Concentration detection method and system based on combination of STM-BJ technology and deep learning, and storage medium

Abstract

The invention discloses a concentration detection method and system based on combination of an STM-BJ technology and deep learning, and a storage medium. The method comprises the following steps: S1, collecting conductivity data of a pure probe molecule solution and a mixed solution of probe molecules and substrate molecules; s2, training a pre-constructed supervised deep learning model by adopting the data; s3, collecting conductance data of a mixed solution under different substrate molecule concentrations, and inputting the conductance data into the trained model to obtain the proportion of the conductance data determined to contain substrate molecules under each concentration to the total conductance data under the concentration; s4, fitting and establishing a standard relation curve between the molecular concentration of the substrate and the proportion; s5, collecting the conductivity data of a solution to be detected, inputting the conductivity data into the model, and substituting the identified proportion of the conductivity data containing the substrate molecules in the total number of the conductivity data of the solution to be detected into the standard relation curve to calculate the concentration of the substrate molecules in the solution to be detected; according to the invention, the concentration of an object to be detected can be detected at a second level and an extremely low concentration.

Description

Technical Field

[0001] This invention relates to the detection of substance concentration, and in particular to a concentration detection method, system, and storage medium based on STM-BJ technology combined with deep learning. Background Technology

[0002] Single-molecule sensing is required for detection at extremely low concentrations. Single-molecule sensing is supported by two fundamental pillars: a methodological dimension, requiring real-time, event-resolved interrogation of individual molecules; and a performance dimension, demanding detection sensitivity close to the ultimate physical limit, typically at sub-anomolar concentrations. However, even if both dimensions are met simultaneously, only sub-anomolar concentrations can be detected, and the signal intensity generated by a single molecule is weak, requiring long signal acquisition and averaging processes, resulting in long detection times.

[0003] Intermolecular interactions, such as hydrogen bonds, π-π stacking, and metal ligand coordination, provide an intrinsic recognition pathway that directly couples molecular identity with measurable electrical signals, enabling detection at lower concentrations. However, at low analyte concentrations, these interactions become sparse, making them difficult to capture using traditional methods and hindering effective detection at even lower concentrations. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide a concentration detection method, system, and storage medium based on STM-BJ technology combined with deep learning, which can achieve analyte concentration detection at the second level and at extremely low concentrations.

[0005] Technical solution: The concentration detection method based on STM-BJ technology combined with deep learning described in this invention includes the following steps:

[0006] S1. The conductivity data of pure probe molecule solutions and mixed solutions of probe molecules and substrate molecules at preset concentrations were collected using the STM-BJ measurement system.

[0007] S2. Use the collected conductivity data as a training set to train a pre-built supervised deep learning model. This model extracts the time-domain and frequency-domain features of the conductivity data to identify whether the solution corresponding to the conductivity data contains substrate molecules.

[0008] S3. Collect conductivity data of mixed solutions at different substrate molecule concentrations, input them into the trained model, and obtain the proportion of conductivity data at each concentration that is determined to contain substrate molecules to the total number of conductivity data at that concentration.

[0009] S4. Based on the known substrate molecule concentration and its corresponding ratio obtained in S3, fit and establish a standard relationship curve between the substrate molecule concentration and the ratio.

[0010] S5. Collect the conductivity data of the test solution and input it into the trained model. Based on the proportion of conductivity data containing substrate molecules to the total conductivity data of the test solution, substitute it into the standard relationship curve to calculate the concentration of substrate molecules in the test solution.

[0011] This method trains a supervised deep learning model using a training set, enabling it to identify the presence of a substrate molecule in a solution based on conductivity data. The trained model then identifies which data points in the conductivity data of multiple mixed solutions with known substrate concentrations contain substrate molecules, calculates the proportion of these substrate-containing data points, and fits a standard relationship curve between substrate molecule concentration and the proportion of substrate-containing data in the conductivity data. The conductivity data of the test solution is then input into the model, and the concentration of the substrate in the test solution is obtained by substituting the data results into the standard relationship curve. Because the supervised deep learning model extracts not only time-domain features from conductivity data but also frequency-domain features, compared to existing clustering algorithms that only perform time-domain analysis, this method's model, which integrates multi-domain features, significantly improves the accuracy of the final classification. This is especially beneficial for situations where molecular interaction events are sparse and transient at extremely low concentrations, making them difficult to detect. The supervised deep learning model can sensitively identify minute changes in conductivity characteristics caused by specific weak interactions between probe and substrate molecules, enabling this method to detect extremely low concentrations. Meanwhile, because STM-BJ technology can acquire data at high speed and deep learning models can perform rapid inference, the entire detection process can be completed in seconds, achieving a major technological leap from "undetectable" extremely low concentrations to "fast and accurate detection", providing a brand-new technical path for real-time online ultra-trace analysis.

[0012] Preferably, there is a weak non-covalent interaction between the probe molecule and the substrate molecule, the probe molecule can measure conductivity, and the change in conductivity before and after the interaction is greater than a preset value.

[0013] Preferably, the supervised deep learning model includes a cascaded dual-channel encoder, a dual-path projector, and a downstream classifier. The dual-channel encoder includes a time-domain encoder and a frequency-domain encoder, and the dual-path projector includes two independent projectors adapted to the two encoders.

[0014] Through a multi-domain feature fusion architecture, the molecular information contained in conductivity data is comprehensively and deeply mined. Traditional single-path models may only analyze time-domain morphology, easily neglecting rhythmic and characteristic frequency information in the frequency domain. The dual-channel encoder (time domain + frequency domain) used in this invention can simultaneously capture the dynamic morphological evolution of molecular events (time domain) and the underlying energy distribution and periodicity features (frequency domain), providing richer feature representations. The dual-path projector is responsible for compressing and mapping high-dimensional features to a lower-dimensional space more suitable for classification, while the downstream classifier ultimately completes high-precision judgment. This cascaded and divide-and-conquer structure enables the model to robustly extract weak features related to molecular binding events from the original conductivity signal with high noise and complex background.

[0015] As a preferred option, the time-domain encoder employs a cascaded multi-layer one-dimensional convolutional neural network to process the original conductance signal, with the size of each convolutional kernel decreasing progressively to capture the dynamic form of molecular events.

[0016] Frequency domain encoder: The frequency domain features are extracted by processing the amplitude spectrum obtained by the fast Fourier transform of the original conductance signal using a cascaded multi-layer one-dimensional convolutional neural network. The size of each convolution kernel decreases progressively, and its initial convolution kernel size is smaller than that of the time domain encoder.

[0017] Dual-path projector: includes a one-dimensional convolutional layer, a normalization layer, and an adaptive average pooling layer to compress features into a projection vector of a preset dimension;

[0018] Downstream classifier: consists of a first fully connected layer, an activation function layer, and a second fully connected layer connected in sequence. The second fully connected layer outputs the classification result.

[0019] Preferably, the downstream separator is provided with a regularization layer before the first fully connected layer for regularizing the spliced ​​time-domain projection vector and frequency-domain projection vector.

[0020] Preferably, the supervised deep learning model training employs a composite loss function:

[0021]

[0022] in, For the total loss function, To monitor the classification of losses, The cross-domain consistency loss is represented by λ, which is a weighting factor.

[0023] The expression is

[0024]

[0025] in, and Let represent the time-domain and frequency-domain projection vectors of the i-th sample, respectively, and B represent the batch size.

[0026] Through the guiding role of the composite loss function, the model is guided to learn more essential and consistent molecular feature representations. This loss function includes not only the supervised classification loss that drives the model to classify correctly. It also innovatively introduces cross-domain consistency loss. The latter approach forces the model to semantically maintain consistency in the feature representations learned from the time and frequency domains by minimizing the distance between the time-domain and frequency-domain projection vectors of the same sample. This means that the model is guided to focus on core features that exist simultaneously in both the time and frequency domains and are strongly correlated with molecular binding events, while filtering out irrelevant features that may be noise and only appear in a single domain. This training strategy significantly improves the quality and robustness of the learned features, making the model's decisions more reliable and further enhancing the accuracy and stability of detection.

[0027] Preferably, the probe molecule is copper porphyrin, the substrate molecule is caffeine, and the solvent is 1,2,4-trichlorobenzene.

[0028] Preferably, step S4 uses least squares fitting to obtain the standard relationship curve.

[0029] The concentration detection system based on STM-BJ technology combined with deep learning described in this invention includes:

[0030] Data acquisition module: used to acquire conductivity data of pure probe molecule solutions and mixed solutions of probe molecules and substrate molecules at preset concentrations using the STM-BJ measurement system;

[0031] Training module: Used to train a pre-built supervised deep learning model using the collected conductivity data as a training set. This model extracts the time-domain and frequency-domain features of the conductivity data to identify whether the solution corresponding to the conductivity data contains substrate molecules.

[0032] Proportion recognition module: used to collect conductivity data of mixed solutions at different substrate molecule concentrations, input them into the trained model, and obtain the proportion of conductivity data that is identified as containing substrate molecules at each concentration to the total number of conductivity data at that concentration;

[0033] Standard Relationship Curve Construction Module: Used to fit and establish a standard relationship curve between the known substrate molecule concentration and its corresponding proportion obtained in the proportion identification module;

[0034] Detection module: Used to collect conductivity data of the test solution and input it into the trained model. Based on the proportion of conductivity data containing substrate molecules to the total conductivity data of the test solution, the concentration of substrate molecules in the test solution is calculated by substituting it into the standard relationship curve.

[0035] The computer-readable storage medium for storing one or more programs according to the present invention includes one or more programs comprising instructions that, when executed by a computing device, cause the computing device to perform any of the methods described above.

[0036] Beneficial effects: This invention utilizes non-covalent intermolecular interactions as a specific recognition mechanism. Through a supervised deep learning model that integrates multi-domain features, it extracts weak molecular event features from complex conduction trajectories to accurately classify them. Based on the statistical proportion of correctly classified analyte at each concentration and the fitting standard curve of the analyte concentration, the final concentration is calculated. The supervised deep learning model significantly improves the lower limit of concentration detection and shortens the response time to the second level, fundamentally solving the problems of insufficient sensitivity and lengthy processing time faced by traditional methods in ultra-trace detection. Attached Figure Description

[0037] Figure 1 The one-dimensional conductivity histograms are for copper porphyrin solution and mixed solutions containing caffeine.

[0038] Figure 2 The following are the conductivity test diagrams for caffeine: (a) is a one-dimensional conductivity diagram of pure caffeine solution, (b) is the corresponding two-dimensional histogram of (a), and the inset is a conductivity step diagram of pure caffeine solution.

[0039] Figure 3 This is a schematic diagram of the architecture of a supervised learning model;

[0040] Figure 4 This is the confusion matrix for all data in the example;

[0041] Figure 5 This is a schematic diagram of the standard relationship curve;

[0042] Figure 6 The diagram shows the standard relationship curves obtained by fitting different numbers of trajectories, where (a) uses 40 trajectories, (b) uses 66 trajectories, (c) uses 79 trajectories, and (d) uses all collected trajectories.

[0043] Figure 7 This is a graph showing the trend of the fitting parameters of the standard relationship curve as a function of response time. Detailed Implementation

[0044] As shown in the figure, the concentration detection method based on STM-BJ technology combined with deep learning described in this invention includes the following steps:

[0045] S1. The conductivity data of pure probe molecule solutions and mixed solutions of probe molecules and substrate molecules at preset concentrations were collected using the STM-BJ measurement system.

[0046] The probe molecule and substrate molecule used in this invention only need to satisfy the following conditions: there is a weak non-covalent interaction between them, the conductivity of the probe molecule can be measured, and the change in conductivity before and after the interaction is greater than a preset value (at least the change can be clearly detected by STM-BJ).

[0047] The probe molecule and substrate molecule can be copper porphyrin and caffeine, respectively, and the solvent of the solution can be 1,2,4-trichlorobenzene. The probe molecule, substrate molecule and solvent can be adaptively adjusted.

[0048] The preset concentration should be relatively high so that the conductivity data of the two solutions have corresponding obvious characteristics. In particular, the concentration of probe molecules and substrate molecules in the mixed solution should be sufficient to ensure that there are enough weak interactions between probe molecules and substrate molecules. These interactions will be reflected in the conductivity data so that the subsequent model can learn to identify the corresponding features.

[0049] S2. The collected conductivity data is used as a training set to train a pre-built supervised deep learning model. This model extracts the time-domain and frequency-domain features of the conductivity data to identify whether the solution corresponding to the conductivity data contains substrate molecules.

[0050] The model was trained to recognize the characteristics corresponding to the conductivity data of a solution with no substrate molecules at all by using conductivity data of a pure probe molecule solution, and to recognize the characteristics of a probe molecule solution after weak interactions with the addition of substrate molecules by using conductivity data of a mixed solution.

[0051] A supervised deep learning model consists of a cascaded dual-channel encoder, a dual-path projector, and a downstream classifier. The dual-channel encoder includes a time-domain encoder and a frequency-domain encoder, and the dual-path projector includes two independent projectors adapted to the two encoders.

[0052] Among them, the temporal encoder uses a cascaded multi-layer one-dimensional convolutional neural network to process the original conductance signal (the original conductance signal mentioned in this paper can be the conductance signal obtained after data augmentation preprocessing of the initially acquired signal), and the size of each convolutional kernel decreases step by step to capture the dynamic form of molecular events.

[0053] Frequency domain encoder: The frequency domain features are extracted by processing the amplitude spectrum obtained by the fast Fourier transform of the original conductance signal using a cascaded multi-layer one-dimensional convolutional neural network. The size of each convolution kernel decreases progressively, and its initial convolution kernel size is smaller than that of the time domain encoder.

[0054] Dual-path projector: includes a one-dimensional convolutional layer, a normalization layer, and an adaptive average pooling layer to compress features into a projection vector of a preset dimension;

[0055] Downstream classifier: consists of a first fully connected layer, an activation function layer, and a second fully connected layer connected in sequence. The second fully connected layer outputs the classification result.

[0056] Specifically, supervised deep learning models consist of three core components: a dual-channel encoder, a dual-path projector, and a downstream classifier.

[0057] The dual-channel encoder, a key component for feature extraction, comprises a temporal encoder and a frequency-domain encoder. While both encoders share a similar architecture, they operate with independent parameters, each composed of cascaded one-dimensional convolutional neural networks following a hierarchical feature extraction strategy. This strategy initially uses larger convolutional kernels to capture macro-features, followed by smaller kernels to analyze finer patterns. Each convolutional layer includes a GELU activation function and a normalization layer, while max-pooling layers are utilized to expand the receptive field. Despite adopting a unified design philosophy, the specific implementations of both encoders are carefully tailored to their respective input data. The temporal encoder directly processes the raw conductance signal, with its three convolutional blocks having kernel sizes set to 9, 5, and 3, respectively, to capture the dynamic morphology of molecular events throughout the time series.

[0058] In contrast, frequency domain encoders focus on analyzing the amplitude spectrum of the signal, specifically the amplitude spectrum obtained by performing a Fast Fourier Transform on the original conductivity signal. This approach not only minimizes redundancy in the spectrum but also aligns more closely with traditional signal processing practices. To accommodate the unique characteristics of frequency domain data, the initial convolution kernel size of the frequency domain encoder differs from that of the time domain encoder, specifically being set to 7, and its kernel size decreases progressively. This adjustment aims to effectively identify the inherent rhythms and characteristic frequencies of the signal derived from the spectrum.

[0059] Both the time-domain and frequency-domain encoders produce 128-dimensional feature vectors. Or, as indicated later, compression. The detailed layer-by-layer architecture of these two encoders, including their output shapes and parameter distributions, are shown in Tables 1 and 2, respectively.

[0060] Table 1. Specific architecture of the time-domain encoder

[0061] Layer (Type) Output shape Number of parameters Conv1d-1 [-1, 64, 1500] 640 GELU-2 [-1, 64, 1500] 0 LayerNorm-3 [-1, 64, 1500] 192,000 MaxPool1d-4 [-1, 64, 750] 0 Conv1d-5 [-1, 128, 750] 41,088 GELU-6 [-1, 128, 750] 0 LayerNorm-7 [-1, 128, 750] 192,000 Conv1d-8 [-1, 128, 750] 49,280 GELU-9 [-1, 128, 750] 0 LayerNorm-10 [-1, 128, 750] 192,000 Flatten-11 [-1, 96000] 0 Linear-12 [-1, 128] 12,288,128 GELU-13 [-1, 128] 0 LayerNorm-14 [-1, 128] 256 Dropout-15 [-1, 128] 0

[0062] Table 2 Detailed architecture of the frequency domain encoder

[0063] Layer (Type) Output shape Number of parameters Conv1d-1 [-1, 64, 751] 512 GELU-2 [-1, 64, 751] 0 LayerNorm-3 [-1, 64, 751] 96,128 MaxPool1d-4 [-1, 64, 375] 0 Conv1d-5 [-1, 128, 375] 41,088 GELU-6 [-1, 128, 375] 0 LayerNorm-7 [-1, 128, 375] 96,000 Conv1d-8 [-1, 128, 375] 49,280 GELU-9 [-1, 128, 375] 0 LayerNorm-10 [-1, 128, 375] 96,000 Flatten-11 [-1, 48000] 0 Linear-12 [-1, 128] 6,144,128 GELU-13 [-1, 128] 0 LayerNorm-14 [-1, 128] 256 Dropout-15 [-1, 128] 0

[0064] Each encoder is backed by an independent projector. This lightweight network consists of one-dimensional convolutional layers, normalization layers, and adaptive average pooling layers. Its main function is to process the high-dimensional features generated by the encoders. Nonlinear mapping to a 64-dimensional low-dimensional space generates more refined projection vectors for downstream classifiers. .

[0065] The model's final decision is presented by an independent downstream classifier, which receives and concatenates the projection vectors from the two projectors. This results in a 128-dimensional fused feature vector. This feature vector is then processed by a module consisting of two fully connected layers and a ReLU activation function. To reduce the risk of overfitting in the classification task, a Dropout layer (regularization layer) is added before the first fully connected layer. This Dropout layer is an additional layer added independently outside the final Dropout layers of the two simulators. Finally, the classifier outputs a binary classification result: Class 0 indicates that the mixed solution corresponding to the conductivity data does not contain substrate molecules, meaning that no features corresponding to the weak interaction between probe molecules and substrate molecules were extracted from either the time or frequency domains; Class 1 indicates that the mixed solution corresponding to the conductivity data contains substrate molecules, meaning that the model extracted and identified features corresponding to the weak interaction between probe molecules and substrate molecules.

[0066] We employ a composite loss function to optimize the model training process end-to-end. First, we address the supervised classification loss by directly optimizing the model's core classification performance using weighted cross-entropy loss. question.

[0067] Secondly, we introduce a cross-domain consistency loss, which forces the model to learn features consistent in time and frequency. We use mean squared error loss to minimize the distance between the time-domain and frequency-domain projection vectors, specifically expressed as:

[0068]

[0069] in, For cross-domain consistency loss, and Let B represent the time-domain and frequency-domain projection vectors of the i-th sample, respectively, and let B represent the batch size. This represents the square of the L2 norm, i.e., the square of the Euclidean distance. The loss value is the average distance across the entire batch.

[0070] Finally, the model's total loss function is expressed as: It is the weighted sum of all the above loss components:

[0071]

[0072] Where λ is the weighting factor, which can be set to 0.2.

[0073] The entire training process used the Adam optimizer with a learning rate of 4e-4, iteratively optimizing over 20 iterations with a batch size of 32. The trained model was saved for subsequent classification tasks on the conductivity dataset in experiments detecting CAF at different concentrations.

[0074] S3. Collect conductivity data of mixed solutions at different substrate molecule concentrations, input them into the trained model, and obtain the proportion of conductivity data at each concentration that is identified as containing substrate molecules to the total conductivity data at that concentration.

[0075] Multiple sets of conductivity data will be measured for each substrate molecule concentration during the mixing and fusion process. It should be noted that the concentration of probe molecules in the mixed solution remains constant for each substrate molecule concentration.

[0076] Because the number of substrate molecules in the mixed solution is small when the substrate molecule concentration is extremely low, the weak interaction between the substrate molecules and the probe molecules is a probabilistic event, not a guaranteed occurrence. In particular, this probability gradually decreases as the concentration decreases. In other words, even if the substrate molecules are present in the mixed solution, it is impossible to detect their presence in every set of conductivity data. Only by detecting multiple sets of data can the proportion (percentage) of substrate molecules in the mixed solution at that substrate molecule concentration be determined.

[0077] Specifically, this involves inputting all conductivity data of the mixed solution at each substrate molecule concentration into the trained model, and calculating the proportion of conductivity data that is determined to contain substrate molecules at that concentration.

[0078] S4. Based on the known substrate molecule concentration and its corresponding ratio obtained in S3, fit and establish a standard relationship curve between the substrate molecule concentration and the ratio.

[0079] Using substrate molecule concentration as the x-axis and the proportion of all conductivity data at that concentration (identified as containing substrate molecules) as the y-axis, multiple data points are constructed. In the linear region of these data points, least squares fitting is used to establish a standard relationship curve between substrate molecule concentration and the aforementioned proportion.

[0080] S5. Collect the conductivity data of the test solution and input it into the trained model. Based on the proportion of conductivity data containing substrate molecules to the total conductivity data of the test solution, substitute it into the standard relationship curve to calculate the concentration of substrate molecules in the test solution.

[0081] Similarly, multiple sets of data are collected. The proportion of substrate molecules in all conductivity data is determined by the trained model. Then, this proportion is substituted into the standard relationship curve obtained in step S4 to calculate the concentration of substrate molecules in the test solution.

[0082] To better illustrate this method, the following example will provide further explanation:

[0083] Example

[0084] Formation of the single-molecule sensor: In the preparation stage, the metal tip and metal substrate undergo rigorous electrochemical cleaning and plasma treatment to obtain atomically clean surfaces. The metal substrate has multiple solutions for holding different solutions; a 5 μL solution of probe molecules, or a mixture of probe molecules and substrate molecules of a specific analyte, can be deposited on the metal substrate surface using a micropipette. When the distance between the metal tip and metal substrate reaches the nanoscale, a DC bias voltage of 0.1 V is applied, and the current monitoring system is activated. The tip-substrate distance is dynamically adjusted through a closed-loop feedback system to precisely stabilize the real-time conductivity value at 10-1. -5 G o Within the target window. Finally, the needle tip is driven to perform periodic stretching-retracting motions (at a constant speed of 0.5 μm / s). During this process, the captured molecules repeatedly form "junctions" between the electrodes, and their conductivity signals are continuously recorded, thus providing high signal-to-noise ratio raw data for subsequent statistical identification and analysis. The resulting sensor can be directly used for highly sensitive, single-molecule-level detection of specific molecules in solution environments.

[0085] Formation and detection of supramolecular junctions: First, the metal tip and the metal substrate are subjected to a periodic "contact-separation-re-contact" mechanical cycle at a speed of 0.5 μm / s for approximately 30 seconds to stabilize the system. Then, 5 μL of probe molecule solution or a mixture of probe molecules and substrate molecules is added to the metal substrate. The relative velocity is then adjusted to 8 nm / s and the cycle continues. During this process, the conductivity signal is recorded at a sampling rate of 20 kHz. When the molecular junction is successfully constructed, the conductivity signal will exhibit jumps or brief fluctuations. At least 2000 valid trajectories are collected in a single detection, with the preset effective conductivity signal range being 10... -6 G o Up to 10 3 G o The lower limit of the characteristic interval used to analyze supramolecular structures is 10. -6 Go Up to 10 -9 G o Within this range, the conductivity signal indicates that the metal tip has separated from the metal substrate; the upper limit is 10. 1 G o Up to 10 3 G o The conductivity signal within this range indicates that the metal needle tip has successfully made contact with the metal substrate to form a circuit.

[0086] Deep learning analysis: The probe molecule solution uses copper porphyrin (CuP) molecules, and the substrate molecule solution can be selected with a concentration of 0.2 mM and 10. -1 mM, 10 -4 mM, 10 -7 mM, 10 -10 mM, 10 -13 mM, 10 -15 mM caffeine solution.

[0087] First, the conductivity data of the probe molecule solution and the mixture of the probe molecule and the substrate molecule were measured. The conductivity data obtained from the probe molecule solution was labeled as label 1, and the conductivity data obtained from the mixture of the probe molecule and the substrate molecule was labeled as label 0. A training set was constructed using these data to train a supervised deep learning model, transforming the recognition task into a classic supervised binary classification task (e.g., ...). Figure 3 ).

[0088] A supervised deep learning model is used to sensitively identify the weak interaction between a probe and caffeine. An intelligent classifier is trained to distinguish between the probe's "bound (label 1)" and "unbound (label 0)" states. The trained model is then used to construct a standard relationship curve, enabling quantitative analysis and sensitivity assessment of caffeine.

[0089] The operation process is explained in detail below:

[0090] 1) First, preprocessing is performed according to the formation process of a single-molecule sensor. Then, 5 μL of 0.1 mM CuP molecular solution is added between the two working electrodes, the metal tip and the metal substrate. The metal tip is then controlled to continuously cycle through "contact-separation-re-contact" relative to the metal substrate at a speed of 8 nm / s, while simultaneously recording conductivity signal data. After successfully recording 20 sets of valid data, a brand new metal tip and metal substrate are replaced, and another 5 μL of 0.1 mM CuP molecular solution is added. The above cycle and data recording process is repeated until another 20 sets of valid data are obtained. Its conductivity histogram is in the range of 10. -5.05 G0 has a conductivity peak (e.g.) Figure 1 ).

[0091] 2) Replace with a new metal needle tip and metal substrate, add 5 μL of a mixed solution of 0.1 mM CuP molecules and 0.2 mM caffeine molecules, and repeat step 1) until more than 2000 valid data points are obtained. Its conductivity histogram is within 10... -5.43 G0 has a conductivity peak (e.g.) Figure 1 The change was found to be 0.4 G0.

[0092] 3) Replace with a new metal needle tip and metal substrate, and add 5 μL of 0.1 mM CuP molecules with 0.1 mM CuP. A mixed solution of mM caffeine molecules was prepared, and step 1 was repeated until more than 2000 valid data points were obtained.

[0093] 4) Replace with a new metal needle tip and metal substrate, and add 5 μL of 0.1 mM CuP molecules with 10 -4 A mixed solution of mM caffeine molecules was prepared, and step 1 was repeated until more than 2000 valid data points were obtained.

[0094] 5) Replace with a new metal needle tip and metal substrate, and add 5 μL of 0.1 mM CuP molecules with 10 μL of water. -7 A mixed solution of mM caffeine molecules was prepared, and step 1 was repeated until more than 2000 valid data points were obtained.

[0095] 6) Replace with a new metal needle tip and metal substrate, and add 5 μL of 0.1 mM CuP molecules with 10 μL of water. -10 A mixed solution of mM caffeine molecules was prepared, and step 1 was repeated until more than 2000 valid data points were obtained.

[0096] 7) Replace with a new metal needle tip and metal substrate, and drop 5 μL of 0.1 mM CuP molecules with 10 μL of water. -13 A mixed solution of mM caffeine molecules was prepared, and step 1 was repeated until more than 2000 valid data points were obtained.

[0097] 8) Replace with a new metal needle tip and metal substrate, and add 5 μL of 0.1 mM CuP molecules with 10 μL of water. -15 A mixed solution of mM caffeine molecules was prepared, and step 1 was repeated until more than 2000 valid data points were obtained.

[0098] 9) After labeling the data in step 1) as 0 and the data in step 2) as label 1, train the supervised deep learning model. When the training accuracy reaches 98% or more, the model training is considered complete.

[0099] The data on different substrate concentrations obtained in steps 3-7) are input into a large model for analysis to obtain a highly accurate linear relationship (e.g., Figure 4 (5) The results also show that as the caffeine concentration decreases, the conductivity data of the mixed solution containing substrate molecules also decreases. Standard relationship curves were constructed based on different substrate molecule concentrations and the proportion of data indicating the presence of substrate molecules.

[0100] Then, the percentage of conductivity data containing substrate molecules obtained from the input model in step 8) is substituted into the standard relationship curve to confirm that the concentration of substrate molecules in the solution is indeed 10. -18 M (that is, 10) -15 The mM indicates that this method can achieve detection of extremely low concentrations.

[0101] In addition to verifying the accuracy of concentration detection, we also validated the efficiency of this detection method:

[0102] For each dataset, we recorded the proportion of traces classified as "1" (i.e., the proportion of conductivity data) (in Figure 6 (This is denoted as the CAF ratio). Subsequently, we plotted the data points of the "1" trace ratio versus caffeine concentration, and obtained the standard relationship curve equation and coefficient of determination (R^2) by performing least-squares fitting on its linear region.

[0103] By controlling the number of conductance trajectories used for analysis (each loop represented as one trajectory data point), we systematically evaluated the model's quantitative analysis capabilities at corresponding detection times with different sample sizes (40, 66, 79, and the entire dataset). The CuP probe system reached its performance threshold within extremely short detection times. When the detection time was only 3 seconds (40 trajectories), the calibration curve had a slope of 0.06, an intercept of 1.17, and a coefficient of determination (R^2) of 0.71, which was very close to the calibration curve obtained from the entire dataset (slope: 0.07, intercept: 1.21, R^2: 0.86). As the detection time increased to 5 seconds (66 trajectories), the slope of the calibration curve became 0.07, the intercept became 1.19, and the R^2 value rapidly increased to 0.81. Notably, at detection times of 6 seconds (79 records), the slope reached 0.06, the intercept stabilized at 1.19, and the R^2 stabilized at 0.83, highly consistent with the final performance obtained from the entire dataset (with a deviation of ±5%). All key metrics defining the calibration curve converged and stabilized rapidly within a very short time. For CuP probe systems, results comparable to longer data acquisition periods were obtained in only about 6 seconds of detection time (response time), highlighting the significant advantage of this model in detection speed for CuP probe systems.

[0104] To verify that the detection accuracy of this invention is higher than that of the prior art, the following comparative examples are used for verification:

[0105] Comparative example:

[0106] This embodiment provides a method for detecting caffeine using CuP and determining its limit through a clustering algorithm, including the following steps:

[0107] 1) Add 5 μL of 0.1 mM CuP molecular solution between the two working electrodes, the metal tip and the metal substrate. Then, control the metal tip to continuously cycle through "contact-separation-re-contact" relative to the metal substrate at a speed of 8 nm / s, and simultaneously record conductivity signal data. After successfully recording 20 sets of valid data, replace the metal tip and metal substrate with a fresh one, and add 5 μL of 0.1 mM CuP molecular solution again, repeating the above cycle and data recording process until another 20 sets of valid data are obtained. Its conductivity histogram is within 10... -5.05 G0 has a conductivity peak (e.g.) Figure 1 ).

[0108] 2) Replace with a new metal needle tip and metal substrate, add 5 μL of a mixed solution of 0.1 mM CuP molecules and 0.2 mM caffeine molecules, and repeat step 1) until more than 2000 valid data points are obtained. Its conductivity histogram is within 10... -5.43 G0 has a conductivity peak (e.g.) Figure 1 The change was found to be 0.4 G0.

[0109] 3) Replace with a new metal needle tip and metal substrate, add 5 μL of a mixed solution of 0.1 mM CuP molecules and 0.1 mM caffeine molecules, and repeat step 1) until more than 2000 valid data are obtained.

[0110] 4) Replace with a new metal needle tip and metal substrate, and add 5 μL of 0.1 mM CuP molecules with 10 -2 A mixed solution of mM caffeine molecules was prepared, and step 1 was repeated until more than 2000 valid data points were obtained.

[0111] 5) Replace with a new metal needle tip and metal substrate, and add 5 μL of 0.1 mM CuP molecules with 10 μL of water. -3 A mixed solution of mM caffeine molecules was prepared, and step 1 was repeated until more than 2000 valid data points were obtained.

[0112] 6) Replace with a new metal needle tip and metal substrate, and add 5 μL of 0.1 mM CuP molecules with 10 μL of water. -4 A mixed solution of mM caffeine molecules was prepared, and step 1 was repeated until more than 2000 valid data points were obtained.

[0113] 7) Replace with a new metal needle tip and metal substrate, and drop 5 μL of 0.1 mM CuP molecules with 10 μL of water.-5 A mixed solution of mM caffeine molecules was prepared, and step 1 was repeated until more than 2000 valid data points were obtained.

[0114] 8) Analyze the data on different substrate concentrations obtained in steps 3-7) using a clustering algorithm. Subdivide the peak values ​​of conductivity changes ( ). As the concentration changes, the caffeine binding rate gradually decreases. When the concentration decreases to 10... -8 At point M, the binding rate reaches below 5%. Therefore, using clustering algorithms, we can detect a minimum caffeine concentration of 10. -8 M, which is 10 compared to deep learning algorithms -18 M is 10 orders of magnitude less.

[0115] Clustering algorithms are existing technologies that can be referenced: Lin, L. et al. Spectral clustering to analyze the hidden events in single-molecule break junctions. J. Phys. Chem. C 125, 3623-3630 (2021). and van Veen, F., Ornago, L., van der Zant, HSJ & El Abbassi, M. A generalized neural network approach for separation of molecular breaking traces. J. Mater. Chem. C 11, 15564–15570 (2023).

[0116] The concentration detection system based on STM-BJ technology combined with deep learning described in this invention includes:

[0117] Data acquisition module: used to acquire conductivity data of pure probe molecule solutions and mixed solutions of probe molecules and substrate molecules at preset concentrations using the STM-BJ measurement system;

[0118] Training module: Used to train a pre-built supervised deep learning model using the collected conductivity data as a training set. This model extracts the time-domain and frequency-domain features of the conductivity data to identify whether the solution corresponding to the conductivity data contains substrate molecules.

[0119] Proportion recognition module: used to collect conductivity data of mixed solutions at different substrate molecule concentrations, input them into the trained model, and obtain the proportion of conductivity data that is identified as containing substrate molecules at each concentration to the total number of conductivity data at that concentration;

[0120] Standard Relationship Curve Construction Module: Used to fit and establish a standard relationship curve between the known substrate molecule concentration and its corresponding proportion obtained in the proportion identification module;

[0121] Detection module: Used to collect conductivity data of the test solution and input it into the trained model. Based on the proportion of conductivity data containing substrate molecules to the total conductivity data of the test solution, the concentration of substrate molecules in the test solution is calculated by substituting it into the standard relationship curve.

[0122] The computer-readable storage medium for storing one or more programs according to the present invention includes one or more programs comprising instructions that, when executed by a computing device, cause the computing device to perform the method described above.

Claims

1. A concentration detection method based on STM-BJ technology combined with deep learning, characterized in that, Includes the following steps: S1. The conductivity data of pure probe molecule solutions and mixed solutions of probe molecules and substrate molecules at preset concentrations were collected using the STM-BJ measurement system. S2. Use the collected conductivity data as a training set to train a pre-built supervised deep learning model. This model extracts the time-domain and frequency-domain features of the conductivity data to identify whether the solution corresponding to the conductivity data contains substrate molecules. S3. Collect conductivity data of mixed solutions at different substrate molecule concentrations, input them into the trained model, and obtain the proportion of conductivity data at each concentration that is determined to contain substrate molecules to the total number of conductivity data at that concentration. S4. Based on the known substrate molecule concentration and its corresponding ratio obtained in S3, fit and establish a standard relationship curve between the substrate molecule concentration and the ratio. S5. Collect the conductivity data of the test solution and input it into the trained model. Based on the proportion of conductivity data containing substrate molecules to the total conductivity data of the test solution, substitute it into the standard relationship curve to calculate the concentration of substrate molecules in the test solution.

2. The method according to claim 1, characterized in that: The probe molecule and the substrate molecule have a weak non-covalent interaction. The conductivity of the probe molecule can be measured, and the change in conductivity before and after the interaction is greater than a preset value.

3. The method according to claim 1, characterized in that: The supervised deep learning model includes a cascaded dual-channel encoder, a dual-path projector, and a downstream classifier. The dual-channel encoder includes a time-domain encoder and a frequency-domain encoder, and the dual-path projector includes two independent projectors adapted to the two encoders.

4. The method according to claim 3, characterized in that: Temporal encoder: The original conductance signal is processed by a cascaded multi-layer one-dimensional convolutional neural network, with the size of each convolutional kernel decreasing progressively to capture the dynamic form of molecular events; Frequency domain encoder: The frequency domain features are extracted by processing the amplitude spectrum obtained by the fast Fourier transform of the original conductance signal using a cascaded multi-layer one-dimensional convolutional neural network. The size of each convolution kernel decreases progressively, and its initial convolution kernel size is smaller than that of the time domain encoder. Dual-path projector: includes a one-dimensional convolutional layer, a normalization layer, and an adaptive average pooling layer to compress features into a projection vector of a preset dimension; Downstream classifier: consists of a first fully connected layer, an activation function layer, and a second fully connected layer connected in sequence. The second fully connected layer outputs the classification result.

5. The method according to claim 4, characterized in that: The downstream separator has a regularization layer before the first fully connected layer for regularizing the spliced ​​time-domain projection vector and frequency-domain projection vector.

6. The method according to claim 1, characterized in that, The supervised deep learning model training uses a composite loss function: ， in, For the total loss function, To monitor the classification of losses, The cross-domain consistency loss is represented by λ, which is a weighting factor. The expression is ， in, and Let represent the time-domain and frequency-domain projection vectors of the i-th sample, respectively, and B represent the batch size.

7. The method according to claim 1, characterized in that: The probe molecule is copper porphyrin, the substrate molecule is caffeine, and the solvent is 1,2,4-trichlorobenzene.

8. The method according to claim 1, characterized in that: Step S4 uses least squares fitting to obtain the standard relationship curve.

9. A concentration detection system based on STM-BJ technology combined with deep learning, characterized in that, include: Data acquisition module: used to acquire conductivity data of pure probe molecule solutions and mixed solutions of probe molecules and substrate molecules at preset concentrations using the STM-BJ measurement system; Training module: Used to train a pre-built supervised deep learning model using the collected conductivity data as a training set. This model extracts the time-domain and frequency-domain features of the conductivity data to identify whether the solution corresponding to the conductivity data contains substrate molecules. Proportion recognition module: used to collect conductivity data of mixed solutions at different substrate molecule concentrations, input them into the trained model, and obtain the proportion of conductivity data that is identified as containing substrate molecules at each concentration to the total number of conductivity data at that concentration; Standard Relationship Curve Construction Module: Used to fit and establish a standard relationship curve between the known substrate molecule concentration and its corresponding proportion obtained in the proportion identification module; Detection module: Used to collect conductivity data of the test solution and input it into the trained model. Based on the proportion of conductivity data containing substrate molecules to the total conductivity data of the test solution, the concentration of substrate molecules in the test solution is calculated by substituting it into the standard relationship curve.

10. A computer-readable storage medium for storing one or more programs, characterized in that: The program includes one or more instructions that, when executed by a computing device, cause the computing device to perform any of the methods according to claims 1 to 8.

Images