Digital twin method and system based on multi-type chromatography unified physical modeling and machine learning

Abstract

The present application relates to the technical field of chromatographic purification, and particularly relates to a digital twin method and system based on unified physical modeling and machine learning of multiple types of chromatography; the present application constructs a unified physical modeling framework containing at least five chromatographic separation mechanisms. Through the base class design based on the unified interface, the framework can be compatible with multiple mainstream separation mechanisms such as affinity, ion exchange, volume exclusion, hydrophobic and reversed phase chromatography, and new types only need to inherit the base class and implement the calculation logic of the unified interface, so that the existing modeling system can be seamlessly accessed, and the cost of system expansion and iteration is greatly reduced; a hybrid architecture decoupling the physical model and the machine learning correction model is adopted, the physical model layer independently outputs a prediction benchmark, and the machine learning model only needs to calculate a correction residual based on a lightweight gradient boosting decision tree, can run at a high speed without GPU, and is very suitable for deployment on an edge side industrial controller or a server.

Description

Technical Field

[0001] This application relates to the field of chromatography purification technology, and in particular to a digital twin method and system based on unified physical modeling and machine learning for multiple types of chromatography. Background Technology

[0002] Chromatographic purification is a core step in the biopharmaceutical process. Currently, the construction of digital twin models of chromatographic processes mainly faces the following two problems: First, there is a lack of a unified physical modeling architecture for multiple types of chromatography, and existing hybrid modeling schemes suffer from strong coupling and high computational overhead. In general, existing chromatography models are typically built for single mechanisms such as affinity, ion exchange, and size exclusion, making it difficult to accommodate multiple separation mechanisms. To address this, existing technologies have proposed hybrid modeling schemes that embed physical equations into machine learning loss functions, such as the PINN architecture. A search reveals that US Patent Publication No. US12682233B2 discloses the optimization and digital twinning of chromatographic purification processes using a neural network guided by physical knowledge. It discloses a system and method for constructing a digital twin of a chromatographic purification process using a mechanistic model and process data for process optimization. Additionally, International Patent Publication No. WO / 2024 / 077271 discloses a digital twin and anomaly detection framework that includes a hybrid mechanistic model of biological processes (mass and energy balance). However, this architecture deeply and strongly couples the physical model with the machine learning model, causing the model training to be extremely dependent on GPU computing power and the single inference time to be long, making it difficult to achieve millisecond-level real-time response on ordinary CPU controllers in industrial fields. At the same time, due to the tight coupling, once the physical equation or actual process parameters are locally adjusted, the entire hybrid model often needs to be retrained globally, which cannot adapt to the needs of rapid iteration of biopharmaceutical processes and efficient deployment at the edge.

[0003] Second, the online accuracy of purely physical models is insufficient, and there is a lack of dynamic compensation methods for systematic biases. Traditional purely physical models are built based on first principles, and their performance is acceptable during offline calibration. However, after long-term online operation, due to factors such as aging of the chromatography medium, slight fluctuations in the mobile phase ratio, and changes in environmental temperature and humidity, the output values ​​of the physical model calibrated offline (such as UV absorbance and column end pressure) will accumulate biases over time. Currently, correction is generally achieved through periodic offline recalibration. Therefore, it is difficult to dynamically compensate for errors in physical reference values ​​based on real-time sensor data during the operation of the chromatography process. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this application provides a digital twin method and system based on unified physical modeling of multiple types of tomography and machine learning. This solves the technical problems that currently exist in constructing digital twin models of tomography processes, such as the lack of a unified physical modeling architecture for multiple types of tomography, strong coupling and high computational overhead of existing hybrid modeling schemes, insufficient online accuracy of pure physical models, and the lack of dynamic compensation methods for systematic deviations.

[0005] To address the aforementioned technical problems, this application provides the following technical solution: a digital twin method based on multi-type tomography unified physical modeling and machine learning, comprising: S1. Obtain the chromatography process parameters and input them into the pre-constructed multi-type unified physical model layer of chromatography. The physical model layer contains sub-models of various chromatography separation mechanisms. Each sub-model calculates the chromatography peak parameters based on the first-principles equations and outputs the UV prediction baseline value and pressure prediction baseline value of the peak segment of the chromatography process. S2. Obtain the process parameter features collected by the sensor in real time, and use them as input features along with the UV prediction baseline value and the pressure prediction baseline value. Input them into the pre-built machine learning correction model that is decoupled from the physical model layer, and output the corresponding UV correction residual and pressure correction residual. S3. Superimpose the UV prediction baseline value and the UV correction residual to obtain the final UV prediction peak segment, and superimpose the pressure prediction baseline value and the pressure correction residual to obtain the final pressure prediction peak segment, and output them.

[0006] Preferably, in step S1, the physical model layer includes a unified interface configured to receive process parameters including at least column volume, flow rate, column bed height and particle size, and uniformly output chromatography peak parameters including at least peak position, peak width, peak symmetry, theoretical plate number and resolution.

[0007] Preferably, the physical model layer is deployed decoupled from the machine learning correction model. The physical model layer runs independently and outputs UV prediction baseline values ​​and pressure prediction baseline values, while the machine learning correction model is trained independently. During inference, the UV prediction baseline values ​​and pressure prediction baseline values ​​are used as input features of the machine learning correction model and concatenated with process parameter features collected by real-time sensors to form a complete feature vector, which is then input into the machine learning correction model.

[0008] Preferably, in step S2, the input features of the machine learning correction model include at least UV detection value, conductivity, pH value, inlet pressure, outlet pressure, flow rate, number of column cycles, step time, column volume multiple, UV trend, conductivity trend, pressure trend, UV prediction baseline value, and pressure prediction baseline value; the trend is the linear regression slope of the corresponding sensor data within the current time window.

[0009] Preferably, in step S1, the sub-models of the at least five chromatographic separation mechanisms are the affinity chromatography sub-model, the ion exchange chromatography sub-model, the size exclusion chromatography model, the hydrophobic chromatography model, and the reversed-phase chromatography model. The affinity chromatography model is based on the Langmuir adsorption isotherm and a pH-dependent dissociation constant model. The ion exchange chromatograph model is based on the SMA model. The volumetric exclusion tomography model is based on the Stokes radius and pore size logarithmic correlation model. The hydrophobic chromatography sub-model is modeled based on a salt-induced binding and decreasing salt concentration elution model. The inverse tomography model is modeled based on the LSS model.

[0010] Preferably, in step S2, the training steps of the machine learning correction model include: calling the unified interface of the physical model layer in step S1, using the physical model to generate multiple batches of virtual tomography running data for each tomography type, and collecting and generating multiple rows of raw data according to a preset sampling interval; extracting and aggregating features from the multiple batches of virtual tomography running data according to a preset time window, filtering out redundant time windows with UV values ​​lower than a preset threshold, and obtaining training samples; using the difference between the real UV signal and the UV prediction benchmark value output by the physical model as the UV training label, and the difference between the real pressure signal and the pressure prediction benchmark value output by the physical model as the pressure training label, and training the machine learning correction model using the training samples.

[0011] This application also provides a digital twin system based on multi-type tomographic unified physical modeling and machine learning, including: The physical model module is used to integrate sub-models of at least five chromatographic separation mechanisms, calculate chromatographic peak parameters based on input chromatographic process parameters, and output UV prediction baseline values ​​and pressure prediction baseline values ​​for the chromatographic process. The machine learning correction module, decoupled from the physical model module, is used to receive the process parameter features collected in real time by the sensor and the predicted baseline value output by the physical model module, and output the UV correction residual and the pressure correction residual. The superposition output module is used to superimpose the UV prediction reference value and the pressure prediction reference value with the corresponding correction residuals, and output the final UV prediction peak segment and the final pressure prediction peak segment.

[0012] Preferably, the digital twin system further includes: The peak segment evaluation module is used to compare and evaluate the final UV prediction peak segment obtained by the overlay output module with the actual collected UV peak segment. Based on the set peak detection threshold, the predicted UV data and the actual UV data are binarized and segmented. A confusion matrix is ​​calculated, which includes at least correctly detected TP, false positive FP, false negative FN, and correctly excluded TN. Precision and recall are calculated based on the confusion matrix, and then the F1 score is calculated. The F1 score is used as the main evaluation index for evaluating the peak segment detection effect of the chromatography, and the mean absolute error of the peak segment and the peak segment determination coefficient are used as auxiliary evaluation indexes. The peak segment determination coefficient is the determination coefficient calculated for samples with UV values ​​greater than the peak detection threshold.

[0013] This application also provides a computer-readable storage medium storing a computer program thereon, characterized in that the computer program, when executed by a processor, implements the steps of the digital twin method.

[0014] By employing the above technical solutions, this application provides a digital twin method and system based on multi-type tomographic unified physical modeling and machine learning, which has at least the following beneficial effects: 1. This application constructs a unified physical modeling framework that includes at least five chromatographic separation mechanisms. Through a base class design based on a unified interface, the framework is compatible with a variety of mainstream separation mechanisms such as affinity, ion exchange, size exclusion, hydrophobicity, and reversed-phase chromatography. New types can be seamlessly integrated into the existing modeling system simply by inheriting the base class and implementing the computational logic of the unified interface, which greatly reduces the cost of system expansion and iteration.

[0015] 2. This application adopts a hybrid architecture that decouples the physical model and the machine learning correction model. The physical model layer independently outputs the prediction benchmark, and the machine learning model only needs to calculate the correction residual based on the lightweight gradient boosting decision tree. The inference speed is extremely fast and it can run at high speed without GPU. It is very suitable for deployment on edge industrial controllers or servers, and effectively solves the pain points of the existing PINN strongly coupled architecture, which has high computing power overhead and is difficult to meet the needs of industrial real-time control.

[0016] 3. Since the global coefficient of determination is extremely sensitive to small shifts in peak position, it often leads to distortion of evaluation results. The F1 score, based on binary segmentation and confusion matrix calculation, can accurately reflect the location, detection and false negative of peak segments during chromatography. Therefore, this invention uses the peak segment F1 score to replace the global coefficient of determination as the main evaluation index. Its evaluation results are highly consistent with the actual process control capabilities in industrial settings, providing a real and reliable quantitative basis for process optimization.

[0017] 4. This application can generate large-scale and diverse virtual tomography running data based on first-principles equations and EMG functions through a unified interface of the physical model layer. For example, it can generate 100 batches of data for each tomography type, covering 346K rows of sampling points. This feature can generate large-scale virtual training data, significantly reducing the dependence on expensive and scarce real batch data, and greatly saving the data acquisition cost and cycle of model training. Attached Figure Description

[0018] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of the digital twin method based on multi-type tomography unified physical modeling and machine learning of the present invention; Figure 2 This is a structural block diagram of the digital twin system based on multi-type tomography unified physical modeling and machine learning, according to the present invention.

[0019] In the diagram: 1. Physical model module; 2. Machine learning correction module; 3. Superimposed output module; 4. Peak segment evaluation module. Detailed Implementation

[0020] To make the above-mentioned objectives, features, and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments. This will allow for a full understanding and implementation of how the application uses technical means to solve technical problems and achieve technical effects.

[0021] Example 1 To address the technical problems currently existing in constructing digital twin models for tomography processes, such as the lack of a unified physical modeling architecture for multiple types of tomography, the strong coupling and high computational overhead of existing hybrid modeling schemes, and the insufficient online accuracy of pure physical models and the lack of dynamic compensation methods for systematic biases, this embodiment provides a digital twin method based on unified physical modeling of multiple types of tomography and machine learning, including the following steps: S1. Obtain the chromatography process parameters and input them into the pre-constructed multi-type unified physical model layer of chromatography. This physical model layer contains sub-models of various chromatography separation mechanisms. Each sub-model calculates the chromatography peak parameters based on the first-principles equations and outputs the UV prediction baseline value and pressure prediction baseline value of the peak segment of the chromatography process.

[0022] To achieve unification across multiple layer types, in step S1, the physical model layer includes a unified interface. This unified interface is configured to receive process parameters including at least column volume, flow rate, column bed height, and particle size, and to uniformly output chromatography peak parameters including at least peak position, peak width, peak symmetry, theoretical plate number, and resolution. Each subclass calls the unified interface compute_peaks(params) defined in the base class, returning parameters including peak position. Peak width Peak symmetry Theoretical plate number Resolution The result is that the base class extension mechanism supports adding new chromatography types through inheritance, such as mixed-mode chromatography (MMC), hydroxyapatite chromatography (HAT), and metal chelate chromatography (IMAC). New types only need to implement the compute_peaks method of the unified interface to be seamlessly integrated into the existing hybrid modeling framework.

[0023] In this embodiment, at least five sub-models of chromatographic separation mechanisms are included, namely, affinity chromatography, ion exchange chromatography, size exclusion chromatography, hydrophobic chromatography, and reversed-phase chromatography. These sub-models employ first-principles calculations as their core physical equations, and their specific parameters are defined as follows: The affinity chromatography model is based on the Langmuir adsorption isotherm and a pH-dependent dissociation constant model. The core physical equation is: in, This refers to the adsorption capacity. For maximum adsorption capacity, The concentration in the liquid phase. is the dissociation constant.

[0024] The corrected equation for the effect of pH on the dissociation constant during pH elution is as follows: in, This represents the corrected dissociation constant. For reference dissociation constant, pH sensitivity coefficient For reference pH.

[0025] The ion-exchange chromatograph model is based on the SMA model, and its core physical equations are: In the above formula, It is the equilibrium constant. Components The concentration of the adsorbed phase, Salt concentration, The characteristic charge number, Components The liquid phase concentration, denoted as the adsorption capacity of the salt phase.

[0026] The relationship between retention volume and salt concentration is as follows: In the above formula, Indicates the volume to be retained. This represents the salt concentration.

[0027] The volumetric exclusion tomography model is based on the Stokes radius versus pore size logarithmic correlation model, and its core physical equations are as follows: The log-correlation equation is: In the above formula, For allocation coefficients, For the volume outside the column, For the volume of the cylinder, The Stokes radius of the solute. The aperture is [not specified].

[0028] The hydrophobic chromatography sub-model is based on a salt-induced binding and decreasing salt concentration elution model, and its core physical equations are as follows: In the above formula, As capacity factor, and These are two empirical constants specific to the solute and the chromatography medium.

[0029] The antiphase tomography model is based on the LSS model, and its core physical equations are as follows: In the above formula, The capacity factor when the organic phase fraction is 0. The elution strength coefficient, This represents the organic phase fraction.

[0030] The physical model layer generates ultraviolet (UV) curves based on the EMG (Exponentially Modified Gaussian) function, and its mathematical expression is as follows: In the above formula, For the independent variable The EMG (Exponentially Modified Gaussian) function, The independent variable is CV, representing time or column volume multiple. The peak area is calculated by the physical model based on the adsorption / elution amount. The retention time for the peak position is calculated by the core equation of each subclass. The standard deviation is Gaussian, reflecting the degree of peak diffusion within the column, and is related to the theoretical plate number. Relevant, satisfying relationship , The exponential decay constant reflects the degree of peak tailing and is related to mass transfer resistance. The co-error function satisfies The UV curve calculated from the physical model layer is the baseline value for UV prediction.

[0031] The UV curve generated by the EMG function in the physical model layer serves as the baseline value for UV prediction. Simultaneously, for the baseline value of pressure prediction in the peak segment of the chromatography process, the physical model layer uses Darcy's Law and its improved Kozeny-Carman equations from fluid mechanics for modeling and calculation. For the mobile phase in the chromatography column, the pressure drop at the column tip... The general physical model is as follows: In the above formula, This is the pressure drop at the column end, which is also the baseline value for pressure prediction. The dynamic viscosity of the mobile phase. The height of the post bed, The apparent velocity of the mobile phase within the column. The permeability of the chromatography column bed.

[0032] In practical implementation, to more accurately describe the influence of stationary phase particles and pore structure on pressure drop, the Köseni-Kaman equation can be used for detailed calculations. The calculation formula is as follows: In the above formula, The column bed porosity, The average particle size of the filler is This refers to the height of the column bed.

[0033] S2. Obtain the process parameter features collected by the sensor in real time, and use them as input features along with the UV prediction baseline value and the pressure prediction baseline value. Input these features into a pre-built machine learning correction model that is decoupled from the physical model layer, and output the corresponding UV correction residual and pressure correction residual.

[0034] In this embodiment, the machine learning correction model adopts the XGBoost gradient boosting decision tree regression model. This model consists of 300 decision trees, each with a maximum depth of 6 layers (i.e., a maximum of 64 leaf nodes). Preferably, the model hyperparameters are configured as follows: number of decision trees n_estimators = 300, maximum depth max_depth = 6, learning rate 0.08, sample sampling rate subsample = 0.8, and feature sampling rate colsample = 0.8. On the target hardware (e.g., an Intel Core i5-8500@3.0GHz single-core CPU), after warm-up, the XGBoost predict() method is called 1000 times in a loop, and the average time is taken, resulting in approximately 1.4ms per inference iteration, which can run without a GPU. The complete end-to-end inference time (including approximately 0.3ms for feature extraction, approximately 0.5ms for EMG calculation of the physical model, and data format conversion, etc.) is approximately 2.2ms, which is still far lower than the control cycle of a typical tomography process (100-1000ms), meeting the requirements of industrial real-time control.

[0035] In this embodiment, the input features of the machine learning correction model include at least 14 dimensions: UV detection value, conductivity, pH value, inlet pressure, outlet pressure, flow rate, number of column cycles, step time, column volume multiple, UV trend, conductivity trend, pressure trend, UV prediction baseline value, and pressure prediction baseline value. The trend is the linear regression slope of the corresponding sensor data within the current time window, which can be the data from the most recent 5 sampling points.

[0036] Furthermore, the training of this machine learning correction model specifically includes the following steps: Calling the unified interface of the physical model layer in step S1, using the physical model to generate multiple batches of virtual tomography running data for each tomography type, and collecting and generating multiple rows of raw data at a preset sampling interval. A "batch" refers to a complete tomography running cycle (including steps such as equilibration, loading, cleaning, elution, regeneration, and storage). Specifically, 100 batches are generated for each tomography type, and 346K rows of raw data are generated at a 0.5-second sampling interval. In this embodiment, the virtual tomography running data is constructed as follows: First, the unified interface of the physical model layer is called, and random sampling is performed according to the process parameter range of each tomography type to generate multiple batches of UV prediction baseline values ​​and pressure prediction baseline values. Then, simulated disturbance terms are superimposed on this data to simulate real system deviations, constructing measured signals. This construction method ensures that the residual correction model learns the systematic deviation pattern between the physical model and the real system, rather than a zero-information label. Subsequently, features were extracted and aggregated from multiple batches of virtual tomography data within preset time windows. Redundant time windows with UV values ​​below a preset threshold were filtered out to obtain training samples. Specifically, aggregation was performed in 10-second time windows (with statistics such as mean and trend slope within the window), and baseline stable segments with UV < 20mAU were removed, resulting in approximately 34K time windows. The obtained samples were divided into a training set (80%) and a test set (20%). The training set was used for model training, and the test set was used for model validation and early stopping. The difference between the actual UV signal and the UV prediction baseline value output by the physical model was used as the UV training label, and the difference between the actual pressure signal and the pressure prediction baseline value output by the physical model was used as the pressure training label. The machine learning correction model was trained using the training samples.

[0037] S3. After completing model training, inference is performed on the real-time collected data. The UV prediction baseline value and the UV correction residual are superimposed to obtain the final UV prediction peak segment, and the pressure prediction baseline value and the pressure correction residual are superimposed to obtain the final pressure prediction peak segment. The results are then output. The specific superposition calculation formula is as follows: In the above formula, Let be the final UV prediction value at time t. The UV prediction baseline value output by the physical model layer at time t. The UV correction residual is the output of the machine learning correction model at time t. Let be the final predicted pressure value at time t. The pressure prediction baseline value output by the physical model layer at time t. The stress correction residual is the output of the machine learning correction model at time t.

[0038] In practical applications, the physical model layer and the machine learning correction model are deployed decoupled. The physical model layer runs independently and outputs UV prediction baseline values ​​and pressure prediction baseline values. The machine learning correction model is trained independently. During inference, the UV prediction baseline values ​​and pressure prediction baseline values ​​are used as input features of the machine learning correction model and concatenated with the process parameter features collected by real-time sensors to form a complete feature vector, which is then input into the machine learning correction model.

[0039] Based on the method of this embodiment, the peak segment F1 values ​​of pure physical models and hybrid models (physical + ML correction) for five types of chromatography were experimentally measured. The F1 values ​​of the pure physical models ranged from 0.582 to 0.823, while the F1 values ​​improved to 0.708 to 0.977 after using the hybrid model of this embodiment, with an average improvement of approximately +0.145. Among them, reverse phase chromatography (RPC) showed the highest dependence on ML correction, with the pure physical model having an F1 value of 0.582 and the hybrid model having an F1 value of 0.708, an improvement of +0.126; while the affinity chromatography hybrid model had an F1 value as high as 0.977. This data verifies the effectiveness of this embodiment in significantly improving the peak segment prediction accuracy.

[0040] In this embodiment, by decoupling the physical model layer from the machine learning calibration model, when the physical parameters of the chromatography process (such as column volume and flow rate) are locally adjusted, only the process input parameters of the physical model layer need to be updated, without re-calibrating the machine learning model. Conversely, when the chromatography system experiences unstructured cumulative deviations in the physical baseline due to media aging, only offline retraining of the machine learning calibration model is required, without altering the intrinsic computational logic of the first-principles equations. The two operate independently, reducing model maintenance costs in actual industrial deployments.

[0041] Example 2 This embodiment provides a digital twin system based on multi-type tomography unified physical modeling and machine learning. The system mainly includes three modules: physical model module 1, machine learning correction module 2, and superposition output module 3. Each module will be described in detail below: The physical model module 1 is used to integrate sub-models of at least five chromatographic separation mechanisms, calculate chromatographic peak parameters based on input chromatographic process parameters, and output UV prediction baseline values ​​and pressure prediction baseline values ​​for the chromatographic process.

[0042] Machine learning correction module 2 is decoupled from physical model module 1. It is used to receive process parameter features acquired in real time by the sensor and the predicted baseline value output by physical model module 1, and output UV correction residual and pressure correction residual. This module is implemented based on the XGBoost algorithm. Its input features cover 14-dimensional features, and the difference between the true value and the physical baseline value is used as the label during training.

[0043] The superposition output module 3 is used to superimpose the UV prediction baseline value and the pressure prediction baseline value with the corresponding correction residuals, and output the final UV prediction peak segment and the final pressure prediction peak segment.

[0044] In addition, this digital twin system also includes: a peak evaluation module 4, which is used to compare and evaluate the final UV prediction peaks output by the overlay output module 3 with the actual collected UV peaks. Based on the set peak detection threshold, the predicted UV data and the actual UV data are binarized and segmented. A confusion matrix is ​​calculated, which includes at least correctly detected TP, false positive FP, false negative FN, and correctly excluded TN. Precision and recall are calculated based on the confusion matrix, and then the F1 score is calculated. The F1 score is used as the main evaluation index for evaluating the effect of tomographic peak detection, and the peak mean absolute error and peak determination coefficient are used as auxiliary evaluation indexes. The peak determination coefficient is the determination coefficient calculated for samples with UV values ​​greater than the peak detection threshold.

[0045] In this embodiment, the specific implementation steps of peak segment evaluation are as follows: Step A: Signal preprocessing and baseline correction. First, a sliding window with a width of 5 sampling points is used to perform a moving average filter on the raw UV signal to remove high-frequency noise. Then, the median of the UV signal at the beginning of the chromatography run (before sample injection) is taken as the baseline value, and this baseline value is subtracted from the entire UV signal.

[0046] Step B: Binarization Segmentation and Peak Segment Merging. Set the peak detection threshold to 50mAU and perform point-by-point binarization of the entire UV signal: if the final predicted UV value at time t is greater than the peak detection threshold, then time t is marked as having a peak; otherwise, it is marked as having no peak. When consecutive segments are marked as having a peak, they are merged into one peak segment interval; if the peakless interval between two intervals is less than 3 sampling points, they can be merged into one peak segment interval.

[0047] Step C: Peak Segment Feature Extraction. For each peak segment interval, extract the peak position. (Ratio of the front half width to the back half width of the peak), peak width (Time / volume span of peak intervals), peak symmetry (Ratio of the first half width to the second half width of the peak), peak height H (maximum UV value), and peak area (Integral of UV over time within the peak segment).

[0048] Step D: Calculate the confusion matrix and F1 score. Statistically, compare the binarized results of predicted UV and true UV. , , and To construct the confusion matrix and calculate the F1 score, the formula is as follows: in, The number of samples that are correctly detected is the number of samples that are predicted to have a peak and actually do have a peak. The number of false alarms is the number of samples that were predicted to have a peak but actually had no peak. This represents the number of missed detections, i.e., the number of samples that were predicted to have "no peak" but actually had "peaks". To correctly exclude the number of samples that are predicted to be "peakless" and are actually "peakless".

[0049] Taking the actual evaluation of affinity chromatography peak segments by this system as an example: Total sampling points N=12,000. Statistically, TP was correctly detected 3,850 times, false positives FP=95 times, false negatives FN=88 times, and correctly excluded TN=7,967 times. According to the above formula, Precision=3850 / (3850+95)=0.976, Recall=3850 / (3850+88)=0.978, and finally F1=2×0.976×0.978 / (0.976+0.978)=0.977. This example shows that the F1 index can accurately reflect the actual effect of chromatographic peak segment detection, overcoming the defect of global R² being affected by baseline interference.

[0050] Ultimately, the F1 score is used as the primary evaluation metric, with the mean absolute error of peak segments (MAE_peak) and the coefficient of determination of peak segments (R²_peak) used as auxiliary metrics. R²_peak is calculated only for valid peak segment regions where the UV value is greater than the detection threshold (e.g., >50mAU), avoiding the global R² being lowered by random bias in the flat baseline segment, thus accurately reflecting the local fitting accuracy of the peak segment. In actual operation, if the F1 score or auxiliary evaluation metric obtained by peak segment evaluation module 4 is lower than the preset threshold, the system does not directly change the parameters of the current model; instead, it marks and temporarily stores the real collected data of the current batch, triggers the background offline training mechanism, retrains the machine learning correction model, and updates the model files deployed in the inference phase after training is completed.

[0051] Example 3 This embodiment provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of a digital twin method.

[0052] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0053] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. Since the above embodiments are substantially similar to the method embodiments, their descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0054] The above embodiments provide a detailed description of this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A digital twin method based on multi-type tomography unified physical modeling and machine learning, characterized in that, include: S1. Obtain the chromatography process parameters and input them into the pre-constructed multi-type unified physical model layer of chromatography. The physical model layer contains sub-models of various chromatography separation mechanisms. Each sub-model calculates the chromatography peak parameters based on the first-principles equations and outputs the UV prediction baseline value and pressure prediction baseline value of the peak segment of the chromatography process. S2. Obtain the process parameter features collected by the sensor in real time, and use them as input features along with the UV prediction baseline value and the pressure prediction baseline value. Input them into the pre-built machine learning correction model that is decoupled from the physical model layer, and output the corresponding UV correction residual and pressure correction residual. S3. Superimpose the UV prediction baseline value and the UV correction residual to obtain the final UV prediction peak segment, and superimpose the pressure prediction baseline value and the pressure correction residual to obtain the final pressure prediction peak segment, and output them.

2. The digital twin method according to claim 1, characterized in that, In step S1, the physical model layer includes a unified interface configured to receive process parameters including at least column volume, flow rate, column bed height and particle size, and uniformly output chromatography peak parameters including at least peak position, peak width, peak symmetry, theoretical plate number and resolution.

3. The digital twin method according to claim 1, characterized in that, The physical model layer operates independently of the machine learning correction model and outputs UV prediction baseline values ​​and pressure prediction baseline values. The machine learning correction model is trained independently. During inference, the UV prediction baseline values ​​and pressure prediction baseline values ​​are used as input features of the machine learning correction model and concatenated with the process parameter features collected by real-time sensors to form a complete feature vector, which is then input into the machine learning correction model.

4. The digital twin method according to claim 1, characterized in that, In step S2, the input features of the machine learning correction model include at least UV detection value, conductivity, pH value, inlet pressure, outlet pressure, flow rate, number of column cycles, step time, column volume multiple, UV trend, conductivity trend, pressure trend, UV prediction baseline value, and pressure prediction baseline value. The trend is the linear regression slope of the corresponding sensor data within the current time window.

5. The digital twin method according to claim 1, characterized in that, In step S1, the sub-models of the at least five chromatographic separation mechanisms are the affinity chromatography sub-model, the ion exchange chromatography sub-model, the size exclusion chromatography sub-model, the hydrophobic chromatography model, and the reversed-phase chromatography model. The affinity chromatography model is based on the Langmuir adsorption isotherm and a pH-dependent dissociation constant model. The ion exchange chromatograph model is based on the SMA model. The volumetric exclusion tomography model is modeled based on the Stokes radius and pore size logarithmic correlation model; The hydrophobic chromatography sub-model is modeled based on a salt-induced binding and decreasing salt concentration elution model. The inverse tomography model is modeled based on the LSS model.

6. The digital twin method according to claim 1, characterized in that, In step S2, the training steps of the machine learning correction model include: calling the unified interface of the physical model layer in step S1, using the physical model to generate multiple batches of virtual tomography running data for each tomography type, and collecting and generating multiple rows of raw data according to a preset sampling interval; performing feature extraction and aggregation on the multiple batches of virtual tomography running data according to a preset time window, filtering out redundant time windows with UV values ​​lower than a preset threshold, and obtaining training samples; using the difference between the real UV signal and the UV prediction benchmark value output by the physical model as the UV training label, and the difference between the real pressure signal and the pressure prediction benchmark value output by the physical model as the pressure training label, and training the machine learning correction model using the training samples.

7. The digital twin method according to claim 1, characterized in that, The method also includes an offline training step, specifically including: comparing and evaluating the final UV prediction peak segment with the actual collected UV peak segment; when the evaluation index is lower than a preset threshold, marking and temporarily storing the actual collected data of the current batch, triggering offline retraining of the machine learning correction model, and updating the model file deployed in the inference stage after training is completed.

8. A digital twin system based on multi-type tomographic unified physical modeling and machine learning, used to implement the digital twin method according to any one of claims 1-6, characterized in that, include: The physical model module (1) is used to integrate at least five sub-models of chromatographic separation mechanisms, calculate chromatographic peak parameters based on input chromatographic process parameters, and output UV prediction baseline values ​​and pressure prediction baseline values ​​for the chromatographic process. The machine learning correction module (2) is decoupled from the physical model module (1) and is used to receive the process parameter features collected by the sensor in real time and the prediction benchmark value output by the physical model module (1), and output the UV correction residual and the pressure correction residual. The superposition output module (3) is used to superimpose the UV prediction baseline value and the pressure prediction baseline value with the corresponding correction residuals respectively, and output the final UV prediction peak segment and the final pressure prediction peak segment.

9. The digital twin system according to claim 8, characterized in that, Also includes: The peak segment evaluation module (4) is used to compare and evaluate the final UV prediction peak segment output by the superposition output module (3) with the actual collected UV peak segment. Based on the set peak detection threshold, the predicted UV data and the actual UV data are binarized and segmented. The confusion matrix is ​​calculated, which includes at least the correctly detected TP, false positive FP, false negative FN, and correctly excluded TN. The precision and recall are calculated based on the confusion matrix, and then the F1 score is calculated. The F1 score is used as the main evaluation index for evaluating the tomographic peak segment detection effect, and the peak segment mean absolute error and peak segment determination coefficient are used as auxiliary evaluation indexes. The peak segment determination coefficient is the determination coefficient calculated for samples with UV values ​​greater than the peak detection threshold. It is also used to trigger offline retraining of the machine learning correction module and update the model file deployed in the inference stage when the evaluation index is lower than the preset threshold.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the digital twin method as described in any one of claims 1 to 7.

Images