Method of aligning a charged particle beam apparatus

Abstract

The present disclosure relates to a method of aligning a charged particle beam device comprising the steps of: providing a charged particle beam device in a first alignment state; implementing, by a processing unit, an alignment transition from the first alignment state to a second alignment state using an alignment algorithm; and providing data related to the alignment transition to a modification algorithm to modify the alignment algorithm in order to implement a modified alignment transition.

Description

[0001] describe

[0002] The invention relates to a method of aligning a charged particle beam device, and to a charged particle beam device having a processing unit for performing such a method.

[0003] A charged particle beam device can, in principle, be any device that uses charged particles, such as electrons, protons, and / or ions, to generate a beam for, for example, irradiating an object of interest. Charged particle beams can be used to study samples, examine objects, and / or remove parts of such objects. Applications include, for example, transmission electron microscopy, scanning electron microscopy, (plasma) focused ion beam technology, and photolithography.

[0004] In many of these charged particle beam instruments, alignment of the charged particle beam is essential. For example, in a transmission electron microscope (TEM), the electron beam needs to be focused so that the object of interest is illuminated with a parallel beam, and then the transmitted beam is focused onto a detector. In a scanning electron microscope, the electron beam is focused onto a small spot on the object, the beam is scanned over the object, and the resulting secondary radiation is detected to examine the object of interest. To obtain high-quality results, charged particle beam instruments require alignment.

[0005] Alignment of charged particle beam equipment has proven to be very difficult and requires a high level of operator expertise to achieve quickly and reliably. For example, in a TEM, alignment involves one or more of the following parameters and operations: setting the gun and condenser system, the condenser aperture, selecting the sample height (concentricity), correctly achieving the beam offset and tilt pivot, accurately tilting the beam down along the optical axis of the objective (center of rotation, objective aperture, if necessary), and refining the focus of the diffraction lenses, appropriately correcting the astigmatism of each lens at each point. Alignment in a TEM can require several iterations to achieve the desired end result, which is time-consuming and error-prone.

[0006] To this end, it is an object of the present invention to provide a method by which the alignment of a charged particle beam device is improved, in particular with respect to ease of use, accuracy and / or speed.It is a further object of the present invention to provide a charged particle beam device of the above mentioned qualities.

[0007] To this end, the present disclosure provides a method of aligning a charged particle beam device as defined in claim 1 .

[0008] The method as defined herein comprises the step of providing a charged particle beam device in a first alignment state. The first alignment state is typically a state in which the charged particle beam device, such as a charged particle microscope, is misaligned, i.e. not (optimally) aligned for its intended use.

[0009] The method as defined herein comprises the step of implementing an alignment conversion using an alignment algorithm. The alignment algorithm may be performed by a processing unit that may be part of the charged particle beam apparatus or may be externally connected to the charged particle beam apparatus. The alignment algorithm causes the charged particle beam apparatus to change from the first alignment state to a second alignment state. Generally speaking, the second alignment state is an improved alignment state, i.e., a state with a higher degree of alignment (optimal alignment) for the intended use of the charged particle beam apparatus, but this is not necessarily required. States with a lower degree of alignment are also conceivable. The conversion from the first alignment state to the second alignment state is defined as an alignment conversion in this article.

[0010] In the method as defined herein, the data relevant to the alignment conversion can be provided to the modification algorithm. The data may include, for example, settings, speed, timing and / or calibration results of the charged particle beam device, but other data relevant to any alignment state and / or configuration of the charged particle beam device are also conceivable. The data may include alignment state, alignment action and corresponding quality parameters.

[0011] The modification algorithm as defined herein is arranged to modify the alignment algorithm. The modified alignment algorithm is arranged in such a way that a modified alignment transformation is performed starting from the first alignment state. The modified alignment transformation may result in a second alignment state that is substantially the same as or at least very similar to the end point of the alignment, but the intermediate steps to reach the second alignment state may be different. Alternatively, the modified alignment transformation may result in a modified second alignment state that is actually different from the initial second alignment state. In the latter case, the modification algorithm is arranged to provide a modified alignment algorithm that results in a modified second alignment that is improved relative to the initial second alignment.

[0012] Therefore, by using the modification algorithm to modify the alignment algorithm, based on the data obtained during the alignment conversion from the first alignment state to the second alignment state, an improved alignment algorithm is obtained, which can be used to perform the alignment of the charged particle beam device in an improved manner, thereby producing a charged particle beam device with an improved second alignment state or a faster alignment conversion. Thus, the purpose as defined herein is achieved.

[0013] Advantageous embodiments of the present disclosure will be described below.

[0014] In one embodiment, the alignment algorithm is a trainable decision algorithm that is arranged to infer the order of alignment actions. The trainable decision algorithm, also referred to as an agent, can be used by the processing unit to infer the order of the alignment actions. In this embodiment, the inferred alignment action sequence is performed to cause the charged particle beam device to enter a second alignment state. The data associated with the alignment conversion, i.e., the change of the alignment state, can be stored for further use, for example, to modify at least one of the agents.

[0015] The alignment transformation may comprise a series of individual alignment time steps. Data corresponding to the individual alignment time steps, such as at least one of the alignment state, alignment action and / or quality parameter, may be stored for further use in order to modify the alignment algorithm.

[0016] In one embodiment, the above-described method steps correspond to a single iteration that may be repeated at least once in order to align the charged particle beam apparatus.

[0017] The method may include the step of modifying the alignment algorithm. Modification of the alignment algorithm may include using data related to the alignment transformation, i.e. data related to the alignment state. It should be noted that additional data may be used to modify the alignment algorithm.

[0018] In this context, modification can be understood as a change in at least one of the hidden states, activations, trainable parameters, or structure of the machine learning model underlying the alignment algorithm's decision-making mechanism. In the first two cases, this means that the agent can use recursive or autoregressive models to infer alignment actions. Such models can be used to perform alignment where action inference does not require a single observation, but rather a history of past observations. However, modifications to the trainable weights and machine learning model structure can be performed offline, meaning that the alignment is performed by the agent and not modified during alignment, but rather after it is completed.

[0019] Note that modifications to the alignment algorithm or agent may include the following steps:

[0020] - Provide alignment algorithms, such as proxies;

[0021] - providing at least data of the alignment transformation performed by said alignment algorithm;

[0022] - determining required modifications of said alignment algorithm, wherein said data of the alignment transformation are used;

[0023] - modifying the alignment algorithm based on said determined modification.

[0024] In one embodiment, described method comprises and uses the alignment algorithm after described modification to realize the additional step of subsequent alignment conversion.Subsequent alignment conversion can be used for described charged particle beam equipment to be brought to the 3rd alignment state from described second alignment state.Compared with described second alignment state, described 3rd alignment state can have improved alignment characteristic.In fact, method as defined herein at least partially repeats so that charged particle beam equipment reaches another improved alignment state.The data relevant with described subsequent alignment conversion can be provided to described modification algorithm to modify described modified alignment algorithm so that realize another modified alignment conversion.If necessary, this allows to check alignment algorithm after each alignment procedure, and improves when necessary.

[0025] The alignment transformation can comprise a plurality of independent alignment actions. In principle, this is applicable to any alignment transformation as indicated herein, such as alignment transformation, modified alignment transformation, subsequent alignment transformation. Each of these alignment transformations can comprise a series of independent alignment steps, which are referred to as independent alignment actions in this article. For example, these independent alignment actions can comprise one or more of the following: arranging the intensity of the components of the charged particle beam device, arranging the position of the object of interest, arranging the beam properties such as beam offset and / or beam tilt, arranging the focus of one or more lenses, correcting the astigmatism of one or more lenses, and selecting the order of any of these independent actions or arrangements.

[0026] In one embodiment, the method includes the step of determining a quality parameter. The quality parameter may relate to a final alignment state, such as a second alignment state, a third alignment state, etc. The quality parameter may additionally or alternatively relate to an alignment transition, or to one or more individual alignment actions thereof. It is contemplated that one or more quality parameters may be determined for a single alignment transition. It is also contemplated that one or more quality parameters may be determined for a single individual alignment action.

[0027] In one embodiment, the step of determining the quality parameter comprises evaluating one or more of the plurality of individual alignment actions. This allows the individual alignment actions to be evaluated and corresponding quality parameters to be associated with them. When modifying the alignment algorithm, the modified algorithm may take these individual quality parameters into account.

[0028] The evaluation may include identifying individual alignment actions that result in desired and / or undesirable alignment states. For example, individual alignment actions that result in improved alignment may be rewarded and promoted in the modified alignment algorithm. However, individual alignment actions that result in degraded alignment may be penalized and prevented in the modified alignment algorithm. Evaluation of individual alignment actions may identify, for example, those actions that result in deviations from desired alignment or those actions that are slow, and prevent these actions from being used in the modified alignment algorithm. Accordingly, the method may include the step of modifying the alignment algorithm to prevent undesirable alignment actions during use of the modified alignment algorithm.

[0029] In one embodiment, the alignment algorithm comprises at least one neural network. As known to those skilled in the art, a neural network (NN)—also known as an artificial neural network (ANN) or simulated neural network (SNN)—is a set of interconnected artificial neurons that uses mathematical or computational models to perform information processing based on computational connectionist methods. As defined herein, an artificial neural network is an adaptive system that changes its structure based on external or internal information flowing through the network. In more practical terms, a neural network is a nonlinear statistical data modeling or decision-making tool that can be used to model complex relationships between inputs and outputs or to find patterns in data. Learning in neural networks is particularly useful in applications where the complexity of the data or task makes manual design of such functionality impractical. Therefore, the use of neural networks is beneficial for optimizing alignment procedures in any charged particle beam device.

[0030] In particular, if the machine learning model is a neural network, the procedure for determining the modification may consist essentially in the calculation of a loss function and subsequently in the calculation of the gradient of the loss function with respect to the trainable weights of the neural network.

[0031] Said modification of the agent may consist essentially in updating the trainable weights of the neural network based on the calculated gradients using any classical optimizer such as ADAM.

[0032] Alternatively, the modification rules used by the modification procedure may not be given in terms of mathematical formulas (such as used by classical neural network optimizers), but may be learned from data in a meta-loop of the training procedure.

[0033] In an embodiment, the method may include the step of providing stored data acquired during the alignment transformation. This allows, for example, the use of on-policy, off-policy, and offline reinforcement learning algorithms.

[0034] The modification algorithm as defined herein can be a training algorithm for training a neural network. In this sense, the modification algorithm can be part of the neural network or an external algorithm that provides the input required to train the neural network. Data related to the alignment transformation can be used as input to train the neural network. Here, quality parameters related to the alignment transformation and / or individual alignment actions can be used as a penalty / reward mechanism for training the neural network. Retraining the at least one neural network can use one or more of the undesirable alignment states as input parameters. In this regard, the neural network can be trained using deep reinforcement learning, which focuses on finding a balance between exploration (unknown territory) and exploitation (current knowledge). Compared to other types of learning (such as supervised learning and unsupervised learning), labeled input / output pairs do not need to be presented, and suboptimal actions do not need to be explicitly corrected.

[0035] Therefore, in one embodiment, said step of modifying said algorithm comprises a step of training or retraining said at least one neural network, in particular using deep reinforcement learning.

[0036] It is conceivable that the charged particle beam device is set to an undesirable alignment state, and the revised alignment algorithm is used to realize another conversion. The influence of the other conversion can then be mainly used to collect other data, and use this data to revise (previously revised) alignment algorithm again. It is also conceivable that the undesirable alignment state is relevant with the alignment state in which it is found that the alignment algorithm performs non-optimal.

[0037] In one embodiment, a training algorithm is provided, and the training algorithm is used to identify and / or set undesirable alignment states. The training algorithm makes it possible to quickly and effectively identify situations or alignment states that the alignment algorithm has difficulty in, and accurately uses these difficult situations to test the alignment algorithm, collects the data of the alignment conversion, and uses the modification algorithm to improve the alignment algorithm. In a sense, a two-step approach is used, wherein the training algorithm is used to identify difficult alignment states, and the alignment algorithm is used to effectively align the charged particle beam device starting from these difficult alignment states. When using training neural network and alignment neural network, they can be used to reinforce each other, to quickly and effectively obtain accurate alignment neural network fast.

[0038] According to one aspect, a method of training an alignment algorithm for aligning a charged particle beam device is provided, comprising the following steps:

[0039] - Provide the alignment algorithm to be trained;

[0040] - providing data relating to an alignment transition of the charged particle beam device, wherein the alignment transition extends from a first alignment state to a second alignment state; and

[0041] - providing said data to a modification algorithm for modifying said alignment algorithm, wherein a processing unit is used.

[0042] The method for training the alignment algorithm can be performed within the charged particle beam device, for example, within a processing unit of the charged particle beam device. Additionally or alternatively, at least a portion of the training method can be performed externally. For example, it is conceivable that the method for training the alignment algorithm is performed in the cloud or on an independent computer device. In this case, the charged particle beam device can transmit the data related to the alignment transformation to the cloud or the computer, where the next step of providing the data to the modification algorithm can be performed.

[0043] In an embodiment, the method may include the step of providing stored data acquired during the alignment transformation. This allows, for example, the use of on-policy, off-policy, and offline reinforcement learning algorithms.

[0044] Where relevant, a method of training an alignment algorithm may include one or more of the embodiments as described herein.

[0045] The method may comprise, for example, the steps of determining at least one quality parameter of the stored data and modifying the alignment algorithm based on the at least one quality parameter.

[0046] In one embodiment, the method may include providing a step of storing data relevant to a plurality of alignment conversions of a plurality of charged particle beam devices. This allows data to be collected from a plurality of charged particle beam devices, and these data can be used to modify a single alignment algorithm. In the cloud or on a stand-alone computer, performing the method for training the alignment algorithm, this embodiment is particularly useful. It effectively allows a large amount of data to be provided to a single training algorithm, through which it is possible to train the algorithm effectively.

[0047] In the embodiment of the method of carrying out described method from charged particle beam equipment outside therein, the revised alignment algorithm can be provided back to the corresponding charged particle beam equipment, as the updating of alignment algorithm.The revised alignment algorithm also can be provided to a plurality of charged particle beam equipment as the updating to the corresponding alignment algorithm.The revised alignment algorithm can comprise a modification specific to a plurality of charged particle beam equipment.Therefore, the method as defined herein can comprise the step of providing a plurality of revised alignment algorithms.

[0048] It should be noted that the charged particle beam device can be a physical device or a virtual device. A virtual device, such as a so-called digital twin, can be advantageously used to test modified alignment algorithms and / or generate a large amount of alignment transformation data that can be used to modify the alignment algorithm, for example by training the alignment neural network.

[0049] According to an aspect, there is provided a charged particle beam apparatus comprising a processing unit arranged for performing at least part of the method as defined herein.

[0050] According to another aspect, a method of dealigning a charged particle beam device is provided, which is similar to the method disclosed herein. According to this aspect, a charged particle beam device is provided in a first alignment state, and a processing unit uses a dealignment algorithm to implement an alignment transition from the first alignment state of the charged particle beam device to a second alignment state, wherein the second alignment state is less optimal than the first alignment state.

[0051] In one embodiment, an alignment method as defined herein may be used to bring a charged particle beam apparatus into a more highly aligned state.

[0052] According to one aspect, the alignment algorithm and the misalignment algorithm are used for so-called curriculum training to improve at least one of the agents, i.e., at least one of the alignment algorithm and the misalignment algorithm. According to this aspect, a method of curriculum training is provided, wherein the following steps correspond to a single iteration performed at least once to improve at least one of the agents participating in the curriculum training:

[0053] - providing at least one charged particle beam device;

[0054] - misaligning the at least one charged particle beam device using a misalignment method such as defined herein, wherein at least a first agent is used;

[0055] - aligning the at least one charged particle beam device using an alignment method such as defined herein, wherein at least a second agent is used;

[0056] - determining at least one quality parameter related to at least a portion of an alignment transition obtained during said misalignment and / or alignment transition; and

[0057] - modifying at least one of said first or said second agent based on said at least one determined quality parameter.

[0058] It is noted that additional data may be used in the step of modifying the first or second agent.

[0059] The curriculum training method maintains two groups (also called teams) of agents (algorithms). A first team of agents (also called the alignment team) is trained to align a charged particle beam device, while a second team (also called the misalignment team) is trained to misalign the device. The curriculum training method is a zero-sum game. In each iteration, the alignment team agent performs alignment of the charged particle beam device. A quality parameter related to the performed alignment is calculated. The misalignment team agent brings the charged particle beam device into a state from which the alignment team agent has difficulty performing alignment. The agent can be modified according to the following rule: the value of the alignment team agent's utility function is calculated based on the quality parameter related to the performed alignment. The value of the misalignment team agent's utility function is the negative value of the alignment team agent's utility function. Therefore, the agents are trained in an adversarial manner. The competition between the two teams of agents causes the quality of the actions performed by the agents to gradually improve.

[0060] The agent's modification procedure can be formulated as a reinforcement learning task. In this case, the calculated quality parameter associated with the performed alignment corresponds to the reward provided to the agent. The trainable parameters of the agent's neural network are updated in a way that makes qualitative (high-reward) actions more likely and flawed (low-reward) actions less likely.

[0061] The value of the utility function of the misaligned team agent can be calculated based on a quality parameter associated with the misaligned actions performed by the agent. In this case, the curriculum training method is no longer a zero-sum game.

[0062] Inaccurate teams of agents can be replaced by hard-coded algorithms that do not contain trainable machine learning models. In this case, the curriculum training method is reduced to a single-team, crowd-based training task. In the case where the crowd consists of only a single agent, the curriculum training method is further reduced to the classic reinforcement learning formulation of the single-agent training procedure.

[0063] The swarm can include agents with the same neural network architecture and agents with different architectures. Each agent has unique values ​​for its trainable weights. Therefore, each agent devises a unique strategy (sequence of actions) to align or misalign the charged particle beam device.

[0064] The best trained agents may be selected at the end of the course training procedure and may be utilized additionally to perform alignment of the user charged particle beam device.

[0065] Embodiments of the present disclosure provide a method for aligning charged particle beam devices using a trainable sequential decision-making algorithm (agent). A population of agents is trained according to an iterative curriculum, resulting in a gradual improvement in the quality of actions inferred by the agents for performing alignment. The best trained agents undergo a deployment phase. The agents can also initiate a retraining process to adapt to the peculiarities of the user's charged particle beam device.

[0066] As described herein, data may include provided data and stored data.

[0067] The data provided to the modification algorithm may include an alignment trajectory drawn by an earlier version (i.e., algorithm) of the provided agent, an alignment trajectory drawn by an agent other than the agent provided to the modification algorithm, or an alignment trajectory drawn by a demonstration algorithm. The alignment trajectory may be drawn by interacting with a plurality of different charged particle beam devices. The earlier version of the agent is the state of the agent before one or more modifications performed by the modification algorithm. The demonstrator algorithm is an algorithm capable of inferring an alignment action that is different (and potentially better) than the alignment action inferred by the current version of the provided agent. The demonstrator algorithm may be an agent that utilizes a machine learning model for action reasoning, or alternatively, the demonstrator algorithm may be a hard-coded algorithm that utilizes oracle knowledge to perform alignment.

[0068] The data provided can include human expert demonstrations. This is a person who performs the alignment of the charged particle beam device and whose actions are recorded and stored. In this case, the human expert alignment trajectory can be used during the training process to incorporate human knowledge into the machine learning model that serves as the basis for the agent's reasoning mechanism.

[0069] The data provided to the modification algorithm may also include data generated by an external algorithm for artificial data generation. The external algorithm for artificial data generation is an algorithm capable of generating data without physical interaction with the charged particle beam device.

[0070] The provided data may further comprise a quality parameter corresponding to the alignment state and / or alignment action. In particular, the quality parameter may be regarded as a reward provided to the agent in response to the alignment action inferred by the agent during alignment.

[0071] The structure of the provided data can determine the structure of the modification algorithm. If the provided data is labeled, the modification algorithm can be structured as a supervised learning task; if the data lacks corresponding labels, the modification algorithm can be structured as an unsupervised learning task; if the provided data is partially labeled, the modification algorithm can be structured as an active learning task (i.e., the agent requires a human to label not all data samples, but only those that cause uncertainty during action reasoning); and if the agent is trained on a downstream task, the modification algorithm can be structured as a self-supervised learning task. If the provided data includes quality parameters, the modification algorithm can be structured as a reinforcement learning task. These quality parameters can be considered as rewards provided to the agent in response to alignment actions. If the provided data includes alignment trajectories drawn by the current version of the agent, an on-policy reinforcement learning algorithm can be used to train the agent. If the provided alignment trajectories were drawn by a previous version of the agent, an off-policy reinforcement learning algorithm can be used. If the provided dataset is static (i.e., the agent is not allowed to interact with the environment to collect more training data), the agent can be trained using offline reinforcement learning.

[0072] The invention will now be explained in more detail based on exemplary embodiments and the accompanying schematic drawings, in which:

[0073] Figure 1 A longitudinal sectional view showing a charged particle microscope according to a first embodiment of the present invention;

[0074] Figure 2 A longitudinal sectional view showing a charged particle microscope according to a second embodiment of the present invention;

[0075] Figures 3a-3c An embodiment of the method as disclosed herein is shown;

[0076] Figure 4 An embodiment of a system as disclosed herein is shown.

[0077] Figure 1 (not to scale) is a highly schematic depiction of an embodiment of a charged particle microscope M according to an embodiment of the present invention. More specifically, it shows one embodiment of a transmission type microscope M, which in this case is a TEM / STEM (although, in the context of the present invention, it may effectively just be a SEM (see Figure 2 ), or for example ion-based microscopy). In Figure 1In FIG, within a vacuum housing 2, an electron source 4 generates an electron beam B that propagates along an electron-optical axis B' and passes through an electron-optical illuminator 6 for directing / focusing the electrons onto a selected portion of a sample S (which may, for example, be (locally) thinned / planarized). A deflector 8 is also depicted, which may be used (among other things) to achieve a scanning motion of the beam B.

[0078] The sample S is held on a sample holder H, which can be positioned with multiple degrees of freedom by a positioning device / stage A that moves a carriage A' to which the holder H is (removably) attached; for example, the sample holder H may include fingers that are movable (especially) in the XY plane (see the Cartesian coordinate system depicted; generally, movement parallel to Z and tilting about X / Y are also possible). This type of movement allows different parts of the sample S to be illuminated / imaged / inspected by the electron beam B traveling along axis B' (in the Z direction) (and / or allows a scanning movement to be performed as an alternative to beam scanning). If desired, an optional cooling device (not depicted) can be brought into close thermal contact with the sample holder H, for example, to maintain it (and the sample S thereon) at a low temperature.

[0079] The electron beam B interacts with the sample S in such a way that various types of "stimulated" radiation are emitted from the sample S, including, for example, secondary electrons, backscattered electrons, X-rays, and optical radiation (cathodoluminescence). If desired, one or more of these radiation types can be detected by means of an analysis device 22, which can be, for example, a combined scintillator / photomultiplier tube or an EDX or EDS (energy dispersive X-ray spectrometer) module; in this case, an image can be constructed using essentially the same principles as in an SEM. Alternatively or additionally, however, electrons that pass through the sample S, exit / emit therefrom, and continue to propagate along the axis B' (essentially, though typically with some deflection / scattering) can be studied. This transmitted electron flux enters an imaging system (projection lens) 24, which typically includes various electrostatic / magnetic lenses, deflectors, correctors (such as an anastigmator), etc. In normal (non-scanning) TEM mode, this imaging system 24 can focus the transmitted electron flux onto a phosphor screen 26, which can be retracted / withdrawn if necessary (as schematically indicated by arrow 26') so as to be away from axis B'. An image (or diffraction pattern) of (a portion of) sample S will be formed by imaging system 24 on screen 26, and this can be viewed through a viewing port 28 located in a suitable portion of the wall of housing 2. The retraction mechanism of screen 26 can, for example, be mechanical and / or electrical in nature and is not depicted here.

[0080] Instead of observing the image on the screen 26, one can instead take advantage of the fact that the depth of focus of the electron flux leaving the imaging system 24 is typically large (e.g., about 1 meter). Therefore, various other types of analytical equipment can be used downstream of the screen 26, such as:

[0081] - TEM camera 30. At camera 30, the electron flux can form a static image (or diffraction pattern), which can be processed by controller / processor 20 and displayed on display device 14, such as a flat-panel display. When not needed, camera 30 can be retracted / withdrawn (as schematically indicated by arrow 30') to move it away from axis B'.

[0082] - STEM camera 32. The output from camera 32 can be recorded as a function of the (X, Y) scanning position of beam B on sample S, and an image can be constructed as a "map" of the output from camera 32 as a function of X, Y. Camera 32 can contain a single pixel of, for example, 20 mm in diameter, as opposed to the matrix of pixels typically found in camera 30, but camera 32 can also be an electron microscope pixel array detector (EMPAD). In addition, the acquisition rate of camera 32 (for example, 10 per second) can be very high. 6 points) will typically be faster than a camera 30 (e.g., 10 per second 2 Again, when not needed, the camera 32 can be retracted / withdrawn (as schematically indicated by arrow 32') so as to be out of the way of the axis B' (although such retraction is not necessary in the case of, for example, a donut-shaped annular dark field camera 32; in such cameras, the central aperture will allow flux to pass when the camera is not in use).

[0083] As an alternative to imaging using cameras 30 or 32 , a spectroscopic device 34 can also be called upon, which can be, for example, an EELS module.

[0084] It should be noted that the order / position of items 30 , 32 and 34 is not critical and many possible variations are envisioned. For example, spectroscopic device 34 may also be integrated into imaging system 24 .

[0085] In the illustrated embodiment, microscope M additionally includes a retractable X-ray computed tomography (CT) module, generally indicated by the reference numeral 40. In computed tomography (also known as tomographic imaging), a source and (diametrically opposed) detectors are used to observe a sample along different lines of sight in order to obtain penetrating views of the sample from a variety of angles.

[0086] Note that a controller (computer processor) 20 is connected to the various illustrated components via control lines (bus) 20'. This controller 20 can provide a variety of functions, such as synchronizing actions, providing setpoints, processing signals, performing calculations, and displaying messages / information on a display device (not depicted). It goes without saying that the (schematically depicted) controller 20 can be (partially) internal or external to the housing 2 and can have a unitary or modular structure as desired. As shown in this embodiment, the controller comprises a data processing device P arranged to carry out the method as defined herein.

[0087] Those skilled in the art will appreciate that the interior of the housing 2 need not be maintained under a strict vacuum; for example, in so-called "ambient TEM / STEMs," a background atmosphere of a given gas is intentionally introduced / maintained within the housing 2. The skilled person will also appreciate that, in practice, it may be advantageous to limit the volume of the housing 2 so that it substantially surrounds the axis B', if possible, in the form of a small tube (e.g., approximately 1 cm in diameter) through which the electron beam used passes, but widened to accommodate structures such as the source 4, sample holder H, screen 26, camera 30, camera 32, spectroscopic equipment 34, and the like.

[0088] Now refer to Figure 2 , shows another embodiment of the device according to the present invention. Figure 2 (not to scale) is a highly schematic depiction of a charged particle microscope M according to the invention; more precisely, it shows an embodiment of a non-transmission type microscope M, which in this case is a SEM (although, in the context of the present invention, it may just be effectively an ion-based microscope, for example). Figure 1 Corresponding parts of the items in the are indicated by the same reference numerals and are not discussed separately here. Figure 1 Among them are (especially) the following components:

[0089] 2a: a vacuum port which can be opened to introduce / remove objects (components, samples) into / from the interior of the vacuum chamber 2 or to which, for example, auxiliary devices / modules can be mounted. If necessary, the microscope M can comprise a plurality of such ports 2a;

[0090] - 10a, 10b: schematic representations of lenses / optical elements in the illuminator 6;

[0091] - 12: voltage source, allowing, if necessary, the sample holder H or at least the sample S to be biased (floated) to a potential relative to ground;

[0092] - 14: Display, such as FPD or CRT;

[0093] 22a, 22b: Segmented electron detector 22a, comprising a plurality of independent detection segments (e.g., quadrants) arranged around a central aperture 22b (which allows the passage of beam B). This type of detector can be used, for example, to study (the angular dependence of) the output (secondary or backscattered) flux of electrons emitted from the sample S.

[0094] Also present herein is a controller 20. The controller is connected to a display 14, and the display 14 can be connected to a data processing device P arranged to carry out the methods as defined herein. In the embodiment shown, the data processing device P is a separate structure that does not form part of the controller, and does not even form part of the microscope P. The data processing device P can be local or cloud-based and is, in principle, not restricted to any location. Note that in all embodiments described herein, the data processing unit P can be part of the charged particle beam device (e.g., a charged particle microscope) or can be located externally.

[0095] Now turn Figure 3a , schematically illustrating an embodiment of a method as defined herein. The method comprises the following steps: providing 101 a charged particle beam device in a first alignment state; implementing an alignment transition from the first alignment state to a second alignment state of the charged particle beam device by a processing unit P using an alignment algorithm 102; and providing 103 data related to the alignment transition to a modification algorithm to modify the alignment algorithm in order to implement a modified alignment transition. In the embodiment shown, the method comprises the step of modifying 104 the alignment algorithm by the processing unit P to obtain a modified alignment algorithm.

[0096] In one embodiment, the step of using 102 the alignment algorithm by the processing unit P may be performed by the processing unit P as part of the charged particle beam device. The step of providing 103 data related to the alignment transformation to the modification algorithm is at least partially performed in the charged particle beam device and may include the step of providing data external to the charged particle beam device. The step of modifying 104 the alignment algorithm may be completed within the charged particle beam device, but may also be performed in the cloud or on an external computing device.

[0097] Now turn Figure 3b , shows another embodiment of the method. Figure 3a The embodiment shown is comparable, but now comprises the additional step of feeding back 105 the modified alignment algorithm (as obtained in step 104 ) to the charged particle beam device and using 102 the (now modified) alignment algorithm to establish another alignment transformation.

[0098] Now turn Figure 3c , shows another embodiment of the method. Figure 3aThe embodiment shown is comparable. However, here the step of using the alignment algorithm 102 is shown to include the sub-step of providing 102a-102c a plurality of individual alignment actions. Here a total of three individual alignment actions 102a-102c are shown, but it will be clear to a person skilled in the art that in principle any number of individual alignment actions is conceivable.

[0099] exist Figures 3a-3c In all the embodiments shown, the method may comprise the step of determining a quality parameter. Figure 3c In the illustrated embodiment, the step of determining the quality parameter includes the step of evaluating one or more of the plurality of individual alignment actions. The evaluation may include identifying individual alignment actions that result in an undesirable alignment state. The method may include the step of modifying the alignment algorithm to prevent undesirable alignment actions during use of the modified alignment algorithm.

[0100] As previously mentioned, the alignment algorithm may include at least one neural network, and the step of modifying 104 the algorithm may include the step of retraining the at least one neural network. This may include the step of retraining the at least one neural network using one or more of the undesired alignment states.

[0101] In the illustrated embodiment, the step of providing a charged particle beam device may include the step of setting the charged particle beam device to an undesirable alignment state. An alignment algorithm (modified or unmodified) may then be used to implement an alignment switch, or another alignment switch. It is contemplated that a training algorithm may be used to identify and / or set the undesirable alignment state. The training algorithm may comprise a neural network.

[0102] Now turn Figure 4 , shows a system 201 as defined herein. The system 201 comprises a charged particle beam device CPBA, such as a charged particle microscope M, having an alignment algorithm AA for aligning the CPBA. The CPBA is connected to a modification algorithm MA. The connection may be local, such as in the device itself, or external, such as a wireless connection to a cloud-based modification algorithm MA. The modification algorithm may modify the alignment algorithm AA of the CPBA based on data received by the CPBA and return the modified algorithm to the CPBA. A training algorithm TA may be connected to the modification algorithm MA and / or the CPBA. Based on the received data, the training algorithm TA is designed to select and / or set alignment states that are relatively difficult for the alignment algorithm to resolve. Another alignment transformation may then be performed and subsequent data may be used again to check whether further modifications are required.

[0103] like Figure 4The CPBA shown may be a system of multiple CPBAs. It is conceivable that multiple CPBAs, each having an (optional) alignment algorithm AA, are used to provide data to the modification algorithm MA. The data can be used to update or modify the existing alignment algorithm AA of one or more CPBAs.

[0104] like Figure 4 The CPBA shown can be a physical device or a virtual device such as a digital twin. Figure 4 The CPBA shown may include at least one physical CPBA and at least one virtual device. The at least one virtual device may be used to accelerate training and / or modification of an alignment algorithm. The alignment algorithm and / or the training algorithm may be a neural network. In particular, the alignment algorithm may be modified based on deep reinforcement learning.

[0105] As will be clear from the above description of the accompanying drawings, the method as defined herein involves training an alignment algorithm for aligning a charged particle beam device, comprising the following steps:

[0106] - Provide the alignment algorithm AA to be trained;

[0107] - providing data relating to an alignment transition of the charged particle beam device, wherein the alignment transition extends from a first alignment state to a second alignment state; and

[0108] - providing said data to a modification algorithm MA for modifying said alignment algorithm, wherein a processing unit P is used.

[0109] From the above description of the drawings, it is clear that a charged particle beam apparatus, such as a charged particle microscope, is arranged to participate in the method described herein.

[0110] The desired protection is afforded by the following claims.

Claims

1. A method for aligning a charged particle beam apparatus, comprising the steps of: - providing a charged particle beam apparatus in a first alignment state; - implementing, by a processing unit, an alignment transition from said first alignment state to a second alignment state of said charged particle beam device using an alignment algorithm; - providing data related to the alignment transformation to a modification algorithm to modify the alignment algorithm so as to implement a modified alignment transformation using the modified alignment algorithm; wherein the first alignment state and the second alignment state are states describing one or more parameters and operations of the charged particle beam device for carrying out an intended use of the charged particle beam device; wherein said data related to said alignment transformation comprises settings, speed, timing and / or calibration results of said charged particle beam device; Wherein the modification algorithm is a training algorithm for training a neural network and the alignment algorithm is a trainable decision algorithm arranged to infer a sequence of alignment actions.

2. The method according to claim 1, comprising the step of using the modified alignment algorithm to bring the charged particle beam device to a further alignment state. The method according to claim 1 , wherein the alignment transformation comprises a plurality of separate alignment actions.

4. The method according to claim 1 or 2, comprising the step of determining a quality parameter, wherein the quality parameter is a penalty / reward mechanism for training the neural network.

5. The method of claim 4, wherein the step of determining the quality parameter comprises evaluating one or more of a plurality of individual alignment actions. The method of claim 5 , wherein the evaluating comprises identifying individual alignment actions that cause an undesired alignment state.

7. The method of claim 6, comprising the step of modifying the alignment algorithm such that undesired alignment actions are prevented during use of the modified alignment algorithm.

8. The method of claim 1 or 2, wherein the alignment algorithm comprises at least one neural network.

9. The method of claim 8, wherein the step of modifying the algorithm comprises the step of retraining the at least one neural network.

10. The method of claim 9, comprising the step of retraining the at least one neural network using one or more of the undesired alignment states.

11. A method according to claim 10, comprising the steps of setting the charged particle beam apparatus to an undesired alignment state and using the modified alignment algorithm to bring the charged particle beam apparatus to another alignment state.

12. The method according to claim 1 or 2, wherein a training algorithm is used to identify and / or set undesired alignment states.

13. The method of claim 12, wherein the training algorithm comprises a neural network.

14. The method of claim 1 or 2, wherein the charged particle beam apparatus is a virtual device.

15. A method for training an alignment algorithm for aligning a charged particle beam device, comprising the steps of: - provide an alignment algorithm to be trained; - providing data relating to an alignment transition of the charged particle beam device, wherein the alignment transition extends from a first alignment state to a second alignment state; as well as - providing said data to a modification algorithm for modifying said alignment algorithm, wherein a processing unit is used; wherein the first alignment state and the second alignment state are states describing one or more parameters and operations of the charged particle beam device for carrying out an intended use of the charged particle beam device; wherein said data related to said alignment transformation of said charged particle beam device comprises settings, speed, timing and / or calibration results of said charged particle beam device; Wherein the modification algorithm is a training algorithm for training a neural network and the alignment algorithm is a trainable decision algorithm arranged to infer a sequence of alignment actions.

16. The method according to claim 15, comprising the steps of determining at least one quality parameter of the data, and modifying the alignment algorithm based on the at least one quality parameter, wherein the quality parameter is a penalty / reward mechanism for training a neural network.

17. A method according to claim 15 or 16, comprising the step of providing stored data relating to a plurality of alignment transitions of a plurality of charged particle beam devices.

18. The method of claim 15 or 16, wherein the charged particle beam apparatus is a virtual device.

19. Charged particle beam apparatus comprising a processing unit arranged to perform the method according to any one of the preceding claims.

Images