A deep learning-based fully-automatic scanning tunneling microscope control method

Abstract

The application discloses a kind of full-automatic scanning tunneling microscope control methods based on deep learning, mainly STM is combined with deep learning, is composed of three parts of probe classification module based on standard sample, probe tip repair module based on reinforcement learning, automatic scanning module, first control probe approaches standard sample, scanning imaging is carried out;Second, the image after scanning is sent into probe classification module, to output the state of probe tip;If state is not good, then send into probe tip repair module, otherwise enter automatic scanning module, scanning imaging is carried out to unknown sample.The application can realize full-automatic scanning to unknown sample, improve scanning efficiency;Full-automatic regional scanning is carried out, and full coverage to unknown sample surface scanning is realized;Probe tip state is judged automatically and repaired probe, reduce repetitive physical labor, reduce the technical requirements of STM to operator.

Description

Technical Field

[0001] This invention relates to scanning tunneling microscopes, and more particularly to a fully automated scanning tunneling microscope control method based on deep learning. Background Technology

[0002] The basic structure of a scanning tunneling microscope (STM) includes a stepper motor, a scanning head, a preamplifier, and related electronic control units and supporting control software. Its working principle is based on the tunneling effect in quantum mechanics. The metal probe and the sample are considered as two electrodes. When the probe tip is very close to the sample surface, under the influence of an applied electric field, electrons pass through the space between them, forming a tunneling current. The changes in the tunneling current between the probe tip and the sample surface are recorded to characterize information about the sample surface.

[0003] Because the scanning area of ​​STM is very small (typically only on the order of hundreds of nanometers), less than one hundred millionth of the entire sample surface, many important areas are often missed during the scanning process. Furthermore, STM imaging relies heavily on operator experience, and when scanned images are poor, the lack of quantitative classification and evaluation criteria easily leads to misjudgments and the blind replacement of samples or probes. Such misjudgments can sometimes be fatal, potentially causing the miss of an important microscopic phenomenon. Based on these considerations, if the traditional STM scanning process could be combined with the currently popular deep learning technology to achieve full-process automation of the STM scanning process, the probability and number of regions of interest scanned by STM could be greatly improved.

[0004] Currently, the integration of deep learning and STM is not yet deep, and it still only combines a certain process in traditional STM scanning with deep learning, which is far from true end-to-end intelligent scanning. Even if some international groups have achieved end-to-end automated scanning of a sample, it is still based on a large amount of labeled data for that sample. For example, a binary classification model trained using CNN and four other fully connected networks can automatically identify single-tip and double-tip states in STM with an accuracy of up to 97%. Based on this, scientists have expanded the range of tip types and conducted automated research on single-crystal surfaces of gold and copper, creating STM imaging work that is closer to end-to-end intelligent scanning. This work can complete all steps autonomously except for the initial collection of 7,600 images of the sample, including scanning area selection, automatic image processing, tip state judgment and modification, and shows 86 hours of fully automated scanning images. However, this work can only be applied to MgPc / Ag(100) samples. After changing to other samples, it is still necessary to collect data and train the model again, which is obviously time-consuming and laborious. Summary of the Invention

[0005] Purpose of the invention: The purpose of this invention is to provide a fully automated scanning tunneling microscope control method based on deep learning.

[0006] Technical solution: The present invention includes the following steps:

[0007] (1) Control the probe to approach the standard sample and perform scanning imaging;

[0008] (2) The scanned image is sent to the probe classification module to output the state of the probe tip;

[0009] (3) If the probe tip is not in good condition, it is sent to the tip repair module;

[0010] (4) If the probe tip is in good condition, it will enter the automatic scanning module to scan and image the unknown sample.

[0011] Further, step (1) uses probes in different states to scan and image the standard sample and collect image data. The probes in different states include blunt probes, multi-tip probes, contaminated probes, intact and sharp probes, and probes with uneven surfaces.

[0012] Furthermore, step (2) uses a convolutional neural network model to classify and train the dataset, thereby automatically judging the probe status.

[0013] Furthermore, step (3) employs a reinforcement learning strategy to repair the needle tip using a field emission method.

[0014] Furthermore, the field emission method considers factors including the depth of penetration into the standard sample and pulse signals of different voltages.

[0015] Furthermore, the field emission method includes:

[0016] (3.1) Test the insertion depth and the bias applied to the sample to obtain the repair results of different needle tips;

[0017] (3.2) Construct an action list based on the repair status. The CNN selects a repair action from the predefined action list based on the needle tip status, and provides feedback and accumulates it after the action is executed.

[0018] (3.3) If the probe is in a bad state, a reward of -1 is accumulated and then the CNN predicts the repair action based on the state to perform the repair. This process is repeated until the probe is in a good state.

[0019] (3.4) If the probe status is good, then the cumulative reward of +1 is fed back, and the repair is terminated.

[0020] Furthermore, in step (3.2), the feedback after the action is completed is obtained by classifying the processed probe by scanning the standard sample to obtain the probe status.

[0021] Further, step (4) includes:

[0022] (4.1) Divide the unknown sample into several regions, and drive the probe to scan from the upper left corner of the unknown sample, scanning sequentially from left to right;

[0023] (4.2) The current state of the needle tip is determined by scanning the standard sample. If the needle tip is not in good condition, it is sent to the needle tip repair module. Otherwise, the image of the upper left corner area of ​​the scanned unknown sample is saved.

[0024] (4.3) Determine whether all areas of the unknown sample from the previous scan have been scanned. If not, replace the unknown sample from the previous scan with the one from the previous scan and continue scanning the next area. Then make a judgment on the needle tip and repeat this process.

[0025] (4.4) After scanning all areas of the surface of an unknown sample, automatically replace the next unknown sample and repeat steps (4.1)-(4.3).

[0026] Furthermore, after the probe in step (4.1) moves to a new line, it still scans sequentially from left to right.

[0027] Furthermore, the condition for determining the needle tip state by scanning the standard sample in step (4.2) is after scanning the upper left corner area twice for the first time.

[0028] Beneficial effects: Compared with the prior art, the present invention has the following advantages: it can achieve fully automatic scanning of unknown samples, improving scanning efficiency; it can achieve full coverage of the surface of unknown samples through fully automatic regional scanning; it can automatically judge the tip status and repair the probe, reducing repetitive manual labor and lowering the technical requirements of STM operators; when the image quality is poor, it can automatically determine whether the problem comes from the tip or the sample. Attached Figure Description

[0029] Figure 1 This is a flowchart of the present invention;

[0030] Figure 2 Flowchart of the probe classification module;

[0031] Figure 3 Here is a flowchart of the needle tip repair module;

[0032] Figure 4 Here is the flowchart for the automatic scanning module;

[0033] Figure 5 A classification structure diagram for the dataset;

[0034] Figure 6 This is a schematic diagram of a multi-sample rapid replacement structure. Detailed Implementation

[0035] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0036] This invention discloses a fully automated scanning tunneling microscope (STM) control method based on deep learning. It primarily combines STM with deep learning and consists of three parts: a probe classification module based on standard samples, a tip repair module based on reinforcement learning, and an automatic scanning module. First, the probe is controlled to approach the standard sample for scanning and imaging. The scanned image is then sent to the probe classification module to output the probe tip status. If the probe status is poor, it is sent to the tip repair module; otherwise, it enters the automatic scanning module to scan and image unknown samples, as shown in the attached diagram. Figure 1 As shown.

[0037] Standard samples were scanned and imaged using probes in various states, including blunt-tipped probes, multi-tipped probes, contaminated probes, intact and sharp probes, and probes with uneven surfaces. A convolutional neural network model was trained to classify the dataset, thereby automatically determining the probe state, as shown in the attached diagram. Figure 2 As shown.

[0038] A reinforcement learning strategy is employed to repair the needle tip using a field emission method. (See attached image.) Figure 3 As shown, field emission probe repair needs to consider factors such as the depth of penetration into the standard or third sample, and pulse signals with different voltages. The probe tip state does not have a clear correspondence with these factors; therefore, reinforcement learning is used to process the probe tip. The penetration depth and bias voltage applied to the sample in the field emission method are tested to obtain the repair results of different repair schemes. Then, an action list is constructed based on the repair results. The CNN selects a repair action from the predefined action list based on the probe tip state. After the action is executed, feedback is provided and accumulated. The feedback is obtained by scanning the standard sample with the processed probe, classifying it, and obtaining the probe state. If the probe state is poor, the accumulated reward is -1. Then, the CNN predicts the repair action based on the state and performs the repair, repeating this cycle until the probe state improves. When the probe state is good (i.e., intact and sharp), the accumulated reward is +1, and the repair is terminated. The probe is then used to scan unknown samples.

[0039] The unknown sample is divided into several regions. The driven probe starts scanning from the upper left corner of the unknown sample, proceeding sequentially from left to right. After each line break, the scanning continues from left to right. (See attached image) Figure 4As shown, after scanning the upper left corner region twice in the first scan, the tip condition needs to be determined by scanning a standard sample. If the tip condition is poor, it is sent to the tip repair module; otherwise, the image of the upper left corner region of the scanned unknown sample is saved. Then, it is determined whether all areas of the previously scanned unknown sample have been scanned. If not, the previously scanned unknown sample is used to continue scanning the next region, and the tip condition is determined again, thus repeating the cycle. After scanning all areas of the surface of an unknown sample, the next unknown sample is automatically selected, and the above steps are repeated.

[0040] One embodiment of the present invention is as follows:

[0041] The first step involved creating a dataset by scanning a standard sample of highly oriented pyrolytic graphite (HOPG) using probes of different states. Specifically, the probes used included intact, sharp probes, blunt probes, multi-point probes, and contaminated probes, resulting in 1937 images. These scanned images were then labeled according to the probe state. The approximate proportion of these four types of probe scans in the dataset was 3:3:2:2. The dataset was then divided into training, validation, and test sets in a 6:2:2 ratio. After the dataset was created, a convolutional neural network was used to train the dataset for classification, as shown in the attached diagram. Figure 5 As shown, this enables the network model to automatically determine the state of the probe based on the scanned image.

[0042] The second step involves selecting a sample switching method. When scanning standard samples, a third sample specifically for repairing needle tips, or unknown samples, rapid sample switching between different samples is required. To address this, a laboratory high-throughput technology (patent number: ZL202010341820.2) is employed to switch between different scanning samples. The specific device is as follows: Figure 6 As shown, five samples are placed on the sample stage: the standard sample in the middle, the third sample in the upper left corner, and the remaining three are unknown samples.

[0043] The third step is to construct a reinforcement learning action library, as shown in Table 1. All actions are defined based on the z-position of the probe above the sample, determined by a set point (V_bias = 1V, I_t = 25pA, applying bias voltage to the sample).

[0044] Table 1 Reinforcement Learning Action Library

[0045]

[0046] Then, based on the Q-learning algorithm, the reinforcement learning used in this method is initially configured, and the Q-table, i.e., probe states and repair actions, is initialized. When faced with the three bad probe states from the first step, actions are randomly selected and executed from the action library, and then verified. The probe scans a standard sample, and the scanned image is input into the classification module to obtain the probe state. If the probe state is still bad, a reward of -1 is given, the Q-table is updated, and repair is repeated. The CNN determines whether this state has been experienced before. If not, a repair action is randomly selected; if so, the optimal action is selected. After repair, the probe state needs to be verified again, and this process is repeated. This continues until the probe state after repair is verified to be a good, sharp state, at which point a reward of +1 is given, the Q-table is updated, and repair is stopped.

[0047] The fourth step involves starting the process. After the probe scans the standard sample, the scanned image is fed into the classification network. If the probe is found to be blunt, the probe enters the tip repair module. The standard sample under the probe is automatically replaced with the third sample specifically designed for tip repair in the upper left corner of the sample stage. Then, through reinforcement learning, field emission tip repair is performed. The probe state is fed into the CNN to predict the optimal action, with an insertion depth of 3nm and a bias voltage of 0.02V applied for repair. After tip repair, the sample is replaced with the standard sample, and the repaired probe is used to scan the standard sample. The scanned image is then used by the classification network to determine the current probe state. If the probe is not yet in a good and sharp state, a reward of -1 is given, and the Q-table of reinforcement learning is updated. The probe then enters the tip repair module again. Since the probe state after the first tip repair has not been experienced, a repair action is randomly selected, with an insertion depth of 0nm and a bias voltage of 2V applied. After tip repair, the probe is replaced with the standard sample again, and the probe after the second tip repair is used to scan the image. If the classification module determines that the probe state is in a good and sharp state, a reward of +1 is given, the Q-table is updated, and the repair is stopped, entering the automatic scanning module.

[0048] In the fifth step, the surface area of ​​the unknown sample in the automatic scanning module has been divided into 1000*1000 blocks. During scanning, the pattern follows a left-to-right and top-to-bottom scanning order. To begin, the unknown sample in the upper right corner of the sample stage is used, and the probe scans the first upper left corner area twice. Then, a standard sample is used for scanning. The current probe status is obtained through the classification module. If the probe status is good, the scanned image is saved. If the probe status is poor, the repair module is entered, and the repair steps in step three are followed. After repair, the next area is scanned twice, and the above steps are repeated. This continues until the entire surface area of ​​an unknown sample has been scanned. Then, the next unknown sample is used, and the automatic scanning steps are repeated to complete the scanning and imaging of the unknown sample. This achieves fully automated processing of the scanning tunneling microscope, including probe identification, repair, and scanning of the sample to be tested.

Claims

1. A fully automated scanning tunneling microscope control method based on deep learning, characterized in that, Includes the following steps: (1) Control the probe to approach the standard sample and perform scanning imaging; (2) The scanned image is sent to the probe classification module to output the state of the probe tip; (3) If the probe tip is not in good condition, it is sent to the tip repair module; (4) If the probe tip is in good condition, it will enter the automatic scanning module to scan and image the unknown sample; Step (3) employs a reinforcement learning strategy to repair the needle tip using a field emission method; The field emission method takes into account factors including the depth of penetration into the standard sample and pulse signals of different voltages. The field emission method includes: (3.1) The insertion depth and the bias applied to the sample were tested to obtain the repair results for different needle tips; (3.2) Construct an action list based on the repair status. The CNN selects a repair action from the predefined action list based on the needle tip status. Feedback is given and accumulated after the action is executed. (3.3) If the probe is in a bad state, a reward of -1 is accumulated and then the CNN predicts the repair action based on the state to perform the repair. This process is repeated until the probe is in a good state. (3.4) If the probe status is good, then the cumulative reward of +1 is fed back, and the repair is terminated.

2. The fully automated scanning tunneling microscope control method based on deep learning according to claim 1, characterized in that, Step (1) involves scanning and imaging a standard sample using probes in different states to collect image data. The probes in different states include blunt probes, multi-tip probes, contaminated probes, intact and sharp probes, and probes with uneven surfaces.

3. The fully automated scanning tunneling microscope control method based on deep learning according to claim 1, characterized in that, Step (2) uses a convolutional neural network model to classify and train the dataset, thereby automatically judging the probe status.

4. The fully automated scanning tunneling microscope control method based on deep learning according to claim 1, characterized in that, In step (3.2), the feedback after the action is completed is obtained by classifying the standard sample after scanning the processed probe to obtain the probe status.

5. The fully automated scanning tunneling microscope control method based on deep learning according to claim 1, characterized in that, Step (4) includes: (4.1) Divide the unknown sample into several regions, and drive the probe to scan from the upper left corner of the unknown sample, scanning sequentially from left to right; (4.2) The current state of the needle tip is determined by scanning the standard sample. If the needle tip is not in good condition, it is sent to the needle tip repair module. Otherwise, the image of the upper left corner area of ​​the scanned unknown sample is saved. (4.3) Determine whether all areas of the unknown sample from the previous scan have been scanned. If not, replace it with the unknown sample from the previous scan and continue scanning the next area. Then make a judgment on the needle tip and repeat this process. (4.4) After scanning all areas of the surface of an unknown sample, automatically replace the next unknown sample and repeat steps (4.1)-(4.3).

6. The fully automated scanning tunneling microscope control method based on deep learning according to claim 5, characterized in that, After the driving probe moves to the next line in step (4.1), it still scans sequentially from left to right.

7. The fully automated scanning tunneling microscope control method based on deep learning according to claim 5, characterized in that, The condition for determining the needle tip state by scanning the standard sample in step (4.2) is after scanning the upper left corner area twice in the first scan.

Images