Method and system for optimizing operating parameters of a protein purification process

Abstract

The application discloses a protein purification process operation parameter optimization method and system, relates to the field of protein purification, and comprises the following steps: processing various process data through a mixed digital twin model to output a protein purification process state prediction value, obtaining a prediction deviation between the protein purification process state prediction value and a measured value, and detecting data distribution drift of the various process data; in response to the prediction deviation or the data distribution drift meeting a preset calibration condition, performing incremental updating on the mixed digital twin model, and performing physical constraint verification on the updated mixed digital twin model; taking the mixed digital twin model passing the physical constraint verification as a proxy model, taking multiple key quality indexes as optimization objectives, and searching for optimal operation parameters meeting a preset process constraint condition in a parameter space of the protein purification process. In the operation parameter optimization process of the protein purification process, physical consistency and data fitting capability are considered, and the application has online self-adaptive calibration capability.

Description

Technical Field

[0001] This application relates to the field of protein purification technology, and in particular to a method and system for optimizing the operating parameters of a protein purification process. Background Technology

[0002] In the development and production of downstream purification processes in biopharmaceuticals, digital twin technology is widely used to simulate and optimize processes. Existing technologies typically construct digital twins using either pure mechanistic models or pure data-driven models. However, pure mechanistic models rely on numerous prior parameters such as mass transfer coefficients and diffusion coefficients, which are difficult to obtain accurately under complex real-world conditions, leading to significant model prediction bias. While pure data-driven models have strong fitting capabilities, they lack the constraints of physical laws, making them prone to unreasonable predictions that violate mass conservation laws. Furthermore, they are highly susceptible to model drift failure under changing operating conditions such as raw material fluctuations or equipment aging. In addition, existing model update mechanisms are mostly offline retraining, which cannot meet the real-time and robustness requirements of continuous biomanufacturing, resulting in long process development cycles and insufficient online control accuracy.

[0003] Therefore, there is an urgent need for a method to optimize the operating parameters of protein purification processes that can balance physical consistency and data fitting ability, and has online adaptive calibration capability. Summary of the Invention

[0004] In view of the above problems, this application provides a method and system for optimizing the operating parameters of a protein purification process, which balances physical consistency and data fitting ability during the optimization of protein purification process operating parameters, and has online adaptive calibration capability. The specific solution is as follows:

[0005] The first aspect of this application provides a method for optimizing the operating parameters of a protein purification process, including:

[0006] Acquire various process data during the protein purification process, including monitoring data collected by sensors and process operation parameter data;

[0007] The various process data are processed by a hybrid digital twin model to output predicted values ​​of the protein purification process status. The hybrid digital twin model includes a mechanism encoding module and a data compensation module. The mechanism encoding module generates basic prediction results based on the physical mechanism model of protein purification, and the data compensation module performs residual compensation on the basic prediction results.

[0008] During the protein purification process, the predicted deviation between the predicted value of the protein purification process status and the measured value of the protein purification process status is obtained, and the data distribution drift of the various process data is detected.

[0009] In response to the prediction deviation or the data distribution drift meeting the preset calibration conditions, the hybrid digital twin model is incrementally updated, and the updated hybrid digital twin model is physically constrained and verified.

[0010] Using the hybrid digital twin model verified by physical constraints as a surrogate model, and with multiple key quality indicators as optimization targets, the optimal operating parameters that satisfy the preset process constraints are searched in the parameter space of the protein purification process.

[0011] In one possible implementation, the mechanism encoding module includes a neural network model that embeds the physical mechanism model used to make the output satisfy physical constraints into the base layer.

[0012] The data compensation module adopts a residual network structure to learn the prediction bias that the physical mechanism model cannot explain;

[0013] The predicted state of the protein purification process is the sum of the basic prediction result and the residual compensation value output by the data compensation module.

[0014] In one possible implementation, the physical mechanism model includes a mass transfer equation and an adsorption isotherm model;

[0015] The loss function of the hybrid digital twin model includes a mean squared error loss term and a physical constraint penalty term, which is used to penalize prediction results that violate the law of conservation of mass.

[0016] In one possible implementation, in response to the prediction bias or the data distribution drift satisfying a preset calibration condition, an incremental update is performed on the hybrid digital twin model, including:

[0017] When the prediction deviation exceeds a preset multiple of the historical deviation standard deviation, or when the detection statistic of the data distribution drift exceeds a preset drift threshold, the parameters of the hybrid digital twin model are updated using newly collected operational data. The weight of the newly collected operational data in the update process decreases exponentially with the time interval between the data and the current time.

[0018] In one possible implementation, the detection statistic of the data distribution drift is calculated using the Page-Hinkley test or the Kolmogorov-Smirnov test;

[0019] The physical constraint verification includes mass conservation equation verification and thermodynamic equilibrium constraint verification;

[0020] If the updated model output violates the mass conservation equation verification or the thermodynamic equilibrium constraint verification, the corresponding prediction result will be marked as unreliable and will not be used for subsequent process parameter optimization.

[0021] In one possible implementation, the search for optimal operating parameters that satisfy preset process constraints in the parameter space of the protein purification process includes:

[0022] A Gaussian process model is constructed based on the predicted input-output relationship of the hybrid digital twin model. The Gaussian process model is used as the evaluation model, and the desired improvement or desired hypervolume improvement is used as the acquisition function to perform multi-objective Pareto optimization in the parameter space of the protein purification process.

[0023] Output the set of non-dominated solutions on the Pareto front, and use the non-dominated solutions that satisfy the preset constraints as the optimal operating parameters.

[0024] In one possible implementation, the optimal operating parameters are also sent to the control system for execution.

[0025] In one possible implementation, sending the optimal operating parameters to the control system for control execution includes:

[0026] Within each control cycle, the hybrid digital twin model is used to predict the process state trajectory within a preset time domain in the future;

[0027] The optimal operating parameter sequence is solved within the preset time domain so that the predicted process state trajectory is close to the target set value;

[0028] Only the first control action in the optimal operating parameter sequence is issued to the control system for execution, and the step of solving the optimal operating parameter sequence is re-executed in the next control cycle to achieve rolling optimization.

[0029] In one possible implementation, the objective function for solving the optimal sequence of operating parameters includes a tracking error weighted norm term and a control action change penalty term, and the expression of the objective function is:

[0030]

[0031] in, For the current control cycle, To predict the number of steps, To predict the length of the time domain, In the current control cycle The predicted first The process status of the step, Set a value for the target. The weight matrix for tracking error, The weight matrix is ​​used to control the changes in actions. For the first The amount of change in the control action of the step.

[0032] A second aspect of this application provides a system for optimizing operating parameters in a protein purification process, comprising: a data acquisition module, a digital twin module, an online calibration module, and an operation optimization module.

[0033] The data acquisition module is configured to acquire various process data during the protein purification process, including monitoring data collected by sensors and process operation parameter data.

[0034] The digital twin module is configured to process the various process data through a hybrid digital twin model to output a predicted value of the protein purification process status. The hybrid digital twin model includes a mechanism encoding module and a data compensation module. The mechanism encoding module generates a basic prediction result based on the physical mechanism model of protein purification, and the data compensation module performs residual compensation on the basic prediction result.

[0035] The online calibration module is configured to, during the operation of the protein purification process, obtain the prediction deviation between the predicted value of the protein purification process status and the measured value of the protein purification process status, and detect the data distribution drift of the various process data; in response to the prediction deviation or the data distribution drift meeting the preset calibration conditions, perform incremental updates on the hybrid digital twin model, and perform physical constraint verification on the updated hybrid digital twin model;

[0036] The operation optimization module is configured to use the hybrid digital twin model, which has passed physical constraint verification, as a proxy model, and search for the optimal operation parameters that meet the preset process constraints in the parameter space of the protein purification process, with multiple key quality indicators as optimization targets.

[0037] A third aspect of this application provides a computer program product including computer-readable instructions that, when executed on an electronic device, cause the electronic device to implement the protein purification process operation parameter optimization method described in the first aspect or any implementation thereof.

[0038] A fourth aspect of this application provides an electronic device, comprising at least one processor and a memory connected to the processor, wherein:

[0039] The memory is used to store computer programs;

[0040] The processor is used to execute the computer program so that the electronic device can implement the method for optimizing the operating parameters of the protein purification process according to the first aspect or any implementation thereof.

[0041] The fifth aspect of this application provides a computer storage medium carrying one or more computer programs, which, when executed by an electronic device, enable the electronic device to implement the protein purification process operation parameter optimization method described in the first aspect or any implementation thereof.

[0042] By employing the aforementioned technical solutions, the protein purification process operation parameter optimization method and system provided in this application can resolve the contradiction between the difficulty in obtaining parameters for pure mechanistic models and the lack of physical constraints in pure data models by constructing a hybrid digital twin model that includes a mechanism encoding module and a data compensation module. This significantly improves the fitting accuracy for complex nonlinear processes while ensuring that the prediction results conform to physical laws such as mass conservation. Furthermore, by monitoring prediction deviation and data distribution drift as dual trigger conditions to perform incremental updates with time-decay weights, and supplemented by rigorous physical constraint verification, the model can adapt to time-varying operating conditions such as raw material fluctuations and equipment aging, avoiding the performance degradation of traditional static models during long-term operation. Simultaneously, this application can also perform incremental updates on the hybrid digital twin model, possessing online adaptive calibration capabilities. Attached Figure Description

[0043] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and the originals and elements are not necessarily drawn to scale.

[0044] Figure 1 This is a flowchart of a method for optimizing operating parameters in a protein purification process, as provided in an embodiment of this application.

[0045] Figure 2 This is a schematic diagram of the structure of an operating parameter optimization system for a protein purification process provided in an embodiment of this application;

[0046] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0047] The embodiments of this application are described below with reference to the accompanying drawings. The terminology used in the implementation section of this application is for explaining specific embodiments only and is not intended to limit the scope of this application.

[0048] As will be known to those skilled in the art, with the development of technology and the emergence of new scenarios, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.

[0049] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms are interchangeable where appropriate; this is merely a way of distinguishing objects with the same attributes in the embodiments of this application. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, so that a process, method, system, product, or apparatus that comprises a series of elements is not necessarily limited to those elements, but may include other elements not explicitly listed or inherent to those processes, methods, products, or apparatuses.

[0050] like Figure 1 As shown, this embodiment provides a method for optimizing the operating parameters of a protein purification process, including steps S100 to S500:

[0051] Step S100: Acquire various process data during the protein purification process, including monitoring data collected by sensors and process operation parameter data.

[0052] Optional, Figure 1 The method shown can be applied to Figure 2 The illustrated protein purification process operation parameter optimization system includes a data acquisition module 100 that communicates with the physical system 600 via an industrial communication protocol to collect dynamic information reflecting the physicochemical state of the process. The monitoring data collected by the sensors may include, but is not limited to, process analytical technology (PAT) signals such as online ultraviolet absorption spectroscopy, conductivity, pH, temperature, pre- and post-column pressure, and transmembrane flux. These signals directly characterize the real-time physical environment inside the chromatography column or membrane assembly. Process operation parameter data may include feedback from the control system, such as the current flow rate setting, elution gradient slope, loading volume, switching time threshold, and buffer ratio.

[0053] It should be understood that although this embodiment uses chromatography and membrane filtration as examples, in other embodiments, the data obtained in step S100 may also cover the corresponding sensor data and operational variables of other purification unit operations such as extraction, precipitation or virus inactivation, as long as these data can be used to reconstruct the physical state of the process site.

[0054] Specifically, the various process data can include at least one of the following: historical process data, online sensor data, offline analysis data, and mechanistic model data. Historical process data can include operating parameters (such as flow rate, temperature, pH, salt concentration gradient, sample loading volume, and switching time), feedstock properties (such as titer, cell density, and impurity profile), and process outputs (such as yield, purity, aggregate content, and host cell protein residue). Historical process data can be collected batch-by-batch through a Laboratory Information Management System (LIMS) and / or an Electronic Batch Record (EBR) system. Online sensor data can include signals such as online UV absorption spectroscopy, conductivity, pH, temperature, pre- and post-column pressure, and transmembrane flux. Offline analysis data can include quality control (QC) laboratory data such as purity detection by High Performance Liquid Chromatography (HPLC), charge variant analysis by Capillary Electrophoresis (CE-CVA), and host cell protein (HCP) residue detection by Enzyme-Linked Immunosorbent Assay (ELISA). Offline analysis data can be collected by batch or by process step. Mechanistic model data can include theoretical predictions based on mass conservation equations, mass transfer equations, and adsorption isotherm models, obtained through offline mechanistic simulation calculations.

[0055] In another embodiment of the present invention, after acquiring the above-mentioned multiple process data, data preprocessing can be performed on the multiple process data. The following is an example of a data preprocessing procedure:

[0056] First, all data are aligned on the time axis using nearest neighbor interpolation to address the issue of inconsistent sampling frequencies from different sensors. Second, outlier detection is performed on the time-aligned data using the Isolation Forest algorithm or the 3σ criterion to remove abnormal data. Then, missing data is handled using K-Nearest Neighbors (KNN) interpolation or forward imputation. Finally, the data is standardized using Z-score standardization or Min-Max normalization.

[0057] After data preprocessing is completed, this application can also perform feature engineering to construct feature vectors for various process data. Optionally, the feature engineering includes constructing at least one of time-series features, ratio features, cumulative features, and cross features.

[0058] The time-series features can include first derivative (rate of change), second derivative (acceleration), moving average, and exponentially weighted moving average (EWMA) to capture the dynamic trends of process variables. Ratio features can include UV280 / UV260 (protein / nucleic acid ratio), conductivity / pH, yield / purity, etc., to characterize the relative relationships between different physical quantities. Cumulative features can include cumulative loading volume, cumulative elution volume, and cumulative run time to reflect the cumulative effects of the process. Cross-feature features can include titer × flow rate, temperature × pH, resin type × target protein, etc., to capture the interactions between different process variables. The constructed feature vectors of various process data... It can be represented as: =[Operating parameters, raw material properties, time series characteristics, ratio characteristics, cumulative characteristics, cross characteristics].

[0059] Step S200: Process various process data through a hybrid digital twin model to output predicted values ​​of protein purification process status. The hybrid digital twin model includes a mechanism encoding module and a data compensation module. The mechanism encoding module generates basic prediction results based on the physical mechanism model of protein purification, and the data compensation module performs residual compensation on the basic prediction results.

[0060] In this embodiment, step S200 can be executed by the digital twin module 200. The digital twin module 200 can receive the feature vector generated in step S100 in real time and perform inference. Figure 2 As shown, the digital twin module 200 includes a hybrid digital twin model comprising a mechanism encoding module 210 and a data compensation module 220. The mechanism encoding module 210 incorporates physical equations describing protein adsorption, mass transfer and diffusion, and fluid dynamics. Utilizing this prior knowledge, it generates basic prediction results that conform to fundamental scientific laws, ensuring that even in data-sparse regions, the output of the physical mechanism model does not violate core laws such as the law of conservation of mass. The data compensation module 220 captures nonlinear dynamic characteristics and unmodeled perturbations not covered by the physical mechanism model. By superimposing and correcting the basic prediction results, it outputs the final predicted values ​​of the protein purification process state, such as elution peak concentration, product purity, or impurity residue. Step S200 overcomes the accuracy bottleneck caused by the difficulty in identifying parameters in pure mechanism models and avoids the shortcomings of pure black-box models, such as poor extrapolation ability and lack of physical interpretability, thus maintaining the physical consistency of the model while ensuring prediction accuracy.

[0061] Step S300: During the protein purification process, obtain the prediction deviation between the predicted value of the protein purification process status and the measured value of the protein purification process status, and detect the data distribution drift of various process data.

[0062] Optionally, step S300 can be performed by the online calibration module 300. Specifically, the online calibration module 300 performs the following two processes during continuous process operation:

[0063] On the one hand, the online calibration module 300 calculates the instantaneous deviation and its statistical distribution characteristics between the output of the hybrid digital twin model and the measured values ​​(such as online UV readings or offline test results) fed back by the physical system 600;

[0064] On the other hand, the online calibration module 300 performs flow cytometry statistical analysis on the input process data to identify the input data distribution shift caused by fluctuations in raw material properties (such as changes in upstream harvest liquid droplet), packing aging, or equipment performance degradation.

[0065] Step S300 can transform model monitoring from offline evaluation to online perception, enabling the system to promptly capture early signs that lead to model failure. Whether it is a decrease in the accuracy of the output or a change in the operating conditions of the input, these can be quantified in a timely manner as the basis for triggering calibration.

[0066] In step S400, in response to the prediction bias or data distribution drift meeting the preset calibration conditions, an incremental update is performed on the hybrid digital twin model, and physical constraint verification is performed on the updated hybrid digital twin model.

[0067] When the online calibration module 300 determines that the monitored indicators exceed the safety threshold, it can automatically perform adaptive adjustments to the hybrid digital twin model. Incremental updates utilize newly collected operational data to locally fine-tune the model parameters, rather than retraining the entire system. This not only significantly reduces computational overhead and ensures the real-time nature of online updates, but also allows the model to quickly adapt to the latest operating conditions. After the parameter updates are completed, the system performs physical constraint verification on the output of the new model, such as checking whether the material balance is closed and whether the thermodynamic adsorption capacity exceeds the limit. If the updated model produces predictions that violate physical laws, the update will be rolled back or marked as invalid, preventing the model from deviating from physical reality due to noisy data or overfitting, and ensuring that the model used for decision-making is always within credible physical boundaries.

[0068] Step S500: Using the hybrid digital twin model verified by physical constraints as a surrogate model, and with multiple key quality indicators as optimization targets, the optimal operating parameters that satisfy the preset process constraints are searched in the parameter space of the protein purification process.

[0069] Optionally, step S500 can be executed by the operation optimization module 400. The operation optimization module 400 can utilize a validated, high-confidence hybrid digital twin model to replace physical experiments, performing efficient optimization within a multi-dimensional parameter space. Optimization objectives typically include multiple interdependent metrics such as maximizing product yield, maximizing purity, and minimizing buffer consumption, while strictly adhering to hard process constraints such as upper pressure limits and lower residence time limits. Because the surrogate model combines the generalization ability of physical mechanisms with the fitting accuracy of data-driven approaches, the optimization algorithm can converge to the Pareto optimal solution set in a relatively small number of iterations. The finally determined optimal operating parameters can be sent to the physical system 600 for execution via the closed-loop control interface 500. It should be understood that the optimization in this embodiment is not only applicable to parameter screening in the new process development stage but also to real-time formulation adjustments to address batch-to-batch differences in the commercial production stage.

[0070] In one implementation, the mechanism encoding module includes a neural network model in which a physical mechanism model for satisfying physical constraints in the output is embedded in the base layer. Specifically, this embedding is not a simple concatenation of modules, but rather the transformation of differential or algebraic equations describing the protein purification process into differentiable operators or custom layers in the neural network. For example, when constructing the mechanism encoding module 210, the Langmuir adsorption isotherm equation or the mass transfer equation in a general rate model can be directly written as the forward propagation layer of the neural network, so that the network's base output naturally follows the fundamental laws of adsorption equilibrium and mass transfer. It should be understood that although the Langmuir model is used as an example here, in other implementations, depending on the interaction characteristics between the target protein and the packing material, the physical mechanism model can also be a Freundlich adsorption isotherm, a Steric Mass Action (SMA) adsorption model, or a set of partial differential equations for column chromatography containing axial diffusion terms, as long as the model can characterize the core physicochemical mechanism of the process. By hard-coding prior physical knowledge into the network structure, the model output will not deviate from basic scientific common sense, even in regions where training data is sparse, thus solving the problem of poor extrapolation ability of purely data-driven models.

[0071] As one implementation method, the data compensation module can employ a residual network structure to learn prediction biases that the physical mechanism model cannot explain. In actual operation, idealized physical models often ignore complex factors such as packing aging, non-specific adsorption, and uneven fluid distribution, making it difficult to achieve industrial-grade accuracy using only the mechanism encoding module. The data compensation module 220 is constructed using multi-layered residual blocks, whose inputs include not only the original process characteristics but also intermediate state variables from the mechanism encoding module. The data compensation module 220 can fit the residual portion after subtracting the mechanism prediction value from the measured value, rather than relearning the entire process mapping relationship. This approach significantly reduces the learning difficulty of the data-driven part, allowing it to focus its limited model capacity on capturing high-frequency nonlinear dynamics not described by the physical equations. Simultaneously, the residual connection structure effectively alleviates the gradient vanishing problem of deep networks, ensuring the stability of compensation learning under complex operating conditions.

[0072] In one implementation, the predicted value of the protein purification process status is the sum of the basic prediction result and the residual compensation value output by the data compensation module. This can be expressed by the formula: ,in This represents the predicted final protein purification process status. The mechanism encoding module is based on physical parameters The generated basic prediction results The data compensation module is based on data parameters. Output residual compensation value, This represents the feature vectors of various process data. Compared to schemes that use the mechanistic model only as a post-processing validator or an independent parallel branch, the scheme that superimposes the basic prediction results with the residual compensation values ​​output by the data compensation module ensures that the data compensation is always a correction based on the physical baseline. This eliminates situations where the prediction results show negative concentrations or purity exceeding 100%, which violate physical reality, and provides a reliable decision-making basis for subsequent optimization control.

[0073] As one implementation method, the physical mechanism model includes mass transfer equations and adsorption isotherm models. In protein purification, especially chromatography, mass transfer equations describe the mass exchange rate of solute between the mobile and stationary phases, while adsorption isotherm models characterize the distribution relationship of solute between the two phases at equilibrium. In this embodiment, these equations are discretized and embedded into a neural network computation graph, so that when the model performs forward inference, it is actually solving a simplified physical model driven by data. It should be understood that, in addition to the two types of models mentioned above, depending on the specific unit operation, the physical mechanism model may also include concentration polarization models in membrane filtration processes, virus inactivation reaction kinetic models, or nucleation and growth models in precipitation processes, etc., and this application does not strictly limit this.

[0074] As one implementation method, the loss function of the hybrid digital twin model includes a mean squared error loss term and a physical constraint penalty term. The physical constraint penalty term is used to penalize predictions that violate the law of conservation of mass. During the model training phase, a composite loss function strategy is adopted to balance data fitting accuracy and physical consistency. Specifically, the total loss function can be expressed as... ,in The mean square error between the predicted and measured values ​​is used to ensure the model's ability to reproduce historical data. This is a physical constraint penalty term. The hyperparameters used to adjust the weights of both are physical constraint penalty terms. The calculation of these penalty terms does not rely on labeled data but is based on real-time evaluation of the model's own output. For example, for a chromatography column system, a material balance equation can be constructed as a constraint. The difference between the total outflow plus the cumulative inflow and the total inflow is calculated. If this difference is not zero, its square is added as a penalty term to the total loss. During backpropagation, the gradient generated by this penalty term forces the model parameters to update in a direction that satisfies mass conservation. Furthermore, physical constraint penalty terms can also include thermodynamic equilibrium constraints (such as adsorption not exceeding saturation capacity) and non-negativity constraints (such as concentration and pressure always being greater than or equal to zero). While learning data patterns, the model is constrained within the feasible region allowed by physical laws, maintaining the physical rationality of its output even when facing unseen abnormal conditions.

[0075] As one implementation method, protein concentration prediction follows the following mechanistic equation: ,in, For the predicted protein concentration, This represents the initial protein concentration. Let be the mass transfer rate constant. For time, The residence time is calculated from the column volume and flow rate. This equation is embedded in the first layer of the mechanism coding module, so that the model's basic output naturally follows the mass transfer laws of proteins in the chromatography column.

[0076] As one implementation method, the quality attribute prediction sub-model employs a multi-task learning framework to simultaneously predict multiple key quality indicators, including monomer purity. Aggregate content Charge Variant (CV) distribution Host cell protein (HCP) residues and DNA residue Specifically, this multi-task learning architecture can be represented as follows: ,in The general feature representation extracted from the shared layer. MTDNN is a multi-task deep neural network designed for dedicated features for each Critical Quality Attribute (CQA). Shared layers extract common features across quality attributes, while task-specific layers perform refined modeling of the dedicated prediction head for each CQA. During training, the weights of the loss function are adjusted based on the importance of each CQA to ensure the model achieves a balance across multiple prediction tasks.

[0077] One implementation method involves predicting the uncertainty interval for each predicted value using an ensemble approach or a Bayesian neural network (BNN). Specifically, a deep ensemble method is employed, independently training five models with different random initializations to predict the same input, obtaining the mean and standard deviation of the predicted values. The uncertainty interval can be represented as... ,in, The mean of the five model predictions. The standard deviation is 1.96, which corresponds to a 95% confidence level. This uncertainty information is used in two ways: first, for active learning sampling, selecting process conditions with high uncertainty for physical experiments to maximize information gain; second, for risk assessment, when the uncertainty of the prediction result exceeds a preset threshold, the system issues an alarm to prompt process engineers to make careful decisions.

[0078] As one implementation method, model training employs a transfer learning strategy. Specifically, it first pre-trains on public protein purification datasets (such as chromatography process data from publicly available literature, standard protein purification curves provided by packing material manufacturers, etc.) to enable the model to learn the general rules and basic feature representations of the protein purification process; then, it fine-tunes on private data specific to the process. The advantage of transfer learning is that even with limited private data, the model can quickly converge to a better performance level by leveraging the general knowledge learned in the pre-training stage, significantly reducing its dependence on large amounts of labeled data. As another implementation method, the training of the digital twin model employs an active learning strategy, including three stages: uncertainty sampling, diversity constraints, and human-machine collaboration. Specifically, in each iteration, the process first selects the process conditions with the highest uncertainty in predicting the digital twin model through uncertainty sampling for physical experiments to maximize information gain and achieve the greatest model improvement with the fewest experiments. Secondly, diversity constraints are added to the uncertainty sampling to ensure that the selected experimental conditions cover the sparse region of the parameter space, avoiding over-concentration of sampling points in a local area that could lead to insufficient predictive ability in other regions. Finally, the machine-recommended candidate experimental conditions are submitted to process experts for review. Experts can then filter or supplement these conditions based on their experience and knowledge, achieving human-machine collaboration. This proactive learning strategy makes the model training process more efficient, enabling the construction of a high-precision digital twin model within a limited experimental budget.

[0079] The following section further elaborates on the specific implementation strategy of online incremental calibration, which can solve the model drift problem caused by raw material fluctuations, packing material aging, or environmental changes during long-term operation of bioprocesses.

[0080] As one implementation method, in response to the prediction deviation or data distribution drift meeting the preset calibration conditions, an incremental update is performed on the hybrid digital twin model. Specifically, when the prediction deviation exceeds a preset multiple of the historical deviation standard deviation, or the detection statistic of the data distribution drift exceeds a preset drift threshold, the parameters of the hybrid digital twin model are updated using newly acquired operational data. The weight of the newly acquired operational data in the update process decreases exponentially with the time interval from the current time.

[0081] Specifically, this application can initiate the calibration process in two situations. The first situation is when the predicted deviation exceeds a preset multiple of the historical deviation standard deviation. When the absolute value of the deviation between the predicted and measured values ​​of the protein purification process exceeds a preset multiple of the historical deviation standard deviation, the accuracy of the model is deemed to have decreased significantly. For example, in a preferred embodiment, this preset multiple can be set to 3 times the standard deviation (3σ), which means that when the deviation exceeds the normal random fluctuation range, it is considered abnormal. It should be understood that 3σ is only a statistically preferred threshold. In practical applications, depending on the process's tolerance for quality attributes, this multiple can also be adjusted to any value between 2σ and 5σ, or a dynamically calculated confidence interval boundary can be used. The second situation is when the detection statistic of data distribution drift exceeds a preset drift threshold. Even if the predicted deviation has not exceeded the limit, if a significant shift in the distribution of the input process data (such as sample droplet size, impurity spectrum characteristics) is detected, it indicates that the applicable domain of the original model may no longer cover the current operating conditions, thus triggering preventive calibration in advance. This application avoids the blind spots of single-index monitoring by initiating the calibration process in both cases. It can correct deviations in a timely manner when the model fails and can proactively adapt in the early stages of operating condition changes, significantly improving the robustness of the system.

[0082] As one implementation method, incremental updates do not treat all historical data the same as new data; instead, a time-forgetting mechanism is introduced. Specifically, the formula for calculating sample weights can be expressed as:

[0083]

[0084] in, Indicates the first The weight of each historical sample during the update. For the current moment, This is the time when the sample was collected. This represents the preset decay time constant. Biopurification is a typical time-varying system; the decay of resin adsorption capacity, the decrease in membrane flux, and the evolution of upstream cell culture states all exhibit time-dependent characteristics. Data closer to the current moment better reflects the system's current true physical state; while older data, although still containing some fundamental patterns, may no longer represent specific operating conditions (such as the high adsorption performance of new resin). By using exponential decay weighting, the model is forced to focus more on the latest features contained in recent data during updates, effectively overcoming the model lag problem caused by old data in traditional full retraining methods.

[0085] As one implementation method, the detection statistic for data distribution drift is calculated using the Page-Hinkley test or the Kolmogorov-Smirnov test. To accurately quantify the degree of drift in the input data, this application employs targeted statistical testing algorithms. The Page-Hinkley test is a cumulative sum test method, particularly suitable for detecting gradual drift in the mean of a signal, such as the slow decrease in column efficiency with increasing usage. The Kolmogorov-Smirnov (KS) test, by comparing the maximum difference between two empirical distribution functions, can effectively identify non-parametric changes in the overall data distribution pattern, such as structural changes in the host cell protein impurity profile in upstream harvest fluid. In practice, these two tests can be run in parallel; if either statistic exceeds a preset threshold, drift is considered to have occurred. It should be understood that, in addition to the above two methods, depending on the data characteristics, alternative solutions such as the CUSUM test, the ADWIN algorithm, or distribution difference measures based on KL divergence can also be used, as long as the function of identifying non-stationary changes from a steady-state data stream can be achieved.

[0086] As one implementation method, physical constraint verification includes mass conservation equation verification and thermodynamic equilibrium constraint verification. If the updated model output violates the mass conservation equation verification or thermodynamic equilibrium constraint verification, the corresponding prediction result is marked as unreliable and will not be used for subsequent process parameter optimization. Since incremental updates rely only on limited new data, they are highly susceptible to interference from sensor noise or local anomalies, causing model parameters to adjust in the wrong direction and resulting in predictions that violate physical principles (such as output exceeding input, adsorption exceeding theoretical saturation capacity, etc.). Therefore, after each model parameter update, the system does not directly adopt the new model but immediately performs strict physical constraint verification on its output. Mass conservation equation verification requires that, within any time window, the total amount of protein entering the purification unit must equal the sum of the total outflow and the cumulative amount within the unit; thermodynamic equilibrium constraint verification requires that the predicted solid-phase adsorption concentration must not exceed the maximum equilibrium capacity determined by the Langmuir or SMA isotherm under those conditions. Once a violation of any constraint is detected, the system marks the prediction result generated by that update as unreliable and removes it or imposes a very high penalty on it in subsequent Bayesian optimization or MPC control solutions. Generating operation instructions based on such false predictions is strictly prohibited. In this way, this embodiment can avoid process runaway due to erroneous model updates, ensuring that even during adaptive learning, the system's decisions are always confined within the safe and feasible region allowed by physical laws.

[0087] As one implementation method, the parameter update formula for incremental learning is: ,in These are the model parameters before the update. For the updated model parameters, For learning rate, The gradient of the loss function with respect to the parameters. The parameters (weights and biases) represent the model. For newly collected operational data, This is a weighted subset of historical data (weights decay exponentially over time). This formula shows that the direction of model parameter updates is jointly determined by new data and weighted historical data, ensuring that the model can quickly adapt to the latest operating conditions while avoiding the catastrophic forgetting problem caused by completely forgetting historical knowledge.

[0088] As one implementation method, the specific triggering conditions for triggered recalibration include one of the following four situations:

[0089] (1) Prediction bias trigger: when Triggered when any key variable is sampled for more than 5 consecutive times, where, This represents the prediction deviation between the predicted value and the measured value of the protein purification process status of the i-th key variable at sampling time t. The standard deviation of the historical bias;

[0090] (2) Statistical test trigger: triggered when the Page-Hinkley statistic exceeds a preset threshold, used to detect asymptotic drift;

[0091] (3) Triggering the change of raw material properties: Triggered when the properties of upstream raw materials change significantly, specifically including a change in titer exceeding 15% or a change in impurity profile exceeding 20%;

[0092] (4) Manual triggering: The process engineer manually triggers the recalibration when he or she judges that recalibration is required based on his or her experience.

[0093] The four triggering conditions mentioned above cover all scenarios from sudden anomalies to gradual drift, and from automatic detection to manual intervention, ensuring that the system can promptly initiate the calibration process in any situation that may lead to model failure.

[0094] As one implementation method, drift root cause analysis identifies key input variables that lead to prediction bias through Shapley value analysis. Specifically, when data distribution drift or prediction bias exceeding limits is detected, the system calculates the contribution of each input variable to the Shapley value of the prediction bias, thereby quantifying the degree of influence of each variable on model failure.

[0095] The following section further elaborates on the process parameter optimization strategy based on the hybrid digital twin model.

[0096] As one implementation method, searching for optimal operating parameters that satisfy preset process constraints in the parameter space of the protein purification process includes: constructing a Gaussian process model based on the predicted input-output relationship of the hybrid digital twin model, and performing multi-objective Pareto optimization in the parameter space of the protein purification process using the Gaussian process model as the evaluation model and the desired improvement or desired hypervolume improvement as the acquisition function; outputting the non-dominated solution set on the Pareto front, and taking the non-dominated solution that satisfies the preset constraints as the optimal operating parameters.

[0097] Specifically, the operation optimization module 400 does not directly perform trial and error on the physical system 600, but instead uses a hybrid digital twin model verified by an online calibration mechanism as a high-confidence virtual evaluator. Based on this, a Gaussian process (GP) is constructed as a probabilistic surrogate model to perform secondary modeling of the input-output mapping relationship of the hybrid digital twin model. The Gaussian process can not only predict the mean process performance under any untested parameter combination, but also provide the uncertainty variance of this prediction. Regarding the selection of the kernel function, although this embodiment preferably uses... The 5 / 2 kernel function balances the smoothness and differentiability of the protein purification response surface. However, in other implementations, depending on the process characteristics, a squared exponential kernel, a Rational Quadratic kernel, or a composite kernel function containing periodic components can be selected, as long as the kernel function can reasonably characterize the correlation structure between process parameters and quality attributes. Through a Gaussian process surrogate model, the system transforms the originally computationally intensive digital twin simulation into millisecond-level probabilistic inference, significantly improving search efficiency.

[0098] In terms of search strategy, the system utilizes acquisition functions to guide the selection of the next set of experimental parameters. When the optimization objective is a single key quality indicator, Expected Improvement (EI) is used as the acquisition function to quantify the potential improvement of candidate points relative to the currently observed best value. However, when facing multiple conflicting optimization objectives such as yield, purity, and cost, Expected Hypervolume Improvement (EHVI) is used. It should be understood that, in addition to EI and EHVI, confidence upper bound (UCB), probabilistic improvement (PI), or entropy search-based acquisition functions can also be used depending on specific needs; this application does not impose strict limitations on these.

[0099] The output of multi-objective Pareto optimization is not a single optimal solution, but rather a set of non-dominated solutions located on the Pareto front. In this set, any solution cannot further improve a specific objective without worsening at least one other objective. This objectively reflects the inherent physical constraints between various quality attributes in protein purification processes. For example, increasing the elution flow rate may shorten the production cycle but reduce resolution, thus creating a trade-off between yield and purity. The system selects solutions from this non-dominated set that simultaneously satisfy all preset process constraints (such as lower purity limits, upper pressure limits, and host cell protein residue thresholds) as candidate optimal operating parameters. If multiple feasible solutions exist, a final decision can be made by combining process expert preferences or dynamic production needs (such as prioritizing yield or quality). Compared to traditional grid search or genetic algorithms, this Bayesian optimization framework based on Gaussian processes and acquisition functions can reduce the number of required physical experiments or high-precision simulations by more than 50% while ensuring the discovery of a globally approximate optimal solution, significantly shortening the process development cycle and reducing R&D costs.

[0100] As one implementation method, the Gaussian process employs... The 5 / 2 kernel function has the following expression: ,in , For length scale hyperparameters, For signal variance, the length-scale hyperparameter is automatically optimized by maximizing marginal likelihood. and These are the two input sample points (input feature vectors) of the Gaussian process. To standardize the distance, for 5 / 2 kernel function. The 5 / 2 kernel function is preferred in protein purification process optimization because it achieves a good balance between the smoothness and differentiability of the response surface. It avoids the over-smoothing and loss of local features seen in square exponential kernels, and also avoids the excessive smoothing and loss of local features seen in square exponential kernels. A 1 / 2 kernel is too coarse and leads to unstable optimization. For multi-objective optimization scenarios, a multi-output Gaussian process (MOGP) is used to model the correlation between objectives, and the covariance structure between objectives is used to improve prediction accuracy.

[0101] As one implementation method, the specific expression of the acquisition function is as follows: For single-objective optimization, Constrained Expected Improvement (CEI) is used, and its expression is: ,in The representative constraint is expected to improve. For the desired improvement value, Let be the probability that a candidate point satisfies the constraints. Representing candidate solutions (i.e., operating parameters), CEI achieves efficient search within the feasible region by multiplying the constraint satisfaction probability with the expected improvement. For multi-objective optimization, Expected Hypervolume Improvement (EHVI) is employed, its expression being: ,in For the current Pareto frontier of hypervolume, To form the supervolume after adding candidate points, Representing candidate solutions (i.e. operating parameters), EHVI can adaptively sample in the trade-off region of multiple objectives by quantifying the expected contribution of candidate points to the Pareto front hypervolume.

[0102] As one implementation method, the candidate solution evaluation process includes the following five steps: (1) The acquisition function recommends the next set of candidate parameters. (2) Rapid prediction using digital twin models and The prediction time is less than 100ms, of which, This indicates that the digital twin model is based on candidate parameters. The estimated key quality indicators (such as predicted purity of 95%). This indicates that the digital twin model is based on candidate parameters. The estimated process constraints (e.g., predicted column pressure of 1.2 MPa). (3) If the prediction results meet the constraints and the expected improvement is significant, then submit a physical experiment for verification; (4) After the physical experiment is completed, the actual results will be submitted. Add to the training set and update the Gaussian process model, where... This represents the actual operating parameters of the physical equipment (chromatographic column). Indicates the parameters to run on the physical device Subsequently, the actual key quality indicators (such as the actual measured purity of 94.5%) were measured using analytical instruments (such as HPLC). This indicates the actual process constraint index collected by the sensor during the actual operation of the physical equipment (e.g., the actual measured column pressure is 1.25 MPa); (5) Repeat steps (1) to (4) until the convergence criterion is met or the experimental budget is exhausted. This process decouples computationally intensive digital twin simulation from expensive physical experiments, and only performs physical verification when the candidate solution shows sufficient potential, thereby significantly reducing experimental costs.

[0103] As one implementation method, the convergence criterion includes one of the following three cases: (1) the expected improvement value in 10 consecutive iterations. ,in (1) The current observed optimal objective value indicates that the optimization is close to convergence; (2) The maximum experimental budget (e.g., 30 iterations) is reached, indicating that the preset experimental resources have been exhausted; (3) The hypervolume increment of the Pareto front is less than 1%, indicating that the frontier of the multi-objective optimization has become stable. When any of the above conditions are met, the system stops optimizing and outputs the current optimal solution set.

[0104] The following further elaborates on the specific implementation process of sending the optimal operating parameters to the control system for execution. After the process parameter optimization based on the hybrid digital twin model is completed in the aforementioned embodiment, the closed-loop control interface 500 is responsible for transforming the abstract optimization decision into real-time instructions that the physical system 600 can execute.

[0105] As one implementation method, the optimal operating parameters are sent to the control system for execution, including: predicting the process state trajectory in a future preset time domain using a hybrid digital twin model in each control cycle; solving the optimal operating parameter sequence in the preset time domain so that the predicted process state trajectory is close to the target setpoint; sending only the first control action in the optimal operating parameter sequence to the control system for execution, and re-executing the steps of solving the optimal operating parameter sequence in the next control cycle to achieve rolling optimization.

[0106] In each control cycle k (preferably 1 to 5 minutes, depending on the response time constant of the chromatography column or membrane filtration unit), the system uses the latest sensor readings as the initial state to drive a hybrid digital twin model validated by an online calibration mechanism. This model extrapolates the process state evolution trajectory over the next N steps (i.e., the prediction time domain, preferably 10 to 30 steps), including but not limited to key variables such as elution peak concentration, product purity, column pressure drop, and transmembrane flux. Subsequently, the optimizer searches for an optimal sequence of operating parameters (such as flow rate setpoint, gradient slope, and switching time) within this prediction time domain, ensuring that the predicted future trajectory approximates the preset target setpoint curve while satisfying various process constraints. However, the system does not distribute the entire parameter sequence; instead, it extracts only the first control action from the sequence. The data is sent to the DCS or PLC controller of the physical system 600 for execution. When entering the next control cycle k+1, the system will collect the latest measured data again, use the new information to correct the prediction of the future state, and resolve for a new optimal sequence. The control system can continuously use the latest feedback information to compensate for model mismatch and external disturbances (such as upstream titer fluctuations, packing performance degradation, etc.), thereby maintaining extremely high control accuracy and robustness in long-cycle continuous production. It should be understood that although this embodiment uses a fixed control cycle as an example, in other embodiments, the control cycle can also be adaptively adjusted according to the dynamic characteristics of the process stage. For example, the cycle can be shortened during the critical window period of elution peak elution to improve resolution, while the cycle can be appropriately extended during the balancing or regeneration stage to reduce the computational load, as long as the basic principles of real-time feedback and rolling optimization are met.

[0107] As one implementation method, the objective function for solving the optimal sequence of operating parameters includes a tracking error weighted norm term and a control action change penalty term. The expression for the objective function is:

[0108]

[0109] Where k is the current control period, i is the prediction step number, and N is the prediction time domain length. This is the process state predicted at step k+i in the current control cycle k. Let Q be the weight matrix for tracking error and R be the weight matrix for changes in control action, where Q is the target setpoint. Let be the change in control action at step k+i.

[0110] The design of this objective function reflects the dual pursuit of precision and stability in bioprocess control. The first term... The weight matrix Q quantifies the degree to which the predicted trajectory deviates from the target setpoint, allowing engineers to differentiate configurations based on the criticality of different quality attributes. For example, it assigns higher weights to product purity to ensure compliance, while granting a relatively lenient tolerance to deviations in intermediate process variables. (Second term) The Q and R matrices are specifically designed to penalize drastic changes in control actions, which is of crucial physical significance in protein purification processes. Specifically, biomolecules are extremely sensitive to fluid shear forces. Frequent and large jumps in pump speed or valve opening can not only cause pressure shocks in pipelines and shorten the lifespan of chromatographic columns or membrane modules, but can also directly lead to protein denaturation, aggregation, or loss of activity. Furthermore, the actuators themselves suffer from mechanical wear and response hysteresis; overly aggressive control commands often cannot be accurately tracked by the physical equipment, instead causing oscillations. Therefore, the weight matrix R allows the algorithm to pursue setpoint tracking accuracy while simultaneously considering the smoothness and continuity of the control input, thus establishing an intrinsic coupling between the mathematical optimization objective and the safety boundaries of the physical equipment. In practical applications, the parameter tuning of the Q and R matrices usually requires a comprehensive trade-off between sensitivity analysis of the specific process and the technical specifications of the equipment manufacturer to achieve optimal overall control performance.

[0111] As one implementation, the objective function also includes a constraint violation penalty term, used to penalize predictions that violate process constraints. Specifically, the complete objective function can be expressed as:

[0112] ;

[0113] in To constrain violations and penalties, For process constraint functions (such as lower purity limit, upper pressure limit, etc.). This is the weight matrix for constraint violations. When the predicted process state violates the constraints, A positive value is generated, which penalizes the objective function, forcing the optimizer to prioritize sequences of operational parameters that satisfy the constraints during the search process. This soft constraint handling complements the aforementioned physical constraint verification: physical constraint verification acts as a hard safety valve, directly rejecting model updates that violate physical laws, while the constraint violation penalty term guides the search direction during the optimization process. Together, they ensure the physical feasibility and process safety of control commands.

[0114] As one implementation method, all issued parameters undergo triple verification of the safety constraint layer: (1) hard constraint verification, i.e., physical limitations of the equipment, including maximum operating pressure, maximum flow rate, upper and lower temperature limits, etc.; (2) soft constraint verification, i.e. process experience limitations, including pH value not exceeding the protein stability range, salt concentration gradient not exceeding the packing tolerance range, etc.; (3) change rate constraint, i.e. parameter adjustment speed not exceeding the equipment response capability, such as flow rate change rate not exceeding 10% per second, pH change rate not exceeding 0.1 units per minute, etc.

[0115] As one implementation method, when this method is implemented in a continuous biomanufacturing mode, the optimization problem further includes dynamic constraints. Specifically, the dynamic constraints include the following three categories: (1) Upstream and downstream coupling constraints: the harvest rate of upstream perfusion culture. Processing rate of downstream purification steps satisfy ,in (1) Buffer coefficient (usually taken as 0.05-0.15); (2) Cumulative material balance constraint: within any time window, the total amount of material entering the system is equal to the sum of the total amount of material flowing out of the system and the cumulative amount in the system; (3) Steady-state deviation constraint: the fluctuation range of the operating parameters of the continuous flow process. This is to ensure batch-to-batch consistency in product quality.

[0116] like Figure 2 As shown, this embodiment provides an operating parameter optimization system for a protein purification process. This system serves as the hardware carrier and execution entity for the methods described in the preceding embodiments, aiming to achieve intelligent closed-loop control of the process through deep interaction between physical modules and physical equipment. The system mainly includes a data acquisition module 100, a digital twin module 200, an online calibration module 300, and an operation optimization module 400. Optionally, the system may also include a closed-loop control interface 500. The modules interact in real time via a high-speed data bus or network interface, collectively forming a control system tightly coupled with the physical system 600.

[0117] The data acquisition module 100 is configured to acquire various process data during the protein purification process, including monitoring data collected by sensors and process operation parameter data. Specifically, this module acts as a hardware gateway between the information system and the physical site. In terms of hardware connectivity, it connects directly to the Process Analysis Technology (PAT) sensor array in the physical system 600 via analog / digital input / output cards or an industrial Ethernet interface. This includes, but is not limited to, online UV detectors, conductivity meters, pH meters, flow meters, and pressure transmitters, acquiring raw signals reflecting the physicochemical state of the process at millisecond-level sampling frequencies. Simultaneously, the module establishes a bidirectional connection with the distributed control system (DCS) or programmable logic controller (PLC) of the physical system 600 via industrial communication protocols, reading real-time feedback of current flow rate settings, valve openings, temperature setpoints, and other operational parameters. Before the data enters subsequent processing stages, the embedded processor or edge computing unit built into the data acquisition module 100 performs timestamp alignment, outlier removal, and standardization on the multi-source heterogeneous data, ensuring that the transmitted data has a unified spatiotemporal reference and quality reliability. It should be understood that although this embodiment uses DCS / PLC as an example of the underlying control system, in other implementations, this module can also be adapted to PC-based control schemes or cloud-native control architectures, as long as reliable data mapping from physical entities to digital space can be achieved.

[0118] The digital twin module 200 is configured to process multiple process data through a hybrid digital twin model to output predicted values ​​for the protein purification process status. The hybrid digital twin model includes a mechanism encoding module 210 and a data compensation module 220. The mechanism encoding module 210 generates basic prediction results based on the physical mechanism model of protein purification, and the data compensation module 220 performs residual compensation on the basic prediction results. In this embodiment, the digital twin module 200 is typically deployed in a high-performance edge server or cloud computing cluster to ensure the real-time performance of complex model inference. The mechanism encoding module 210 and the data compensation module 220 are not independent software plugins, but rather collaborative computing units integrated within the same computing framework. The mechanism encoding module 210 internally encapsulates discretized mass transfer equations and adsorption isotherm operators, using GPU acceleration or dedicated AI chips for parallel solving; the data compensation module 220 employs a lightweight neural network structure, receiving intermediate states and original input features from the mechanism module and quickly outputting residual correction values. Data fusion is completed at the memory level, avoiding cross-process communication latency, thereby meeting the stringent requirements of online control for low response time.

[0119] The online calibration module 300 is configured to obtain the prediction deviation between the predicted and measured values ​​of the protein purification process status during the protein purification process operation, and to detect data distribution drift of various process data. In response to the prediction deviation or data distribution drift meeting preset calibration conditions, it performs incremental updates to the hybrid digital twin model and performs physical constraint verification on the updated hybrid digital twin model. As the system's adaptive adjustment center, the online calibration module 300 continuously monitors the output performance and input data flow characteristics of the digital twin module 200. When calibration conditions are triggered, the online calibration module 300 automatically calls the incremental learning algorithm to fine-tune the model parameters in the digital twin module 200 and immediately performs physical consistency verification. This process is completely automated within the system without manual intervention, ensuring that the model remains synchronized with the actual state of the physical system 600 throughout long-term operation.

[0120] The operation optimization module 400 is configured to use a hybrid digital twin model validated by physical constraints as a surrogate model, and to search for optimal operating parameters that satisfy preset process constraints within the parameter space of the protein purification process, targeting multiple key quality indicators. The operation optimization module 400 utilizes Gaussian process regression and acquisition function strategies to efficiently explore parameter combinations within the virtual space. Because the surrogate model it relies on has undergone real-time calibration and physical constraint filtering by the online calibration module 300, the recommended parameters output by the operation optimization module 400 not only possess statistical optimality but also physical feasibility and safety, significantly reducing the risk of invalid searches and unauthorized operations.

[0121] The closed-loop control interface 500 is configured to send optimal operating parameters to the control system for execution. This is the key physical interface for the implementation of the technology and compliance of the system in this application. Specifically, the closed-loop control interface 500 establishes a secure and reliable command transmission channel with the physical system 600 through a standardized industrial communication protocol. In a preferred embodiment, this interface supports the OPC UA (Open Platform Communications Unified Architecture) protocol to achieve cross-platform, semantic data exchange and secure authentication; it is also compatible with traditional industrial protocols such as Modbus TCP / RTU to adapt to the access needs of devices from different eras and manufacturers. Before sending control commands, the closed-loop control interface 500 also performs a final safety verification logic, including rate of change limits, hard limit checks, and interlock protection, to prevent equipment damage or process accidents caused by algorithm anomalies or communication interference. It should be understood that although the figure shows the closed-loop control interface 500 directly connected to the physical system 600, in actual industrial deployments, an independent Safety Instrumented System (SIS) or firewall can be set up between the two as an isolation layer, as long as it does not affect the final execution of the optimization commands and the real-time acquisition of status feedback.

[0122] Through the aforementioned modular architecture and hardware interaction design, this application's system solidifies abstract optimization algorithms into engineering techniques that can directly interact with physical devices. The data acquisition module 100 ensures the physical authenticity of the perception layer, the closed-loop control interface 500 guarantees the physical security of the execution layer, while the digital twin module 200, online calibration module 300, and operation optimization module 400 operate under strict physical constraints and real-time feedback mechanisms. Together, they constitute a complete technical solution that possesses both intelligent decision-making capabilities and complies with biomanufacturing safety standards, effectively overcoming the challenges of object suitability and engineering implementation faced by pure software algorithms in industrial applications.

[0123] This application also provides an electronic device in its embodiments. (See reference...) Figure 3 The diagram illustrates a structural schematic of an electronic device suitable for implementing the protein purification process operation parameter optimization method in the embodiments of this application. The electronic device in the embodiments of this application may include, but is not limited to, fixed terminals such as mobile phones, laptops, PDAs (personal digital assistants), PADs (tablet computers), desktop computers, etc. Figure 3 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0124] like Figure 3As shown, the electronic device may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) 601, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 602 or a program loaded from a storage device 608 into a random access memory (RAM) 603. When the electronic device is powered on, the RAM 603 also stores various programs and data required for the operation of the electronic device. The processing unit 601, ROM 602, and RAM 603 are interconnected via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0125] Typically, the following devices can be connected to I / O interface 605: input devices 606 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 607 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 608 including, for example, memory cards, hard drives, etc.; and communication devices 609. Communication device 609 allows electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 3 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown. More or fewer devices may be implemented or have alternatively.

[0126] This application also provides a computer program product, including computer-readable instructions, which, when executed on an electronic device, cause the electronic device to implement any of the protein purification process operation parameter optimization methods provided in this application.

[0127] This application also provides a computer-readable storage medium carrying one or more computer programs. When the one or more computer programs are executed by an electronic device, the electronic device can implement any of the protein purification process operation parameter optimization methods provided in this application.

[0128] It should also be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. In addition, in the device embodiment drawings provided in this application, the connection relationship between modules indicates that they have a communication connection, which can be implemented as one or more communication buses or signal lines.

[0129] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware, or it can be implemented by special-purpose hardware including application-specific integrated circuits, special-purpose CPUs, special-purpose memory, special-purpose components, etc. Generally, any function performed by a computer program can be easily implemented by corresponding hardware, and the specific hardware structure used to implement the same function can also be diverse, such as analog circuits, digital circuits, or special-purpose circuits. However, for this application, software program implementation is more often the preferred implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a readable storage medium, such as a computer floppy disk, USB flash drive, mobile hard disk, ROM, RAM, magnetic disk, or optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, training equipment, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0130] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product.

[0131] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, training device, or data center to another website, computer, training device, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a training device or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)).

Claims

1. A method for optimizing operating parameters in a protein purification process, characterized in that, include: Acquire various process data during the protein purification process, including monitoring data collected by sensors and process operation parameter data; The various process data are processed by a hybrid digital twin model to output predicted values ​​of the protein purification process status. The hybrid digital twin model includes a mechanism encoding module and a data compensation module. The mechanism encoding module generates basic prediction results based on the physical mechanism model of protein purification, and the data compensation module performs residual compensation on the basic prediction results. During the protein purification process, the predicted deviation between the predicted value of the protein purification process status and the measured value of the protein purification process status is obtained, and the data distribution drift of the various process data is detected. In response to the prediction deviation or the data distribution drift meeting the preset calibration conditions, the hybrid digital twin model is incrementally updated, and the updated hybrid digital twin model is physically constrained and verified. Using the hybrid digital twin model verified by physical constraints as a surrogate model, and with multiple key quality indicators as optimization targets, the optimal operating parameters that satisfy the preset process constraints are searched in the parameter space of the protein purification process.

2. The method according to claim 1, characterized in that, The mechanism encoding module includes a neural network model that embeds the physical mechanism model used to make the output satisfy physical constraints into the base layer. The data compensation module adopts a residual network structure to learn the prediction bias that the physical mechanism model cannot explain; The predicted state of the protein purification process is the sum of the basic prediction result and the residual compensation value output by the data compensation module.

3. The method according to claim 2, characterized in that, The physical mechanism model includes mass transfer equations and adsorption isotherm models; The loss function of the hybrid digital twin model includes a mean squared error loss term and a physical constraint penalty term, which is used to penalize prediction results that violate the law of conservation of mass.

4. The method according to claim 1, characterized in that, In response to the prediction bias or the data distribution drift meeting a preset calibration condition, an incremental update is performed on the hybrid digital twin model, including: When the prediction deviation exceeds a preset multiple of the historical deviation standard deviation, or when the detection statistic of the data distribution drift exceeds a preset drift threshold, the parameters of the hybrid digital twin model are updated using newly collected operational data. The weight of the newly collected operational data in the update process decreases exponentially with the time interval between the data and the current time.

5. The method according to claim 4, characterized in that, The detection statistics of the data distribution drift were calculated using the Page-Hinkley test or the Kolmogorov-Smirnov test. The physical constraint verification includes mass conservation equation verification and thermodynamic equilibrium constraint verification; If the updated model output violates the mass conservation equation verification or the thermodynamic equilibrium constraint verification, the corresponding prediction result will be marked as unreliable and will not be used for subsequent process parameter optimization.

6. The method according to claim 1, characterized in that, The process of searching for optimal operating parameters that satisfy preset process constraints in the parameter space of the protein purification process includes: A Gaussian process model is constructed based on the predicted input-output relationship of the hybrid digital twin model. The Gaussian process model is used as the evaluation model, and the desired improvement or desired hypervolume improvement is used as the acquisition function to perform multi-objective Pareto optimization in the parameter space of the protein purification process. Output the set of non-dominated solutions on the Pareto front, and use the non-dominated solutions that satisfy the preset constraints as the optimal operating parameters.

7. The method according to claim 1, characterized in that, Also includes: The optimal operating parameters are sent to the control system for execution.

8. The method according to claim 7, characterized in that, The step of sending the optimal operating parameters to the control system for execution includes: Within each control cycle, the hybrid digital twin model is used to predict the process state trajectory within a preset time domain in the future; The optimal operating parameter sequence is solved within the preset time domain so that the predicted process state trajectory is close to the target set value; Only the first control action in the optimal operating parameter sequence is issued to the control system for execution, and the step of solving the optimal operating parameter sequence is re-executed in the next control cycle to achieve rolling optimization.

9. The method according to claim 8, characterized in that, The objective function for solving the optimal sequence of operating parameters includes a tracking error weighted norm term and a control action change penalty term. The expression of the objective function is as follows: in, For the current control cycle, To predict the number of steps, To predict the length of the time domain, In the current control cycle The predicted first The process status of the step, Set a value for the target. The weight matrix for tracking error, The weight matrix is ​​used to control the changes in actions. For the first The amount of change in the control action of the step.

10. A system for optimizing operating parameters in a protein purification process, characterized in that, include: Data acquisition module, digital twin module, online calibration module, and operation optimization module. The data acquisition module is configured to acquire various process data during the protein purification process, including monitoring data collected by sensors and process operation parameter data. The digital twin module is configured to process the various process data through a hybrid digital twin model to output a predicted value of the protein purification process status. The hybrid digital twin model includes a mechanism encoding module and a data compensation module. The mechanism encoding module generates a basic prediction result based on the physical mechanism model of protein purification, and the data compensation module performs residual compensation on the basic prediction result. The online calibration module is configured to, during the operation of the protein purification process, obtain the prediction deviation between the predicted value of the protein purification process status and the measured value of the protein purification process status, and detect the data distribution drift of the various process data; in response to the prediction deviation or the data distribution drift meeting the preset calibration conditions, perform incremental updates on the hybrid digital twin model, and perform physical constraint verification on the updated hybrid digital twin model; The operation optimization module is configured to use the hybrid digital twin model, which has passed physical constraint verification, as a proxy model, and search for the optimal operation parameters that meet the preset process constraints in the parameter space of the protein purification process, with multiple key quality indicators as optimization targets.

Images