AI-based clindamycin phosphate processing hydrolysis optimization method and system

Abstract

The invention discloses an AI (artificial intelligence)-based clindamycin phosphate processing hydrolysis optimization method and an AI-based clindamycin phosphate processing hydrolysis optimization system. According to the method, a reaction process prediction model capable of representing a process bifurcation phenomenon and fusing a chemical reaction mechanism and process data is constructed, and the model is utilized to train a multi-target decision-making agent. In the training process, causal relationship analysis is utilized to guide exploration of the intelligent agent, dynamic security constraints based on a chemical mechanism are introduced, and finally a strategy set covering the Pareto optimal frontier is generated. An operator can select a strategy according to production requirements, and closed-loop self-adaptive control is carried out by the system. The method aims to solve the problem of multi-target dynamic optimization caused by process bifurcation in the hydrolysis process, balance of multiple conflict targets such as yield, impurities and energy consumption is achieved by actively controlling the process bifurcation, and the robustness, safety and stability of the process are improved.

Description

Technical Field

[0001] The present invention relates to the field of chemical pharmaceutical process control through specific computational models, and more specifically, to an AI-based clindamycin phosphate processing hydrolysis optimization method and system. Background Art

[0002] Clindamycin phosphate is a clinically important lincosamide antibiotic, typically prepared by semi-synthesis of lincomycin. In its industrial production process, the hydrolysis step is a critical step in converting the intermediate into the final active pharmaceutical ingredient. The reaction performance of this step directly determines the final yield and purity of the product. The core of pharmaceutical quality control lies in the strict management of impurities. During the hydrolysis of clindamycin phosphate, a variety of process impurities are inevitably generated, such as clindamycin-B-phosphate and 7-epi-clindamycin phosphate. Pharmacopoeias worldwide have set strict upper limits for the content of these impurities.

[0003] Existing techniques, such as those described in Chinese patent publication CN107652332B, primarily improve yield and purity by optimizing chemical formulations, such as using specific hydrolyzing agents. However, these methods provide a static, fixed process parameter window and fail to address dynamic disturbances in actual production, such as raw material batch fluctuations and catalyst activity decay over time.

[0004] In complex chemical reactions, particularly in multivariable, nonlinear systems like the hydrolysis of clindamycin phosphate, critical, unstable critical regions exist. When small changes in process parameters such as temperature, pH, and reactant concentrations cause the system state to enter or pass through these critical regions, the main reaction pathway undergoes qualitative changes, leading to significant, nonlinear differences in final product properties such as yield and impurity profile. This phenomenon is referred to as "process bifurcation" in this paper.

[0005] Specifically in the hydrolysis process of clindamycin phosphate, there are at least the following typical process bifurcation phenomena: One is the yield-impurity bifurcation, which means that under certain process conditions, increasing the reaction temperature to increase the rate of main product formation may cause the system to cross a bifurcation point, resulting in a disproportionately sharp increase in the formation rate of a specific impurity, such as 7-epi-clindamycin phosphate; The second is bifurcation between impurity spectra, that is, adjusting the pH value may effectively inhibit the formation of impurity A, but at the same time may trigger another side reaction pathway, resulting in a significant increase in the concentration of impurity B; The third is kinetic-degradation bifurcation, that is, in the later stage of the reaction, in order to pursue a higher conversion rate, prolonging the reaction time or maintaining a higher temperature may cause the system to cross a degradation bifurcation point, resulting in the decomposition of the target product clindamycin phosphate and the generation of new degradation impurities.

[0006] The locations of these process bifurcation points are not fixed but will drift dynamically with changes in hidden variables such as raw material batches and catalyst activity. Traditional control methods, such as PID (proportional-integral-derivative) controllers, are limited by their single-input and single-output nature and are unable to foresee and manage bifurcations caused by multivariable coupling. Even advanced control strategies, such as model predictive control (MPC), struggle to effectively address these dynamically drifting, nonlinear bifurcations because they rely on a fixed process model that often cannot accurately describe complex reaction kinetics. In recent years, AI technology has been attempted to be applied to process optimization, but both predictive models based on supervised learning and standard single-objective reinforcement learning can only learn surface correlations in the data or optimize a single objective, failing to address the multi-objective trade-off decision-making problems faced at multiple process bifurcations. Summary of the Invention

[0007] To solve the above technical problems, one aspect of the present invention provides an AI-based clindamycin phosphate processing and hydrolysis optimization method, which is implemented by a computer and comprises the following steps: Step S1: Acquire real-time and offline data from the reaction process and perform state characterization. This step acquires data through a data acquisition layer, which includes conventional sensors such as temperature sensors, pH meters, and mass flow meters, as well as Process Analytical Technology (PAT) sensors for online, real-time monitoring of key component concentrations in the reaction system. Simultaneously, offline analytical data such as final yield and impurity content from historical batches tested using methods such as high-performance liquid chromatography (HPLC) are integrated. After preprocessing, the collected data forms a high-dimensional state vector that characterizes the state of the reaction system at any given moment.

[0008] Step S2: Construct a reaction process prediction model that can characterize the process bifurcation phenomenon. The model can depict the entire reaction kinetics including the process bifurcation point. This step specifically includes: Build a simulation module that integrates chemical reaction mechanisms and process data to generate a high-fidelity dynamic model that can predict the evolution trajectory of the system under different process conditions; and constructing a causal relationship analysis module for identifying and quantifying the causal transmission paths between process parameters and final product attributes from the data.

[0009] Step S3: Based on the prediction model, a multi-objective decision-making agent is trained to generate a strategy set for actively controlling the process bifurcation. This step specifically includes: The optimization problem is formalized as a multi-objective decision process, wherein the high-dimensional state vector is defined as the process state, the adjustment amount of the controllable process parameter is defined as the control operation, and multiple process performance indicators including yield, impurities, and energy consumption are defined as a multi-dimensional reward vector; In the virtual environment constructed by the simulation module, the output of the causal relationship analysis module is used to guide the agent to conduct exploratory training to improve learning efficiency; During the training process, dynamic safety constraints based on known chemical reaction mechanisms are introduced to limit the range of optional control operations of the agent to ensure the safety and robustness of the generated strategy. Through a multi-objective optimization algorithm, a strategy set covering the Pareto optimal frontier is finally generated, where each strategy in the set represents a specific trade-off between multiple conflicting objectives.

[0010] Step S4, online deployment and adaptive control. This step applies the trained model to the actual production process, specifically including: The Pareto optimal strategy set is visualized through a human-computer interaction interface, so that an operator can select an operation strategy according to current production needs; According to the selected strategy, real-time closed-loop control is performed during the production process, and after each production batch is completed, the reaction process prediction model is updated online using the full process data of the batch to adapt to changes in process characteristics.

[0011] Another aspect of the present invention provides an AI-based clindamycin phosphate processing and hydrolysis optimization system, which includes: a data acquisition and execution unit for acquiring process data and executing control instructions; a central computing processing unit, which is deployed with modules for executing the algorithms described in steps S2, S3 and S4 of the above method; and a visual display unit for human-computer interaction and strategy selection. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 4 is a structural block diagram of an AI-based clindamycin phosphate processing and hydrolysis optimization system according to an embodiment of the present invention.

[0013] Figure 2 4 is a flow chart of a method for optimizing the processing and hydrolysis of clindamycin phosphate based on AI according to one embodiment of the present invention.

[0014] Figure 3 is a schematic diagram for explaining the core concept of the present invention, wherein: Figure 3 A schematically shows the process bifurcation phenomenon to be solved by the present invention. Figure 3 B schematically shows the Pareto optimal strategy set generated by the method of the present invention.

[0015] Figure 4 3 is a comparison chart of the prediction performance of the PI-DT model according to one embodiment of the present invention and the standard neural network model.

[0016] Figure 5 is a process parameter causal relationship diagram learned by a structural causal model (SCM) according to an embodiment of the present invention.

[0017] Figure 6 2 is a comparison chart of the exploration trajectories of a CI-MORL agent and a standard RL agent according to one embodiment of the present invention.

[0018] Figure 7 1 is a Pareto optimal strategy set and its corresponding process control curve diagram according to an embodiment of the present invention. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the scope of protection of the present invention.

[0020] Example 1 This embodiment provides an AI-based clindamycin phosphate processing and hydrolysis optimization method, which is implemented by computer, referring to Figure 2 The detailed steps of the process shown are as follows: Step S1: Acquire real-time and offline data of the reaction process and perform state characterization.

[0021] In this embodiment, this step is implemented using an integrated data acquisition layer. This layer includes multiple sensors installed on the hydrolysis reactor, such as a Pt100 platinum resistance temperature sensor for measuring the temperature of the material within the reactor, an online pH meter for online pH monitoring, and a mass flow meter for precisely controlling the rate of addition of acid and alkali solutions. Specifically, to obtain real-time concentration information of key components, this embodiment utilizes a process analytical technology (PAT) system, specifically a Raman spectrometer equipped with an immersion fiber optic probe, with an excitation wavelength of 785 nm. This Raman spectrometer scans the reaction solution every 60 seconds. Using a pre-established chemometric correction model, the collected Raman spectra are interpreted in real time as the concentrations of key substances such as clindamycin phosphate, clindamycin-B-phosphate, and 7-epi-clindamycin phosphate. Furthermore, the system database also integrates data on final product yield and precise impurity content from historical production batches, obtained through offline high-performance liquid chromatography (HPLC) analysis after the reaction. All collected data, including real-time sensor data and offline analysis data, undergoes preprocessing before entering the subsequent model. The preprocessing process includes: first, smoothing high-frequency noisy data (such as temperature readings) using a moving average method; second, aligning data with different sampling frequencies (such as 1-second temperature data and 60-second concentration data) using timestamps; and finally, Z-score normalization of all data dimensions to a mean of 0 and a standard deviation of 1 to eliminate the influence of different physical dimensions. After this processing, the system process state at any time t can be represented as a high-dimensional state vector x(t), for example, x(t) = [temperature(t), pH(t), acid solution droplet velocity(t), base solution droplet velocity(t), clindamycin phosphate concentration(t), impurity A concentration(t), ...].

[0022] Step S2: construct a reaction process prediction model that can characterize the process bifurcation phenomenon.

[0023] This step aims to establish a predictive model that can accurately depict the overall picture of the clindamycin phosphate hydrolysis process. This model can not only predict the future process state of the reaction, but also explain its internal causal mechanism, especially the ability to identify and quantify process bifurcation phenomena. Figure 1 , the model is completed by the collaborative work of the physical information digital twin module 101 and the structural causal model engine 102.

[0024] Preferably, this step specifically includes: Sub-step S201: constructing a physical information digital twin (PI-DT) module 101 for generating a panoramic model of reaction dynamics.

[0025] In this embodiment, the core of the module is a physical information neural network (PINN). This network is a deep feedforward neural network, for example, containing 5 hidden layers, 128 neurons per layer, and the activation function is the hyperbolic tangent function (tanh). The training process of the PINN aims to minimize a mixed loss function L_total = L_data + λ L_phys. L_data is the data fitting loss term, calculated using mean squared error (MSE), which measures the degree to which the network output (such as predicted concentration and temperature) matches the historical and real-time data collected in step S1. L_phys is the physical law residual term, which embeds the core chemical reaction kinetics equations for the clindamycin phosphate hydrolysis reaction (including partial differential equations for mass conservation and energy conservation, such as dC / dt = f(C, T, pH) and dT / dt = g(T, Q)) into the loss function. C represents the concentration of key substances, T represents temperature, pH represents acidity, and Q represents the heating power of the system. This term calculates the extent to which the network output violates these physical laws. λ is a hyperparameter that balances the importance of data fitting with physical constraints and can be set to 0.1, for example. By minimizing L_total using optimizers such as Adam, the trained PINN not only fits the available data but also ensures that its internal structure adheres to the fundamental laws of chemical reactions. Therefore, the PI-DT module 101 can generate a high-fidelity digital twin, which can serve as a simulation environment for a reaction process. Given any sequence of process parameters, it can accurately simulate the dynamic evolution trajectory of the entire reaction process, including those areas that have never appeared in historical data but are physically possible, thereby effectively identifying process bifurcation phenomena in the state space, such as Figure 3 As shown in A, the system state may evolve along different bifurcation paths after passing through a critical region. Figure 4 Compared with the standard neural network based solely on data-driven methods, the PI-DT model shows significant predictive advantages in data-sparse or extrapolated areas, can provide prediction results that are more consistent with physical and chemical laws, and avoid the numerical instability oscillations caused by overfitting in traditional models.

[0026] In another embodiment, the construction of the physical information digital twin module can be replaced by using a traditional first-principles model based on the Arrhenius formula, etc. as a basis, and then training a residual neural network to learn and compensate for the deviation between the first-principles model and the actual data.

[0027] Sub-step S202 , constructing a structural causal model (SCM) engine 102 for identifying and quantifying bifurcation paths.

[0028] In this embodiment, the module aims to identify the causal relationship between process parameters and process results from the data. First, based on known chemical engineering knowledge, an initial causal graph skeleton is constructed. For example, the graph contains directed edges pointing from "temperature" to "main reaction rate" and "impurity A generation rate". Then, using historical production data, the PC-stable causal discovery algorithm is used to learn and refine the skeleton. Through a series of conditional independence tests, the algorithm can discover hidden causal relationships in the data that are not included in prior knowledge, such as discovering that "the specific properties of a batch of raw materials" are a hidden common cause affecting "impurity B generation". Finally, a refined directed acyclic graph (DAG) is obtained, refer to Figure 5 , which shows the causal transmission path between controllable variables, intermediate state variables and the final goal. For each causal edge in the graph, the system learns a structural equation. For example, for the concentration C_A of impurity A, its structural equation may be C_A = f(T,pH) + N_A, where f is a function (which can be represented by a small neural network or linear regression model), T and pH are its direct causes, and N_A is a noise term representing random disturbances. The SCM engine 102 is capable of causal inference. For example, it can quantify the direct causal effect of "raising the temperature set point by 0.5°C under the current state" on the "final yield" and "7-epi-clindamycin phosphate concentration".

[0029] Step S3: Based on the prediction model, a multi-objective decision-making agent is trained to generate a strategy set for actively controlling the process bifurcation.

[0030] This step aims to train an intelligent agent that can generate optimal timing control strategies so that it can learn to actively control in a complex, multi-objective reaction environment.

[0031] In sub-step S301 , the optimization problem is formalized as a multi-objective Markov decision process (MOMDP).

[0032] Specifically, the process state (State) S_t is the real-time state vector x(t) defined in step S1, supplemented by the PI-DT module 101, for example, to provide short-term predictions of future process states. The control action (Action) A_t is defined as a discrete set of control instructions executable by the underlying programmable logic controller (PLC). For example, a set of control actions might be the Cartesian product of {temperature setpoint +0.2°C, temperature setpoint -0.2°C, maintain constant} and {acid solution dripping rate +0.5 ml / min, acid solution dripping rate -0.5 ml / min, maintain constant}. The process performance indicator (Reward) R_t is defined as a four-dimensional vector R_t = [r_yield, r_impurity1, r_impurity2, r_energy]. Among them, r_yield is positively correlated with the formation rate of clindamycin phosphate; r_impurity1 and r_impurity2 are penalty items related to the concentration of key impurities. When the impurity concentration exceeds the warning threshold, a large negative value is given; r_energy is the energy consumption cost item related to heating power and stirring power, which is a negative value.

[0033] Sub-step S302: performing causal guidance exploration and training.

[0034] The training of the agent 103 is mainly carried out offline in the PI-DT virtual environment constructed in step S201 to ensure safety and high efficiency. During the exploration phase of training, the agent does not select control operations completely randomly. Instead, before selecting a tentative control operation, it queries the SCM engine 102. The SCM engine evaluates the causal effects of each optional control operation on multiple process performance indicator dimensions under the current process state. The agent will give priority to those control operations that the SCM predicts have a strong positive causal effect on the desired target (such as yield) and a weak negative causal effect on the undesirable target (such as impurities). For example, if the SCM engine indicates that increasing the temperature in the current reaction stage mainly leads to an increase in yield and has little effect on impurity formation, then the agent will choose the control operation related to temperature increase with a higher probability.

[0035] Furthermore, in order to improve the robustness and safety of the system, dynamic taboo constraints based on reaction mechanisms are introduced when performing causal-guided trial and training. In this embodiment, a dynamic taboo list generation module 104 is constructed. Dangerous operation rules based on chemical mechanisms are pre-encoded in this module, for example, Rule 1: "When the pH value in the kettle is lower than 4.5 and the temperature is higher than 55°C, the risk of degradation of the target product is high." During the training process, the PI-DT module 101 will evaluate the current simulation process state in real time. If it is detected that the current process state meets the conditions of Rule 1, that is, the system state enters a bifurcation area that may cause the degradation of the target product, the dynamic taboo list generation module 104 will immediately generate a taboo control operation list, for example, all "heating" or "acid addition" related control operations are prohibited in the next 15 minutes. When the intelligent agent 103 makes a decision, its optional control operation set will be dynamically constrained by this list, thereby forcing it to choose from a safe subset of control operations. Reference Figure 6 By comparing the exploration trajectories of a CI-MORL (Causal-Informed Multi-Objective Reinforcement Learning) agent with a standard reinforcement learning agent, we clearly see that the proposed agent effectively avoids pre-set dynamic taboo regions, while the standard agent's exploration is blind and can potentially lead to dangerous process states. This adds a strict safety constraint to the AI's exploration process, ensuring that the resulting strategy is physically safe and feasible.

[0036] Alternatively, the step of introducing dynamic taboo constraints can also be achieved in the following way: the degree of violation of safety rules is used as a negative process performance indicator dimension and incorporated into the multi-objective process performance indicator vector, thereby guiding the intelligent agent to avoid dangerous areas in a soft constraint manner.

[0037] Sub-step S303: generating a Pareto optimal strategy set.

[0038] This embodiment adopts a known multi-objective reinforcement learning algorithm based on a conditional network (such as C-MORL). The policy network trained by this algorithm takes not only the process state S as input, but also a preference vector w as an additional input. The preference vector w represents the degree of emphasis on different objectives. For example, w=[0.7, 0.1, 0.1, 0.1] represents a greater focus on yield. By sampling a variety of preference vectors w during the training process, the single policy network finally trained can be generalized to any preference. Therefore, after the training is completed, by inputting different w, a series of different optimal control strategies can be generated from the policy network. These strategies constitute the Pareto optimal frontier, as shown in the reference Figure 7Each point in the figure (such as P1, P2, P3) represents a complete, end-to-end control strategy, corresponding to Figure 7 The right subplot shows different dynamic control curves. For example, strategy P1 might represent a high-yield solution with slightly higher impurities, while strategy P3 represents a slightly lower yield solution with extremely high purity. No single strategy outperforms another in all objectives. This set of strategies provides multiple optimal options for practical production.

[0039] Step S4: online deployment and adaptive control.

[0040] This step applies the trained model to the closed-loop control of the actual production process.

[0041] Sub-step S401: providing a visual decision interface and selecting a strategy.

[0042] The Pareto optimal strategy set generated in step S303 is visualized through the human-machine interface (HMI) 105. Figure 7 , the interface will be the Pareto optimal frontier (such as Figure 7 The interface is a two-dimensional scatter plot with the horizontal axis representing the "total impurity content" and the vertical axis representing the "final yield". The production supervisor or process engineer can intuitively click on the graph to select a strategy point that best meets their needs, such as selecting P2 as a balance point, based on the specific business objectives of the current batch (for example, whether to pursue the fastest reaction speed to meet orders, or to pursue the highest purity to produce high-specification drugs). When a strategy point is selected, the system can further display the specific process control curve corresponding to the strategy (such as Figure 7 right sub-figure) for reference by decision makers.

[0043] Sub-step S402 , executing real-time closed-loop control and model adaptation.

[0044] Once a strategy is selected, the system enters automatic execution mode. Throughout the hydrolysis batch, the data acquisition layer continuously transmits the real-time process state vector S_t to the central processing unit 100. The selected CI-MORL strategy outputs the optimal control action A_t based on the current process state S_t. This command is sent via the control network to the underlying distributed control system (DCS) or programmable logic controller (PLC) for precise execution, such as adjusting the steam valve opening of the heating jacket or the speed of the acid and alkali dosing pump. This "perception-decision-execution" closed loop continues at a set time step (e.g., every 60 seconds) until the reaction endpoint. After a production batch is completed, the entire process data (including all-time process states, control operations, and process performance indicators) is automatically collected and used to incrementally update the PI-DT module 101 and SCM engine 102 online. For example, a small learning rate is used to fine-tune the model parameters. This adaptive mechanism enables the system to continuously learn and effectively address slow drift in process characteristics caused by factors such as catalyst activity decay and equipment aging, thereby maintaining the effectiveness of the control strategy.

[0045] Example 2 This embodiment provides an AI-based clindamycin phosphate processing and hydrolysis optimization system, the structure of which is shown in the attached figure. Figure 1 This system is the physical carrier of the method described in Example 1.

[0046] Reference Figure 1 The system includes: a data acquisition and execution unit 110, a central computing processing unit and a visualization display unit 105.

[0047] The data acquisition and execution unit 110 is responsible for interacting with the physical production process. It includes various sensors deployed on the reactor, such as temperature sensors, pH meters, and Raman spectrometers, for performing data acquisition in step S1. It also includes various actuators, such as control valves connected to the heating system and frequency converters connected to the feed pumps, for executing the control instructions issued by the central processing unit 100 in step S402.

[0048] The central processing unit 100 is the core of the system of the present invention and is typically a high-performance industrial computer or server. The functional modules that implement the core algorithm of the present invention are deployed within the unit, specifically including: A physical information digital twin (PI-DT) module 101 is used to perform the function described in step S201 and establish a high-fidelity process model.

[0049] A structural causal model (SCM) engine 102 is used to execute the function described in step S202 to perform causal relationship mining and quantification.

[0050] A multi-objective decision-making agent 103, which integrates training and reasoning functions, is used to perform the strategy generation described in step S3 and the online control instruction generation in step S402.

[0051] A dynamic taboo list generation module 104 is used to execute the security constraint function described in step S302.

[0052] An online adaptive module is used to coordinate and call the collected new data to update the PI-DT module 101 and the SCM engine 102 after the batch is completed, thereby realizing the adaptive function described in step S402.

[0053] The visualization display unit 105 , typically a touch screen or computer display running human-machine interface (HMI) software, is used to perform step S401 , graphically displaying the Pareto optimal strategy set generated by the agent 103 to the operator and receiving the operator's strategy selection instructions.

[0054] During actual operation, the data acquisition and execution unit 110 continuously uploads process data to the central processing unit 100. The various modules within the central processing unit 100 work together to calculate optimal control instructions in real time based on the strategy selected by the operator on the visual display unit 105. These instructions are then issued to the actuators in the data acquisition and execution unit 110 for operation, thus forming a complete, intelligent closed-loop control system and achieving dynamic, multi-objective optimization of the clindamycin phosphate hydrolysis process.

[0055] Example 3 This example describes an implementation method for a specific application scenario. The preconditions for this scenario are: the hydrolysis catalyst used has been run for 45 batches; and the lincomycin hydrochloride raw material contains a precursor impurity content of 0.15%. The control objectives for this scenario are: the content of 7-epi-clindamycin phosphate in the final product does not exceed 1.0%, while maximizing the yield.

[0056] Before adding the raw materials, the system of the present invention first performed pre-analysis and strategy generation. The system input offline HPLC analysis data for new raw material batch YL-20230815 and a record of 45 catalyst uses. Based on this new information, the PI-DT module 101 and SCM engine 102 performed a rapid online fine-tuning. Subsequently, the PI-DT module 101 simulated the standard operating procedure (SOP) in a virtual environment. The simulation results clearly identified a shifted "yield-impurity" bifurcation point: due to the presence of high levels of precursor impurities and reduced catalyst activity, the side reaction pathway for the formation of 7-epi-clindamycin phosphate was significantly activated within the temperature (52°C) and pH (4.8-5.2) control ranges of the standard SOP. The predicted final content of 7-epi-clindamycin phosphate reached 1.35%, failing to meet quality requirements. The SCM engine 102 also quantified this causal effect, indicating that under these specific conditions, the positive causal effect of temperature on the formation of this impurity increased by 40% compared to normal.

[0057] The system invoked the multi-objective decision-making agent 103, which generated a new Pareto-optimal strategy set for this specific batch, using "final 7-epi-clindamycin phosphate content ≤ 0.9%" and "maximum final yield" as primary objectives. Combined with secondary objectives such as energy consumption, it generated a new Pareto-optimal strategy set for this specific batch. On the visualization display unit 105, the process engineer observed that, compared to the Pareto front for conventional batches, the newly generated front shifted downward to the left (low yield, high impurities), intuitively reflecting the complexity of this production run. The engineer selected a preset strategy point, designated EHP-45. This strategy predicted a slightly lower yield of 2.5% compared to the standard SOP, but only predicted an impurity value of 0.88%.

[0058] After receiving authorization, the system began executing EHP-45 closed-loop control. This strategy is highly dynamic: During the initial reaction period (0-2 hours), the system maintained a low temperature of 46°C, 6°C below the SOP, to suppress the initial rate of epimerization. At this point, the SCM engine analysis indicated that the negative causal impact of low temperature on the main reaction rate was not yet significant. In the middle of the reaction period (2-5 hours), when the concentration of the main product reached a critical value, the strategy began to slowly increase the temperature at a rate of 0.1°C / 10 minutes. Meanwhile, the hydrochloric acid addition rate was precisely controlled to maintain the pH within a narrower range of 4.9 ± 0.05, seeking a dynamic optimal equilibrium point. At 4.5 hours, the Raman spectrometer detected a transient increase in the concentration of an intermediate product in the system, prompting the PI-DT module 101 to predict that the system process state was approaching a "kinetic-degradation" bifurcation point. Dynamic taboo list generation module 104 was immediately triggered, temporarily disabling all control options for "increasing temperature" and "increasing acid droplet acceleration" for the next 30 minutes, forcing the system to operate within a safe zone and effectively preventing degradation of the target product. At the end of the reaction, the strategy stabilized the temperature at 50°C, slightly below the SOP, and extended the reaction time by approximately 90 minutes to ensure conversion.

[0059] Ultimately, the batch reaction took 9.5 hours to complete (1.5 hours longer than the SOP). Sampling and HPLC endpoint analysis revealed a final yield of 89.2% for clindamycin phosphate (typically 91.5%-92.5% for the SOP), while the content of the key impurity, 7-epi-clindamycin phosphate, was 0.92%, successfully controlled within the strict limit of 1.0%. The results of this example demonstrate that the present invention can effectively address the dynamic changes in the states of raw materials and catalysts. By accurately identifying and actively controlling process bifurcations, it generates and executes a dynamic, globally optimal, and safe and reliable control strategy for a specific production task under multiple constraints and conflicting objectives, achieving precise process control.

[0060] The above description is only a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included in the scope of protection of the present invention.

Claims

1. An AI-based clindamycin phosphate processing and hydrolysis optimization method, characterized in that: The method is implemented by a computer and comprises the following steps: Step S1: Acquire real-time and offline data of the reaction process and perform state characterization to construct a high-dimensional state vector that can characterize the state of the reaction system at any moment; Step S2: Constructing a reaction process prediction model that can characterize process bifurcation phenomena. This step specifically includes constructing a simulation module that integrates chemical reaction mechanisms and process data, as well as a causal relationship analysis module. Step S3: Based on the prediction model, a multi-objective decision-making agent is trained to generate a strategy set covering the Pareto optimal frontier for actively controlling the process bifurcation, wherein the training process combines guidance using the output of the causal relationship analysis module and dynamic safety constraints based on known chemical reaction mechanisms; Step S4: Online deployment and adaptive control, which includes visually displaying the strategy set covering the Pareto optimal front for selection, performing real-time closed-loop control according to the selected strategy, and using the batch data to update the prediction model online after the production batch is completed.

2. The method according to claim 1, characterized in that The simulation module in step S2 is a physical information digital twin (PI-DT) module, the core of which is a physical information neural network (PINN). The loss function of the physical information neural network includes a data fitting loss term and a physical law residual term based on the partial differential equation of chemical reaction kinetics.

3. The method according to claim 1, characterized in that The causal relationship analysis module in step S2 is a structural causal model (SCM) engine, and its construction steps include: constructing an initial causal graph skeleton based on prior knowledge, refining it using historical data and causal discovery algorithms, and learning structural equations for the causal edges in the refined graph.

4. The method according to claim 1, wherein The steps of using the output of the causal relationship analysis module to guide the intelligent agent to train in step S3 are specifically as follows: in the exploration phase of training, before selecting a control operation, the intelligent agent queries the causal relationship analysis module to evaluate the causal effects of each optional operation on multiple process performance indicators, and preferentially selects a control operation that has a stronger positive causal effect on a preset desired process performance indicator and a weaker negative causal effect on a preset undesirable process performance indicator.

5. The method according to claim 1 or 4, characterized in that The step of introducing dynamic security constraints in step S3 is specifically implemented by at least one of the following methods: Method 1: Construct a dynamic taboo list generation module. Based on preset dangerous operation rules, during training, when the simulation module assesses that the current simulation state has entered a high-risk area, it generates a set of taboo control operations to constrain the agent's optional operation set. Method 2: The degree of violation of the dangerous operation rules is used as a negative process performance indicator dimension and incorporated into the multi-objective reward, guiding the agent to avoid high-risk areas in a soft constraint manner.

6. The method according to claim 1, characterized in that Before training the multi-objective decision-making agent in step S3, the optimization problem is formalized into a multi-objective Markov decision process, which specifically includes: defining the high-dimensional state vector as a process state; The adjustment instruction set of the controllable process parameters is defined as a control operation; And a multidimensional vector including those related to product generation rate, key impurity concentration and energy consumption is defined as the process performance index.

7. The method according to claim 1, characterized in that The step of executing real-time closed-loop control in step S4 specifically includes: during the production process, the intelligent agent outputs the optimal control operation based on the high-dimensional state vector obtained in real time, and executes it through the control system to form an ongoing perception-decision-execution closed loop.

8. The method according to claim 1, characterized in that The step of updating the prediction model online in step S4 specifically includes: after a production batch is completed, using the full process data of the batch to incrementally update the model parameters of the simulation module and the causal relationship analysis module online.

9. The method according to claim 1, characterized in that Acquiring real-time data in step S1 is achieved through a process analytical technology (PAT) sensor, which is specifically a Raman spectrometer or a near-infrared spectrometer.

10. An AI-based clindamycin phosphate processing and hydrolysis optimization system, characterized in that: include: A data acquisition and execution unit, used to obtain data of the clindamycin phosphate hydrolysis process and execute control instructions; A visual display unit for running human-computer interaction interface software; and A central computing and processing unit connected to the data acquisition and execution unit and the visual display unit, wherein the central computing and processing unit is configured to execute the method according to any one of claims 1 to 9. Specifically, the central computing and processing unit is configured to have: a simulation module; a causal relationship analysis module; A multi-objective decision-making agent; In which, the central computing processing unit is configured to: graphically display the Pareto optimal strategy set generated by the multi-objective decision-making intelligent agent on the visual display unit and receive the operator's strategy selection instructions, then generate real-time control instructions according to the selected strategy, and perform closed-loop control of the hydrolysis process through the data acquisition and execution unit.

Images