs; and the connection relationships between the physical reservoirs, and that are calibrated to have equal physical properties.

[0072] Then, an AI device including the processing unit 303 shown in Figure 7 is installed as a calibration-verified AI device together with the I / O control device 201, the voltage modulation device 202, and the prediction device, thereby generating a first prediction system.

[0073] (3) Details of the Prediction Device Learning Process Next, details of the prediction device learning process 501 will be described. Fig. 8 is a diagram showing a specific example of the prediction device learning process in the first prediction system. As shown in Fig. 8, a specific first prediction system that is learned in the prediction device learning process 501 includes: a calibration-verified AI device 203; and a prediction device 204'.

[0074] Of these, the processing unit 303 included in the calibration-verified AI device 203 is a processing unit mass-produced in the first prediction system generation process 404. In the prediction device learning process 501, first, a learning signal is input to the D / A conversion unit 301, and an M-sequence noise signal is added by the noise addition unit 302 to the time-series analog voltage signal output from the D / A conversion unit 301, and the signal is input to the processing unit 303. As a result, the current signal output from the processing unit 303 is read out by the reservoir feature output unit 304, converted into voltage data, and then output from the calibration-verified AI device 203 as a reservoir feature.

[0075] The prediction device 204′ has a FORCE learning unit 800 using the recursive least squares method before learning is performed. An initial value of a read weight is set in the FORCE learning unit 800 using the recursive least squares method before learning is performed. The FORCE learning unit 800 using the recursive least squares method reads out the reservoir feature amount by adding a read weight, and sequentially updates the read weight by comparing it with previously input process state data (correct data).

[0076] As a result, according to the prediction device learning process 501, the read weights are optimized in the FORCE learning unit 311 using the recursive least squares method (that is, the learned prediction device 204 is generated).

[0077] (4) Details of Readout Weight Sharing Process Next, details of the readout weight sharing process 502 will be described. Fig. 9 is a diagram showing a specific example of the readout weight sharing process of the prediction device in the first prediction system. For convenience of explanation, the example in Fig. 9 shows only the prediction device among the devices included in a mass-produced first prediction system.

[0078] Among these, the trained prediction device 204 indicates a trained prediction device that is generated by performing learning on one specific prediction device in the prediction device learning process 501. On the other hand, the multiple prediction devices 204′ indicate prediction devices other than the one specific prediction device that are not trained in the prediction device learning process 501.

[0079] The read weight sharing process 502 copies the read weights of the trained prediction device 204 to multiple prediction devices 204'. As a result, the read weights of the trained prediction device 204 are set to multiple prediction devices 204', and multiple trained prediction devices are generated. In other words, the read weight sharing process 502 allows multiple trained first prediction systems to be mass-produced.

[0080] (5) Details of the Trained First Prediction System Verification Process Next, details of the trained first prediction system verification process 503 will be described. Fig. 10A is a diagram showing a specific example of the verification process for a trained first prediction system. For ease of explanation, the example in Fig. 10A shows only the calibrated and verified AI device 203 and the trained prediction device 204 among the devices included in the mass-produced trained first prediction system.

[0081] In the trained first prediction system verification process 503, first, a system verification signal is input to the D / A conversion unit 301, and an M-sequence noise signal is added by the noise addition unit 302 to the time-series analog voltage signal output from the D / A conversion unit 301. The time-series analog voltage signal with the M-sequence noise signal added is input to the processing unit 303. As a result, the current signal output from the processing unit 303 is read out by the reservoir feature output unit 304, converted into voltage data, and then output from the calibration-verified AI device 203 as a reservoir feature.

[0082] The trained prediction device 204 reads the reservoir feature, adds weights to it, and reads it out to output a prediction value.

[0083] In the example of Figure 10A, for the sake of convenience of explanation, only one of the mass-produced trained first prediction systems is shown, but the system verification signal is input to the calibrated and verified AI device 203 of all mass-produced trained first prediction systems.

[0084] As a result, predicted values ​​are output from the trained prediction devices 204 of all mass-produced trained first prediction systems. As a result, the trained first prediction system verification process 503 can verify whether the respective predicted values ​​are equal (or within a predetermined error range).

[0085] As a result of the verification, if the predicted value is determined to be equal (or within a predetermined error range), the trained first prediction system is permitted to be installed in the substrate processing apparatus 120.

[0086] (6) Details of the Trained Prediction Device Re-learning Process Next, details of the trained prediction device re-learning process 505 will be described. Fig. 10B is a diagram showing a specific example of the re-learning process of the trained prediction device in the trained first prediction system. For ease of explanation, the example of Fig. 10B shows only the calibration-verified AI device 203 and the trained prediction device 204 among the devices included in the trained first prediction system installed in the substrate processing apparatus 120.

[0087] In the learned prediction device relearning process 505, first, the time-series sensor data a output from the sensor a 122a is notified from the I / O control device 201 to the voltage modulation device 202, and is converted into voltage data in the voltage modulation device 202. Next, in the learned prediction device relearning process 505, the voltage data converted in the voltage modulation device 202 is input to the D / A conversion unit 301 as a relearning signal.

[0088] As a result, the time-series analog voltage signal output from the D / A conversion unit 301 has an M-sequence noise signal added by the noise addition unit 302, and is then input to the processing unit 303. The current signal output from the processing unit 303 is read out by the reservoir feature output unit 304, converted into voltage data, and then output as a reservoir feature. The output reservoir feature is then input to the trained prediction device 204, which has a FORCE learning unit 311 using a recursive least squares method with optimized readout weights.

[0089] Next, in the trained prediction device re-learning process 505, the FORCE learning unit 311 using the recursive least squares method reads out the reservoir feature amount after adding readout weights, and compares the readout weights with the process state data (correct answer data) input in advance. As a result, the trained prediction device re-learning process 505 can sequentially re-update the readout weights.

[0090] As a result, according to the learned prediction device re-learning process 505, the read weights are re-learned in the FORCE learning unit 311 using the recursive least squares method according to the machine differences of the substrate processing apparatus 120 (i.e., a re-learned prediction device is generated).

[0091] <Summary> As is clear from the above description, the trained first prediction system 128 includes: a processing unit 303 to which multiple physical reservoirs that process time-series analog voltage signals are connected; a trained prediction device 204 that adds readout weights to reservoir features that are outputs from the multiple physical reservoirs of the processing unit 303, reads them, and outputs predicted values; the processing unit 303 includes multiple physical reservoirs that have been verified to be calibrated to have equal physical characteristics by adding an M-sequence noise signal to the time-series analog voltage signals; and the trained prediction device 204 has readout weights copied from another trained prediction device. The other trained prediction device is a device that includes multiple other physical reservoirs that have been verified to be calibrated to have equal physical characteristics with the multiple physical reservoirs by adding a common M-sequence noise signal. The other trained prediction device is a device that adds readout weights updated by a learning process to outputs (reservoir features) from the multiple other physical reservoirs, reads them, and outputs predicted values.

[0092] As a result, the trained first prediction system 128 does not need to start learning from scratch at startup, thereby reducing the workload at startup.

[0093] The mass production method according to the first embodiment also includes: executing an AI device calibration verification process 403 in which a time-series analog voltage signal to which a common M-sequence noise signal has been added is input to each of the multiple physical reservoirs to verify that the multiple physical reservoirs are calibrated to have equal physical characteristics; when multiple processing units 303 are generated in which the verified multiple physical reservoirs are connected to each other, learning is performed on a prediction device that adds read weights to and reads out reservoir features that are outputs from the multiple physical reservoirs of one of the processing units; that is, executing a prediction device learning process 501 in which one trained prediction device is generated, and updating the read weights of the prediction device to generate one trained prediction device; and executing a read weight sharing process 502 in which the updated read weights are copied and set in each prediction device that adds read weights to and reads out reservoir features that are outputs from the multiple physical reservoirs of processing units other than the one processing unit.

[0094] As a result, according to the mass production method of the first embodiment, there is no need to perform learning from scratch at the time of start-up, and the workload at the time of start-up can be reduced.

[0095] In the first embodiment, a common M-sequence noise signal is added to mass-produced physical reservoirs to equalize the physical characteristics of the physical reservoirs, thereby suppressing individual differences in AI devices. When the same voltage data is input, the trained prediction device 204 can read out the same reservoir feature values.

[0096] In contrast, in the second embodiment, a correction device is placed after the AI ​​device. As a result, in the second embodiment, the correction device suppresses the influence of individual differences in the AI ​​device, and when the same voltage data is input to the AI ​​device, the trained prediction device 204 reads out corrected reservoir features as the same reservoir features. The second embodiment will be described below, focusing on the differences from the first embodiment.

[0097] <System Configuration of Trained Second Prediction System> First, a system configuration of a trained second prediction system according to a second embodiment, which is generated by training a prediction system (hereinafter referred to as the second prediction system), will be described. Note that the second embodiment will also be described assuming that the trained second prediction system is applied to the same application destination as in the first embodiment.

[0098] Fig. 11 is a diagram showing an example of the system configuration of a trained second prediction system. As shown in Fig. 11, the trained second prediction system 1100 has a system configuration similar to that of the trained first prediction system 128 shown in Fig. 2, and differs from the trained first prediction system 128 in that: an AI device 1101 is provided instead of the calibration-verified AI device 203; a trained correction device 1102 is provided downstream of the AI ​​device 1101; and the trained prediction device 204 reads corrected features instead of reservoir features (corrected reservoir features when the AI ​​device 1101 has a physical reservoir; the same applies hereinafter in the description of Fig. 11).

[0099] The AI ​​device 1101 is a device that outputs a feature (a reservoir feature if the AI ​​device 1101 has a physical reservoir; the same applies hereinafter in the description of FIG. 11 ), and is one of the mass-produced AI devices. However, individual differences that arise between mass-produced AI devices are not calibrated. In addition, in the case of the second embodiment, the AI ​​device is, for example, a device that has one physical reservoir, converts a current signal output from the one physical reservoir into voltage data when a voltage signal based on voltage data is input, and outputs the reservoir feature.

[0100] The trained correction device 1102 is a device that corrects the feature values ​​output by the AI ​​device 1101 and outputs the corrected feature values. As described above, mass-produced AI devices vary from one another. For this reason, even if the same voltage data as that input to the AI ​​device 1101 is input to another mass-produced AI device, the feature values ​​output from the AI ​​device 1101 will be different from the feature values ​​output by the other AI device.

[0101] The trained correction device 1102 is trained so that the corrected features are equal to the features output from other AI devices.

[0102] The trained prediction device 204 is a prediction device that reads the corrected reservoir features output from the trained correction device 1102, adds read weights, and outputs predicted values, and is a prediction device in which the read weights have been optimized by a learning process.

[0103] <Details of the process state prediction system> Next, the detailed configuration of each device (here, the AI ​​device 1101, the learned correction device 1102, and the learned prediction device 204) in the trained second prediction system 1100 shown in Fig. 11 will be described. Fig. 12 is a diagram showing an example of the detailed configuration of the AI ​​device, the trained correction device, and the trained prediction device in the trained second prediction system. Here, for convenience of explanation, it is assumed that the AI ​​device 1101 has a physical reservoir.

[0104] (1) Detailed Configuration of AI Device First, we will explain the detailed configuration of the AI ​​device 1101. As shown in Fig. 12 , the AI ​​device 1101 has a D / A conversion unit 1201, a physical reservoir 1202, and a reservoir feature output unit 1203.

[0105] The D / A conversion unit 1201 D / A converts the time-series voltage data input by the voltage modulation device 202 to generate a time-series analog voltage signal, and inputs it to the physical reservoir 1202 .

[0106] The physical reservoir 1202 has a plurality of interconnected reservoir nodes that process time-series analog voltage signals.

[0107] As in the first embodiment, in the second embodiment, the physical reservoir includes at least one of a nanomolecule reservoir, a spin reservoir, an optical reservoir, a CMOS reservoir, and a memristor reservoir. Also, as in the first embodiment, in the second embodiment, a nanomolecule reservoir is used as the physical reservoir 1202.

[0108] In addition, in the AI ​​device 1101, the physical reservoir 1202 is arranged without being calibrated.

[0109] The reservoir feature output unit 1203 reads out current signals from each of the plurality of reservoir nodes of the physical reservoir 1202, converts them into voltage signals, and then converts them into digital voltage data, which is then output as reservoir features.

[0110] (2) Detailed Configuration of Trained Correction Device Next, a description will be given of the detailed configuration of the trained correction device 1102. As shown in FIG.

[0111] The correction reservoir 1211 is a replicable digital reservoir realized by a CPU (Central Processing Unit), FPGA (Field Programmable Gate Array), etc. Here, the number of reservoir nodes n possessed by the correction reservoir 1211 of the learned correction device 1102, and the number of reservoir nodes (an example of a variable) N possessed by the physical reservoir 1202 of the AI ​​device 1101 have a relationship of N>n.

[0112] The correction readout unit 1212 reads out reservoir features from each of the plurality of reservoir nodes of the correction reservoir 1211, and adds correction readout weights to the reservoir features, thereby outputting corrected reservoir features.

[0113] As described above, the correction readout unit 1212 is trained so that, when the same voltage data as that input to another mass-produced AI device is input to the AI ​​device 1101, the corrected reservoir feature output by the correction readout unit 1212 is equal to the reservoir feature output by the other AI device (here, one specific AI device). In other words, when outputting the corrected reservoir feature, the correction readout unit 1212 adds an appropriately updated correction readout weight.

[0114] (3) Detailed Configuration of Trained Prediction Device Next, we will explain the detailed configuration of the trained prediction device 204. As shown in Fig. 12, the trained prediction device 204 has a FORCE learning unit 311 that uses the recursive least squares method.

[0115] As in the first embodiment, in the second embodiment, the FORCE learning unit 311 using the recursive least squares method is set with read weights calculated by performing FORCE learning processing using the recursive least squares method. Alternatively, the FORCE learning unit 311 using the recursive least squares method is set with read weights that are duplicated and calculated by another FORCE learning unit using the recursive least squares method by performing FORCE learning processing using the recursive least squares method. The FORCE learning unit 311 using the recursive least squares method adds read weights to the corrected reservoir features that are output from the learned correction device 1102, reads them, and outputs a predicted value.

[0116] The read weights set in the FORCE learning unit 311 using the recursive least squares method may be updated again by performing re-learning.

[0117] <Flow from generation to start-up of second prediction system> Next, the flow from generation to start-up of the second prediction system will be described with reference to Fig. 13 and Fig. 14. Fig. 13 and Fig. 14 are first and second diagrams showing the flow from generation to start-up of the second prediction system.

[0118] As in the first embodiment, the trained second prediction system 1100 is applied to the substrate processing apparatus 120, etc., but the substrate processing apparatus to which the trained second prediction system is applied is generally mass-produced and undergoes start-up processing at the delivery destination. For this reason, it is desirable that the trained second prediction system be mass-produced using a method suitable for mass production and start-up.

[0119] On the other hand, mass-producing a trained second prediction system requires mass-producing physical reservoirs, but generally, when physical reservoirs are mass-produced, individual differences arise between the mass-produced physical reservoirs.

[0120] Therefore, when a second prediction system is formed using an AI device equipped with a mass-produced physical reservoir, individual differences arise in the second prediction system, and the substrate processing apparatus to which the second prediction system is applied cannot use the learning results of other second prediction systems. As a result, the substrate processing apparatus to which the second prediction system is applied must perform learning from scratch at startup. In other words, the workload at startup of the substrate processing apparatus increases.

[0121] In contrast, if the influence of individual differences in the AI ​​devices equipped with mass-produced physical reservoirs can be suppressed and the individual differences in the second prediction systems can be eliminated, there will be no need to perform learning from scratch at startup. For example, if learning is performed using one specific second prediction system and readout weights are calculated, trained second prediction systems can be generated for other second prediction systems simply by duplicating the calculated readout weights.

[0122] As a result, when starting up the substrate processing apparatus, instead of performing learning from scratch, it is sufficient to re-update the read weights by performing re-learning on the trained second prediction system according to the machine differences of the substrate processing apparatus, thereby reducing the workload at the time of start-up. Hereinafter, a flow from generation to start-up of the second prediction system, including the mass production method according to the second embodiment, will be described with reference to FIGS.

[0123] 13, in the second embodiment, first, a physical reservoir generation process 1301 is executed. Here, physical reservoirs (physical reservoirs with N reservoir nodes) to be mounted on the AI ​​device are generated. As described above, the physical reservoirs mass-produced in the physical reservoir generation process 1301 have different physical properties and individual differences between the physical reservoirs mounted on the AI ​​device.

[0124] Next, an AI device generation process 1302 is executed. As described above, the AI ​​device includes a D / A conversion unit 1201, a physical reservoir 1202, and a reservoir feature output unit 1203, and the physical reservoir 1202 includes N reservoir nodes. Since there are individual differences in the physical reservoir 1202, there are also individual differences in the multiple AI devices mass-produced in the AI ​​device generation process 1302.

[0125] Subsequently, a correction device generation process 1303 is executed. The correction device generated by executing the correction device generation process 1303 includes a correction reservoir 1211 and a correction readout unit 1212, and the correction reservoir 1211 includes n reservoir nodes. The correction reservoir 1211 is a replicable digital reservoir such as a CPU or FPGA.

[0126] Next, a correction device learning process 1304 is executed. In the correction device learning process 1304, when a correction signal is input to one specific AI device among a plurality of mass-produced AI devices, the reservoir feature output from the one specific AI device is acquired as a teacher signal.

[0127] In addition, the correction device learning process 1304 inputs the same correction signal as that input to the specific AI device to AI devices other than the specific AI device among the multiple mass-produced AI devices.

[0128] In addition, the correction device learning process 1304 inputs reservoir feature amounts output from AI devices other than the specific one AI device to the corresponding correction device, thereby outputting corrected reservoir feature amounts from the corresponding correction device.

[0129] Furthermore, the correction device learning process 1304 updates the correction readout weights of the corresponding correction devices by learning the corresponding correction devices so that the corrected reservoir features output from the corresponding correction devices are equal to the acquired teacher signals. As a result, the correction device learning process 1304 can generate a learned correction device having appropriate correction readout weights.

[0130] Next, the trained correction device verification process 1305 is executed. The trained correction device verification process 1305 inputs the same verification signal to multiple mass-produced AI devices. This makes it possible to verify that each of the multiple trained correction devices outputs a corrected reservoir feature that is the same as (or within a predetermined error range) the output (reservoir feature) of a specific AI device.

[0131] Next, as shown in Figure 14, a second prediction system generation process 1401 is executed. The second prediction system generated by executing the second prediction system generation process 1401 includes an I / O control device 201, a voltage modulation device 202, an AI device 1101, a learned correction device 1102, and a prediction device. However, for a second prediction system equipped with a specific AI device selected in the correction device learning process 1304, the learned correction device 1102 is not included. Note that the prediction device is a device equipped with a FORCE learning unit using the recursive least squares method before learning of the readout weights is performed.

[0132] As a result, the second prediction system generation process 1401 makes it possible to mass-produce second prediction systems in which individual differences between AI devices are reduced.

[0133] Next, a prediction device learning process 1402 is executed. In the prediction device learning process 1402, a learning signal is input to a specific one of the mass-produced second prediction systems, thereby learning the specific one second prediction system. As a result, the read weight of the specific one second prediction system is optimized, and a learned second prediction system is generated.

[0134] Next, a read weight sharing process 1403 is executed. The read weight sharing process 1403 copies the read weights of the trained second prediction system to a plurality of second prediction systems other than the specific one second prediction system. As a result, the read weight sharing process 1403 makes it possible to mass-produce trained second prediction systems for a plurality of second prediction systems other than the specific one second prediction system.

[0135] Next, a trained second prediction system verification process 1404 is executed. In the trained second prediction system verification process 1404, the same verification signal is input to multiple mass-produced trained second prediction systems. This makes it possible to verify that the trained second prediction system verification process 1404 outputs the same predicted value (or within a predetermined error range).

[0136] Next, a substrate processing apparatus mounting process 1405 is executed. In the substrate processing apparatus mounting process 1405, a plurality of trained second prediction systems that have been verified to output identical predicted values ​​(or within a predetermined error range) as a result of the verification by the trained second prediction system verification process 1404 are mounted on mass-produced substrate processing apparatuses. The substrate processing apparatuses equipped with the trained second prediction systems are then delivered to their respective destinations and installed.

[0137] After the installation work is completed, the substrate processing apparatus undergoes a start-up process. At this point, the prediction system installed in the substrate processing apparatus has already been trained, so the start-up process involves executing a trained prediction device re-learning process 1406 to re-learn the prediction system according to differences between the substrate processing apparatuses.

[0138] Specifically, the learned prediction device re-learning process 1406 performs re-learning by inputting a re-learning signal to the learned second prediction system installed in the substrate processing apparatus, and re-updates the read weights.

[0139] In this way, according to the flow from generation to start-up of the second prediction system, including the mass production method according to the second embodiment, it is possible to reduce the workload at the time of start-up.

[0140] Next, details of each process from generation to startup of the second prediction system will be described. Here, details of the correction device learning process 1304, the learned correction device verification process 1305, the prediction device learning process 1402 to the learned second prediction system verification process 1404, and the learned prediction device relearning process 1406 will be described.

[0141] (1) Details of Correction Device Learning Process First, details of the correction device learning process 1304 will be described. Fig. 15 is a diagram showing a specific example of the correction device learning process. As shown in Fig. 15, in the correction device learning process 1304, first, a correction signal is input to one specific AI device 1101_1 out of multiple mass-produced AI devices. As a result, a reservoir feature is output from the one specific AI device 1101_1. Note that in the correction device learning process 1304, the reservoir feature output from the one specific AI device 1101_1 is used as a teacher signal.

[0142] 15, AI device 1101_2 is one of the multiple mass-produced AI devices other than the specific AI device 1101_1. In correction device learning process 1304, the same correction signal as the correction signal input to the specific AI device 1101_1 is input to AI device 1101_2. As a result, reservoir features are output from AI device 1101_2.

[0143] Next, in the correction device learning process 1304, the reservoir feature output from the AI ​​device 1101_2 is input to the correction reservoir 1211_2 of the correction device 1102_2.

[0144] Next, in the correction device learning process 1304, the correction readout unit 1212_2 of the correction device 1102_2 adds a correction readout weight and reads out the reservoir feature, thereby outputting the corrected reservoir feature. As a result, the corrected reservoir feature output from the correction readout unit 1212_2 is compared with a teacher signal in the comparison change unit 1501, and the correction readout weight of the correction readout unit 1212_2 is updated according to the comparison result.

[0145] In this way, according to the correction device learning process 1304, learning is performed on the correction readout unit 1212_2, and a learned correction device can be generated in which the correction readout weights of the correction readout unit 1212_2 are optimized.

[0146] 15 shows the case where learning is performed on the correction device 1102_2 corresponding to the AI ​​device 1101_2. However, learning in the correction device learning process 1304 is not limited to the correction device 1102_2 corresponding to the AI ​​device 1101_2. Similar learning is performed in parallel on all correction devices corresponding to all mass-produced AI devices other than one specific AI device 1101_1.

[0147] (2) Learned Correction Device Verification Process Next, the learned correction device verification process 1305 will be described in detail. Fig. 16 is a diagram showing a specific example of the verification process for a learned correction device. For ease of explanation, the example in Fig. 16 shows only one specific AI device 1101_1, an AI device 1101_2 other than the specific AI device, and a learned correction device 1102_2 corresponding to the AI ​​device 1101_2, among the mass-produced AI devices.

[0148] In the learned correction device verification process 1305, first, a verification signal is input to the AI ​​device 1101_1, which then outputs a reservoir feature.

[0149] Next, in the learned correction device verification process 1305, the same verification signal as the verification signal input to the specific AI device 1101_1 is input to the AI ​​device 1101_2.

[0150] As a result, reservoir features are output from the AI ​​device 1101_2. In the trained correction device verification process 1305, the reservoir features output from the AI ​​device 1101_2 are input to the correction reservoir 1211_2 of the corresponding trained correction device 1102_2.

[0151] Subsequently, in a learned correction device verification process 1305, the correction readout unit 1212_2 adds a correction readout weight to the output of the correction reservoir 1211_2 and reads it out, thereby outputting a corrected reservoir feature.

[0152] As a result, according to the trained correction device verification process 1305, it is possible to verify whether the reservoir feature output from a specific AI device 1101_1 and the corrected reservoir feature output from the trained correction device 1102_2 are equal (or within a specified error range).

[0153] (3) Details of Prediction Device Learning Process Next, details of the prediction device learning process 1402 will be described. FIG. 17 is a diagram showing a specific example of the learning process of the prediction device in the second prediction system. As shown in FIG. 17, a specific second prediction system in which learning is performed in the prediction device learning process 1402 includes: an AI device 1101; a learned correction device 1102; and a prediction device 204′. Of these, the correction readout unit 1212 included in the learned correction device 1102 has a correction readout weight that was optimized in the correction device learning process 1304 and verified in the learned correction device verification process 1305.

[0154] In the prediction device learning process 1402, first, a learning signal is input to the AI ​​device 1101_1, which then outputs a reservoir feature.

[0155] Next, in the prediction device learning process 1402, the reservoir feature output from the AI ​​device 1101 is input to the correction reservoir 1211 of the learned correction device 1102.

[0156] Next, in a prediction device learning process 1402, the correction readout unit 1212 adds a correction readout weight to the output of the correction reservoir 1211 and reads it out, thereby outputting a corrected reservoir feature.

[0157] The prediction device 204' has a FORCE learning unit 800 using the recursive least squares method before learning is performed. Initial values ​​of read weights are set in the FORCE learning unit 800 using the recursive least squares method before learning is performed.

[0158] Next, in the prediction device learning process 1402, the FORCE learning unit 800 using the recursive least squares method reads the corrected reservoir features by adding read weights, and sequentially updates the read weights by comparing them with the correct data input in advance.

[0159] In this way, the prediction device learning process 1402 can generate a FORCE learning unit 311 (that is, a learned prediction device 204) using the recursive least squares method, with optimized read weights.

[0160] (3) Details of Read Weight Sharing Process Next, details of the read weight sharing process 1403 will be described. Fig. 18 is a diagram showing a specific example of the read weight sharing process of a prediction device in the second prediction system. For convenience of explanation, the example of Fig. 18 shows only the prediction device among the devices included in a mass-produced second prediction system.

[0161] Among these, the trained prediction device 204 indicates a trained prediction device (a source prediction device) generated by performing training on a specific prediction device in the prediction device training process 1402. On the other hand, the multiple prediction devices 204′ indicate prediction devices other than the specific prediction device that were not trained in the prediction device training process 1402.

[0162] In the read weight sharing process 1403, the read weights of the trained prediction device 204 are copied to multiple prediction devices 204'. As a result, the read weights of the trained prediction device 204 are set to multiple prediction devices 204', and multiple trained prediction devices are generated. In other words, the read weight sharing process 1403 allows multiple trained second prediction systems to be mass-produced.

[0163] (4) Details of the Trained Second Prediction System Verification Process Next, details of the trained second prediction system verification process 1404 will be described. Fig. 19A is a diagram showing a specific example of the verification process for a trained second prediction system. For ease of explanation, the example of Fig. 19A shows only the AI ​​device 1101, the trained correction device 1102, and the trained prediction device 204 among the devices included in a mass-produced trained second prediction system.

[0164] In the trained second prediction system verification process 1404, first, a verification signal is input to the AI ​​device 1101. As a result, the AI ​​device 1101 outputs reservoir features.

[0165] Next, in the trained second prediction system verification process 1404, the reservoir features output from the AI ​​device 1101 are input to the correction reservoir 1211 of the trained correction device 1102.

[0166] Next, in the trained second prediction system verification process 1404, the correction readout unit 1212 adds a correction readout weight to the output of the correction reservoir 1211 and reads it out, thereby outputting a corrected reservoir feature.

[0167] Next, in the trained second prediction system verification process 1404, the trained prediction device 204 reads the corrected reservoir features, adds weights to them, and reads them out to output a predicted value.

[0168] In the example of Figure 19A, for the sake of convenience of explanation, only one of the mass-produced trained second prediction systems is shown, but the verification signal is input to the AI ​​device 1101 of all mass-produced trained second prediction systems.

[0169] As a result, predicted values ​​are output from the trained prediction devices 204 of all mass-produced trained second prediction systems. As a result, the trained second prediction system verification process 1404 can verify whether the respective predicted values ​​are equal (or within a predetermined error range).

[0170] As a result of the verification, if the predicted value of the trained second prediction system is determined to be equal (or within a predetermined error range), the trained second prediction system is permitted to be installed in the substrate processing apparatus 120 .

[0171] (5) Details of the Trained Prediction Device Re-learning Process Next, details of the trained prediction device re-learning process 1406 will be described. Fig. 19B is a diagram showing a specific example of the re-learning process of the trained prediction device in the trained second prediction system. For ease of explanation, the example of Fig. 19B shows only the AI ​​device 1101, the trained correction device 1102, and the trained prediction device 204 among the devices included in the trained second prediction system installed in the substrate processing apparatus 120.

[0172] In the learned prediction device relearning process 1406, first, the time-series sensor data a output from the sensor a 122a is notified from the I / O control device 201 to the voltage modulation device 202, and the data is converted into voltage data in the voltage modulation device 202. Next, in the learned prediction device relearning process 1406, the voltage data converted by the voltage modulation device 202 is input to the AI ​​device 1101 as a relearning signal. As a result, the AI ​​device 1101 outputs a reservoir feature.

[0173] Next, in the trained prediction device re-learning process 1406, the reservoir feature output from the AI ​​device 1101 is input to the correction reservoir 1211 of the trained correction device 1102.

[0174] Next, in the trained prediction device re-learning process 1406, the correction readout unit 1212 adds a correction readout weight to the output of the correction reservoir 1211 and reads it out, thereby outputting a corrected reservoir feature. The output reservoir feature is then input to the trained prediction device 204, which has a FORCE learning unit 311 using the recursive least squares method with an optimized readout weight.

[0175] Next, in the learned prediction device re-learning process 1406, the FORCE learning unit 311 using the recursive least squares method reads out the reservoir feature amount after adding a readout weight and compares it with the process state data (correct answer data) input in advance. As a result, the learned prediction device re-learning process 1406 can sequentially re-update the readout weight.

[0176] As a result, according to the learned prediction device re-learning process 1406, it is possible to generate a FORCE learning unit 311 (i.e., a re-learned prediction device) using the recursive least squares method, which has been re-learned according to the machine differences of the substrate processing apparatus 120.

[0177] <Summary> As is clear from the above explanation, the trained second prediction system 1100 includes: an AI device 1101 that processes time-series signals, and a trained correction device 1102 that corrects reservoir features that are output by the AI ​​device 1101. a trained prediction device that reads out the reservoir features corrected by the trained correction device 1102, adds read weights, and outputs predicted values.

[0178] Furthermore, in the trained second prediction system 1100, the trained correction device 1102 has a correction reservoir 1211 and a correction readout unit 1212 that adds a correction readout weight to the output of the correction reservoir 1211 and reads it out. The correction readout unit 1212 adds a correction readout weight to the output from the correction reservoir 1211 when the reservoir feature, which is the output of the AI ​​device 1101 when processing a time-series signal, is input to the correction reservoir 1211 and reads it out. As a result, the correction readout unit 1212 outputs the corrected reservoir feature. The correction readout unit 1212 has a correction readout weight that has been updated by a learning process based on the error between the corrected reservoir feature and the reservoir feature, which is the output of another AI device when the other AI device processes the correction signal. - When the learned prediction device is a learned prediction device whose read weights have been updated by performing a learning process on another prediction device, the learned prediction device has read weights copied from the learned prediction device.

[0179] As a result, the trained second prediction system 1100 does not need to start learning from scratch at startup, thereby reducing the workload at startup.

[0180] Furthermore, the mass production method according to the second embodiment includes: - When multiple AI devices that process time-series signals are generated, mass-producing correction devices that correct reservoir features, which are the output of each of the AI ​​devices except for one specific AI device, so that they match the reservoir features, which are the output of the specific AI device; - Mass-producing prediction devices that output predicted values ​​by adding read weights copied from the source prediction device to the reservoir features corrected by the correction device and reading them out; - When mass-producing the correction devices, acquiring, as a teacher signal, the reservoir features, which are the output of the specific AI device when a correction signal is input; - When mass-producing the correction devices, inputting, to each corresponding correction device, the reservoir features, which are the output of each of the AI ​​devices except for the specific AI device when a correction signal is input to each of the AI ​​devices except for the specific AI device; thereby, each corresponding correction device outputs a corrected reservoir feature. - When mass-producing correction devices, a learning process is performed on the correction devices based on each error between the teacher signal and each corrected reservoir feature, and each corrected readout weight is updated. - When mass-producing prediction devices, a learning process is performed on a specific prediction device that adds a readout weight to the output of a specific AI device, reads it, and outputs a predicted value, and the readout weight is updated. The updated readout weight is copied from the specific prediction device to each of the other prediction devices except for the specific AI device.

[0181] As a result, according to the mass production method of the second embodiment, there is no need to perform learning from scratch at the time of start-up, and the workload at the time of start-up can be reduced.

[0182] [Other Embodiments] In the second embodiment, the AI ​​device 1101 is described as a device having a physical reservoir. However, the target device whose output is corrected by the trained correction device 1102 is not limited to a device having a physical reservoir. A device that is a high-order nonlinear system and cannot be corrected by a linear optimization method may also be the target device.

[0183] In the second embodiment, the relationship between the number N of reservoir nodes of the physical reservoir possessed by the AI ​​device 1101 and the number n of reservoir nodes of the correction reservoir possessed by the learned correction device 1102 is described as N>n. However, the relationship between the number N of reservoir nodes of the physical reservoir possessed by the AI ​​device 1101 and the number n of reservoir nodes of the correction reservoir possessed by the learned correction device 1102 may be set to an optimal ratio based on prediction accuracy or memory capacity.

[0184] In the first and second embodiments, the case where the presence or absence of an abnormality or sensor data is output as a predicted value of the process state has been described. However, the predicted value output by the trained prediction device 204 is not limited to the presence or absence of an abnormality or sensor data, and may be, for example, a level indicating the state of the process or the presence or absence of a failure in the equipment that executes the manufacturing process.

[0185] The present invention is not limited to the configurations described in the above embodiments, but may be combined with other elements, etc. These aspects can be changed without departing from the spirit of the present invention, and can be appropriately determined depending on the application form.

[0186] This application claims priority based on Japanese Patent Application No. 2023-188667, filed on November 2, 2023, the entire contents of which are incorporated herein by reference.

[0187] 120: Substrate processing apparatus 121: Chamber 122: Sensor 123: Management apparatus 124: Management unit 125: Control device 126: Control unit 127: Actuator 128: Trained first prediction system 201: I / O control device 202: Voltage modulation device 203: Calibration verified AI device 204: Trained prediction device 301: D / A conversion unit 303: Processing unit 304: Reservoir feature output unit 311: FORCE learning unit using recursive least squares method 401: Physical reservoir generation processing 402: AI device generation processing 403: AI device calibration verification processing 404: First prediction system generation processing 501: Prediction device learning processing 502: Readout weight sharing processing 503: Trained first prediction system verification processing 504: Substrate processing apparatus mounting process 505: Learned prediction device relearning process 800: FORCE learning unit using recursive least squares method 1100: Learned second prediction system 1101: AI device 1102: Learned correction device 1201: D / A conversion unit 1202: Physical reservoir 1203: Reservoir feature output unit 1211: Correction reservoir 1212: Correction readout unit 1301: Physical reservoir generation process 1302: AI device generation process 1303: Correction device generation process 1304: Correction device learning process 1305: Learned correction device verification process 1401: Second prediction system generation process 1402: Prediction device learning process 1403: Readout weight sharing process 1404: Learned second prediction system verification process 1405: Substrate processing apparatus mounting process 1406: Learned prediction device relearning process

Claims

1. A prediction system comprising: a calibration-verified AI device in which multiple physical reservoirs for processing time-series signals are connected to each other; and a trained prediction device that reads out the output from the multiple physical reservoirs possessed by the calibration-verified AI device, adds weights to the output, and outputs a predicted value, wherein the calibration-verified AI device has the multiple physical reservoirs that have been verified to be calibrated to have equal physical characteristics by adding M-series noise to the time-series signals; and the trained prediction device is another trained prediction device that reads out the output from multiple other physical reservoirs that have equal physical characteristics to the multiple physical reservoirs, adds weights to the output, and outputs a predicted value, and has readout weights copied from the other trained prediction device whose readout weights have been updated through learning.

2. The prediction system of claim 1, wherein the multiple physical reservoirs connected to each other are configured such that the output of the connected physical reservoir is averaged over a predetermined time period and a weight is added between the connected physical reservoir and the output of the connected physical reservoir, and then the output is input to the connected physical reservoir.

3. A prediction system as described in claim 1, wherein the number and connection relationships of the multiple physical reservoirs possessed by the calibration-verified AI device and the multiple physical reservoirs possessed by the other calibration-verified AI device are equal to each other.

4. The prediction system of claim 1, wherein the trained prediction device is re-trained using a time-series signal acquired in an apparatus in which the prediction system is installed, thereby re-updating the readout weights.

5. A mass production method comprising the steps of: verifying that a plurality of physical reservoirs are calibrated to have equal physical characteristics by inputting a time series signal to which a common M-sequence noise has been added to each of the plurality of physical reservoirs; when a plurality of calibration-verified AI devices are generated by generating a plurality of processing units in which the verified plurality of physical reservoirs are connected to each other, updating the readout weights by learning a prediction device that reads out the output from a specific calibration-verified AI device and adds a readout weight; and duplicating and setting the updated readout weights for each prediction device that reads out the output from each of the calibration-verified AI devices other than the specific calibration-verified AI device and adds a readout weight.

6. The mass production method described in claim 5, wherein in the step of updating the readout weights, the multiple physical reservoirs connected to each other of the specific one calibration-verified AI device are configured so that the output of the source physical reservoir is averaged over a predetermined time period, a weight is added between the source physical reservoir and the output of the destination physical reservoir, and then the output is input to the destination physical reservoir.

7. The mass production method of claim 5, wherein the step of updating the readout weights includes calculating an error by comparing a predicted value output from the prediction device by adding readout weights to the outputs from the multiple physical reservoirs possessed by the specific one calibration-verified AI device and reading the predicted value with ground truth data, and updating the readout weights based on the calculated error.

8. The mass production method according to claim 7, wherein the step of updating the read weights updates the read weights by performing FORCE learning based on the calculated errors.

9. The mass production method described in claim 5, wherein the physical reservoirs each possessed by the multiple calibration-verified AI devices are equal in number and connection relationship among the multiple calibration-verified AI devices.

10. The mass production method according to claim 5, further comprising a step of verifying that when a time series signal to which a common M-sequence noise has been added is input to a plurality of the calibration-verified AI devices, the predicted values ​​output from the plurality of prediction devices by adding the updated readout weights and reading out are equal to each other.

11. A prediction system having: a target device that processes a time-series signal and outputs a feature; a correction device that corrects the feature output by the target device; and a prediction device that adds readout weights of an original prediction device to the feature corrected by the correction device, reads it out, and outputs a predicted value, wherein the correction device has: a correction reservoir to which the feature output by the target device is input; and a correction readout unit that, when the feature output by the target device is input to the correction reservoir, corrects the feature by adding a correction readout weight and reading out the output of the correction reservoir, wherein the correction readout unit has the correction readout weights updated by learning based on an error between the feature output by the other target device when a learning signal is processed by the other target device and the feature output by the target device when the learning signal is processed by the target device and corrected by the correction device.

12. The prediction system of claim 11, wherein the prediction device is a source prediction device that reads the output of the other target device, adds read weights, reads it, and outputs a predicted value, and has read weights copied from the source prediction device whose read weights have been updated by learning.

13. The prediction system of claim 12, wherein the target device is an AI device having N variables, and the correction reservoir is a digital reservoir having n nodes (N>n).

14. The prediction system according to claim 13, wherein the correction readout unit adds the correction readout weight to the output of the correction reservoir, reads it out, and outputs N corrected features.

15. A prediction system as described in claim 14, wherein the prediction device is re-learned using a time-series signal acquired in an apparatus in which the prediction system is installed, thereby re-updating the read weights of the source prediction device.

16. A mass-production method for mass-producing a correction device that corrects features output by each of the target devices other than one specific target device so that they match the features output by the one specific target device when a plurality of target devices for processing a time-series signal are generated, and a prediction device that outputs a predicted value by adding a readout weight of an original prediction device and reading out the corrected features, the mass-production method comprising the steps of: acquiring, as a teacher signal, the feature output by the one specific target device when a time-series signal is input; inputting, into each correction reservoir possessed by each of the target devices other than the one specific target device when the time-series signal is input to each of the target devices other than the one specific target device, the feature output by each of the target devices other than the one specific target device when the time-series signal is input to each of the target devices other than the one specific target device, so that each correction readout unit possessed by each of the corresponding correction devices reads out the output from each correction reservoir after adding a correction readout weight; and updating each correction readout weight possessed by each correction readout unit by learning based on each error between the teacher signal and each of the corrected features.

17. The mass production method according to claim 16, further comprising a step of duplicating read weights from the source prediction device, the read weights of which have been updated through learning, to the prediction devices other than the source prediction device, the source prediction device adding read weights to the feature quantities output by the specific one target device, and outputting a predicted value.

18. The mass production method according to claim 17, wherein the target device is an AI device having N variables, and the correction reservoir is a digital reservoir having n nodes (N>n).

19. The mass production method according to claim 18, wherein each of the correction readout units adds the correction readout weight to the output of each of the correction reservoirs and reads out the output, thereby obtaining N corrected feature quantities, respectively.

20. The mass production method according to claim 16, further comprising a step of verifying that, when signals of equal time series are input to a plurality of the target devices, the predicted value output from the source prediction device is equal to the predicted value output from the prediction devices other than the source prediction device to which duplicated read weights are set.

Images