Single-molecule charge transport unsupervised identification method, device and equipment and storage medium

Abstract

This invention relates to the field of single-molecule electronics technology, and more particularly to an unsupervised identification method, apparatus, device, and storage medium for single-molecule charge transport. The method standardizes the format of raw conductivity data, generates global features that conform to the physical properties of single molecules through bi-branch contrastive learning and feature fusion, and employs a progressive strategy of single-class pre-screening and binary fine-screening to extract effective conductivity traces. Unsupervised classification is completed by combining data standardization and an optimal cluster number determination mechanism. This solves the problems of low data utilization and weak generalization ability caused by training with only single-molecule data in existing technologies, and avoids the interference of subjective errors on the analysis results. The resulting classification results are classified according to conductivity step height characteristics, which can adapt to the analysis needs of multi-molecule systems and large-scale single-molecule conductivity measurement data, and makes up for the deficiencies of existing technologies in terms of feature extraction targeting, effective trace screening accuracy, automation level, and classification refinement.

Description

Technical Field

[0001] This invention relates to the field of single-molecule electronics technology, and in particular to an unsupervised identification method, apparatus, device, and storage medium for single-molecule charge transport. Background Technology

[0002] Single-molecule electronics is a core research direction in nanoscience and technology, mainly exploring electronic transport properties and related physicochemical processes at the single-molecule scale. With the iteration of precision measurement technologies such as mechanically controlled break-bonding (MCBJ) and scanning tunneling microscopy break-bonding (STM-BJ), precise measurement of conductivity at the single-molecule scale has become a reality. This type of technology fixes the target molecule between two poles by anchoring groups, and can capture the electrical transport behavior signal of a single molecule in real time. Due to the strong randomness of the single-molecule connection process, accurately extracting effective conductivity information from massive experimental measurement data has become a core technical bottleneck in the field of single-molecule charge transport research.

[0003] Current single-molecule charge transport data classification techniques all revolve around the conductance-distance curves obtained from split-junction techniques such as STM-BJ and MCBJ for event identification and classification. The overall evolution is from traditional statistical methods to machine learning and deep clustering methods. Traditional statistical methods, using one-dimensional conductance histograms and two-dimensional conductance-distance statistical graphs as core tools, can only qualitatively present data distribution patterns. They are easily obscured by dominant data, and also suffer from a lack of quantitative analysis capabilities and poor anti-interference performance. Traditional machine learning methods, such as multi-parameter vector classification and K-means++, further demonstrate this. Methods such as spectral clustering combined with the CH index require manual feature selection, which not only limits the applicable scenarios but also makes it difficult to capture the nonlinear structural features within the data. Deep clustering methods, such as the DAK algorithm, use autoencoders to complete feature extraction and clustering. Although they can adapt to high-dimensional and large-volume data, feature learning lacks physical targeting for single-molecule data. The model design is highly subjective and lacks interpretability, making it prone to failure in experimental scenarios with a large proportion of tunneled data and a small effective sample size. Specialized algorithms such as multi-prototype clustering have the drawbacks of over-splitting clustering results and only being suitable for a single experimental scenario.

[0004] In summary, existing technologies all employ single-molecule data to drive model training, failing to uncover the commonalities and differences in multi-molecule experimental data. They also exhibit weak data utilization and model generalization capabilities, while suffering from insufficient feature extraction accuracy, susceptibility of effective conduction traces to interference from tunneling data, lack of a universal end-to-end automated analysis framework, and low classification precision. Consequently, they struggle to meet the precise analysis requirements of complex single-molecule charge transport data. Therefore, existing technologies require further improvement and enhancement. Summary of the Invention

[0005] In order to overcome the shortcomings of the prior art, the present invention aims to provide an unsupervised identification method for single-molecule charge transport, which solves the problems of low data utilization and weak generalization ability caused by the prior art using only single-molecule data for training, and avoids the interference of subjective error on the analysis results.

[0006] The first aspect of this invention provides an unsupervised identification method for single-molecule charge transport, comprising: acquiring at least two sets of experimental datasets, the experimental datasets including a blank substrate experimental dataset and a molecular substrate experimental dataset; standardizing the two sets of experimental datasets respectively to obtain blank dimension feature vectors and molecular dimension feature vectors; using the blank dimension feature vectors as reference samples, sequentially performing bi-branch contrastive learning processing and feature fusion processing on the blank dimension feature vectors and molecular dimension feature vectors to obtain blank global features and molecular global features; using the blank global features as a reference, sequentially performing single-classifier pre-screening processing and binary classifier fine-screening processing on the molecular global features to obtain positive molecular samples, the positive molecular samples including multiple conductivity trajectory lines with conductivity step features; standardizing the positive molecular samples to obtain standard positive molecular samples and determining the optimal number of clusters; based on the optimal number of clusters, performing unsupervised clustering processing on the standard positive molecular samples to obtain classified positive samples, the classified positive samples being the classification results of multiple conductivity trajectory lines classified according to conductivity step height.

[0007] Optionally, in a first implementation of the first aspect of the present invention, obtaining at least two sets of experimental datasets includes: constructing a single-molecule conductivity measurement experimental system based on scanning tunneling microscopy split junction technology, and setting the core parameters of the experimental system, including stretching rate, stretching stroke, and sampling frequency; using a gold needle tip as the upper electrode and a blank gold substrate or a molecularly functionalized gold substrate as the lower electrode, repeatedly performing multiple stretching experiments using the experimental system to obtain multiple stretching experimental results; and performing baseline correction processing, outlier removal processing, and effective segment truncation preprocessing on the multiple stretching experimental results to construct the blank substrate experimental dataset and the molecular substrate experimental dataset.

[0008] Optionally, in a second implementation of the first aspect of the present invention, the standardization of the two sets of experimental datasets to obtain blank dimension feature vectors and molecular dimension feature vectors includes: extracting the conductivity value sequence of each conductivity trajectory line in the blank substrate experimental dataset and the molecular substrate experimental dataset and performing logarithmic transformation; using histogram statistics, dividing the logarithmically transformed conductivity value sequence into a preset number of equidistant conductivity intervals, and counting the frequency of conductivity values ​​of each conductivity trajectory line in each conductivity interval to construct a one-dimensional frequency feature blank distribution vector and a molecular distribution vector; and performing normalization and principal component analysis dimensionality reduction on the blank distribution vector and the molecular distribution vector to obtain the blank dimension feature vector and the molecular dimension feature vector.

[0009] Optionally, in a third implementation of the first aspect of the present invention, the step of using the blank dimension feature vector as a reference sample and sequentially performing dual-branch contrastive learning processing and feature fusion processing on the blank dimension feature vector and the molecular dimension feature vector to obtain blank global features and molecular global features includes: constructing a contrastive learning neural network with dual-branch shared weights; inputting the blank dimension feature vector as a reference sample into the first branch of the contrastive learning neural network and inputting the molecular dimension feature vector into the second branch of the contrastive learning neural network; training the contrastive learning neural network using the InfoNCE loss function to extract its output blank high-level semantic features and molecular high-level semantic features; performing concatenated fusion processing on the blank high-level semantic features and the blank dimension feature vector to obtain the blank global features; and performing concatenated fusion processing on the molecular high-level semantic features and the molecular dimension feature vector to obtain the molecular global features.

[0010] Optionally, in the fourth implementation of the first aspect of the present invention, the step of using the blank global features as a benchmark to sequentially perform single-classifier pre-screening and binary-classifier fine-screening on the molecular global features to obtain molecular positive samples, wherein the molecular positive samples include multiple conductivity trajectory lines with conductivity step features, includes: training a single-classifier model using the blank global features as a training set; inputting the molecular global features into the trained single-classifier model, removing pure tunneling samples that are highly similar to the blank global features, and retaining potentially abnormal samples including conductivity steps as pre-screening results; constructing a binary-classification training set using the blank global features as samples with label 0 and the pre-screened abnormal samples as samples with label 1, and training a binary-classifier model based on the binary-classification training set; inputting the molecular global features into the trained binary-classifier model, and separating molecular positive samples including conductivity step features by cross-validation of a preset probability threshold.

[0011] Optionally, in a fifth implementation of the first aspect of the present invention, the standardization of the molecular positive samples to obtain standard molecular positive samples and the determination of the optimal number of clusters includes: performing standardization on the molecular positive samples using the Z-Score standardization method to obtain standard molecular positive samples; calculating the sample distortion coefficient and sample silhouette coefficient under different preset numbers of clusters based on the standard molecular positive samples, and plotting the distortion coefficient curve and the silhouette coefficient curve for correlating the distortion coefficient and the number of clusters; determining the range of candidate optimal number of clusters by combining the elbow inflection point of the distortion coefficient curve and the maximum value point of the silhouette coefficient curve; obtaining experimental prior knowledge, and determining the optimal number of clusters from the range of candidate optimal number of clusters based on the experimental prior knowledge.

[0012] Optionally, in a sixth implementation of the first aspect of the present invention, the step of performing unsupervised clustering processing on the standard molecular positive samples based on the optimal number of clusters to obtain classified positive samples includes: obtaining a pre-selected unsupervised clustering model; inputting the standard molecular positive samples into the unsupervised clustering model to divide the molecular positive samples into multiple cluster categories corresponding to the optimal number of clusters; for each cluster category, extracting the conductivity step height feature in the corresponding conductivity trajectory line; based on the extracted conductivity step height feature, labeling the multiple cluster categories respectively; and integrating the multiple labeled cluster categories to obtain the classified positive samples.

[0013] A second aspect of the present invention provides an unsupervised identification device for single-molecule charge transport, comprising: an acquisition module for acquiring at least two sets of experimental datasets, the experimental datasets including a blank substrate experimental dataset and a molecular substrate experimental dataset; a processing module for standardizing the two sets of experimental datasets respectively to obtain a blank dimension feature vector and a molecular dimension feature vector; and a learning module for using the blank dimension feature vector as a reference sample to sequentially perform bi-branch contrastive learning processing and feature fusion processing on the blank dimension feature vector and the molecular dimension feature vector to obtain blank global features and molecular global features. The module includes a screening module, which uses the blank global features as a benchmark to perform single-classifier pre-screening and binary-classifier fine-screening on the molecular global features to obtain positive molecular samples. The positive molecular samples include multiple conductivity trajectory lines with conductivity step features. The module determines the positive molecular samples by standardizing them to obtain standard positive molecular samples and determines the optimal number of clusters. The module classifies the standard positive molecular samples by performing unsupervised clustering based on the optimal number of clusters to obtain classified positive samples. The classified positive samples are the classification results of multiple conductivity trajectory lines classified according to conductivity step height.

[0014] A third aspect of the present invention provides a single-molecule charge transport unsupervised identification device, the single-molecule charge transport unsupervised identification device comprising: a memory and at least one processor, the memory storing instructions; the at least one processor calling the instructions in the memory to cause the single-molecule charge transport unsupervised identification device to perform the various steps of the single-molecule charge transport unsupervised identification method described in any of the preceding claims.

[0015] A fourth aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed by a processor, implement the steps of the unsupervised identification method for single-molecule charge transport described in any of the preceding claims.

[0016] The technical solution of this invention standardizes the format of variable-length raw conductivity data, generates global features that conform to the physical properties of single molecules through bi-branch contrastive learning and feature fusion, and achieves high-purity extraction of effective conductivity traces by employing a progressive strategy of single-class pre-screening and binary fine-screening. It combines data standardization and an optimal cluster number determination mechanism to complete unsupervised fine-grained classification, solving the problems of low data utilization and weak generalization ability caused by training with only single-molecule data in existing technologies, and avoiding the interference of subjective errors on the analysis results. The resulting classification results are classified according to conductivity step height characteristics, which can adapt to the analysis needs of multi-molecule systems and large-scale single-molecule conductivity measurement data, making up for the deficiencies of existing technologies in terms of feature extraction targeting, effective trace screening accuracy, automation level, and classification fineness. Attached Figure Description

[0017] Figure 1 The flowchart is shown below for the unsupervised identification method for single-molecule charge transport provided in the embodiments of the present invention.

[0018] Figure 2 This is a schematic diagram of the structure of the unsupervised recognition device for single-molecule charge transport provided in an embodiment of the present invention;

[0019] Figure 3 This is a schematic diagram of the structure of the unsupervised recognition device for single-molecule charge transport provided in an embodiment of the present invention;

[0020] Figure 4 This is a schematic diagram illustrating the specific process of the unsupervised identification method for single-molecule charge transport based on multiple molecular data-driven and feature fusion methods provided in this embodiment of the invention. Detailed Implementation

[0021] This invention provides a method, apparatus, device, and storage medium for unsupervised identification of single-molecule charge transport. In this invention, the terms "first," "second," "third," "fourth," etc. (if present)," in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0022] This invention discloses an unsupervised identification method for single-molecule charge transport. For ease of understanding, the specific process of the embodiments of this invention is described below. Please refer to [link / reference]. Figure 1 One embodiment of the unsupervised identification method for single-molecule charge transport in this invention includes:

[0023] 101. Obtain at least two sets of experimental datasets, including blank substrate experimental datasets and molecular substrate experimental datasets;

[0024] In this embodiment, the blank substrate experimental dataset consists of a pure gold electrode substrate without any target molecule modification. Single-molecule conductivity measurements are performed using scanning tunneling microscopy-batch junction (STM-BJ) technology. All generated conductivity trajectory data includes only quantum tunneling signals and has no conductivity step characteristics. The molecular substrate experimental dataset consists of a functionalized gold substrate modified with the target molecule to be tested. Under the same measurement conditions as the blank substrate, all conductivity trajectory data generated using the same STM-BJ technology includes conductivity signals generated during the gold-molecule-gold junction formation process, which may show obvious conductivity steps. Both sets of experimental data must be acquired under the same experimental environment and the same core parameters to ensure the validity of the data comparison. This overcomes the core problems of existing technologies, such as signal recognition deviation, low data utilization, and weak generalization ability, caused by the lack of a unified benchmark and the use of only single-molecule data for training. This provides a reliable physical benchmark for subsequently distinguishing between tunneling trajectories and molecular conductivity trajectories.

[0025] 102. Standardize the two sets of experimental datasets to obtain blank dimension feature vectors and molecular dimension feature vectors;

[0026] In this embodiment, to address the issues of random length, discrete numerical distribution, and inconsistent format of the original conduction trajectories caused by experimental perturbations, the blank substrate experimental dataset and the molecular substrate experimental dataset are standardized respectively. This standardizes the unstructured, variable-length original conduction trajectories into dimension-fixed, numerically regular, low-redundancy, dimensionless feature vectors. Specifically, blank substrate experimental data generates blank dimension feature vectors, and molecular substrate experimental data generates molecular dimension feature vectors. This ensures that the two types of feature vectors have a unified format, meeting the input requirements of subsequent dual-branch contrastive learning neural networks and classification models, and significantly improving the standardization and computational efficiency of data processing.

[0027] 103. Using the blank dimension feature vector as the reference sample, perform dual-branch contrast learning and feature fusion processing on the blank dimension feature vector and the molecular dimension feature vector in sequence to obtain blank global features and molecular global features.

[0028] In this embodiment, the nonlinear physical characteristics of single-molecule data are automatically captured through bi-branch contrastive learning, which enhances the feature representation capability of molecular conductance signals and improves the robustness and physical specificity of global features. This effectively solves the core defects of existing deep clustering methods, such as lack of physical specificity in feature learning, inability to match the physical characteristics of single-molecule data, and low feature discrimination. It provides high-quality feature support for subsequent trace screening and significantly improves the subsequent screening accuracy.

[0029] 104. Based on the blank global features, the molecular global features are subjected to single-classifier pre-screening and binary-classifier fine screening in sequence to obtain molecular positive samples. The molecular positive samples include multiple conductivity trajectory lines with conductivity step features.

[0030] In this embodiment, the pure tunneling signal features represented by the blank global features are used as the sole criterion for judgment. A progressive strategy of single-class pre-screening and two-class fine screening is adopted to achieve high-purity extraction of effective conduction traces. This solves the core pain points in the prior art, such as effective conduction traces being covered by massive tunneling data, low screening accuracy, and invalid data interfering with subsequent classification.

[0031] 105. Standardize the positive molecular samples to obtain standard positive molecular samples, and determine the optimal number of clusters;

[0032] In this embodiment, standardization eliminates the impact of differences in feature dimensions on clustering stability and avoids clustering bias caused by different feature magnitudes. By determining the optimal number of clusters, the problem of subjective cluster selection and mismatch with the physical laws of single-molecule experiments in existing technologies, which leads to large classification errors and meaningless clustering results, is solved, thus ensuring the physical rationality and stability of the clustering results.

[0033] 106. Based on the optimal number of clusters, perform unsupervised clustering on the standard molecular positive samples to obtain classified positive samples, wherein the classified positive samples are the classification results of multiple conductivity trajectory lines classified according to conductivity step height;

[0034] In this embodiment, a refined conductivity classification without manual annotation is achieved, overcoming the shortcomings of existing technologies that can only distinguish between valid and invalid traces, have insufficient classification refinement, and have ambiguous physical meanings in the classification results. This enables the classification results to accurately correspond to the molecular conductivity step height, and each classification result can correspond to a stable gold-molecule-gold junction connection configuration. This provides direct data support for the study of single-molecule charge transport mechanisms and the correlation analysis of molecular structure and conductivity performance, while improving classification efficiency and objectivity, and adapting to the analysis needs of large-scale single-molecule conductivity data.

[0035] The unsupervised identification method for single-molecule charge transport disclosed in this application standardizes the format of variable-length raw conductivity data, generates global features that fit the physical properties of single molecules based on bi-branch contrastive learning and feature fusion, and achieves high-purity extraction of effective conductivity traces by employing a progressive strategy of single-class pre-screening and binary fine-screening. Combined with data standardization and an optimal cluster number determination mechanism, it completes unsupervised fine-grained classification, solving the problems of low data utilization and weak generalization ability caused by using only single-molecule data for training in existing technologies, and avoiding the interference of subjective errors on the analysis results. The resulting classification results are classified according to the conductivity step height feature, which can adapt to the analysis needs of multi-molecule systems and large-scale single-molecule conductivity measurement data, making up for the deficiencies of existing technologies in terms of feature extraction targeting, effective trace screening accuracy, automation level, and classification fineness.

[0036] In other embodiments, there can be multiple molecular substrate experimental datasets, such as experimental datasets corresponding to molecules A, B, and C. Each molecular substrate experimental dataset independently completes the entire process according to steps 101-106, that is, each molecular substrate experimental dataset undergoes standardization, bi-branch contrastive learning and feature fusion, single-class pre-screening and binary fine-screening, standardization and determination of the optimal number of clusters, and unsupervised clustering, ultimately obtaining its corresponding positive samples. Figure 4 The process shown, taking the identification and classification of molecule A as an example, is divided into three core stages:

[0037] Figure 4 The multi-source data acquisition and global feature generation in (a) can realize the acquisition of raw conductivity data of blank substrates and multi-molecular substrates, and generate highly robust global features through reinforcement learning and feature fusion.

[0038] In this diagram, the yellow spherical array represents gold (Au) electrodes, including a gold needle tip at the top and a gold substrate at the bottom; T represents a blank substrate, indicating an experimental scenario with a blank gold substrate unmodified with any target molecules. In this case, there is only a vacuum / air between the two electrodes, and electron transport is mainly achieved through quantum tunneling; A, B, and C (molecular substrates) represent experimental scenarios with functionalized gold substrates modified with different target molecules, where A refers to molecular system A, B refers to molecular system B, and C refers to molecular system C; the colored spheres / chains in the middle represent target molecules, and different colors / structures (such as A, B, and C) correspond to different chemical molecular structures (such as thiols, pyridines, etc.), which constitute the core conductive channel of the gold-molecule-gold junction; the current symbol ⊕ and the ammeter characterize the single-molecule electrical transport measurement system, used to collect the change signal of conductivity (G) during stretching in real time;

[0039] , , and These represent the original histogram statistical feature vectors of the blank substrate, molecule A, molecule B, and molecule C, respectively. They are the basic statistical features without deep learning processing and reflect the distribution pattern of the original conductivity data. , , and The graphs correspond to four original conductance-displacement trajectory diagrams. The horizontal axis represents displacement / nm, indicating the stretching distance (nm) between the electrodes, and the vertical axis represents the relative value of the conductance value (G) with respect to the conductance quantum. The curve exhibits a single exponential decay curve with no obvious plateau, corresponding to a pure tunneling signal. , and These are the sets of original conductivity trajectories corresponding to molecules A, B, and C, respectively. The curves show obvious descending plateaus (conductivity steps), corresponding to the fracture process of gold-molecule-gold junction.

[0040] The input to the reinforcement learning module is , , and This refers to the original histogram statistical features. This module corresponds to the two-branch contrastive learning network in step 401, using blank features. As a benchmark, learn molecular characteristics , and Commonalities and differences between the features and the baseline features; through reinforcement learning strategy, the feature extractor is automatically optimized so that the learned features can maximize the distinction between tunneling signals and molecular conductance signals, and strengthen the conductance step features unique to molecular structures. The final output is: the deep features extracted by reinforcement learning, which correspond to high-level semantic features.

[0041] The feature fusion module corresponds to step 403. This module combines the high-level semantic features extracted by reinforcement learning with the original histogram statistical features. , , and By performing serial fusion to integrate basic statistical information and deep physical features, a global feature with high robustness and high physical specificity is generated, and the final output is... , , and ,in, This represents the robust global feature vector corresponding to the blank substrate, which is the highest quality feature representation of the pure tunneling signal. Let be the robust global eigenvector corresponding to molecule A, representing the highest quality feature expression of the junction conductance signal of molecule A. and These correspond to the robust global eigenvectors of molecules B and C, respectively.

[0042] Figure 4 The two-step conductivity trace selection method in (b) accurately screens the blank global features. Based on the global features of molecule A In the process of precisely eliminating pure tunneling samples, effective positive samples containing conductivity steps are selected. ;

[0043] The input to a single classifier is blank global features. and molecular global features The single classifier model corresponds to the OC-SVM model in step 501, and is used for learning. The boundary representing the characteristic distribution of pure tunneling signals will and Samples with highly similar features were identified as invalid tunneling samples. And remove potential positive samples. Retain and output the set of potentially valid samples whose feature distribution deviates from the tunneling benchmark;

[0044] The input to the binary classifier is and blank global features The binary classifier mode is the Multilayer Perceptron (MLP) of step 503, used to classify... For label 0, with For label 1, a precise classification model is trained, which ultimately distinguishes the effective samples containing electrical conductivity steps. ;

[0045] Below is a magnified view of the trajectory. The corresponding black curve is a typical pure tunneling attenuation curve. The conductivity drops rapidly to an extremely low value with distance, without plateau, which verifies that this part of the sample is invalid data. The orange / red curves represent typical effective curves containing conductivity steps. During the decrease in conductivity, one or more distinct horizontal plateaus (conductivity steps) appear, corresponding to the fracture configuration of the molecular structure, thus verifying the effectiveness of the screening.

[0046] Figure 4 The unsupervised refined clustering classification in (c) is used to classify the selected valid positive samples. Unsupervised clustering is performed, and classes are automatically classified according to the electrical conductivity step height.

[0047] The input to the unsupervised clustering model is the standardized positive molecular sample features. This model can automatically classify samples based on the similarity of their features. Divided into N cluster categories , , , ..., N is the optimal number of clusters;

[0048] The figure below shows the conductance trajectory corresponding to the cluster categories. The curve on the left corresponds to the category. For categories with low conductivity, the trajectory characteristics are represented by a plateau of low conductivity values, and the curve on the right corresponds to this category. For categories with equal high conductivity, the trajectory characteristics are characterized by a plateau of high conductivity values.

[0049] Through such Figure 4 The process shown transforms raw, complex conductivity data into interpretable physical classification results grouped by conductivity step height, ultimately achieving accurate identification and end-to-end automated analysis of single-molecule charge transport characteristics.

[0050] Multiple molecular substrate experimental datasets share the same blank substrate experimental dataset as a benchmark, eliminating the need for repeated collection of blank substrate data and repeated training of single classifiers and dual-branch contrastive learning neural networks. The trained models and blank global features can be directly reused. Through the collaborative processing of multi-molecule data, the commonalities and differences in the electrical conductivity properties of different molecules can be explored, further improving the generalization ability of the model. This makes the method disclosed in this application adaptable to the analysis needs of multi-molecule systems, breaking through the limitations of existing technologies that can only adapt to single molecules and have weak generalization ability. It provides an efficient and unified technical solution for comparative studies of multi-molecule charge transport properties.

[0051] Furthermore, in this embodiment of the invention, obtaining at least two sets of experimental datasets includes:

[0052] 201. Construct a single-molecule conductivity measurement experimental system based on scanning tunneling microscopy cracking technology, and set the core parameters of the experimental system, including stretching rate, stretching stroke and sampling frequency;

[0053] In this embodiment, a single-molecule conductivity measurement experimental platform was constructed using scanning tunneling microscopy split junction (STM-BJ) technology. The assembly and debugging of the gold needle tip electrode, displacement control module, conductivity signal acquisition module, and temperature control module were completed to ensure coordinated operation and accurate signal acquisition. Based on industry standards and experimental requirements for single-molecule conductivity measurement, the core parameters of the experimental system were set as follows: stretching rate 0.01~0.1 nm / ms, stretching stroke 10~100 nm, and sampling frequency 10~100 kHz. This ensured that the measurement conditions for the blank substrate and all molecular substrates were completely consistent, avoiding data deviations caused by parameter differences.

[0054] 202. Using a gold needle tip as the upper electrode and a blank gold substrate or a molecularly functionalized gold substrate as the lower electrode, the experimental system described above was used to repeatedly perform multiple tensile experiments to obtain multiple tensile test results.

[0055] In this embodiment, a gold needle tip with a purity ≥99.99% was selected as the upper electrode, and a blank gold substrate without any modification and a molecularly functionalized gold substrate modified with the target molecule were used as the lower electrodes. The target molecule was a common anchoring molecule in single-molecule electronics, including thiols, pyridines, and amines. The modification method adopted was self-assembled monolayer (SAM) technology. The target molecule was dissolved in anhydrous ethanol to prepare a solution with a concentration of 1~10 mmol / L. The gold substrate was immersed in the solution for 12~24 hours, so that the target molecule was firmly adsorbed onto the surface of the gold substrate through the anchoring group to form a uniform monolayer.

[0056] Repeat the experiment under the set experimental parameters. ~ Multiple tensile fracture experiments were conducted, and the original conductance-displacement signals were collected in real time for each tensile fracture experiment to form multiple sets of tensile test results. This ensured that the dataset had sufficient statistical validity to meet the data volume requirements of subsequent bi-branch comparative learning and classification models.

[0057] 203. Perform baseline correction, outlier removal, and effective segment truncation preprocessing on the multiple tensile test results to construct the blank substrate experimental dataset and the molecular substrate experimental dataset.

[0058] In this embodiment, a fifth-order polynomial fitting combined with a moving average subtraction method is first used to fit the tensile test results to obtain a baseline drift curve. Then, the baseline drift curve is subtracted from the tensile test results to eliminate baseline shifts caused by temperature fluctuations and vibration interference, ensuring the authenticity of the conductivity signal. Next, a combination of the 3σ criterion and the IQR interquartile range method is used. First, the mean μ and standard deviation σ of all conductivity values ​​are calculated, and extreme outliers where |G-μ|>3σ are removed. Then, the interquartile range (IQR) is calculated using a box plot, and outliers below Q1-1.5IQR and above Q3+1.5IQR are removed to eliminate invalid data with abnormal fluctuations in conductivity values. Finally, by identifying conductivity abrupt change points, the effective displacement range from electrode contact to electrode breakage is extracted. Invalid signal segments where conductivity values ​​tend to stabilize after the electrode moves away are removed, and effective conductivity segments with a length of 500~5000 sampling points are retained. Finally, a clean blank substrate experimental dataset and a molecular substrate experimental dataset are obtained through screening.

[0059] Furthermore, in this embodiment of the invention, the standardization process performed on the two sets of experimental datasets to obtain blank dimension feature vectors and molecular dimension feature vectors includes:

[0060] 301. For each conductivity trajectory line in the blank substrate experimental dataset and the molecular substrate experimental dataset, extract its conductivity value sequence and perform logarithmic transformation.

[0061] In this embodiment, for each conductivity trajectory line in the blank substrate experimental dataset and the molecular substrate experimental dataset, a continuous conductivity value time series is extracted point by point according to the sampling order. Considering the characteristics of single-molecule conductivity data that spans orders of magnitude and has a nonlinear distribution, a logarithmic transformation is performed on the extracted conductivity value sequence. Through the logarithmic transformation, the nonlinear conductivity values ​​spanning orders of magnitude are converted into linearly distributed data, which is suitable for the feature extraction requirements of subsequent histogram statistics. At the same time, the relative difference characteristics of conductivity values ​​are preserved, avoiding the feature overwhelming problem caused by the large numerical range.

[0062] 302. Using histogram statistics, the conductivity value sequence after logarithmic transformation is divided into a preset number of equidistant conductivity intervals, and the frequency of conductivity values ​​of each conductivity trajectory line in each conductivity interval is counted to construct a one-dimensional frequency feature blank distribution vector and molecular distribution vector.

[0063] In this embodiment, histogram statistics are used to perform feature transformation on the logarithmically transformed conductivity value sequence. The preset number of equidistant conductivity intervals is fixed at 64. The conductivity value of each conductivity trajectory line after logarithmic transformation is assigned to the corresponding equidistant conductivity interval. The frequency of conductivity value occurrence for each trajectory line within each conductivity interval is counted, forming a one-dimensional frequency feature vector, which is the distribution vector. A blank distribution vector is generated corresponding to the blank substrate data, representing the frequency of conductivity value occurrence for a single trajectory line on the blank substrate within the 64 conductivity intervals. A molecular distribution vector is generated corresponding to the molecular substrate data, representing the frequency of conductivity value occurrence for a single trajectory line on the molecular substrate within the 64 conductivity intervals. Each conductivity trajectory line corresponds to a 64-dimensional one-dimensional frequency feature distribution vector.

[0064] 303. Normalize the blank distribution vector and perform principal component analysis dimensionality reduction on the molecular distribution vector to obtain the blank dimension feature vector and the molecular dimension feature vector, respectively.

[0065] In this embodiment, the blank distribution vector and the molecular distribution vector are normalized and subjected to principal component analysis (PCA) dimensionality reduction. First, the min-max normalization method is used to uniformly reduce each frequency value in the distribution vector to the [0, 1] interval to eliminate the interference caused by the difference in the magnitude of the frequency values ​​of different conductance trajectories and avoid the imbalance of feature weights caused by some trajectory lines having excessively high frequency values. Then, principal component analysis is performed on the normalized distribution vector to calculate the covariance matrix, extract eigenvalues ​​and eigenvectors, and retain the principal components with a variance contribution rate ≥95%, reducing the 64-dimensional distribution vector to a fixed 32-dimensional dimension to reduce data redundancy and the computational complexity of subsequent model training, and improve the efficiency of feature expression. Finally, the blank dimension feature vector and the molecular dimension feature vector are obtained. The blank dimension feature vector is a 32-dimensional real vector used to characterize the core conductance distribution characteristics of the blank substrate trajectory line. The molecular dimension feature vector is also a 32-dimensional real vector used to characterize the core conductance distribution characteristics of the molecular substrate trajectory line.

[0066] Further, in this embodiment of the invention, the step of using the blank dimension feature vector as a reference sample and sequentially performing dual-branch contrastive learning processing and feature fusion processing on the blank dimension feature vector and the molecular dimension feature vector to obtain blank global features and molecular global features includes:

[0067] 401. Construct a contrastive learning neural network with dual-branch shared weights, and input the blank dimension feature vector as a reference sample into the first branch of the contrastive learning neural network, and input the molecular dimension feature vector into the second branch of the contrastive learning neural network;

[0068] In this embodiment, a contrastive learning neural network with shared weights in both branches is constructed. The network structure of the contrastive learning neural network consists of three fully connected layers combined with the ReLU activation function. The input layer has a dimension of 32, the first hidden layer has a dimension of 64, the second hidden layer has a dimension of 32, and the output layer has a dimension of 32. The input, hidden, and output layer structures of the two branches are completely identical, and they share all network weights, ensuring that the feature extraction rules and parameter settings of the two branches are completely consistent, thus avoiding feature extraction deviations caused by differences in network structure.

[0069] The blank dimension feature vectors are used as benchmark samples and are input into the first branch of the contrastive learning neural network in batches. The molecular dimension feature vectors are input into the second branch of the neural network in batches to ensure that feature extraction is performed in both branches simultaneously, laying the foundation for subsequent contrastive learning.

[0070] 402. The contrastive learning neural network is trained using the InfoNCE loss function to extract its output blank high-level semantic features and molecular high-level semantic features.

[0071] In this embodiment, the InfoNCE loss function is selected as the training optimization target of the contrastive learning neural network. During the training process, by minimizing the loss function, the similar sample features inside the blank substrate are forced to aggregate in the high-dimensional space, the similar sample features inside the molecular substrate are aggregated, and the dissimilar sample features between the blank substrate and the molecular substrate are separated in the high-dimensional space. This allows the contrastive learning neural network to learn to distinguish the core differences between pure tunneling signals and molecular conductance signals, and greatly improves the feature discrimination.

[0072] After training, the 32-dimensional feature vectors of the output layer of the contrastive learning neural network are extracted, which are the high-level semantic features: the 32-dimensional feature vectors output from the first branch, corresponding to the blank dimension feature vectors, are blank high-level semantic features, used to characterize the deep physical nature of pure tunneling signals, such as the absence of conductance steps and the exponential decay of conductance distribution characteristics; the 32-dimensional feature vectors output from the second branch, corresponding to the molecular dimension feature vectors, are molecular high-level semantic features, used to characterize the deep physical nature of molecular conductance signals, such as the presence of conductance steps and the multi-peak distribution of conductance characteristics.

[0073] 403. The high-level semantic features of the blank space are concatenated and fused with the feature vector of the blank space dimension to obtain the global feature of the blank space;

[0074] 404. The high-level semantic features of the molecule and the dimensional feature vector of the molecule are concatenated and fused to obtain the global features of the molecule;

[0075] In this embodiment, a feature concatenation approach is adopted to deeply fuse the blank high-level semantic features with the blank dimensional feature vector, and to deeply fuse the molecular high-level semantic features with the molecular dimensional feature vector. The resulting blank global features and molecular global features both possess the basic statistical features (from the blank dimensional feature vector) and deep semantic features (from the blank high-level semantic features) of the blank substrate trajectory line. This approach can comprehensively and accurately characterize the conductivity characteristics of the conductivity trajectory line, and has strong robustness and physical specificity. It solves the problems of insufficient characterization ability and poor anti-interference of single feature types in the prior art.

[0076] Further, in this embodiment of the invention, the molecular global features are subjected to single-classifier pre-screening and binary-classifier fine-screening sequentially based on the blank global features to obtain positive molecular samples. The positive molecular samples include multiple conductivity trajectory lines with conductivity step characteristics, including:

[0077] 501. Train a single classifier model using blank global features as the training set;

[0078] In this embodiment, a single-class support vector machine (OC-SVM) is selected as the single classifier model. The kernel function of the OC-SVM model is the RBF kernel, the kernel parameter γ=0.1, and the penalty coefficient C=1.0. During training, only blank global features are used as the only training samples, without introducing any molecular substrate samples. The single classifier model learns the distribution law of blank global features and fits a hyperplane boundary that can wrap all blank global features. This hyperplane boundary is the feature distribution boundary of pure tunneling signals. The inside of the boundary is the feature of pure tunneling signals, and the outside of the boundary is the feature of non-pure tunneling signals, thereby establishing an accurate judgment benchmark for tunneling signals. The trained single classifier model can quickly and accurately identify pure tunneling invalid samples, realize the initial rapid screening of invalid samples, and greatly improve the efficiency of subsequent processing.

[0079] 502. Input the global molecular features into the trained single classifier model, remove pure tunneling samples that are highly similar to the blank global features, and retain potential abnormal samples including conductivity steps as pre-screening results.

[0080] In this embodiment, molecular global features are input in batches into the trained single classifier model. The model determines the sample attributes by calculating the distance between each molecular global feature and the distribution boundary of the pure tunneling signal feature: if the molecular global feature falls inside the hyperplane boundary, it is determined to be a pure tunneling invalid sample and is directly removed; if the molecular global feature falls outside the hyperplane boundary, it is determined to be a non-pure tunneling sample, that is, an abnormal sample that may contain conductivity steps, and is retained as a pre-screening result, completing the initial enrichment of effective samples. This can reduce the data processing volume of subsequent binary classification fine screening, greatly improve the efficiency of the entire process analysis, and at the same time avoid the interference of invalid samples on subsequent fine screening.

[0081] 503. Construct a binary classification training set using blank global features as samples with label 0 and pre-screened abnormal samples as samples with label 1, and train a binary classifier model based on the binary classification training set.

[0082] In this embodiment, blank global features are labeled as negative samples without conductivity steps (label 0), and abnormal samples obtained from single-class pre-screening are labeled as positive samples with conductivity steps (label 1). A binary classification training set is constructed by mixing them in a 1:1 ratio, with 80% used as the training set and 20% as the validation set.

[0083] Multilayer perceptron (MLP) was selected as the binary classifier model. The model structure consists of an input layer (64-dimensional, consistent with the global feature dimension), two hidden layers (64-dimensional and 32-dimensional respectively), and an output layer (2-dimensional, corresponding to the two categories). The activation function is ReLU, the loss function is cross-entropy loss, and the optimizer is Adam (learning rate 0.001, decay coefficient 0.9).

[0084] During training, a 5-fold cross-validation method was used, dividing the training set into 5 parts. Four parts were used as the training set and one part as the validation set, and the training was repeated 5 times. After each training, the classification accuracy of the validation set was calculated. The classification boundary was optimized by adjusting the network parameters to ensure that the model's classification accuracy on the validation set was ≥95%, thus completing the training of the binary classifier model.

[0085] By constructing a binary classification discrimination model, the problems of inaccurate discrimination and high misclassification rate of a single classifier are effectively solved, and the accuracy of distinguishing between valid and invalid samples is greatly improved.

[0086] 504. Input the global molecular features into the trained binary classifier model, and separate positive molecular samples including conductivity step features by cross-validation with a preset probability threshold.

[0087] In this embodiment, the global molecular features are batch-input into the trained binary classifier model, and the model outputs the probability value of each sample belonging to the positive sample. On the validation set of the binary classification training, the classification accuracy, recall, and F1 score corresponding to different probability thresholds are tested. The probability threshold with the highest F1 score is selected as the optimal threshold, i.e., the preset probability threshold. Samples with probability values ​​higher than the optimal threshold are determined as molecular positive samples containing conductivity step features, thus completing the accurate extraction of effective conductivity traces.

[0088] By optimizing the classification probability threshold through cross-validation, the accuracy and recall of classification are balanced, maximizing the screening accuracy of effective samples. This solves the problems of low accuracy and high false positive rate in the screening of effective traces in existing technologies, and achieves accurate extraction of effective conductivity traces.

[0089] Furthermore, in this embodiment of the invention, the standardization of the molecular positive samples to obtain standard molecular positive samples and the determination of the optimal number of clusters includes:

[0090] 601. The positive molecular samples are standardized using the Z-Score standardization method to obtain standard positive molecular samples;

[0091] In this embodiment, the mean and standard deviation of each feature dimension of the positive molecular sample are calculated, and each feature value is converted into standard data with a mean of 0 and a standard deviation of 1 by using the Z-Score formula, thereby eliminating the differences in dimensions and values ​​between feature dimensions.

[0092] 602. Based on the standard molecular positive samples, calculate the sample distortion coefficient and sample contour coefficient under different preset cluster numbers, and plot the distortion coefficient curve for correlating the distortion coefficient with the number of clusters and the contour coefficient curve for correlating the contour coefficient with the number of clusters.

[0093] In this embodiment, a range of consecutive cluster numbers is first set. Based on the conventional rules of single-molecule experiments, the range of cluster numbers is set to 2~10. For each preset cluster number, the corresponding sample distortion coefficient and sample profile coefficient are calculated. Specifically, the sample distortion coefficient is obtained by calculating the sum of squared Euclidean distances from all samples within each cluster to the cluster center, and then summing the squared distances of all clusters. The smaller the sample distortion coefficient, the higher the aggregation degree of the samples within the cluster. For each sample, the average distance between it and all other samples within the cluster is calculated, and then the average distance between it and all samples in the nearest other cluster is calculated to obtain the sample profile coefficient, which ranges from [-1, 1]. The closer the sample profile coefficient is to 1, the better the sample clustering effect. The average of the profile coefficients of all samples is taken to obtain the overall profile coefficient corresponding to that cluster number.

[0094] After the calculation is completed, the distortion coefficient curve and the contour coefficient curve are plotted with the number of clusters on the horizontal axis and the coefficient value on the vertical axis, respectively, to intuitively present the variation of the coefficient with the number of clusters.

[0095] 603. Combine the elbow inflection point of the distortion coefficient curve and the maximum value point of the profile coefficient curve to determine the range of candidate optimal cluster numbers;

[0096] In this embodiment, the distortion coefficient curve is analyzed to find the point where the downward trend of the curve changes abruptly, namely the elbow inflection point. The elbow inflection point indicates that the degree of aggregation of samples within a cluster has reached a high level. If the number of clusters is further increased, the decrease in the sample distortion coefficient will be significantly reduced, and the clustering gain will not be significant. The silhouette coefficient curve is analyzed to find the point with the largest value in the curve. The maximum value point indicates that the clustering effect is optimal at this time, with high similarity of samples within a cluster and large difference between samples between clusters.

[0097] Using the number of clusters corresponding to these two points as the core, the range of candidate optimal cluster numbers is determined. Typically, the range of optimal cluster numbers is 2 to 5. By combining two quantitative indicators to screen candidate cluster numbers, both the aggregation degree of samples within the cluster and the optimal clustering effect are guaranteed, which greatly improves the scientific nature of cluster number selection. At the same time, it narrows the judgment range of the optimal cluster number, avoids the blindness when combining experimental priors in subsequent selection, and provides a clear direction for finally determining the optimal cluster number.

[0098] 604. Obtain prior experimental knowledge, and based on the prior experimental knowledge, determine the optimal number of clusters from the range of candidate optimal cluster numbers;

[0099] In this embodiment, prior knowledge of the single-molecule experiment is obtained, specifically including the number of adsorption configurations of the target molecule, the stable connection mode of gold-molecule-gold junction, etc. Based on the prior knowledge of the experiment, a value matching the physical law is selected from the range of candidate optimal cluster numbers as the final optimal cluster number. If multiple values ​​in the candidate cluster number range match the physical law, the value with the largest silhouette coefficient is selected as the optimal cluster number. If the silhouette coefficients are the same, the value that is completely consistent with the number of adsorption configurations in the prior experiment is selected to ensure that the cluster number conforms to the physical law and ensures the optimal clustering effect.

[0100] Further, in this embodiment of the invention, the step of performing unsupervised clustering on the standard molecular positive samples based on the optimal number of clusters to obtain classified positive samples includes:

[0101] 701. Obtain a pre-selected unsupervised clustering model, input the standard positive molecular samples into the unsupervised clustering model, so as to divide the positive molecular samples into multiple cluster categories corresponding to the optimal number of clusters;

[0102] In this embodiment, a pre-selected unsupervised clustering model is obtained, and an appropriate model is selected based on the data distribution characteristics of the standard positive molecular samples:

[0103] If the sample distribution is uniform and there is no obvious noise, the K-Means model should be selected first. The model parameters are set as follows: maximum number of iterations 300, and the initial cluster centers are selected using the K-Means++ algorithm to avoid clustering bias caused by random initial centers.

[0104] If the sample distribution is irregular and there is a small amount of noise, the DBSCAN model is selected, and the model parameters are set as follows: neighborhood radius = 0.5, minimum number of samples in the neighborhood = 5.

[0105] To improve the interpretability of clustering results, a hierarchical clustering model can be selected, and the Ward linking method can be used to aggregate samples layer by layer according to the principle of minimizing intra-cluster variance, which can clearly present the hierarchical relationship between samples.

[0106] Standard positive molecular samples are input in batches into the selected unsupervised clustering model. The model automatically groups the samples based on the similarity of their features. The number of groups matches the optimal number of clusters, ensuring that the conductivity characteristics of samples within each cluster are highly consistent and that the conductivity characteristics of samples between different clusters are significantly different. Finally, the positive molecular samples are divided into multiple clusters corresponding to the optimal number of clusters, and each cluster corresponds to a potential conductivity step height.

[0107] 702. For each cluster category, extract the conductivity step height feature from the corresponding conductivity trajectory line;

[0108] In this embodiment, for all conductivity trajectories within each cluster category, a peak detection combined with statistical averaging method is used to extract conductivity step height features. Specifically, for each conductivity trajectory, an adaptive threshold peak detection algorithm is used, setting the threshold to 1.5 times the mean conductivity of the trajectory, to identify plateau segments where the conductivity value remains stable, i.e., conductivity steps, and to eliminate pseudo-steps with short-term fluctuations. Then, the mean conductivity value of each plateau segment is taken as the conductivity step height value of that trajectory. For the conductivity step height values ​​of all trajectories within the same cluster category, their arithmetic mean is calculated as the conductivity step center height value of that cluster category. At the same time, the standard deviation is calculated to ensure the consistency of step height within the category. The calculated center height value is used as the core physical feature of that cluster category.

[0109] 703. Based on the extracted conductivity step height features, the multiple cluster categories are labeled respectively, and the multiple labeled cluster categories are integrated to obtain the positive classification samples;

[0110] In this embodiment, multiple cluster categories are uniquely labeled based on the center height value of the conductivity step extracted from each cluster category; all labeled cluster categories are integrated, and key information such as the number of samples, conductivity step height range, and standard deviation of each category are summarized to form a complete positive classification sample; the positive classification sample can be directly used for single-molecule charge transport characteristic research, molecular structure-conductivity performance correlation analysis, and other scenarios.

[0111] The above describes the unsupervised identification method for single-molecule charge transport in embodiments of the present invention. The following describes the unsupervised identification device for single-molecule charge transport in embodiments of the present invention. Please refer to [link to relevant documentation]. Figure 2 One embodiment of the unsupervised identification device for single-molecule charge transport in this invention includes:

[0112] The acquisition module 801 is used to acquire at least two sets of experimental datasets, the experimental datasets including blank substrate experimental datasets and molecular substrate experimental datasets;

[0113] Processing module 802 is used to standardize the two sets of experimental datasets respectively to obtain blank dimension feature vectors and molecular dimension feature vectors;

[0114] The learning module 803 is used to perform bi-branch comparative learning and feature fusion processing on the blank dimension feature vector and the molecular dimension feature vector in sequence, using the blank dimension feature vector as the reference sample, to obtain blank global features and molecular global features.

[0115] The screening module 804 is used to perform single-classifier pre-screening and binary-classifier fine screening on the molecular global features based on the blank global features to obtain positive molecular samples. The positive molecular samples include multiple conductivity trajectory lines with conductivity step features.

[0116] The determination module 805 is used to standardize the molecular positive samples to obtain standard molecular positive samples and determine the optimal number of clusters;

[0117] The classification module 806 is used to perform unsupervised clustering on the standard molecular positive samples based on the optimal number of clusters to obtain classified positive samples, wherein the classified positive samples are the classification results of multiple conductivity trajectory lines classified according to conductivity step height.

[0118] Based on the same ideas as the methods in the above embodiments, the apparatus provided in this application can implement the methods in the above embodiments.

[0119] above Figure 2 The unsupervised identification device for single-molecule charge transport in this embodiment of the invention is described in detail from the perspective of modular functional entities. The unsupervised identification device for single-molecule charge transport in this embodiment of the invention is described in detail below from the perspective of hardware processing.

[0120] Figure 3This is a schematic diagram of the structure of a single-molecule charge transport unsupervised identification device 900 provided in an embodiment of the present invention. The single-molecule charge transport unsupervised identification device 900 can vary significantly due to different configurations or performance. It may include one or more central processing units (CPUs) 910 and a memory 920, and one or more storage media 930 (e.g., one or more mass storage devices) storing application programs 933 or data 932. The memory 920 and storage media 930 can be temporary or persistent storage. The program stored in the storage media 930 may include one or more modules (not shown in the diagram), each module may include a series of instruction operations on the single-molecule charge transport unsupervised identification device 900. Furthermore, the processor 910 may be configured to communicate with the storage media 930 and execute the series of instruction operations in the storage media 930 on the single-molecule charge transport unsupervised identification device 900 to implement the steps of the single-molecule charge transport unsupervised identification method provided in the above-described method embodiments.

[0121] The unsupervised identification device 900 for single-molecule charge transport may also include one or more power supplies 940, one or more wired or wireless network interfaces 950, one or more input / output interfaces 960, and / or one or more operating systems 931, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 3 The illustrated unsupervised identification device structure for single-molecule charge transport does not constitute a limitation on unsupervised identification devices for single-molecule charge transport. It may include more or fewer components than illustrated, or combine certain components, or have different component arrangements.

[0122] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when executed on a computer, cause the computer to perform the steps of the unsupervised identification method for single-molecule charge transport.

[0123] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system, device, or unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0124] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0125] Finally, it should be noted that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for unsupervised identification of single-molecule charge transport, characterized in that, include: Obtain at least two sets of experimental datasets, including a blank substrate experimental dataset and a molecular substrate experimental dataset; The two sets of experimental datasets were standardized to obtain blank dimension feature vectors and molecular dimension feature vectors. Using the blank dimension feature vector as a reference sample, the blank dimension feature vector and the molecular dimension feature vector are sequentially subjected to dual-branch contrastive learning and feature fusion processing to obtain blank global features and molecular global features. Specifically, a contrastive learning neural network with shared weights is constructed. The blank dimension feature vector is used as a reference sample and input into the first branch of the contrastive learning neural network, while the molecular dimension feature vector is input into the second branch. The contrastive learning neural network is trained using the InfoNCE loss function to extract its output blank high-level semantic features and molecular high-level semantic features. The blank high-level semantic features and the blank dimension feature vector are then concatenated and fused to obtain the blank global features. The molecular high-level semantic features and the molecular dimensional feature vectors are concatenated and fused to obtain the molecular global features; Based on the blank global features, the molecular global features are subjected to single classifier pre-screening and binary classifier fine screening in sequence to obtain positive molecular samples. The positive molecular samples include multiple conductivity trajectory lines with conductivity step features. The positive molecular samples are standardized to obtain standard positive molecular samples, and the optimal number of clusters is determined. Based on the optimal number of clusters, unsupervised clustering is performed on the standard molecular positive samples to obtain classified positive samples. The classified positive samples are the classification results of multiple conductivity trajectory lines classified according to conductivity step height.

2. The method according to claim 1, characterized in that, The acquisition of at least two sets of experimental datasets includes: A single-molecule conductivity measurement experimental system based on scanning tunneling microscopy crack junction technology was constructed, and the core parameters of the experimental system were set, including stretching rate, stretching stroke and sampling frequency. Using a gold needle tip as the upper electrode and a blank gold substrate or a molecularly functionalized gold substrate as the lower electrode, the experimental system was used to repeatedly perform multiple tensile experiments to obtain multiple tensile test results. Baseline correction, outlier removal, and effective segment truncation preprocessing are performed on the multiple tensile test results to construct the blank substrate test dataset and the molecular substrate test dataset.

3. The method according to claim 1, characterized in that, The standardization process for the two sets of experimental datasets yields blank dimension feature vectors and molecular dimension feature vectors, including: For each conductivity trajectory line in the blank substrate experimental dataset and the molecular substrate experimental dataset, extract its conductivity value sequence and perform logarithmic transformation; Histogram statistics are used to divide the conductivity value sequence after logarithmic transformation into a preset number of equidistant conductivity intervals, and the frequency of conductivity values ​​of each conductivity trajectory line in each conductivity interval is counted to construct a blank distribution vector and a molecular distribution vector of one-dimensional frequency features. The blank distribution vector and the molecular distribution vector are normalized and subjected to principal component analysis for dimensionality reduction, respectively, to obtain the blank dimension feature vector and the molecular dimension feature vector.

4. The method according to claim 1, characterized in that, Based on the blank global features, the molecular global features are sequentially subjected to single-classifier pre-screening and binary-classifier fine-screening to obtain positive molecular samples. These positive molecular samples include multiple conductivity trajectory lines with conductivity step characteristics, including: A single classifier model is trained using blank global features as the training set. The molecular global features are input into the trained single classifier model, and pure tunneling samples that are highly similar to the blank global features are removed, while potential abnormal samples including conductivity steps are retained as pre-screening results. A binary classification training set is constructed using samples labeled 0 with blank global features and samples labeled 1 with pre-screened abnormal samples, and a binary classifier model is trained based on the binary classification training set. The global molecular features are input into a trained binary classifier model, and positive molecular samples, including conductivity step features, are separated by cross-validation with a preset probability threshold.

5. The method according to claim 1, characterized in that, The standardization process for the positive molecular samples to obtain standard positive molecular samples and the determination of the optimal number of clusters include: The positive molecular samples were standardized using the Z-Score standardization method to obtain standard positive molecular samples. Based on the standard molecular positive samples, the sample distortion coefficient and sample contour coefficient under different preset cluster numbers are calculated respectively, and the distortion coefficient curve and the contour coefficient curve are plotted to correlate the distortion coefficient and the number of clusters. By combining the elbow inflection point of the distortion coefficient curve and the maximum point of the profile coefficient curve, the range of candidate optimal cluster numbers is determined. Obtain prior experimental knowledge, and based on the prior experimental knowledge, determine the optimal number of clusters from the range of candidate optimal cluster numbers.

6. The method according to claim 1, characterized in that, The step of performing unsupervised clustering on the standard molecular positive samples based on the optimal number of clusters to obtain classified positive samples includes: Obtain a pre-selected unsupervised clustering model, input the standard positive molecular samples into the unsupervised clustering model, so as to divide the positive molecular samples into multiple cluster categories corresponding to the optimal number of clusters; For each cluster category, extract the conductivity step height feature from the corresponding conductivity trajectory line; Based on the extracted conductivity step height features, the multiple cluster categories are labeled separately, and the labeled cluster categories are integrated to obtain the positive classification samples.

7. A single-molecule charge transport unsupervised identification device, characterized in that, include: The acquisition module is used to acquire at least two sets of experimental datasets, including a blank substrate experimental dataset and a molecular substrate experimental dataset. The processing module is used to standardize the two sets of experimental datasets respectively to obtain blank dimension feature vectors and molecular dimension feature vectors; The learning module is used to perform bi-branch contrastive learning and feature fusion processing on the blank dimension feature vector and the molecular dimension feature vector, respectively, using the blank dimension feature vector as a reference sample, to obtain blank global features and molecular global features. Specifically, a contrastive learning neural network with shared weights is constructed. The blank dimension feature vector is input as a reference sample into the first branch of the contrastive learning neural network, and the molecular dimension feature vector is input into the second branch. The contrastive learning neural network is trained using the InfoNCE loss function to extract its output blank high-level semantic features and molecular high-level semantic features. The blank high-level semantic features and the blank dimension feature vector are fused together to obtain the blank global features. The molecular high-level semantic features and the molecular dimension feature vector are fused together to obtain the molecular global features. The screening module is used to perform single-classifier pre-screening and binary-classifier fine screening on the molecular global features based on the blank global features to obtain positive molecular samples. The positive molecular samples include multiple conductivity trajectory lines with conductivity step features. The determination module is used to standardize the molecular positive samples to obtain standard molecular positive samples and determine the optimal number of clusters; The classification module is used to perform unsupervised clustering on the standard molecular positive samples based on the optimal number of clusters to obtain classified positive samples, wherein the classified positive samples are the classification results of multiple conductance trajectory lines classified according to the conductance step height.

8. A single-molecule charge transport unsupervised identification device, characterized in that, The unsupervised identification device for single-molecule charge transport includes: a memory and at least one processor, wherein the memory stores instructions; At least one of the processors invokes the instructions in the memory to cause the unsupervised identification device for single-molecule charge transport to perform the steps of the unsupervised identification method for single-molecule charge transport as described in any one of claims 1-6.

9. A computer-readable storage medium storing instructions thereon, characterized in that, When the instructions are executed by the processor, they implement the steps of the unsupervised identification method for single-molecule charge transport as described in any one of claims 1-6.

Images