Method and system for predicting metrology data using a neural network.

Abstract

A method is presented for generating, using a neural network, predicted metrology data of a patterned portion. The neural network comprises an input network and an output network. The input network comprises a common input model and a specific input model, wherein the common input model has been trained by a process in which the common input model processes data based on training metrology data generated by a plurality of metrology systems, and the specific input model is specific to a first metrology system.

Description

METHOD AND SYSTEM FOR PREDICTING METROLOGY DATA USING A NEURAL NETWORKCROSS-REFERENCE TO RELATED APPLICATION

[0001] This application claims priority of EP application 24178687.0 which was filed on 29 May 2024 and which is incorporated herein in its entirety by reference.FIELD

[0002] The present invention relates to a method and a system for generating predicted metrology data using a neural network.BACKGROUND

[0003] A lithographic apparatus is a machine constructed to apply a desired pattern onto a substrate. A lithographic apparatus can be used, for example, in the manufacture of integrated circuits (ICs). A lithographic apparatus may, for example, project a pattern (also often referred to as “design layout” or “design”) at a patterning device (e.g., a mask) onto a layer of radiation-sensitive material (resist) provided on a substrate (e.g., a wafer).

[0004] To project a pattern on a substrate a lithographic apparatus may use electromagnetic radiation. The wavelength of this radiation determines the minimum size of features which can be formed on the substrate. Typical wavelengths currently in use are 365 nm (i-line), 248 nm, 193 nm and 13.5 nm. A lithographic apparatus, which uses extreme ultraviolet (EUV) radiation, having a wavelength within the range 4-20 nm, for example 6.7 nm or 13.5 nm, may be used to form smaller features on a substrate than a lithographic apparatus which uses, for example, radiation with a wavelength of 193 nm.

[0005] Low-ki lithography may be used to process features with dimensions smaller than the classical resolution limit of a lithographic apparatus. In such process, the resolution formula may be expressed as CD = k iX / . / NA, where / . is the wavelength of radiation employed, NA is the numerical aperture of the projection optics in the lithographic apparatus, CD is the “critical dimension” (generally the smallest feature size printed, but in this case half-pitch) and ki is an empirical resolution factor. In general, the smaller ki the more difficult it becomes to reproduce the pattern on the substrate that resembles the shape and dimensions planned by a circuit designer in order to achieve particular electrical functionality and performance. To overcome these difficulties, sophisticated fine-tuning steps may be applied to the lithographic projection apparatus and / or design layout. These include, for example, but not limited to, optimization of NA, customized illumination schemes, use of phase shifting patterning devices, various optimization of the design layout such as optical proximity correction (OPC, sometimes also referred to as “optical and process correction”) in the design layout, or other methods generally defined as “resolution enhancement techniques” (RET).Alternatively, tight control loops for controlling a stability of the lithographic apparatus may be used to improve reproduction of the pattern at low kl.

[0006] After a substrate has been patterned using the lithographic apparatus, it is often desirable to inspect the substrate (i.e. to measure properties of the patterned structure), e.g. to determine whether the produced pattern accurately resembles the intended design. To this end, a metrology system (e.g. optical metrology systems, scanning electron microscopes (SEM), soft x-ray metrology systems, and the like) can be used to acquire metrology data (e.g. an image) of the patterned substrate. The acquired metrology data is often used to adjust process parameters (e.g. scanner properties, etch process parameters, and the like) which specify the manufacturing process. Frequently it is desirable to use a large number of metrology systems in parallel to control the fabrication process in a semiconductor fabrication plant. However, in practice, different instances of nominally identical metrology systems generate (slightly) different metrology data for the same target (this phenomenon is also referred to as “tool-mismatch”). Tool-mismatch is caused, for example, by different instances of the metrology tool suffering from different aberrations, different degrees of drift, different degradation during their lifetime, and so on. Conventionally this problem is addressed by imposing tight tolerances on the hardware components of the metrology systems. Such an approach however results in increased cost.

[0007] It is therefore desirable to provide techniques that can mitigate variations in metrology data generated by different metrology systems without relying on hardware -based solutions.SUMMARY

[0008] The present invention aims to provide new and useful methods for predicting metrology data using a neural network which receives, as input, metrology data generated by inspecting a target using a first metrology system (and information about the configuration (e.g. the type of the tool; and / or its settings at the time that the metrology data was created) of the first metrology system) and determines, from the input, predicted metrology data which would have been generated by inspecting the target using a second metrology system. This means that the output data of the neural network “looks” like metrology data generated by the second metrology system. The output of the neural network can then be compared to metrology data that has actually been generated by the second metrology system without being affected by tool-mismatch.

[0009] In broad terms, the present invention proposes a neural network that processes metrology data to effectively remove characteristics which are specific to the metrology system used to generate the input metrology data and to add the characteristics specific to a target metrology system. One way of implementing this is for the neural network to encode the input metrology data using an input network that takes into account the characteristics of the metrology system used to generate the input metrology data (i.e. the input network has been trained using training metrology data generated by this metrology system). An output network that takes into account the characteristics of the targetmetrology system (e.g. the output network may have been trained using training metrology data generated by the target metrology system) can then generate the predicted metrology data from the encoded metrology data.

[0010] According to a first aspect, there is provided a method generating, using a neural network, predicted metrology data of a patterned portion of a substrate patterned using a lithographic apparatus. The neural network comprises an input network and an output network. The input network comprises a common input model and a specific input model, wherein the common input model has been trained by a process in which the common input model processes data based on training metrology data generated by a plurality of metrology systems, and the specific input model is specific to a first metrology system. The method comprises: receiving metrology data of the patterned portion generated using the first metrology system; obtaining configuration data specifying a configuration of the first metrology system used to generate the metrology data (e.g. configuration data may comprise data characterizing the first metrology system and / or setting data specifying operating parameters of the first metrology system at the time the metrology data was captured); processing, using the common input model and the specific input model, the metrology data and the configuration data to generate an encoding of the patterned portion; and processing, using the output network of the neural network, the encoding of the patterned portion to generate the predicted metrology data, the predicted metrology data being a prediction of metrology data which would have been generated by inspecting the patterned portion using a second metrology system.

[0011] The inventors have found that it is advantageous to use a neural network architecture that uses an input network comprising a common input model (trained on training metrology data generated by a plurality of metrology systems) and a specific input model (e.g. trained only on training metrology data generated by the first metrology system). One of the advantages of this architecture is that training metrology data from at least one other metrology system (e.g. a third metrology system different from the first metrology system) may be used to train parts of the neural network, e.g. the common input model. The common input model makes the task of the specific input model easier. Thus, the specific input model can have fewer variable parameters (e.g. fewer than the common input model), so training it is less computationally expensive and can be performed with reduced training data relating to the first metrology system. Furthermore, the neural network can conveniently be adapted to work for a new first metrology system by simply obtaining weights for a corresponding specific input model (e.g. via directly training a corresponding specific input model or via a generator model, as described below in more detail) while the common input model and the output network do not need to be re-trained. Note that the performance of the specific input model for the new first metrology system would typically be as high as that of the specific input models of first metrology systems used in the training of the common input model, provided that the new firstmetrology system is similar to (e.g. not out of the distribution of) the first metrology systems used in the training of the common input model.

[0012] Many possibilities of obtaining suitable weights for the specific input model exist. The term “weights” of a model is used herein to refer to the numerical values of the parameters defining the respective model. The specific model may be specific to the first metrology system in that it is generated by processing data measured using, or data describing, the first metrology system, e.g. by a process in which data measured using, or data describing, other metrology systems is not processed. In a first possibility, the specific input model has been trained by a process in which the specific input model processes data based on training metrology data generated only by the first metrology system. In another possibility, the weights of the specific input model have been determined by processing first metrology system data specifying the first metrology system using an input model generator network. The input model generator network is configured to receive, as input, metrology system data specifying a respective metrology system and to determine, as output, weights of the specific input model for the respective metrology system.

[0013] In some implementations, the metrology system data may comprise one or more parameters which define a distribution of a set of latent variables, and the input model generator network may be configured to determine the weights of the specific input model for a respective metrology system by sampling from the distribution defined by the input metrology system data to determine values of the set of latent variables, and determine the weights of the specific input model for a respective metrology system based on the set of latent variables.

[0014] In some implementations, a first one of the common input model and the specific input model may be configured to process the metrology data to generate an encoding of the metrology data, and a second one of the common input model and the specific input model being configured to process the encoding of the metrology data. For example, the first one of the common input model and the specific input model may be the common input model, and the second one of the common input model and the specific input model may be the specific input model.

[0015] In some implementations, processing the metrology data and the configuration data to generate an encoding of the patterned portion may comprise: processing, using the first one of the common input model and the specific input model, the metrology data to generate an encoding of the metrology data; and processing, using the second one of the common input model and the specific input model, the encoding of the metrology data and the configuration data to generate the encoding of the patterned portion.

[0016] In some implementations, the plurality of metrology systems used in the process to train the common input model may comprise the first metrology system, the second metrology system and at least one further metrology system.

[0017] In some implementations, the method may further comprise a step of obtaining target configuration data specifying a configuration of the second metrology system. In this case, the predicted metrology data may be a prediction of metrology data which would have been generated by inspecting the patterned portion using the second metrology system configured according to the target configuration data.

[0018] In some implementations, the output network may comprise a common output model trained by a process in which the common output model processes data based on training metrology data generated by a plurality of metrology systems. The plurality of metrology systems used in the process to train the common output model may comprise the first metrology system, the second metrology system and at least one further metrology system.

[0019] In some implementations, the output network may further comprise a specific output model that is specific to the second metrology system. The specific model may be specific to the second metrology system in that it is generated by processing data measured using, or data describing, the second metrology system, e.g. by a process in which data measured using, or data describing, other metrology systems is not processed. In some implementations, the specific output model may have been trained by a process in which the weights of the specific output model are adjusted based on a discrepancy between the predicted metrology data and training metrology data generated only by the second metrology system. Alternatively, the weights of the specific output model may have been determined by processing second metrology system data specifying the second metrology system using an output model generator network. The output model generator network is configured to receive, as input, metrology system data specifying a respective metrology system and to determine, as output, weights of the specific output model for the respective metrology system.

[0020] It is advantageous to use a neural network architecture that uses a common output model and a metrology system specific input model since this architecture means the network can conveniently be adapted to work for different target metrology systems by simply training a corresponding specific output model while the common output model does not need to be re-trained.

[0021] In some implementations, a first one of the common output model and the specific output model may be configured to process data generated by the input network, and a second one of the common output model and specific output model may be configured to process the output of the first one of the common output model and specific output model.

[0022] In some implementations, the first one of the common output model and the specific output model may be the specific output model, and the second one of the common output model and the specific output model may be the common output model.

[0023] In some implementations, the plurality of metrology systems used in the process to train the common output model may comprise the first metrology system, the second metrology system and at least one further metrology system.

[0024] In some implementations, processing the encoding of the patterned portion to generate the predicted metrology data may comprise processing, using the specific output model, the encoding of the patterned portion and the target configuration data to generate the encoding of the predicted metrology data, and processing, using the common output model, the encoding of the predicted metrology data to generate the predicted metrology data.

[0025] In some implementations, processing the metrology data and the configuration data of the first metrology system to generate an encoding of the patterned portion may comprise: processing, using the common input model, the metrology data to generate an encoding of the metrology data; processing, using the specific input model, the encoding of the metrology data and the configuration data of the first metrology system to generate first encoded data; processing, using the specific output model, the encoding of the metrology data and the target configuration data to generate second encoded data; and combining the first and second encoded data to generate the encoding of the patterned portion. For example, the first and second encoded data may be combined by concatenating the first and second encoded data.

[0026] In some implementations, at least one of the first metrology system and the second metrology system is a real-world metrology system.

[0027] In some implementations, at least one of the first metrology system and the second metrology system may be a simulated metrology system (e.g. a computational model based on physics).

[0028] In some implementations, the neural network may further comprise an applicationspecific model corresponding to an application type, from a set of application types, associated with the patterned portion, the application-specific model being configured to process data from the input network and to provide data to the output network data. The predicted metrology data may be generated by processing the encoding of the patterned portion using the output network of the neural network and the data provided by the application model. The application- specific model has been trained by a process in which the application-specific model processes data based on training metrology data of patterned portions associated with the application type. Here the term “application” refers to a difference in the characteristics of the patterned substrate used to produce the training metrology data for the first metrology system, and the training data for the second metrology system. For example, the two sets of training data may be produced by inspecting substrates patterned according to different respective uDBO (micro-diffraction based overlay) target designs.

[0029] In some implementations, the configuration data of the first metrology system may comprise reference metrology data of a reference target generated using the first metrology system configured in the same configuration as used to generate the received metrology data.

[0030] In some implementations, the metrology data may comprise image data.

[0031] In some implementations, the different metrology systems may be different instances of an optical dark-field microscope.

[0032] In some implementations, the configuration data of the first metrology system comprises information specifying a spectral density of light used to generate the metrology data.

[0033] In some implementations, the different metrology systems may be different beams of a multi-beam scanning electron microscope.

[0034] In some implementations, the metrology data may comprise a pupil representation defining a distribution of radiation in a pupil plane of the metrology system.

[0035] In some implementations, the neural network model may comprise a U-Net architecture.

[0036] In some implementations, the neural network model may comprise one or more skip connections. The one or more skip connections may comprise at least one skip connection from the common input model to the common output model.

[0037] In some implementations, the first metrology system and the second metrology system may be the same metrology system, and wherein the target configuration data specify a configuration of the metrology system that is different from the configuration used to generate the metrology data.

[0038] In some implementations, the proposed method may be used to iteratively optimize the target configuration data (e.g. over one or iterations). For example, the method may further comprise processing the predicted metrology data to determine one or more performance parameters (e.g. an uncertainty associated with an determined overlay value), processing the one or more performance parameters to generate updated target configuration data (e.g. with updated setting data; for example the updated target configuration data may specify a different (e.g. more suitable) spectral density for inspecting the pattern portion which is likely to have resulted in metrology data from which improved performance parameters could have been obtained (i.e. a reduced uncertainty of the overlay value)), and processing, using the neural network, the metrology data, the configuration data of the first metrology system and the updated target configuration data to generate updated predicted metrology data. This means that improved performance parameters can be obtained from the updated predicted metrology data (i.e. a reduced uncertainty of the overlay value). This may be particularly useful in cases where the first metrology system and the second metrology system may be the same metrology system.

[0039] According to a second aspect, there is provided a method of training a neural network (e.g. the neural aspect according to the first aspect) for generating predicted metrology data of a patterned portion of a substrate patterned using a lithographic apparatus. The neural network comprising an input network and an output network, the input network comprising a common input model and a specific input model, wherein specific input model is associated with a first metrology system. The neural network is configured to: receive metrology data of the patterned portion generated using the first metrology system;receive configuration data specifying a configuration of the first metrology system used to generate the metrology data; process, using the common input model and the specific input model, the metrology data and the configuration data to generate an encoding of the patterned portion; and process, using the output network of the neural network, the encoding of the patterned portion to generate the predicted metrology data, the predicted metrology data being a prediction of metrology data which would have been generated by inspecting the patterned portion using a second metrology system.

[0040] In some implementations, the neural network system may comprise at least one supplementary processing branch. A first of the supplementary processing branches may comprise (i) a first supplementary common input model configured to receive the output of the common input model, (ii) optionally a first supplementary specific input model obtained using training metrology data generated only by the first metrology system, and (iii) a first supplementary output network, the output network being configured to receive the output of the first supplementary output network. Optionally, when there are a plurality of the supplementary processing branches in a sequence, each of the supplementary processing branches, except the first, may comprise (i) a corresponding supplementary common input model configured to receive the output of the supplementary common input model of the preceding supplementary processing branch, (ii) optionally a corresponding supplementary specific input model obtained using training metrology data generated only by the first metrology system, and (iii) a corresponding supplementary output network, the supplementary output network of the preceding supplementary processing branch being configured to receive the output of the corresponding supplementary output network. The supplementary processing branches improve training speed and performance of the neural network.

[0041] In some implementations, the method may further comprise a step of obtaining inspection data comprising a first index value indicating a number of inspections which had been carried out on the patterned portion prior to the generation of the metrology data, and a second index value. In these implementations, the predicted metrology data is a prediction of metrology data which would have been generated by inspecting the patterned portion using a second metrology system if the patterned portion had been subject to a number of prior inspections indicated by the second index value. The neural network may further comprise a transfer model configured to process i) the inspection data, and ii) the encoding of the patterned portion generated by the input network, to generate data, and to provide said data to the output network. The predicted metrology data may then be generated by the output model processing the data provided by the transfer model. In some implementations, the inspection data may be comprised in the configuration data or the metrology system data.

[0042] In some implementations, the inspection data may further comprise data descriptive of the inspection process used in the previous inspections, e.g. data indicating a time duration associated with the previous inspections, and / or an energy level used during the previous inspections.

[0043] The use of the transfer model and the inspection data enables the system to take into account modifications of the sample caused by the metrology systems, i.e. the fact that, for certain types of metrology systems and for certain types of samples, the act of measuring a sample inevitably changes the sample. For example, inspecting a sample that includes a resist pattern with a SEM metrology system can degrade the resist and result in shrinking of critical dimensions of the printed patterns. In such situations, it is important to track the number and conditions of the performed measurements of the target.

[0044] In some implementations, the method may comprise obtaining a training dataset comprising first training metrology data generated using the first metrology system, configuration data specifying a configuration of the first metrology system used to generate the first training metrology data, and a second training metrology data generated using the second metrology system, wherein the first and second training metrology data have been generated by inspecting the same patterned portion of a training substrate patterned by the, or a further, lithographic apparatus.

[0045] In some implementations, the method may comprise processing, using the common input model and the specific input model, the first training metrology data and the corresponding configuration data to generate an encoding of the patterned portion.

[0046] In some implementations, the method may comprise processing, using the output network of the neural network, the encoding of the patterned portion to generate predicted second training metrology data, the predicted second training metrology data being a prediction of metrology data which would have been generated by inspecting the patterned portion using the second metrology system.

[0047] In some implementations, the method may comprise updating weights of the neural network to reduce a measure of the discrepancy between the predicted second training metrology data and the second training metrology data.

[0048] In some implementations, the input network may comprise a plurality of specific input models, each specific input model being associated with a different respective metrology system of a plurality of metrology systems comprising at least the first and second metrology systems and a third metrology system.

[0049] In some implementations, the neural network may be configured to process the metrology data and the configuration data to generate an encoding of the patterned portion by: processing, using the common input model, the metrology data to generate an encoding of the metrology data; and processing, using the specific input model of the first metrology system, the encoding of the metrology data and the configuration data to generate the encoding of the patterned portion,wherein the training dataset further comprises third training metrology data generated using the third metrology system, configuration data specifying a configuration of the third metrology system used to generate the third training metrology data, and fourth training metrology data generated using the second metrology system, wherein the third and fourth training metrology data have been generated by inspecting the same patterned portion of a respective training substrate patterned by the, or a further, lithographic apparatus; the method further comprising: processing, using the common input model, the third training metrology data to generate an encoding of the third training metrology data; and processing, using the specific input model of the third metrology system, the encoding of the third training metrology data and the corresponding configuration data of the third metrology system to generate the encoding of the patterned portion, processing, using the output network of the neural network, the encoding of the patterned portion to generate predicted fourth training metrology data, the predicted fourth training metrology data being a prediction of metrology data which would have been generated by inspecting the patterned portion using the second metrology system; and updating weights of the neural network to reduce: a measure of discrepancy between the predicted fourth training metrology data and the fourth training metrology data.

[0050] In some implementations, the output network may comprise a common output model and a plurality of specific output models, each specific output model associated with a different metrology system of the plurality of metrology systems, wherein the neural network is configured to process the encoding of the patterned portion to generate the predicted metrology data by: processing, using the specific output model of the second metrology system, the encoding of the patterned portion and target configuration data to generate the encoding of the predicted metrology data; and processing, using the common output model, the encoding of the predicted metrology data portion to generate the predicted metrology data, wherein the training dataset further comprises fifth training metrology data generated using the first metrology system, configuration data specifying a configuration of the first metrology system used to generate the fifth training metrology data, a sixth training metrology data generated using the third metrology system, and configuration data specifying a configuration of the third metrology system used to generate the sixth training metrology data, wherein the fifth and sixth training metrology data have been generated by inspecting the same patterned portion of a respective training substrate patterned by the, or a further, lithographic apparatus; the method further comprising:processing, using the common input model, the fifth training metrology data to generate an encoding of the fifth training metrology data; and processing, using the specific input model of the first metrology system, the encoding of the fifth training metrology data and the corresponding configuration data of the first metrology system to generate an encoding of the patterned portion, processing, using the specific output model of the third metrology system, the encoding of the patterned portion and the corresponding configuration data of the third metrology system to generate an encoding of predicted sixth training metrology data; and processing, using the common output model, the encoding of the predicted sixth metrology data to generate the predicted sixth training metrology data, updating weights of the neural network to reduce: a measure of discrepancy between the predicted sixth training metrology data and the sixth training metrology data.

[0051] In some implementations, the first metrology system and the third metrology system may be the same metrology system, and the corresponding configuration data may specify the same configuration of the metrology system. Such training may improve the neural network (in particular the specific models) because it encourages self-consistency.

[0052] In some implementations, the plurality of metrology systems may comprise at least one simulated metrology system.

[0053] According to a third aspect, there is provided a computer-implemented method of training an input model generator network to determine, from metrology system data specifying a respective metrology system, respective weights of a specific input model for the respective metrology system, the specific input model being comprised in a neural network for generating predicted metrology data of a patterned portion of a substrate patterned using a lithographic apparatus. The neural network comprises an input network and an output network. The input network comprises a common input model and the specific input model. The neural network may be the described above, which is configured to: receive metrology data of the patterned portion generated using the first metrology system; receive configuration data specifying a configuration of the first metrology system used to generate the metrology data; process, using the common input model and the specific input model, the metrology data and the configuration data to generate an encoding of the patterned portion; and process, using the output network of the neural network, the encoding of the patterned portion to generate the predicted metrology data, the predicted metrology data being a prediction of metrology data which would have been generated by inspecting the patterned portion using a second metrology system. The method comprises obtaining a training dataset comprising first training metrology data generated using the first metrology system, configuration data specifying a configuration of the firstmetrology system used to generate the first training metrology data, and second training metrology data generated using the second metrology system. The first and second training metrology data have been generated by inspecting the same patterned portion of a training substrate patterned by the, or a further, lithographic apparatus. The method further comprises: obtaining candidate metrology system data specifying the first metrology system; processing, using the input model generator network, the obtained metrology system data to generate training weights of the specific input model for the first metrology system; processing, using the common input model and the specific input model in accordance with the generated training weights, the first training metrology data and the corresponding configuration data to generate an encoding of the patterned portion; processing, using the output network of the neural network, the encoding of the patterned portion to generate predicted second training metrology data, the predicted second training metrology data being a prediction of metrology data which would have been generated by inspecting the patterned portion using the second metrology system; and updating weights of the input model generator network and one or more parameters of the metrology system data to reduce a measure of the discrepancy between the predicted second training metrology data and the second training metrology data.

[0054] In some implementations, to reduce the measure of the discrepancy between the predicted second training metrology data and the second training metrology data, the weights of the input model generator network and one or more parameters defining the metrology system data are jointly adjusted with weights of the common input model and / or weights of the output network.

[0055] In some implementations, when the output network comprises a common output model (trained by a process in which the common output model processes data based on training metrology data generated by a plurality of metrology systems) and a specific output model that is specific to the second metrology system, the method may also involve training of a corresponding output model generator model configured to determine, from metrology system data specifying a respective metrology system, respective weights of a specific output model for the respective metrology system. In this case, the method may further comprise steps of obtaining candidate metrology system data specifying the second metrology system, and processing, using the output model generator network, the obtained metrology system data for the second metrology system to generate training weights of the specific output model for the second metrology system. Thus, to reduce the measure of the discrepancy between the predicted second training metrology data and the second training metrology data, the weights of the input model generator network, the weights of the output model generator network, one or more parameters defining the metrology system data for the first metrology system and one or more parameters defining the metrology system data for the second metrology system are jointly adjusted (optionally also jointly with weights of the common input model and / or weights of the common output model). Note that this concept of an output model generator network can beconsidered as providing an independent aspect of the invention, in which the specific output network is generated by such a generator network in the way described here, but the specific input model is not necessarily produced by a generator model; for example, it may trained by a process in which the specific input model processes data based on training metrology data generated only by the first metrology system.

[0056] According to a fourth aspect, there is provided a method of using a (trained) neural network and a (trained) input model generator network to determine metrology system data comprising one or more parameters and specifying a first metrology system. The input model generator network is configured to determine, from metrology system data specifying a respective metrology system, respective weights of a specific input model for the respective metrology system, the specific input model being comprised in a neural network for generating predicted metrology data of a patterned portion of a substrate patterned using a lithographic apparatus, the neural network comprising an input network and an output network, the input network comprising a common input model and the specific input model. The neural network is configured to: receive metrology data of the patterned portion generated using the first metrology system; receive configuration data specifying a configuration of the first metrology system used to generate the metrology data; process, using the common input model and the specific input model, the metrology data and the configuration data to generate an encoding of the patterned portion; and process, using the output network of the neural network, the encoding of the patterned portion to generate the predicted metrology data, the predicted metrology data being a prediction of metrology data which would have been generated by inspecting the patterned portion using a second metrology system. The method comprises: obtaining a dataset comprising first metrology data generated using the first metrology system, configuration data specifying a configuration of the first metrology system used to generate the first metrology data, and second metrology data generated using a second metrology system, wherein the first and second training metrology data have been generated by inspecting the same patterned portion of a substrate patterned by the, or a further, lithographic apparatus; obtaining candidate metrology system data specifying the first metrology system; processing, using the input model generator network, the candidate metrology system data to generate weights of the specific input model for the first metrology system; processing, using the common input model and the specific input model in accordance with the generated weights, the first training metrology data and the corresponding configuration data to generate an encoding of the patterned portion; processing, using the output network of the neural network, the encoding of the patterned portion to generate predicted second metrology data, the predicted second metrology data being aprediction of metrology data which would have been generated by inspecting the patterned portion using the second metrology system; and updating one or more parameters of the metrology system data to reduce: a measure of the discrepancy between the predicted second metrology data and the second metrology data.

[0057] In broad terms, the proposed technique enables searching for the best metrology system data such that the mapping between metrology systems achieves a low error. Thus, when a new metrology system becomes available and is to be added to the population of metrology systems, the proposed technique enables learning only the metrology system data of the new tool while the weights of the models (which have already been trained to a sufficient accuracy) are fixed. This is beneficial because it reduces the amount of data that needs to be measured or remeasured. Further, the proposed technique has very low data requirements because the metrology system data is typically low dimensional.

[0058] According to a fifth aspect, there is provided a method of using a (trained) neural network and a (trained) output model generator network to determine metrology system data comprising one or more parameters and specifying a second metrology system. The output model generator network is configured to determine, from metrology system data specifying a respective metrology system, respective weights of a specific input model for the respective metrology system, the specific output model being comprised in a neural network for generating predicted metrology data of a patterned portion of a substrate patterned using a lithographic apparatus, the neural network comprising an input network and an output network, the input network comprising a common input model and a specific input model, the output network comprising a common output model and the specific output model. The neural network is configured to: receive metrology data of the patterned portion generated using a first metrology system; receive configuration data specifying a configuration of the first metrology system used to generate the metrology data; process, using the common input model and the specific input model, the metrology data and the configuration data to generate an encoding of the patterned portion; and process, using the common output model and the specific output model of the neural network, the encoding of the patterned portion to generate the predicted metrology data, the predicted metrology data being a prediction of metrology data which would have been generated by inspecting the patterned portion using the second metrology system. The method comprises: obtaining a dataset comprising first metrology data generated using the first metrology system, configuration data specifying a configuration of the first metrology system used to generate the first metrology data, and second metrology data generated using a second metrology system, wherein the first and second training metrology data have been generated by inspecting the same patterned portion of a substrate patterned by the, or a further, lithographic apparatus;obtaining candidate metrology system data specifying the second metrology system; processing, using the output model generator network, the candidate metrology system data to generate weights of the specific input model for the first metrology system; processing, using the common input model and the specific input model, the first training metrology data and the corresponding configuration data to generate an encoding of the patterned portion; processing, using the common output model and the specific output model in accordance with the generated weights, the encoding of the patterned portion to generate predicted second metrology data, the predicted second metrology data being a prediction of metrology data which would have been generated by inspecting the patterned portion using the second metrology system; and updating one or more parameters of the metrology system data to reduce: a measure of the discrepancy between the predicted second metrology data and the second metrology data.

[0059] According to a sixth aspect, there is provided a computer program product comprising machine -readable instructions which, when the program is executed by a computer, cause the computer to carry out the method of any preceding aspect.

[0060] According to a seventh aspect, there is provided a computer system comprising a processor and a data storage device, the data storage device storing program instructions which, when executed by the processor, cause the processor to carry out the method of any one of the first to sixth aspect.BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Embodiments of the invention will now be described, by way of example only, with reference to the accompanying schematic drawings, in which:Figure 1 depicts a schematic overview of a lithographic apparatus;Figure 2 depicts a schematic overview of a lithographic cell;Figure 3 depicts a schematic representation of holistic lithography, representing a cooperation between three key technologies to optimize semiconductor manufacturing;Figure 4 depicts a schematic of an example neural network which is an embodiment of the invention;Figure 5 depicts a flow diagram of an example process for generating predicted metrology data using the neural network of Figure 4;Figure 6 depicts a schematic of a training system for training the neural network of Figure 4;Figure 7 illustrates the training the neural network of Figure 4;Figures 8 and 9 further illustrate the training of the neural network of Figure 4;Figure 10 depicts a schematic of a system that uses generator models to determine weights of the example neural network of Figure 4;Figure 11 depicts a flow diagram of an example process for determining metrology system data;Figures 12 to 15 depict schematics of further example neural networks which are embodiments of the invention;Figure 16 illustrates changes of the sample caused by a metrology tool;Figure 17 depict schematics of a further example neural network which is an embodiment of the invention, andFigures 18 to 22 depict experimental data.DETAILED DESCRIPTION

[0062] In the present document, the terms “radiation” and “beam” are used to encompass all types of electromagnetic radiation, including ultraviolet radiation (e.g. with a wavelength of 365, 248, 193, 157 or 126 nm) and EUV (extreme ultra-violet radiation, e.g. having a wavelength in the range of about 5-100 nm).

[0063] The term “reticle”, “mask” or “patterning device” as employed in this text may be broadly interpreted as referring to a generic patterning device that can be used to endow an incoming radiation beam with a patterned cross-section, corresponding to a pattern that is to be created in a target portion of the substrate. The term “light valve” can also be used in this context. Besides the classic mask (transmissive or reflective, binary, phase-shifting, hybrid, etc.), examples of other such patterning devices include a programmable mirror array and a programmable LCD array.

[0064] Figure 1 schematically depicts a lithographic apparatus LA. The lithographic apparatus LA includes an illumination system (also referred to as illuminator) IL configured to condition a radiation beam B (e.g., UV radiation, DUV radiation or EUV radiation), a mask support (e.g., a mask table) MT constructed to support a patterning device (e.g., a mask) MA and connected to a first positioner PM configured to accurately position the patterning device MA in accordance with certain parameters, a substrate support (e.g., a wafer table) WT constructed to hold a substrate (e.g., a resist coated wafer) W and connected to a second positioner PW configured to accurately position the substrate support in accordance with certain parameters, and a projection system (e.g., a refractive projection lens system) PS configured to project a pattern imparted to the radiation beam B by patterning device MA onto a target portion C (e.g., comprising one or more dies) of the substrate W.

[0065] In operation, the illumination system IL receives a radiation beam from a radiation source SO, e.g. via a beam delivery system BD. The illumination system IL may include various types of optical components, such as refractive, reflective, magnetic, electromagnetic, electrostatic, and / or other types of optical components, or any combination thereof, for directing, shaping, and / or controlling radiation. The illuminator IL may be used to condition the radiation beam B to have a desired spatial and angular intensity distribution in its cross section at a plane of the patterning device MA.

[0066] The term “projection system” PS used herein should be broadly interpreted as encompassing various types of projection system, including refractive, reflective, catadioptric, anamorphic, magnetic, electromagnetic and / or electrostatic optical systems, or any combination thereof, as appropriate for the exposure radiation being used, and / or for other factors such as the use of an immersion liquid or the use of a vacuum. Any use of the term “projection lens” herein may be considered as synonymous with the more general term “projection system” PS.

[0067] The lithographic apparatus LA may be of a type wherein at least a portion of the substrate may be covered by a liquid having a relatively high refractive index, e.g., water, so as to fill a space between the projection system PS and the substrate W - which is also referred to as immersion lithography. More information on immersion techniques is given in US6952253, which is incorporated herein by reference.

[0068] The lithographic apparatus LA may also be of a type having two or more substrate supports WT (also named “dual stage”). In such “multiple stage” machine, the substrate supports WT may be used in parallel, and / or steps in preparation of a subsequent exposure of the substrate W may be carried out on the substrate W located on one of the substrate support WT while another substrate W on the other substrate support WT is being used for exposing a pattern on the other substrate W.

[0069] In addition to the substrate support WT, the lithographic apparatus LA may comprise a measurement stage. The measurement stage is arranged to hold a sensor and / or a cleaning device. The sensor may be arranged to measure a property of the projection system PS or a property of the radiation beam B. The measurement stage may hold multiple sensors. The cleaning device may be arranged to clean part of the lithographic apparatus, for example a part of the projection system PS or a part of a system that provides the immersion liquid. The measurement stage may move beneath the projection system PS when the substrate support WT is away from the projection system PS.

[0070] In operation, the radiation beam B is incident on the patterning device, e.g. mask, MA which is held on the mask support MT, and is patterned by the pattern (design layout) present on patterning device MA. Having traversed the mask MA, the radiation beam B passes through the projection system PS, which focuses the beam onto a target portion C of the substrate W. With the aid of the second positioner PW and a position measurement system IF, the substrate support WT can be moved accurately, e.g., so as to position different target portions C in the path of the radiation beam B at a focused and aligned position. Similarly, the first positioner PM and possibly another position sensor (which is not explicitly depicted in Figure 1) may be used to accurately position the patterning device MA with respect to the path of the radiation beam B. Patterning device MA and substrate W may be aligned using mask alignment marks Ml, M2 and substrate alignment marks Pl, P2. Although the substrate alignment marks Pl, P2 as illustrated occupy dedicated target portions, they may be located in spaces between target portions. Substrate alignment marks Pl, P2 are known as scribe-lane alignment marks when these are located between the target portions C.

[0071] As shown in Figure 2 the lithographic apparatus LA may form part of a lithographic cell LC, also sometimes referred to as a lithocell or (litho)cluster, which often also includes apparatus to perform pre- and post-exposure processes on a substrate W. Conventionally these include spin coaters SC to deposit resist layers, developers DE to develop exposed resist, chill plates CH and bake plates BK, e.g. for conditioning the temperature of substrates W e.g. for conditioning solvents in the resist layers. A substrate handler, or robot, RO picks up substrates W from input / output ports I / Ol, I / O2, moves them between the different process apparatus and delivers the substrates W to the loading bay LB of the lithographic apparatus LA. The devices in the lithocell, which are often also collectively referred to as the track, are typically under the control of a track control unit TCU that in itself may be controlled by a supervisory control system SCS, which may also control the lithographic apparatus LA, e.g. via lithography control unit LACU.

[0072] In order for the substrates W exposed by the lithographic apparatus LA to be exposed correctly and consistently, it is desirable to inspect substrates to measure properties of patterned structures, such as overlay errors between subsequent layers, line thicknesses, critical dimensions (CD), etc. For this purpose, inspection tools (not shown) may be included in the lithocell LC. If errors are detected, adjustments, for example, may be made to exposures of subsequent substrates or to other processing steps that are to be performed on the substrates W, especially if the inspection is done before other substrates W of the same batch or lot are still to be exposed or processed.

[0073] An inspection apparatus, which may also be referred to as a metrology apparatus, is used to determine properties of the substrates W, and in particular, how properties of different substrates W vary or how properties associated with different layers of the same substrate W vary from layer to layer. The inspection apparatus may alternatively be constructed to identify defects on the substrate W and may, for example, be part of the lithocell LC, or may be integrated into the lithographic apparatus LA, or may even be a stand-alone device. The inspection apparatus may measure the properties on a latent image (image in a resist layer after the exposure), or on a semi-latent image (image in a resist layer after a post-exposure bake step PEB), or on a developed resist image (in which the exposed or unexposed parts of the resist have been removed), or even on an etched image (after a pattern transfer step such as etching).

[0074] Typically the patterning process in a lithographic apparatus LA is one of the most critical steps in the processing which requires high accuracy of dimensioning and placement of structures on the substrate W. To ensure this high accuracy, three systems may be combined in a so called “holistic” control environment as schematically depicted in Fig. 3. One of these systems is the lithographic apparatus LA which is (virtually) connected to a metrology tool MT (a second system) and to a computer system CL (a third system). The key of such “holistic” environment is to optimize the cooperation between these three systems to enhance the overall process window and provide tight control loops to ensure that the patterning performed by the lithographic apparatus LA stays within a process window. The process window defines a range of process parameters (e.g. dose, focus,overlay) within which a specific manufacturing process yields a defined result (e.g. a functional semiconductor device) - typically within which the process parameters in the lithographic process or patterning process are allowed to vary.

[0075] The computer system CL may use (part of) the design layout to be patterned to predict which resolution enhancement techniques to use and to perform computational lithography simulations and calculations to determine which mask layout and lithographic apparatus settings achieve the largest overall process window of the patterning process (depicted in Fig. 3 by the double arrow in the first scale SCI). Typically, the resolution enhancement techniques are arranged to match the patterning possibilities of the lithographic apparatus LA. The computer system CL may also be used to detect where within the process window the lithographic apparatus LA is currently operating (e.g. using input from the metrology tool MT) to predict whether defects may be present due to e.g. sub-optimal processing (depicted in Fig. 3 by the arrow pointing “0” in the second scale SC2).

[0076] The metrology tool MT may provide input to the computer system CL to enable accurate simulations and predictions, and may provide feedback to the lithographic apparatus LA to identify possible drifts, e.g. in a calibration status of the lithographic apparatus LA (depicted in Fig. 3 by the multiple arrows in the third scale SC3).

[0077] It is often desirable to inspect the patterned substrate using a metrology system, e.g. to determine whether the patterned substrate was fabricated correctly, and / or whether the lithographic apparatus is operating correctly or needs adjusting. The term metrology system (or just “tool” for short hereafter) may refer to any suitable instrument that is capable of determining relevant properties of the patterned substrate. Examples of metrology systems include image-based metrology instruments (e.g. micro diffraction based overlay metrology (uDBO) instruments, phase based overlay metrology (pDBO) instruments, optical dark-field microscopes and the like), pupil-based metrology instruments (e.g. scatterometers, soft x-ray metrology instruments), scanning electron microscopes (e.g. single beam or multi-beam SEMs), and the like.

[0078] As noted above, in practice, different instances of the same type of metrology system generate, for the same target, metrology data which differ from each other in one or more aspects. Such “tool-mismatch” can occur for different reasons. As a first example, tool-mismatch between single-beam SEMs may be caused by tool-to-tool variations of the scanning beam angle, the contrast, the noise level, and so forth. As a second example, in a multi-beam SEM system (which can have an array of 9 to 2800 electron beams), images produced by different electron beams may vary because the design of the system may inherently result in variations between the electron beams (i.e. in a multi-beam SEM, each electron beam may be considered to be an individual metrology system). For instance, electron beam aberrations may be stronger for electron beams located away from the centre of the beam array than for those near the centre. Hardware -based solutions for reducing electron beam variations (e.g. by tight hardware specifications or by modifications to specific hardware elements) are often undesirable (e.g. costly, problematic to adapt when further beam-to-beam variations occurdue to aging of the system, and so forth). As a third example, tool-mismatch between image-based metrology instruments (e.g. uDBO, or continuous bias DBO (cDBO)) used for inspecting scribe lane targets (i.e. targets positioned in areas between semiconductor dies) may be caused by tool-induced asymmetries, biases, image distortions and inter / intra pad cross-talk. As a fourth example, toolmismatch between pupil-based metrology instruments (e.g. scatterometers using soft x-ray and / or visible to near-IR light) may be caused by tool-dependent optical aberrations that corrupt the generated signal and can introduce tool-induced biases.

[0079] Figure 4 shows a neural network 400 which, once trained, can process input metrology data of a target generated using a first metrology system (denoted tool k, where “k” is an index that identifies the first metrology system from a plurality of N metrology systems of the same type (i.e. k 6 1 ... N)) to generate a prediction of metrology data which would have been generated by inspecting the same target using a second metrology system (denoted tool j, j E l ... N). Thus, using the neural network 400, metrology data acquired using tool k can effectively be “mapped” to metrology data acquired using the tool j (thus, the neural network 400 may be considered to implement a “tool-to-tool” mapping, e.g. a tool k to tool j mapping). The mapped metrology data can then be compared to other metrology data generated using tool j (e.g. metrology data of other targets) without the aforementioned tool-mismatch issues.

[0080] More specifically, the neural network 400 of Figure 4 is configured to receive, as network input, metrology data 401, configuration data 402, and configuration data 403. The metrology data 401 has been generated by inspecting a target generated using tool k. In the embodiment of Figure 4, the inspection target is a patterned portion of a substrate (e.g. a portion of a semiconductor wafer) patterned using a lithographic apparatus (e.g. the above described lithographic apparatus). Tool k is an instance of a type of a metrology system that is suitable for inspecting the patterned portion (e.g. using image -based metrology, pupil-based metrology, scanning electron microscopy, or the like). The metrology data comprises an array of numerical values specifying aspects of the inspected target. As one example, the metrology data 401 of Figure 1 has been generated using image-based metrology, and the metrology data 401 comprises image data. More specifically, the metrology data 401 that is provided to the neural network 400 may comprise a two-dimensional array of intensity values corresponding to the (raw or pre-processed) image data obtained by an image sensor of tool k. As another example, in implementations where tool k uses pupil-based metrology, tool k may generate metrology data comprising an array of numerical values specifying a pupil representation defining a distribution of radiation in a pupil plane of the tool.

[0081] The input configuration data 402 include “immutable” data defining a type of the first metrology system (e.g. data defining a manufacturer of the first metrology system, and / or model type), and / or setting data to specify a configuration of tool k during the acquisition of the metrology data 401. The input configuration data 402 may specify the configuration of tool k in any suitable way (the specified configuration of tool k may be an estimated configuration of tool k). As one example,the input configuration data 402 may comprise setting data specifying operating parameters of the tool k (e.g. exposure time, scanning speed, a spectral density of light used to inspect the target, and the like). The operating parameters may comprise parameters set by a user of the tool before using the tool. As another example, the input configuration data 402 may be an estimate of the configuration of tool k (as discussed below, in some examples of the use of the present system, there may be at least one step of improving this estimate based on the output of the neural network 400, e.g. iteratively). As another example, the input configuration data 402 may comprise reference metrology data to define the configuration of tool k indirectly. The reference metrology data may be metrology data of a reference target generated using the tool k configured in the same configuration as used to generate the metrology data 401. As described below, in this case, the neural network 400 may also be configured to determine the operating parameters of the tool k from the reference metrology data.

[0082] The configuration data 403 specifies a configuration of tool j in a similar way in which the configuration data 402 specifies the configuration of tool k.

[0083] The neural network 400 is configured to process the metrology data 401, the configuration data 402, and the configuration data 403 to generate predicted metrology data 404. The predicted metrology data 404 is a prediction of metrology data which would have been generated by inspecting the target (i.e. the same target for which the input metrology data 401 have been generated) using tool j configured according to the configuration specified in the configuration data 403. In general, tool j may be a different instance of the metrology system (e.g. a different apparatus or a different electron beam in case of a multi -beam SEM) or tool j may be the same instance as tool k but, for example, in a different configuration (i.e. the configurations specified in the configuration data 402, 403 are different).

[0084] To achieve the aforementioned tool-to-tool mapping, a neural network comprising an encoder-decoder architecture may be used (i.e. a variation of the known autoencoder architecture). Such an encoder-decoder architecture may comprise an encoder for generating data in a latent space and a decoder. The encoder is configured to compress input data into a (typically lower dimensional) latent space representation, and then the decoder (instead of being configured to reconstruct the original input) may be configured to predict a modified version of the original input from the encoded data (i.e. from the latent space representation). Thus, a neural network for tool-to-tool mapping may comprise an encoder-decoder architecture. In this case, the encoder-decoder architecture may be used to encode the input metrology data and the input configuration data into a latent space representation and generate the desired predicted metrology data from the latent representation and the configuration data. The inventors have realized that a problem with this network architecture is that the training of such a network is resource-inefficient, in particular when multiple networks are to be trained, e.g. to enable mapping in both directions (e.g. one network for mapping tool k to tool j, and one network for mapping tool j to tool k), or when the “fleet” of metrology systems is large (N>2). For example, to implement a specific tool mapping (say tool j to tool k) the entire network needs to be trained withoutbenefiting from a previously trained network that implements a different mapping within the fleet (say tool k to tool j).

[0085] The inventors realized that it is advantageous to employ a network architecture in which the encoding is split into multiple stages which include a stage that is implemented by a (large) model that is “common” (i.e. a model that can be (re-)used for any mapping within the plurality of N metrological systems) and a stage that is implemented by a (small) model that is “tool- specific” (i.e. a model that can be used for any mapping that maps the tool associated with the tool-specific model to another of the N metrological systems). Here “small” may refer to the number of trained numerical parameters, e.g. the tool-specific model has fewer trained numerical parameters than the common model. Similarly, the decoding stage may also be split into “common” and “tool-specific” models. As described in more detail further below, the proposed architecture reduces the resources needed to train networks that implement new mappings within the N metrological systems, and makes it possible to the use any combination of data pairs, coming from two metrological systems, thereby reducing the number of measurements needed.

[0086] The neural network 400 of Figure 4 is an example of a neural network that implements the above described concept of using common and tool-specific models. More specifically, neural network 400 comprises a common input model 405, a tool re-mapping section 406 comprising a toolspecific input model 407 and tool-specific output model 408, and common output model 409. The common input model 405 is configured to process the metrology data 401 to generate an encoding of the metrology data (the generated encoding is typically lower dimensional than the metrology data 401). The common input model 405 is defined by a plurality of learnable network parameters (denoted “£”’) which values have been obtained by a training process in which the common input model 405 is presented with metrology data from multiple (e.g. all) of the N tools (i.e. not only data from tool k), as further described below with reference to Figures 6 to 8. In broad terms, because the common input model 405 has been trained using data generated from different tools, the common input model 405 has learned to disentangle effects that are common to the tools and effects that are tool-specific. The tool-specific input model 407 is configured to process the encoding generated by the common input model 405 and the input configuration data 402 to generate an encoding of the patterned portion. The tool-specific input model 407 is specific to tool k, i.e. the tool-specific input model 407 is defined by a plurality of adjustable network parameters (denoted “0^”). As described below in more detail, the values of the adjustable network parameters 6^ can be obtained in different ways, e.g. by a training process in which a “tool k to tool i” mapping is implemented (where “tool i” denotes any one of the N tools) or from the output of a generator model trained to generate suitable weights 6 . In broad terms, the tool-specific input model 407 generates the encoding of the patterned portion so that information specifying aspects of the metrology data 401 which are specific to tool k isremoved. For example, the tool-specific input model 407 may generate the encoding of the patterned portion so as to remove information specifying aberrations of tool k.

[0087] Note that in principle the tool-specific input model 407 might be trained in a training process which starts from an initial state (a starting realization of its trainable parameters) which was obtained by a training process in which, during the training, the input model encoded data from tools other than k. However, the final version of its numerical parameters was obtained in a training process in which, in multiple successive learning iterations, it only encoded training data obtained from tool k.

[0088] The tool-specific output model 408 is configured to process the encoding generated by the tool-specific output model 407 and the configuration data 403 to generate an encoding of the predicted metrology data. The tool-specific output model 408 is specific to tool j, i.e. the tool-specific output model 408 is defined by a plurality of learnable network parameters (denoted “0^”). Similar to the weights 0^of the tool- specific input model 407, the values of the adjustable network parameters Qj' can be obtained in different ways, e.g. by a training a process in which a “tool i to tool j” mapping is implemented (i.e. any tool to tool j) or from the output of a corresponding generator model trained to generate suitable weights 0^ . In other words, in some cases, the final version of the numerical parameters of the tool-specific output model 408 was obtained in a training process in which (irrespective of the initial state of the tool-specific output model), in multiple successive learning iterations, it only predicted training data obtained from tool j. In broad terms, the tool-specific output model 408 generates the encoding of the predicted metrology data 404 so as to add information specifying aspects which are specific to tool j. For example, the tool- specific output model 408 may generate the encoding of the predicted metrology data so as to incorporate information specifying aberrations of tool j.

[0089] The common output model 409 is configured to process the encoding of the predicted metrology data to generate the predicted metrology data 404. Similar to the common input model 405, the common output model 409 is defined by a plurality of learnable network parameters (denoted “D”) which values have been obtained by a training process in which the common output model 409 generates predicted metrology data for multiple (e.g. all) of the N tools (i.e. not only for tool j), as further described below with reference to Figures 6 to 8.

[0090] The common input model 405, the tool-specific input model 406, the tool-specific output model 407 and the common output model 409 can have any suitable architecture and can include, e.g., one or more convolutional neural network layers. In some embodiments, the neural network 400 may be implement as U-Net architecture.

[0091] The common input model 405 may be a larger model than the tool-specific input model 406 (e.g. the number of learnable parameters of the common input model E may be larger (e.g. at least 10 times larger) than the number of learnable parameters of the tool-specific input model 0^). Similarly, the common output model 409 may be a larger model than the tool-specific output model407 (e.g. the number of learnable parameters of the common input model D may be larger (e.g. at least 10 times larger) than the number of learnable parameters of the tool-specific input model 0^).

[0092] While the neural network 400 has been described as comprising a common input model for processing the input metrology data and a tool-specific input model for processing the output of the common input model, it is to be understood that in other embodiments, the order of common and tool-specific models may be reversed, i.e. the tool-specific input model may be configured to process the input metrology data, and the common input model may be configured to process the output of the tool-specific input model. Similarly, in some embodiments, the order of the tool-specific output model and the common output model may be reversed, i.e. the common output model may process the encoding of the pattern portion and the tool-specific output model may process the output of the common output model.

[0093] While the neural network 400 has been described as receiving configuration data 402, in other embodiments, the neural network 400 may not directly receive configuration data 402 but may implicitly or explicitly infer the configuration data from the metrology data 401. In this case, the metrology data 401 may comprise data of the same target under multiple conditions (e.g. wafer rotation mode). In this case, the neural network may be configured to process the multiple measurements to implement a multi-channel to multi-channel mapping. Alternatively, the neural network may be configured to process each of the multi-condition acquisitions separately and determine the configuration data from the acquisitions.

[0094] An example process for generating predicted metrology data using the (trained) neural network 400 of Figure 4 is described with reference to Figure 5. In broad terms, a typical scenario in which the process of Figure 5 is applied may be that a substrate is patterned by a lithographic apparatus as part of a manufacturing process, and, once the patterning is completed, the substrate is transferred so that tool k and tool j can inspect (in parallel or sequentially) different (possibly overlapping) portions of the newly patterned substrate. The process of Figure 5 is then performed to achieve a “tool k to tool j” mapping of the metrology data generated by tool k. The mapped metrology data (and optionally and the metrology data generated by tool j) can then be further processed using known algorithm without being affected by tool-mismatch. For example, a person wishing to diagnose a fault in a substrate inspected using tool j, can map metrology data collected using tool j to data generated tool k, and then analyse the result using diagnostic algorithms developed for, e.g. calibrated with reference to, tool k.

[0095] In an initial step S501, metrology data 401 of the patterned portion generated using tool k is received. In step S502, configuration data 402 specifying the configuration of tool k during acquisition of the metrology data 401 and configuration data 403 specifying a configuration of tool j are received. In step S403, the common input model 405 processes the metrology data 401 to generate an encoding of the metrology data. Then, the tool-specific input model 407 processes the encoding ofthe metrology data to generate an encoding (or “latent representation”) of the patterned portion. In step S504, the encoding of the patterned portion is processed to generate the predicted metrology data 404 (so that the predicted metrology data 404 is a prediction of metrology data which would have been generated by inspecting of the patterned portion using tool j configured according to the configuration specified in the configuration data 403). More specifically, the tool-specific output model 408 processes the encoding of the patterned portion and the configuration data 403 of tool j to generate an encoding of the predicted metrology data. Then, the common output model 409 processes the encoding of the predicted metrology data to generate the predicted metrology data 404.

[0096] A training system for training the neural network 400 is now described with reference to Figure 6. The training system comprises training data 601 comprising a plurality of training items 602. The training data 601 is used together with a training engine 603 to iteratively update / optimise the values of a plurality of parameters 604 defining the neural network 400.

[0097] Each training item may be associated with a pair of tools and comprise corresponding training metrology and configuration data. For example, training items associated with a tools k and j comprise training metrology data (denoted mfe) generated using tool k, configuration data of tool k, training metrology data (denoted m7) generated using tool j, and configuration data of tool j. The training metrology data belonging to the same training item have been generated by inspecting the same target (e.g. patterned portion).

[0098] The plurality of training items may be used to train, over a plurality of iterations, the common input model, the common output model, as well as tool-specific input and output models for each of the N tools. In each iteration, an update for the parameters 604 is generated by processing a training item. Figure 7 illustrates training iterations in which training items associated with different tools are processed. Iteration 701 illustrates the case for a training item associated with tools k and j. The training metrology data generated using tool k, the configuration data of tool k, and the configuration data of tool j are processed using the neural network 400 (as described above with reference to Figures 4 and 5) to generate predicted training metrology data. This may be expressed aswhere M(.. .) denotes the neural network 400, Pj denotes the predicted training metrology data (the configuration data are omitted for clarity). The training metrology data generated using tool j and the predicted training metrology data are then processed by the training engine 603 to generate an update for the parameters 604. This generally involves b ackpropagating gradients of an objective function to update the learnable parameters E, D, 6k, 6^ using any appropriate gradient descent optimization algorithm, e.g. Adam or another optimization algorithm. Generally, the parameters E, D, 6k,maybe updated to reduce a discrepancy (or mismatch) between the input training metrology data mkand the predicted training metrology data pj .

[0099] As illustrated in iteration 702, the same training item as used in iteration 701 can be reused to further train the neural network by reversing the role of the metrology data mkand m.j. In this case, a tool-specific input model for tool j (defined by parameters 0?) is used to process the output of the common input model, and a tool-specific output model for tool k (defined by parameters 9k) is used to process the process the encoding of the patterned portion. Thus, in iteration 702, training metrology data generated using tool j, configuration data of tool j, and configuration data of tool k are processed to generate the predicted training metrology data for the respective iteration.Similar to iteration 701, the training engine may update the learnable parameters E, D, 9? , 9kbased on the predicted training metrology data and the metrology data generated using tool k.[000100] Iteration 703 illustrates an iteration where a training item associated with tool 1 and tool j is processed using a corresponding tool-specific input model for tool 1 (defined by parameters 0 ) and the tool-specific output model for tool j. Iteration 704 illustrates an iteration where a training item associated with tool h and tool 1 is processed using a corresponding tool-specific input model for tool h (defined by parameters 9k) and the tool-specific output model for tool 1 (defined by parameters 0;d). Iteration 705 illustrates an iteration in which the training item is only associated with one tool (here tool g). In this case, the tool-specific input model and the tool-specific output model are both specific to tool g, and the neural network attempts to reconstruct the input metrology data mg.[000101] After performing the plurality of training iterations, the training system may update the parameters E, D, 9f , 9^ (for all i, j G 1 .... IV) to optimize a cost function, argwhere C(... ) is a data distance term (i.e. C(Mis a cost term that penalizes any difference between m7and the prediction M.[000102] In some implementations, a variational inference scheme may be employed, i.e. parameters 0(eand 9^ may be sampled from a distribution regularized using Kullback-Leibler (KL) divergence priors. In this case, the distribution modes (e.g. mean and valiance for each 0) may be modeled with a value resulting from drawing a sample from the distribution. To this end, a regularization term r(-) may be included in the cost function, e.g. in some implementations, the cost function may read arg[000103] The illustrated training of the proposed neural network is resource-efficient due the novel architecture of network. As one example, the architecture of the proposed neural network enables using training data from the entire population of tools. As another example, an instance of the neural network 400 may be trained to implement a mapping within the N tools, say tool k to tool h, without having access to corresponding training items (i.e. without requiring tool k and tool h to measure the same target).[000104] Another advantage of the proposed neural network is that new tools can be added by simply training the corresponding specific input and out models. As mentioned above, the specific input and out models may be smaller than the common input and output models, and may be trained using only a few training items. Further, new tools may even be added without requiring to measure “old” targets (i.e. target which have been measured by one or more of the N tools) with the new tool. This may be achieved by performing a “two-pass” process, as illustrated in Figure 8. In the training iteration 801 of Figure 8, a training item associated with tool k and tool j is used to train the specific input and output models for the new tool m. To this end, metrology data generated by tool k is processed to generate predicted metrology data for tool m. The predicted metrology data for tool m is then processed to generate predicted metrology data for tool j which can be compared to metrology data generated by tool j.[000105] The “two-pass” training process illustrated in Figure 8 may also be used to add simulated tools to the plurality of tools, i.e. the process of Figure 8 may be used to train corresponding toolspecific input and output models for the simulated tool. A simulated tool may refer to a computational model (i.e. and not to a real-world device) that generates simulated metrology data. A simulated tool may by comprise physics-based models (e.g. sets of differential equations modelling the imaging process). Adding simulated tools enables transferring domain knowledge (e.g. knowledge of the physics underlying the inspection process) into models to the neural network. In some cases, a simulated tool may represent an “ideal” tool, e.g. a metrology system without aberrations.[000106] In addition or alternative to the two-pass” training process illustrated in Figure 8, simulated tools may be trained using training items associated with pairs of simulated tools (i.e. training items comprising pairs of simulated metrology data (and respective configuration data). For example, Figure 9 illustrates the case where a training item associated with simulated tool k and simulated tool j is processed using a corresponding tool-specific input model for simulated tool k (defined by parameters 9^) and the tool-specific output model for simulated tool j (defined by parameters f ). In this case, the training system may generate parameter updates of a cost function that reads argAdvantageously, both the process of Figure 8 and the process of Figure 9 avoid a need to compare simulated metrology data directly with measured metrology data.[000107] With reference to Figure 10, a variation of the system of Figure 4 will now be described. In addition to the above-described neural network 400, the system of Figure 10 includes an input model generator 1010 and an output model generator 1011. In the system of Figure 10, the weights E, D of the common input model 405 and the common output model 409 have been trained (e.g. as described above with reference to Figures 6 to 9, or jointly with the generators 1010, 1011 as described below). In contrast, the weights 6^, 6^ of the specific models 407, 408 have not been obtained via the training process described above with reference to Figures 6 to 9. Instead, the weights are respectively generated by the input model generator 1010 and the output model generator1011.[000108] The input model generator 1010 is a learnable model (e.g. a neural network) that is configured to receive, as input, tool data (“metrology system data”) specifying a respective tool and to determine, as output, weights of the specific input model for the respective tool. Thus, the input model generator 1010 can generate weights of the specific input model for any metrology system if the respective tool data is available. Similarly, the output model generator 1011 is a learnable model (e.g. a neural network) that is configured to receive, as input, tool data specifying a respective tool and to determine, as output, weights of the specific output model for the respective tool.[000109] The tool data is defined by a set of parameters r|r£(the index indicates the respective tool, denotes the tool data for tool i). The parameters r|r£may include parameters for which the values have been determined via a training process (i.e. based processing training data). In some implementations, the parameters r|r£include both learnt and non-learnt parameters. The non-learnt parameters may specify known and / or measured tool settings or configurations. In some cases, the tool data r|r£may include the above-described configuration data 402, 403. In these cases, it is not necessary to provide the respective configuration data to the tool-specific model as separate input 402, 403.[000110] One or more of the learnt parameters of the tool data may be implemented via a variation scheme that is similar to a variational autoencoder, i.e. where a probability distribution of the configuration is learnt (e.g. in terms of the mean and variance) and sampled from, and priors are added over the distribution (e.g. via a KL loss term during model training). In these implementations, the tool data comprises one or more parameters which define a distribution of a set of latent variables, and the model generator 1010, 1011 is configured to determine the weights of the tool-specific model by sampling from the distribution defined by the input tool data to determine values of the set of latent variables, and determining the weights of the tool-specific model based on the set of latent variables.[000111] Further, in the embodiment of Figure 10, the input model generator 1010 and the output model generator 1011 may employ the same parametrization of the tool data, i.e. the same tool data I|J(for tool i may be used as input to the input model generator 1010 or as input to the output model generator 1011 (depending on the desired tool mapping). However, it is to be understood that this is not necessarily the case and that in other implementations, the input model generator 1010 and the output model generator 1011 may employ different tool parametrizations (i.e. in this case, the input to the input model generator 1010 and the output model generator 1011 are different for the same tool i). [000112] The system of Figure 10 may be used to implement a tool k to took j mapping. To this end, tool data r|rfeis provided as input to the input model generator 1010 and tool data i]ryis provided as input to the output model generator 1011. Then, the input model generator 1010 processes the tool data to generate suitable values 9 (the so-generated values for 9 are generally similar to the ones obtained via the training process described above with reference to Figures 6 to 9). The generated parameters 9 are used to implement the tool specific input model 407 for tool k. Similarly, the output model generator 1011 processes the tool data r|r7to generate suitable values 9^. The generated parameters 9^ are used to implement the tool specific output model 408 for tool j . The neural network 400 is then ready to perform the tool k to tool j mapping as described above with reference to Figure 4, i.e. the neural network 400 receives, as input, tool k metrology data 401 and generates predicted tool j metrology data 404.[000113] It is to be understood that when the generators 1010, 1011 of Figure 10 are used to determine the weights of the tool specific models, the architecture (e.g. number of parameters, number of layers, and the like) of the tool specific input models is the same across the “fleet” of metrology systems (and similarly, the architecture of the tool specific output models is the same).[000114] The generators 1010, 1011 are jointly trained with the learnable parameters of the tool data. For the training, a training dataset is used. The training dataset comprises a plurality of data items. Each training item may be associated with a pair of tools and comprise corresponding training metrology and configuration data, as described above with reference to Figure 6. The training comprises a plurality of training steps to iteratively adjust the parameters of the generators 1010, 1011 and the learnable parameters of the tool data. In each training step, a training item is processed by the system of Figure 10. For example, for a training item associated with tools k and j, a training step includes an initial step of obtaining candidate tool dataand IJJ / (an initial guess forand IJJ / may be based on random values). Then, the generators 1010, 1011 process the respective candidate tool data and IJJ / to generate candidate weights 9^ and 9f for the corresponding tool-specific input and output models. The training metrology data for tool k is then processed by the neural network to generate predicted tool j metrology data. The weights of the generators 1010 and 1011 are then updated to reduce a measure of the discrepancy between the predicted tool j metrology data and the tool j training metrology data of the training item. In practice, the training is repeated for plurality oftraining items which are associated with different pairs of tools. The common input model 405 and the common output model 409 may be trained jointly with the generators 1011, 1011 and the tool data. [000115] A benefit of using the generators 1010, 1011 is that a new tool can be added to the population of tools with a reduced amount of data that needs to be measured or remeasured. This is because, only the relevant tool data needs to be determined which is typically low dimensional.Finding suitable tool data for the new tool can be considered as an “inverse problem”, i.e. the tool data can be found by adjusting candidate tool data such that the mapping between pair of tools (which include the new tool) achieves a low error. The other (already sufficiently trained) model parts are fixed during this process.[000116] An example process of obtaining tool data i|»mfor a new tool m will now be described with reference to Figure 11. At step S 1101 , a training item is obtained. The training item is associated with a pair of tools including the new tool m and comprises corresponding training metrology and configuration data for both tools. Assuming that the training item, is associated with tools m and k, a training step implementing a tool m to tool k mapping may be performed. To this end, candidate tool data i|»m(e.g. randomly initialized) and previously determined tool data r|rfefor tool k are obtained. At step SI 102, the generators 1010, 1011 respectively process tool data i|»mand r|rfeto generate weights for the tool-specific models. At step SI 103, the training metrology data generated by the new tool m is processed by the neural network to generate predicted metrology data associated with tool k. The learnable parameters of i|»mare then updated based on a discrepancy between the training metrology data of tool k and the predicted metrology data. The training may be repeated for plurality of training items. It is to be understood that the roles of tools k and m can be switched, i.e. the tool data i|»mfor the new tool m can also be determined by performing a tool k to tool m mapping instead of the above-described tool m to tool k mapping.[000117] With reference to Figure 12, a variation of the proposed neural network will now be described. The neural network 1200 is generally similar to the neural network 400 of Figure 4, in particular in that the neural network 1200 is configured to process the metrology data 401, the configuration data 402, and the configuration data 403 to generate predicted metrology data 404. As before, the predicted metrology data 404 is a prediction of metrology data which would have been generated by inspecting the target (i.e. the same target for which the input metrology data 401 have been generated) using tool j configured according to the configuration specified in the configuration data 403. The neural network 1200 comprises a common input model 1205, a tool-specific input model 1207, a tool-specific output model 1208, a common output model 1209 , and a combiner module 1210. The common input model 1205 is configured to process the metrology data 401 to generate an encoding of the metrology data. The tool-specific input model 1207 is configured to process the encoding of the metrology data and the configuration data 402 to generate first encoded data. The tool-specific output model 1208 is configured to process the encoding of the metrology data and the configuration data 403 to generate second encoded data. The combiner module 1210 isconfigured to combine the first and second encoded data to generate the encoding of the patterned portion. Many ways of combining the first and second encoded data to generate the encoding of the patterned portion exist. For example, the combiner module 1210 may be configured to combine the first and second encoded data by concatenating the first and second encoded data. Similar to the common output model 409 of Figure 4, the common output model 1209 is configured to process the encoding of the patterned portion to generate the predicted metrology data.[000118] Optionally, the neural network 1200 may comprise one or more skip connections. More specifically, the neural network 1200 may comprise one or more skip connections from the common input model 1205 to the common output model 1209 (i.e. the output of the common input model 1205 is fed into the common output model 1209).[000119] With reference to Figure 13, a further variation of the proposed neural network will now be described. The neural network 1300 is generally similar to the neural network 400 of Figure 4, in particular in that the neural network 1300 is configured to process the metrology data 401, the configuration data 402, and the configuration data 403 to generate predicted metrology data 404. As before, the predicted metrology data 404 is a prediction of metrology data which would have been generated by inspecting the target (i.e. the same target for which the input metrology data 401 have been generated) using tool j configured according to the configuration specified in the configuration data 403. In addition to a “main processing branch” formed by the common input model 405, the tool re-mapping section 406 and the common output model 409, the neural network 1300 also comprises N “supplementary processing branches” 1301 I...N (with N > 1, e.g. N = 1, 2, ...). The supplementary processing branches speed up the training and improve the performance of the neural network. [000120] The first supplementary processing branch 13011 is configured to receive, as input, the output of the common input model 405, process the input, and provide data to the common output model 409. Each of the subsequent supplementary processing branches 13012 . N is configured to receive, as input, data from the preceding supplementary processing branch, process the input, and provide data to the preceding supplementary processing branch. The supplementary processing branches 1301 I...N may include tool-specific blocks / models and / or common (i.e. tool-unspecific) blocks / models. For example, in the embodiment of Figure 13, the first supplementary processing branch 13011 has a similar architecture as the main branch, i.e. in this example, the first supplementary processing branch 13011 includes a first supplementary common input model 13051 configured to receive the output of the common input model 405, a first supplementary tool remapping section 13061 comprising a first supplementary specific input model 13071 (specific to tool k) and a first supplementary specific output model 13081 (specific to tool j), and a first supplementary common output model 13091 configured to provide data to the common output model 409. The subsequent supplementary branches 13012...N may include respective tool-specific blocks / models and / or common blocks / models. The supplementary tool-specific models may have the same or a different architecture as the corresponding tool-specific model in the main branch. Further, theweights of the supplementary tool-specific models may be obtained via training (i.e. analogous to the process described above with reference to Figures 6 to 10), or via corresponding generators (i.e. analogous to the process described above with reference to Figures 10 and 11). Some implementations may include at least one supplementary processing branch that does not include any tool-specific blocks / models but only common blocks / models.[000121] A further variation of the proposed neural network will now be described with reference to Figure 14. The neural network 1400 of Figure 14 has a simpler architecture than the neural network 400 and may be used in cases where all tools are to be mapped to the same reference tool (for example, when the N tools are N beams of a multi-beam SEM, mapping all beams to one reference beam may be desirable). The “reference tool” may be a real- world tool, a simulated tool or an “average” tool of the population of the N tools. In implementations in which the reference tool is the “average” tool of the population, the neural network may be implemented using variational models and a realistic “average” tool may be the determined from most likely sample out of the distribution used as priors for the variational sampling.[000122] Since all tools are to be mapped to the same reference tool, the tool-specific output model is not needed (or can be incorporated into the common output model). More specifically, the neural network 1400 of Figure 14 is generally similar to the neural network 400 of Figure 4, except that the neural network 1400 does not include the tool-specific output model 408 and does not receive the configuration data 403. Similar to the neural network 400, the metrology data 401 and the configuration data 402 are processed by the common input model 405 and the tool-specific input model 407. The common output model 1409 of the neural network 1400 is configured processes the output of the tool-specific input model 407 to generate the predicted metrology data 1404. The predicted metrology data 1404 is a prediction of metrology data which would have been generated by inspecting the target (i.e. the same target for which the input metrology data 401 have been generated) using the reference tool configured according to a “fixed” configuration.[000123] A further variation of the proposed neural network will now be described with reference to Figure 15. The neural network 1500 is identical to the neural network 400 of Figure 4, except that the neural network 1500 further comprises a trained application-specific model 1501. The applicationspecific model 1501 is specific to an application type (i.e. a specification of a type of the patterned substrate used to produce the tool k data 401, e.g. a particular uDBO target design), from a set of application types (e.g. a set of uDBO target designs). The application specific model 1201 assists the neural network 1500 in generating predicted tool j data 404 for this application. Note that the purpose of the application specific model 1501 is not to translate between different applications; rather, the input tool k data 401 and the predicted tool j data 404 both relate to the same application (patterned substrate type). In particular, the application-specific model 1501 is specific to an application type associated with the patterned portion of the input metrology data 401 which means that the application-specific model 1501 has been trained only on training metrology data of thecorresponding application type (i.e. patterns substrates produced according to the application). In embodiments, the neural network may use a training dataset comprising metrology data associated with different applications to train a plurality of application-specific models (each specific to a different application) similar to the above described training of the plurality of specific-input / output models. The use of application-specific models enables the neural network 1500 to be easily trainable across multiple applications (e.g. across multiple uDBO target designs).[000124] More specifically, the application model 1501 is configured to process data from the toolspecific input model based on an application type, from a set of application types, associated with the patterned portion, and to provide the processed data to the tool-specific output model. The toolspecific output model processes the output data of the tool-specific input model and the output data of the application model 1501. The output of the tool-specific output model is processed by the common output model as described above with reference to Figure 4. Thus, because the application model 1501 takes application specific effects into account, the architecture of neural network 1500 removes the necessity to collect training data for a particular application (i.e. a particular target design) from multiple tools.[000125] In some cases, the act of inspecting a sample (e.g. a patterned portion of a substrate) with a metrology system can modify the sample because of the interaction of the metrology system with the sample during the inspection. The induced change of the sample depends on the details of the metrological process (e.g. the wavelength, intensity and exposure time of the employed light or electron beam). Accordingly, when the same sample is repeatedly inspected, each inspection changes the sample and generates different metrology data. For example, inspecting a resist pattern with a SEM tool leads to a degradation of the resist and so to a shrinking of critical dimensions of the printed pattern. This is illustrated in Figure 16 which shows a resist pattern 1600 shrinking over the course of ‘n’ SEM measurements.[000126] It may be valuable, e.g. for additional accuracy, to take the tool-induced sample changes into account. For example, the above-described training process of the proposed neural network requires that the same target is measured multiple times because the process requires metrology data generated by different metrology systems for the same target. In this case, keeping track of the performed measurements and taking them into account improves the accuracy of the trained models. [000127] To this end, a further variation of the proposed neural network will now be described with reference to Figure 17. The neural network 1700 is identical to the neural network 400 of Figure 4, except that the neural network 1700 further comprises a transfer model 1701. The purpose of the transfer model 1701 is to take into account the effects of metrological processes previously performed on the sample, i.e. the transfer model 1701 exploits that the tool-induced changes are normally predictable given the details of the metrological process.[000128] In addition to the metrology data 401, the neural network also receives as input the inspection data comprising a first index value nkindicating a number of inspections which had beencarried out on the patterned portion prior to the generation of the metrology data 401, and a second index value n7. The transfer model 1701 is configured to receives as input the encoding of the patterned portion generated by the tool specific input model 407, and the inspection data. The transfer model 1701 processes the input to generate data which is then provided to the tool specific output model 408 so that the predicted metrology data 404 generated by the common output model 409 is a prediction of metrology data which would have been generated by inspecting the patterned portion using tool j if the patterned portion had been subject to a number of prior inspections indicated by the second index value n7.[000129] The inspection data can include further details of the previously performed inspections. For example, the inspection data may further comprise data specifying a time duration associated with the previous inspections, and / or an energy deposited on the patterned portion during the previous inspections (e.g. a total amount of energy accumulatively deposited by all previous measurements). [000130] The transfer model 1701 is defined by a plurality of parameters. Suitable values of these parameters can be obtained by training (e.g. the parameters may be adjusted during a training process which is analogous to the above-described training process for the neural network 400) or can be obtained / derived based on the details of the metrological process (e.g. based on the physics of the metrological process).[000131] With reference to Figures 18 to 22, experimental data obtained by using a first and a second example implementation of the neural network 1200 of Figure 12 will now be described. Both example implementations of the neural network 1200 are implemented having a U-Net architecture with skip connections between the common input model and the common output model. Both example implementations are fully convolutional with the tool-specific input and output model being also formed by convolutional layers.[000132] In general, the proposed neural network includes at least one tool-specific model. While the above-described embodiments of the proposed neural network comprise a tool-specific input model and, in some cases, also a tool-specific output model, it is to be understood that also other embodiments are envisaged, e.g. embodiments in which the neural network includes a tool-specific output model but does not include a tool-specific input model. Thus, disclosure in this document describing a system comprising a tool specific input model, should also be understood as disclosure of a corresponding system corresponding tool specific output model (not necessarily in combination with a tool specific input model).[000133] Figures 18 to 20 illustrate aberration correction for a multi-beam SEM having an array of 5x5 electron beams which is used to inspect a portion of a substrate that includes a repetitive pattern of a contact array. Figure 18 shows an array of 5x5 panels which show synthetic (i.e. simulated) images “generated” by the multi-beam SEM. As indicated by the dashed lines in Figure 18, nearby beams overlap in terms of measured area. For example, the reference numerals 1800 and 1801 indicate an overlap area between two beams. The overlap areas can be used as to form training items.[000134] Figure 19 illustrates the generation of the synthetic SEM images of Figure 18. Panel 1900 shows the “ideal” pattern of a contact array where, to simulate defects, CD variability and a probability of missing contacts have been included. Then aberrations are added by modelling each beam as a Gaussian convolution with a different covariance matrix (panel 1901). Finally, the noise is added and the images are down-sampled (panel 1902). In this example application, the aim is to correct beam aberration towards the properties of a single beam, namely the center beam (because the center beam is considered to have the best properties).[000135] The first example implementation of the neural network 1200 has 5 convolution layers with 16 filters of size 5x5 for both of the common input model and the common input model. Two skip connections corresponding to the second and forth layer outputs are implemented. Before the skip connections a stride of 2 is used, and otherwise a stride of 1. To map the 24 beams towards the center beam, 24 tool-specific input models are implemented. Each tool-specific input model is implemented by a single convolution layer with the same configuration added over the lowest representation layer in the U-Net. This layer is specific to a particular beam while the encoder decoder U-Net layers are shared across all beams. Figure 20 shows the predicted data generated by the first example implementation when all other beams are mapped towards the center beam. Comparing the panel of Figure 20 to the ground truth of panel 1900 of Figure 19, it can be seen that the pattern of a contact array is faithfully reproduced by the neural network. Further, comparing the panel of Figure 20 to the input data of panel 1902 of Figure 19, it can be seen that the neural network successfully removed the tool mismatch (i.e. the different images of Figure 20 show a similar amount of aberrations while the aberrations in the images of panel 1902 vary strongly).[000136] With reference to Figures 21 and 22, experimental data is shown which was obtained by using the second example implementation of the proposed neural network 1200 on uDBO images. [000137] To obtain training and validation, 60 targets have been measured with each of 6 tools and for a nominal focus setting (in both wafer rotation modes and normal) and for complementary acquisition settings. The data set also includes data from 10 targets measured using 5 focus offsets: +0.6, +0.3, 0 nm. All measured images have been pre-registered. Figure 21 shows uDBO images 2100, 2101 from two different tools (tool A and tool B) as well as the difference between these images 2102. The strong difference signal 2102 indicates strong tool mismatch.[000138] The second example implementation of the neural network includes 10 convolution layers with 32 filters of size 7 x 7 for the common input model and common output model. Two skip connections corresponding to the third and sixth layer outputs are implemented. Before the skip connections a stride of 2 is used, and otherwise a stride of 1. The common input / output models have a residual network structure for the U-Net blocks, with the residual connections corresponding with the skip connection locations. The tool-specific input / output models each include 3 convolution layers with 16 filters. The input and output of the network have 4 channels, namely for the normal andcomplementary measurements and for two wafer rotations. The focus level acts as an input to the toolspecific models and is added as a simple bias term after the convolution layers.[000139] 55 of the nominal focus targets and 8 of the through focus images have been used for the training (the remaining images have been used for validation). No specific parameter tuning and extensive model size search was performed.[000140] The trained neural network is used to map the image 2101 generated by tool B towards tool A. Figure 22 shows that the mapped image 2200 as well as the difference signal 2201 between the image generated by tool A 2100 and the mapped image 2200. It can be seen that the difference signal is strongly reduced compared to the unmapped signal difference 2102.[000141] Further, a matching improvement is also observed with respect to the overlay. To quantify this, for 5 targets (which have not been used in the training) the overlay spread over tools is obtained, then the mean over these 5 targets is taken. Using a standard overlay extraction algorithm an overlay is determined. When the second example neural network is used to map the images before applying the standard overlay extraction algorithm, then the overlay for the 6 tool is obtained, which is better than the standard overlay matching.[000142] Further embodiments are disclosed in the subsequent numbered clauses:1. A method of generating, using a neural network, predicted metrology data of a patterned portion of a substrate patterned using a lithographic apparatus, the neural network comprising an input network and an output network, the input network comprising a common input model and a specific input model, wherein the common input model has been trained by a process in which the common input model processes data based on training metrology data generated by a plurality of metrology systems, and the specific input model has been trained by a process in which the specific input model processes data based on training metrology data generated only by a first metrology system (or a specific input model which is specific to the first metrology system in another way, such as that the weights of the specific input model have been determined by processing first metrology system data specifying the first metrology system using an input model generator network), the method comprising: receiving metrology data of the patterned portion generated using the first metrology system; obtaining configuration data specifying a configuration of the first metrology system used to generate the metrology data; processing, using the common input model and the specific input model, the metrology data and the configuration data to generate an encoding of the patterned portion; and processing, using the output network of the neural network, the encoding of the patterned portion to generate the predicted metrology data, the predicted metrology data being a prediction of metrology data which would have been generated by inspecting the patterned portion using a second metrology system.2. The method of clause 1, wherein a first one of the common input model and the specific input model is configured to process the metrology data to generate an encoding of the metrology data, and a second one of the common input model and the specific input model being configured to process the encoding of the metrology data.3. The method of clause 2 in which the first one of the common input model and the specific input model is the common input model, and the second one of the common input model and the specific input model is the specific input model.4. The method of clause 2 or 3, wherein processing the metrology data and the configuration data to generate an encoding of the patterned portion comprises: processing, using the first one of the common input model and the specific input model, the metrology data to generate an encoding of the metrology data; and processing, using the second one of the common input model and the specific input model, the encoding of the metrology data and the configuration data to generate the encoding of the patterned portion.5. The method of any of clauses 2 to 4, wherein the plurality of metrology systems used in the process to train the common input model comprise the first metrology system, the second metrology system and at least one further metrology system.6. The method of any preceding clause further comprising a step of obtaining target configuration data specifying a configuration of the second metrology system, and wherein the predicted metrology data is a prediction of metrology data which would have been generated by inspecting the patterned portion using the second metrology system configured according to the target configuration data.7. The method of clause 6, wherein the output network comprises a common output model trained by a process in which the common output model processes data based on training metrology data generated by a plurality of metrology systems.8. The method of clause 7, wherein the plurality of metrology systems used in the process to train the common output comprise the first metrology system, the second metrology system and at least one further metrology system.9. The method of clause 7 or 8, wherein the output network further comprises a specific output model that is specific to the second metrology system, and optionally wherein i) the specific output model has been trained by a process in which the weights of the specific output model are adjusted based on a discrepancy between the predicted metrology data and training metrology data generated only by the second metrology system, or ii) weights of the specific output model have been determined by processing second metrology system data specifying the second metrology system using an output model generator network, the output model generator network configured to receive, as input, metrology system data specifying arespective metrology system and to determine, as output, weights of the specific output model for the respective metrology system10. The method of clause 9, wherein a first one of the common output model and the specific output model is configured to process data generated by the input network, and a second one of the common output model and specific output model is configured to process the output of the first one of the common output model and specific output model.11. The method of clause 10 in which the first one of the common output model and the specific output model is the specific output model, and the second one of the common output model and the specific output model is the common output model.12. The method of clause 11, wherein processing the encoding of the patterned portion to generate the predicted metrology data comprises: processing, using the specific output model, the encoding of the patterned portion and the target configuration data to generate the encoding of the predicted metrology data; and processing, using the common output model, the encoding of the predicted metrology data to generate the predicted metrology data.13. The method of clause 9, wherein processing the metrology data and the configuration data of the first metrology system to generate an encoding of the patterned portion comprises: processing, using the common input model, the metrology data to generate an encoding of the metrology data; processing, using the specific input model, the encoding of the metrology data and the configuration data of the first metrology system to generate first encoded data; processing, using the specific output model, the encoding of the metrology data and the target configuration data to generate second encoded data; and combining the first and second encoded data to generate the encoding of the patterned portion.14. The method of clause 13, wherein combining the first and second encoded data to generate the encoding of the patterned portion comprises concatenating the first and second encoded data to generate the encoding of the patterned portion.15. The method of any preceding clause, wherein at least one of the first metrology system and the second metrology system is a real-world metrology system.16. The method of any preceding clause, wherein at least one of the first metrology system and the second metrology system is a simulated metrology system.17. The method of any preceding clause, wherein the neural network further comprises an application-specific model corresponding to an application type, from a set of application types, associated with the patterned portion, the application-specific model being configured to process data from the input network and to provide data to the output network data, and wherein the predicted metrology data is generated by processing the encoding of the patterned portion using the outputnetwork of the neural network and the data provided by the application model, wherein the application-specific model has been trained by a process in which the application-specific model processes data based on training metrology data of patterned portions associated with the application type.18. The method of any preceding clause, wherein the configuration data of the first metrology system comprises setting data specifying operating parameters of the first metrology system.19. The method of any preceding clause, wherein the configuration data of the first metrology system comprises data characterizing the first metrology system.20. The method of any preceding clause, wherein the configuration data of the first metrology system comprises reference metrology data of a reference target generated using the first metrology system configured in the same configuration as used to generate the received metrology data.21. The method of any preceding clause, wherein the metrology data comprises image data.22. The method of clause 21, wherein the different metrology systems are different instances of an optical dark-field microscope.23. The method of clause 22, wherein the configuration data of the first metrology system comprises information specifying a spectral density of light used to generate the metrology data.24. The method of any preceding clause, wherein the different metrology systems are different beams of a multi-beam scanning electron microscope.25. The method of any preceding clause, wherein the metrology data comprises a pupil representation defining a distribution of radiation in a pupil plane of the metrology system.26. The method of any preceding clause, wherein the neural network model comprises a U-Net architecture.27. The method of any preceding clause, wherein the neural network model comprises one or more skip connections.28. The method according to clause 27, when dependent on clause 9, wherein the one or more skip connections comprise at least one skip connection from the common input model to the common output model.29. The method according to any preceding clause, when dependent on clause 6, wherein the first metrology system and the second metrology system are the same metrology system, and wherein the target configuration data specify a configuration of the metrology system that is different from the configuration used to generate the metrology data.30. The method according to any preceding clause, when dependent on clause 6, further comprising processing the predicted metrology data to determine one or more performance parameters.31. The method according to clause 30 further comprising: processing the one or more performance parameters to generate updated target configuration data, andprocessing, using the neural network, the metrology data, the configuration data of the first metrology system and the updated target configuration data to generate updated predicted metrology data.32. The method of any preceding clause further comprising a step of obtaining inspection data comprising a first index value indicating a number of inspections which had been carried out on the patterned portion prior to the generation of the metrology data, and a second index value, the predicted metrology data being a prediction of metrology data which would have been generated by inspecting the patterned portion using a second metrology system if the patterned portion had been subject to a number of prior inspections indicated by the second index value; wherein the neural network further comprises a transfer model configured to: process i) the inspection data, and ii) the encoding of the patterned portion generated by the input network, to generate data, and provide said data to the output network, wherein the predicted metrology data is generated by the output model processing the data provided by the transfer model.33. The method of clause 32, wherein the inspection data further comprises data specifying a time duration associated with the previous inspections, and / or an energy deposited on the patterned portion during the previous inspections.34. A method of training a neural network for generating predicted metrology data of a patterned portion of a substrate patterned using a lithographic apparatus, the neural network comprising an input network and a output network, the input network comprising a common input model and a specific input model, wherein specific input model is associated with a first metrology system, the neural network configured to: receive metrology data of the patterned portion generated using the first metrology system; receive configuration data specifying a configuration of the first metrology system used to generate the metrology data; process, using the common input model and the specific input model, the metrology data and the configuration data to generate an encoding of the patterned portion; and process, using the output network of the neural network, the encoding of the patterned portion to generate the predicted metrology data, the predicted metrology data being a prediction of metrology data which would have been generated by inspecting the patterned portion using a second metrology system.35. The method of clause 34, wherein the method comprises obtaining a training dataset comprising first training metrology data generated using the first metrology system, configuration data specifying a configuration of the first metrology system used to generate the first training metrology data, and a second training metrology data generated using the second metrology system,wherein the first and second training metrology data have been generated by inspecting the same patterned portion of a training substrate patterned by the, or a further, lithographic apparatus.36. The method according to clause 35, wherein the method comprises processing, using the common input model and the specific input model, the first training metrology data and the corresponding configuration data to generate an encoding of the patterned portion.37. The method according to clause 35 or 36, wherein the method comprises processing, using the output network of the neural network, the encoding of the patterned portion to generate predicted second training metrology data, the predicted second training metrology data being a prediction of metrology data which would have been generated by inspecting the patterned portion using the second metrology system.38. The method according to clause 37, wherein the method comprises updating weights of the neural network to reduce a measure of the discrepancy between the predicted second training metrology data and the second training metrology data.39. The method of any of clauses 34 to 38, wherein the input network comprises a plurality of specific input models, each specific input model being associated with a different respective metrology system of a plurality of metrology systems comprising at least the first and second metrology systems and a third metrology system.40. The method of clause 39, wherein the neural network is configured to process the metrology data and the configuration data to generate an encoding of the patterned portion by: processing, using the common input model, the metrology data to generate an encoding of the metrology data; and processing, using the specific input model of the first metrology system, the encoding of the metrology data and the configuration data to generate the encoding of the patterned portion, wherein the training dataset further comprises third training metrology data generated using the third metrology system, configuration data specifying a configuration of the third metrology system used to generate the third training metrology data, and fourth training metrology data generated using the second metrology system, wherein the third and fourth training metrology data have been generated by inspecting the same patterned portion of a respective training substrate patterned by the, or a further, lithographic apparatus; the method further comprising: processing, using the common input model, the third training metrology data to generate an encoding of the third training metrology data; and processing, using the specific input model of the third metrology system, the encoding of the third training metrology data and the corresponding configuration data of the third metrology system to generate the encoding of the patterned portion, processing, using the output network of the neural network, the encoding of the patterned portion to generate predicted fourth training metrology data, the predicted fourth training metrologydata being a prediction of metrology data which would have been generated by inspecting the patterned portion using the second metrology system; and updating weights of the neural network to reduce: a measure of discrepancy between the predicted fourth training metrology data and the fourth training metrology data.41. The method of clause 40, wherein the output network comprises a common output model and a plurality of specific output models, each specific output model associated with a different metrology system of the plurality of metrology systems, wherein the neural network is configured to process the encoding of the patterned portion to generate the predicted metrology data by: processing, using the specific output model of the second metrology system, the encoding of the patterned portion and target configuration data to generate the encoding of the predicted metrology data; and processing, using the common output model, the encoding of the predicted metrology data portion to generate the predicted metrology data, wherein the training dataset further comprises fifth training metrology data generated using the first metrology system, configuration data specifying a configuration of the first metrology system used to generate the fifth training metrology data, a sixth training metrology data generated using the third metrology system, and configuration data specifying a configuration of the third metrology system used to generate the sixth training metrology data, wherein the fifth and sixth training metrology data have been generated by inspecting the same patterned portion of a respective training substrate patterned by the, or a further, lithographic apparatus; the method further comprising: processing, using the common input model, the fifth training metrology data to generate an encoding of the fifth training metrology data; and processing, using the specific input model of the first metrology system, the encoding of the fifth training metrology data and the corresponding configuration data of the first metrology system to generate an encoding of the patterned portion, processing, using the specific output model of the third metrology system, the encoding of the patterned portion and the corresponding configuration data of the third metrology system to generate an encoding of predicted sixth training metrology data; and processing, using the common output model, the encoding of the predicted sixth metrology data to generate the predicted sixth training metrology data, updating weights of the neural network to reduce: a measure of discrepancy between the predicted sixth training metrology data and the sixth training metrology data.42. The method according of clause 41, wherein the first metrology system and the third metrology system are the same metrology system, and the corresponding configuration data specify the same configuration of the metrology system.43. The method of any of clams 39 to 42, wherein the plurality of metrology systems comprises at least one simulated metrology system.44. A method of training an input model generator network to determine, from metrology system data specifying a respective metrology system, respective weights of a specific input model for the respective metrology system, the specific input model being comprised in a neural network for generating predicted metrology data of a patterned portion of a substrate patterned using a lithographic apparatus, the neural network comprising an input network and an output network, the input network comprising a common input model and the specific input model, the method comprising: obtaining a training dataset comprising first training metrology data generated using the first metrology system, configuration data specifying a configuration of the first metrology system used to generate the first training metrology data, and second training metrology data generated using the second metrology system, wherein the first and second training metrology data have been generated by inspecting the same patterned portion of a training substrate patterned by the, or a further, lithographic apparatus; obtaining candidate metrology system data specifying the first metrology system; processing, using the input model generator network, the obtained metrology system data to generate training weights of the specific input model for the first metrology system; processing, using the common input model and the specific input model in accordance with the generated training weights, the first training metrology data and the corresponding configuration data to generate an encoding of the patterned portion; processing, using the output network of the neural network, the encoding of the patterned portion to generate predicted second training metrology data, the predicted second training metrology data being a prediction of metrology data which would have been generated by inspecting the patterned portion using the second metrology system; and updating weights of the input model generator network and one or more parameters of the metrology system data to reduce: a measure of the discrepancy between the predicted second training metrology data and the second training metrology data.45. The method of 44 wherein, to reduce the measure of the discrepancy between the predicted second training metrology data and the second training metrology data, the weights of the input model generator network and one or more parameters defining the metrology system data are jointly adjusted with weights of the common input model and / or weights of the output network.46. A method of generating, using a neural network, predicted metrology data of a patterned portion of a substrate patterned using a lithographic apparatus, the neural network comprising an input network and an output network, the input network comprising a common input model and a specific input model, wherein the common input model has been trained by a process in which the common input model processes data based on training metrology data generated by a plurality of metrology systems, and the specific input model is specific to a first metrology system, the method comprising: receiving metrology data of the patterned portion generated using the first metrology system; obtaining configuration data specifying a configuration of the first metrology system used to generate the metrology data; processing, using the common input model and the specific input model, the metrology data and the configuration data to generate an encoding of the patterned portion; and processing, using the output network of the neural network, the encoding of the patterned portion to generate the predicted metrology data, the predicted metrology data being a prediction of metrology data which would have been generated by inspecting the patterned portion using a second metrology system.47. The method of clause 46 in which (i) the specific input model has been trained by a process in which the specific input model processes data based on training metrology data generated only by the first metrology system or ii) weights of the specific input model have been determined by processing first metrology system data specifying the first metrology system using an input model generator network, the input model generator network being configured to receive, as input, metrology system data specifying a respective metrology system and to determine, as output, weights of the specific input model for the respective metrology system48. The method clause 47, wherein the metrology system data comprises one or more parameters which define a distribution of a set of latent variables, and the input model generator network is configured to determine the weights of the specific input model for a respective metrology system by: sampling from the distribution defined by the input metrology system data to determine values of the set of latent variables; and determining the weights of the specific input model for a respective metrology system based on the set of latent variables.49. A computer program product comprising machine-readable instructions which, when the program is executed by a computer, cause the computer to carry out the method of any preceding clause.50. A computer system comprising a processor and a data storage device, the data storage device storing program instructions which, when executed by the processor, cause the processor to carry out the method of any one of clauses 1 to 48.[000143] Although specific reference may be made in this text to the use of lithographic apparatus in the manufacture of ICs, it should be understood that the lithographic apparatus described herein may have other applications. Possible other applications include the manufacture of integrated optical systems, guidance and detection patterns for magnetic domain memories, flat-panel displays, liquidcrystal displays (LCDs), thin-film magnetic heads, etc.[000144] Although specific reference may be made in this text to embodiments of the invention in the context of a lithographic apparatus, embodiments of the invention may be used in other apparatus. Embodiments of the invention may form part of a mask inspection apparatus, a metrology apparatus, or any apparatus that measures or processes an object such as a wafer (or other substrate) or mask (or other patterning device). These apparatus may be generally referred to as lithographic tools. Such a lithographic tool may use vacuum conditions or ambient (non-vacuum) conditions.[000145] Although specific reference may have been made above to the use of embodiments of the invention in the context of optical lithography, it will be appreciated that the invention, where the context allows, is not limited to optical lithography and may be used in other applications, for example imprint lithography.[000146] While specific embodiments of the invention have been described above, it will be appreciated that the invention may be practiced otherwise than as described. The descriptions above are intended to be illustrative, not limiting. Thus it will be apparent to one skilled in the art that modifications may be made to the invention as described without departing from the scope of the claims set out below.

Claims

CLAIMS:

1. A method of generating, using a neural network, predicted metrology data of a patterned portion of a substrate patterned using a lithographic apparatus, the neural network comprising an input network and an output network, the input network comprising a common input model and a specific input model, wherein the common input model has been trained by a process in which the common input model processes data based on training metrology data generated by a plurality of metrology systems, and the specific input model is specific to a first metrology system, the method comprising: receiving metrology data of the patterned portion generated using the first metrology system; obtaining configuration data specifying a configuration of the first metrology system used to generate the metrology data; processing, using the common input model and the specific input model, the metrology data and the configuration data to generate an encoding of the patterned portion; and processing, using the output network of the neural network, the encoding of the patterned portion to generate the predicted metrology data, the predicted metrology data being a prediction of metrology data which would have been generated by inspecting the patterned portion using a second metrology system.2 The method of claim 1 in which (i) the specific input model has been trained by a process in which the specific input model processes data based on training metrology data generated only by the first metrology system or ii) weights of the specific input model have been determined by processing first metrology system data specifying the first metrology system using an input model generator network, the input model generator network being configured to receive, as input, metrology system data specifying a respective metrology system and to determine, as output, weights of the specific input model for the respective metrology system.

3. The method of claim 1 or 2, wherein a first one of the common input model and the specific input model is configured to process the metrology data to generate an encoding of the metrology data, and a second one of the common input model and the specific input model being configured to process the encoding of the metrology data.

4. The method of claim 3 in which the first one of the common input model and the specific input model is the common input model, and the second one of the common input model and the specific input model is the specific input model.

5. The method of claim 3 or 4, wherein processing the metrology data and the configuration data to generate an encoding of the patterned portion comprises: processing, using the first one of the common input model and the specific input model, the metrology data to generate an encoding of the metrology data; and processing, using the second one of the common input model and the specific input model, the encoding of the metrology data and the configuration data to generate the encoding of the patterned portion.

6. The method of any of claims 3 to 5, wherein the plurality of metrology systems used in the process to train the common input model comprise the first metrology system, the second metrology system and at least one further metrology system.

7. The method of any preceding claim further comprising a step of obtaining target configuration data specifying a configuration of the second metrology system, and wherein the predicted metrology data is a prediction of metrology data which would have been generated by inspecting the patterned portion using the second metrology system configured according to the target configuration data.

8. The method of any preceding claim, wherein at least one of the first metrology system and the second metrology system is a real-world metrology system.

9. The method of any preceding claim, wherein at least one of the first metrology system and the second metrology system is a simulated metrology system.

10. The method of any preceding claim, wherein the configuration data of the first metrology system comprises one or more of setting data specifying operating parameters of the first metrology system, data characterizing the first metrology system, and reference metrology data of a reference target generated using the first metrology system configured in the same configuration as used to generate the received metrology data.

11. The method of any preceding claim, wherein the metrology data comprises image data.

12. The method of any preceding claim, wherein the different metrology systems are different beams of a multi-beam scanning electron microscope.

13. The method of any preceding claim, wherein the metrology data comprises a pupil representation defining a distribution of radiation in a pupil plane of the metrology system.

14. A method of training a neural network for generating predicted metrology data of a patterned portion of a substrate patterned using a lithographic apparatus, the neural network comprising an input network and a output network, the input network comprising a common input model and a specific input model, wherein specific input model is associated with a first metrology system, the neural network configured to: receive metrology data of the patterned portion generated using the first metrology system; receive configuration data specifying a configuration of the first metrology system used to generate the metrology data; process, using the common input model and the specific input model, the metrology data and the configuration data to generate an encoding of the patterned portion; and process, using the output network of the neural network, the encoding of the patterned portion to generate the predicted metrology data, the predicted metrology data being a prediction of metrology data which would have been generated by inspecting the patterned portion using a second metrology system.

15. The method of any preceding claim, wherein the neural network system comprises at least one supplementary processing branch, a first of the supplementary processing branches comprising (i) a first supplementary common input model configured to receive the output of the common input model, (ii) optionally a first supplementary specific input model that is specific to the first metrology system, and (iii) a first supplementary output network, the output network being configured to receive the output of the first supplementary output network, and optionally in which there are a plurality of the supplementary processing branches in a sequence, each of the supplementary processing branches, except the first, comprising (i) a corresponding supplementary common input model configured to receive the output of the supplementary common input model of the preceding supplementary processing branch, (ii) optionally a corresponding supplementary specific input model that is specific to the first metrology system, and (iii) a corresponding supplementary output network, the supplementary output network of the preceding supplementary processing branch being configured to receive the output of the corresponding supplementary output network.

16. The method of any preceding claim further comprising a step of obtaining inspection data comprising a first index value indicating a number of inspections which had been carried out on the patterned portion prior to the generation of the metrology data, and a second index value, the predicted metrology data being a prediction of metrology data which would have been generated byinspecting the patterned portion using a second metrology system if the patterned portion had been subject to a number of prior inspections indicated by the second index value; wherein the neural network further comprises a transfer model configured to: process i) the inspection data, and ii) the encoding of the patterned portion generated by the input network, to generate data, and; provide said data to the output network, wherein the predicted metrology data is generated by the output model processing the data provided by the transfer model.

17. A method of training an input model generator network to determine, from metrology system data specifying a respective metrology system, respective weights of a specific input model for the respective metrology system, the specific input model being comprised in a neural network for generating predicted metrology data of a patterned portion of a substrate patterned using a lithographic apparatus, the neural network comprising an input network and an output network, the input network comprising a common input model and the specific input model, the method comprising: obtaining a training dataset comprising first training metrology data generated using the first metrology system, configuration data specifying a configuration of the first metrology system used to generate the first training metrology data, and second training metrology data generated using the second metrology system, wherein the first and second training metrology data have been generated by inspecting the same patterned portion of a training substrate patterned by the, or a further, lithographic apparatus; obtaining candidate metrology system data specifying the first metrology system; processing, using the input model generator network, the obtained metrology system data to generate training weights of the specific input model for the first metrology system; processing, using the common input model and the specific input model in accordance with the generated training weights, the first training metrology data and the corresponding configuration data to generate an encoding of the patterned portion; processing, using the output network of the neural network, the encoding of the patterned portion to generate predicted second training metrology data, the predicted second training metrology data being a prediction of metrology data which would have been generated by inspecting the patterned portion using the second metrology system; and updating weights of the input model generator network and one or more parameters of the metrology system data to reduce: a measure of the discrepancy between the predicted second training metrology data and the second training metrology data.

18. A method of using a neural network and an input model generator network to determine metrology system data comprising one or more parameters and specifying a first metrology system, the input model generator network configured to determine, from metrology system data specifying a respective metrology system, respective weights of a specific input model for the respective metrology system, the specific input model being comprised in a neural network for generating predicted metrology data of a patterned portion of a substrate patterned using a lithographic apparatus, the neural network comprising an input network and an output network, the input network comprising a common input model and the specific input model, the method comprising: obtaining a dataset comprising first metrology data generated using the first metrology system, configuration data specifying a configuration of the first metrology system used to generate the first metrology data, and second metrology data generated using a second metrology system, wherein the first and second training metrology data have been generated by inspecting the same patterned portion of a substrate patterned by the, or a further, lithographic apparatus; obtaining candidate metrology system data specifying the first metrology system; processing, using the input model generator network, the candidate metrology system data to generate weights of the specific input model for the first metrology system; processing, using the common input model and the specific input model in accordance with the generated weights, the first training metrology data and the corresponding configuration data to generate an encoding of the patterned portion; processing, using the output network of the neural network, the encoding of the patterned portion to generate predicted second metrology data, the predicted second metrology data being a prediction of metrology data which would have been generated by inspecting the patterned portion using the second metrology system; and updating one or more parameters of the metrology system data to reduce: a measure of the discrepancy between the predicted second metrology data and the second metrology data.

19. A computer program product comprising machine-readable instructions which, when the program is executed by a computer, cause the computer to carry out the method of any preceding claim.

20. A computer system comprising a processor and a data storage device, the data storage device storing program instructions which, when executed by the processor, cause the processor to carry out the method of any one of claims 1 to 18.

Images