Food-grade calcium carbonate continuous production device and intelligent control method

Abstract

The invention discloses a food-grade calcium carbonate continuous production device and an intelligent control method, and relates to the technical field of calcium carbonate preparation, a feeding pipeline is provided with a venturi ejector and is connected with a carbon dioxide inlet pipe, and carbon dioxide is powerfully sucked and shredded into bubbles by utilizing instantaneous negative pressure generated by high-speed flowing lime milk, so that micro-mixing is realized; the crystal growth unit comprises a horizontal tank body, the interior of the horizontal tank body is divided into serially-connected compartments through a plurality of staggered overflow baffles, slurry is forced to advance in a piston flow mode, and a low-shear stirring paddle is arranged; and the crystal form regulation and control unit comprises a vertical continuous stirring kettle, and a heating and cooling jacket is arranged on the outer wall to control the aging temperature. The intelligent control method comprises the steps of multi-dimensional data synchronous acquisition, reaction field domain modeling, crystal evolution prediction, collaborative prediction control and collaborative diagnosis optimization, a reaction field domain twinborn model and a crystal evolution twinborn model are constructed, and a field domain separation attention mechanism and a double-path prediction architecture are adopted to realize a field domain driven feedforward control strategy.

Description

Technical Field

[0001] This invention relates to the field of calcium carbonate preparation technology, specifically to a continuous production apparatus and intelligent control method for food-grade calcium carbonate. Background Technology

[0002] Calcium carbonate, as an important functional inorganic powder material, has wide applications in the food, pharmaceutical, and cosmetic industries. Food-grade calcium carbonate requires high purity, uniform particle size distribution, stable crystal form, and good dispersibility, which places extremely high demands on the production process and quality control.

[0003] Traditional calcium carbonate production primarily employs a batch process, involving the chemical precipitation reaction of lime slurry and carbon dioxide in a reactor. In existing technologies, the mixing of lime slurry and carbon dioxide typically utilizes direct aeration and stirring or mechanical dispersion. During the reaction, nucleation, growth, and aging stages occur simultaneously within the same reactor, making independent and optimized control of each stage difficult. Regarding control methods, traditional processes rely mainly on operator experience and simple PID feedback control, lacking a deep understanding of the inherent laws governing the reaction process and the ability to accurately predict its progression.

[0004] Existing technologies improve product quality by optimizing process parameters such as reaction temperature, pH value, and stirring speed, but they still have certain limitations. These include uneven particle size distribution, large batch-to-batch quality fluctuations, poor crystal form consistency, low production efficiency, low equipment utilization, high reliance on manual labor, difficulty in achieving continuous production, and inability to meet the urgent needs of the modern food industry for high-quality, stable calcium carbonate products. Particularly in the nucleation stage, traditional gas-liquid mixing methods cannot achieve uniform distribution at the microscale, leading to inconsistencies in crystal nucleus size and distribution. In the growth stage, crystal particles of various sizes compete for growth in the same environment, causing severe mutual interference. In the aging stage, temperature and residence time are difficult to control precisely, affecting the crystal form stability and surface properties of the final product.

[0005] Therefore, there is an urgent need for a new type of production equipment and control method that can realize continuous production of calcium carbonate, has intelligent predictive control functions, and can significantly improve product quality stability and production efficiency. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and to propose a continuous production device and intelligent control method for food-grade calcium carbonate to solve the above-mentioned problems.

[0007] The objective of this invention is achieved through the following technical solution: a continuous production apparatus for food-grade calcium carbonate, the apparatus comprising, in sequence according to the flow order of the slurry: The crystal nucleus induction unit includes a vertical cylindrical tank. A Venturi injector is installed on the feed pipe of the vertical cylindrical tank. An inlet pipe for inputting carbon dioxide gas is connected to the Venturi injector. The Venturi injector uses the instantaneous negative pressure generated by the high-speed flowing lime slurry in the throat contraction section to forcefully draw in the carbon dioxide gas and tear it into bubbles in the diffusion section, thereby realizing the micro-mixing of the reactants. The crystal growth unit, whose inlet is fluidly connected to the outlet of the crystal nucleus induction unit, includes a horizontal tank. The interior of the horizontal tank is divided into multiple series compartments by multiple highly staggered overflow baffles, which force the slurry to move forward in a piston flow manner of upward and downward flow. Low-shear stirring paddles are installed in the series compartments to maintain the suspension of crystal particles and avoid violent collisions. The crystal form control unit, whose inlet is fluidly connected to the outlet of the crystal growth unit, includes a vertical continuous stirred tank. The outer wall of the vertical continuous stirred tank is equipped with a heating and cooling jacket. The aging temperature is controlled by circulating medium to achieve crystal form stabilization and surface modification.

[0008] The device also includes a sensor network and an actuator network. The sensor network includes temperature sensors, pH sensors, conductivity sensors and flow sensors deployed at key locations in the crystal nucleation unit, crystal growth unit and crystal form control unit, as well as an online image analyzer deployed at the outlet of the crystal growth unit or the crystal form control unit. The online image analyzer provides a continuous real-time video stream for monitoring dynamic changes in crystal morphology.

[0009] In the crystal nucleation induction unit, lime slurry enters the vertical cylindrical tank through a tangential inlet to form a rotating vortex with a residence time of 5-30 seconds. The throat contraction ratio and diffusion angle of the Venturi injector are optimized to achieve micro-mixing of gas and liquid.

[0010] The residence time of the crystal growth unit is 15-40 minutes, the length of the horizontal tank is 3-5 meters, the number of overflow baffles is 8-12, and multiple distributed feed ports are set at the top of the horizontal tank along the slurry flow direction to add low concentrations of reactants to the series compartments to maintain a stable supersaturation.

[0011] The intelligent control method for the device includes the following steps: Multidimensional data synchronous acquisition steps: Synchronously acquire process parameter data such as temperature, pH value, conductivity, and flow rate, which characterize the macroscopic state of the production process, as well as real-time video streams of crystal morphology from an online image analyzer, which characterize the microscopic state of the product; Reaction field modeling steps: Input process parameter data into the reaction field twin model. The reaction field twin model divides the calcium carbonate reaction system into three mutually coupled physical subdomains: nucleation field, growth field, and aging field. The field separation attention mechanism is used to model the spatiotemporal evolution process of calcium ion enrichment, carbonate diffusion, and crystal nucleation. Differential prediction of the physical field in different regions inside the reactor is achieved through localized attention windows. Crystal evolution prediction steps: Input the real-time video stream of crystal morphology into the crystal evolution twin model. The crystal evolution twin model constructs a dual-path prediction architecture based on the principle of crystal growth dynamics. One path predicts the overall morphological evolution trajectory of the crystal group, and the other path identifies and tracks the growth hotspots of individual crystals. Through the growth dynamics perception mechanism, it automatically identifies active crystal regions that are undergoing lattice rearrangement, surface adsorption, and agglomeration. Collaborative predictive control steps: The model predictive controller receives the future states of the three subdomains predicted by the reaction field twin model and the crystal evolution trajectory predicted by the crystal evolution twin model, constructs a field-crystal collaborative optimization objective function, adjusts the mass transfer rate between each field through the inner controller to maintain the ideal concentration gradient distribution, and corrects the reaction driving force of each field according to the deviation of the crystal evolution trajectory through the outer controller, so as to realize the whole process predictive control from the shaping of the reaction environment to the guidance of product form.

[0012] In the crystal evolution prediction step, the local supersaturation gradient distribution predicted in the reaction field modeling step is also used as the driving force input of the crystal evolution twin model. Through field-crystal coupling prediction, the macroscopic mass transfer environment and microscopic crystal growth dynamics are integrated and modeled. When the proportion of the active area of ​​the crystal exceeds the preset threshold, the prediction focus is automatically switched to optimize the allocation of computing resources.

[0013] The reaction field twin model adopts a spatial token partitioning mechanism characterized by chemical processes. Based on the characteristics of calcium carbonate reaction, the reactor space is divided into sub-regions with different flow characteristics, such as turbulent mixing zone, laminar mass transfer zone, and wall boundary zone. Each sub-region uses an independent physical quantity encoding channel. Through a cross-region attention mechanism, the material and energy transfer between different regions is modeled, thereby achieving a refined prediction of the complex flow field of the chemical reactor.

[0014] The crystal evolution twin model adopts a growth kinetics-guided sequence modeling method. Based on the multi-stage evolution mechanism of calcium carbonate crystal nucleation-growth-agglomeration, a stage-aware video prediction model is constructed. The model identifies the location of new crystal nuclei through a motion detector during the nucleation stage, monitors crystal boundary expansion through a deformation tracker during the growth stage, and warns of crystal collisions and aggregation through a contact detector during the agglomeration stage, thereby achieving accurate prediction of the entire crystal life cycle.

[0015] The collaborative predictive control step adopts a field-driven feedforward control strategy. When the crystal evolution twin model predicts that the crystal evolution trajectory deviates from the target morphology, the model predictive controller calculates the optimal inter-field driving force redistribution scheme based on the mass transfer constraints provided by the reaction field twin model. It actively corrects the crystal evolution path by adjusting the supersaturation peak of the nucleation field, the concentration gradient distribution of the growth field, and the temperature stability of the aging field.

[0016] The method also includes field-crystal co-diagnosis and adaptive optimization steps. By continuously comparing the prediction results of the reaction field twin model and the crystal evolution twin model with the actual data collected by the sensor network, when the reaction field prediction is accurate but the crystal evolution prediction is inaccurate, it is determined to be a deviation of the crystal growth kinetic model. When the reaction field prediction is inaccurate, it is determined to be a deviation of the mass transfer or heat transfer model. Based on the deviation type, the coupling parameters between the fields and the crystal growth kinetic parameters are calibrated in a targeted manner through an online learning algorithm.

[0017] The beneficial effects of this invention are: This invention, through a three-stage segmented reactor design, achieves physical separation and independent optimized control of the nucleation, growth, and aging processes of calcium carbonate, fundamentally solving the technical problem of mutual interference and difficulty in precise control of various processes in traditional batch processes. The application of a Venturi injector enables microscopic mixing of lime slurry and carbon dioxide, fundamentally improving the uniformity and controllability of the nucleation process. This device constructs a dual-twin model collaborative predictive control system, combining a reaction field twin model with a crystal evolution twin model, achieving for the first time integrated predictive modeling of macroscopic physicochemical fields and microscopic product morphology in chemical processes, changing the passive response mode of traditional chemical control.

[0018] This invention utilizes a spatial token partitioning mechanism characteristic of chemical processes to divide the reactor space into different functional regions such as turbulent mixing zones, laminar mass transfer zones, and wall boundary zones based on fluid dynamics characteristics, achieving refined modeling of complex chemical reactors. The growth kinetics-guided sequential modeling method enables unprecedented accuracy and temporal resolution in crystal evolution prediction. Collaborative diagnostics and adaptive optimization techniques endow the control system with self-learning and self-evolution capabilities, allowing it to continuously optimize model parameters through online learning algorithms, adapting to various disturbances such as equipment aging, raw material changes, and environmental fluctuations, ensuring long-term stable system operation.

[0019] This invention reduces the particle size distribution coefficient, improves the uniformity of particle size distribution, and decreases the standard deviation of crystal nuclei size through precise nucleation control and a stable growth environment. Morphological guidance is achieved through precise temperature control and an intelligent control system in the aging reactor.

[0020] Compared to traditional intermittent processes, continuous production significantly improves production efficiency and equipment utilization. Continuous operation eliminates the time losses associated with stoppages, cleaning, and restarts inherent in intermittent processes, resulting in a substantial increase in overall capacity. Automated and intelligent control systems reduce the need for manual operation and shorten control system response time, offering faster response compared to traditional PID control and ensuring timely correction of process disturbances. Attached Figure Description

[0021] Figure 1 This is an overall structural diagram of the present invention; Figure 2 This is an exploded view of the entire invention; Figure 3 The local explosion of the present invention Figure 1 ; Figure 4 The local explosion of the present invention Figure 2 ; Figure 5 This is a front view of the present invention; Figure 6 For the present invention Figure 5 Sectional view of AA; Figure 7 For the present invention Figure 6 BB section view; Figure 8 This is a structural diagram of the present invention; Figure 9 This is a flowchart of the present invention.

[0022] Explanation of the labels in the diagram 1. Vertical cylindrical tank; 2. Venturi injector; 3. Air inlet pipe; 4. Horizontal tank; 5. Overflow baffle; 6. Vertical continuous stirring vessel; 7. Heating and cooling jacket. Detailed Implementation

[0023] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] It should be noted that the directional concepts of "left", "right", "up", "down", "front", "back", "inner", and "outer" in the following scheme are all relative directions, and will not be listed one by one here.

[0025] Example 1: like Figures 1 to 8As shown in the figure, this embodiment provides a continuous production device for food-grade calcium carbonate. The device adopts a modular design concept and realizes the physical separation and precise control of crystal nucleation, growth and morphology regulation through a segmented three-stage reactor system. It fundamentally solves the technical problems of wide particle size distribution, inconsistent crystal form and low production efficiency in traditional batch production.

[0026] This continuous production unit is arranged in sequence according to the slurry flow order, consisting of a crystal nucleation induction unit, a crystal growth unit, and a crystal form control unit. These units are interconnected via a piping system, forming a complete continuous production line. The entire unit is skid-mounted, facilitating factory manufacturing, on-site installation, and routine maintenance.

[0027] The nucleation induction unit, as the first-stage reactor in the entire production plant, plays a crucial role in the explosive generation of a large number of uniform and tiny calcium carbonate nuclei in a very short time, laying a key foundation for obtaining the final product with a narrow particle size distribution. This unit includes a vertical cylindrical tank 1, preferably made of 316L stainless steel, with a 60-degree conical bottom design to facilitate complete material discharge. A Venturi ejector 2 is installed on the feed pipe of the vertical cylindrical tank 1. The Venturi ejector 2 employs a fluid dynamics design, comprising three key areas: a contraction section, a throat section, and a diffuser section. The half-angle of the contraction section is 15-25 degrees, the length of the throat section is 1.5-2.5 times the throat diameter, and the half-angle of the diffuser section is 6-8 degrees. An inlet pipe 3 for inputting carbon dioxide gas is connected to the Venturi ejector 2. This inlet pipe 3 is vertically connected to the throat section, ensuring that carbon dioxide gas can be forcefully drawn in by the instantaneous negative pressure generated by the high-speed flowing lime slurry.

[0028] The Venturi injector 2 operates based on the Venturi effect. When lime slurry enters the throat at high speed through the contraction section, the flow velocity increases dramatically. According to Bernoulli's law, the hydrostatic pressure decreases significantly, generating a strong instantaneous negative pressure in the contraction section of the throat. This negative pressure forcefully draws carbon dioxide gas into the throat through the inlet pipe 3. Subsequently, in the diffusion section, the fluid velocity gradually decreases, the pressure recovers, and strong turbulence and shearing are generated, tearing the drawn-in carbon dioxide gas into micron-sized fine bubbles, achieving microscopic mixing of lime slurry and carbon dioxide. Compared to traditional bubbling or mechanical stirring methods, this microscopic mixing technology can achieve uniform dispersion of reactants within milliseconds, avoiding uneven nucleation caused by excessively high local concentrations.

[0029] In the nucleation induction unit, lime slurry enters the vertical cylindrical tank 1 through a tangential inlet at a speed of 2-5 m / s, forming a stable rotating vortex within the tank. This tangential feeding method not only enhances the mixing effect but also prolongs the residence time of the reactants within the tank. The residence time is controlled within the range of 5-30 seconds, preferably 10-20 seconds. During this time period, the nucleation process of calcium carbonate mainly occurs, at which point the supersaturation is highest, and the nucleation rate is much greater than the crystal growth rate, ensuring the formation of a large number of tiny and uniform crystal nuclei. The throat contraction ratio (the ratio of the throat diameter to the inlet diameter) of the Venturi injector 2 is designed to be 0.3-0.6, and the diffusion angle (the half-angle of the diffusion section) is designed to be 6-8 degrees, achieving optimal gas-liquid micro-mixing.

[0030] The crystal growth unit, as the second-stage reactor, functions to provide a stable and uniform growth environment for the tiny crystal nuclei emerging from the nucleation induction unit, inhibiting secondary nucleation while promoting orderly growth of the nuclei in a predetermined direction. The inlet of this unit is fluidly connected to the outlet of the nucleation induction unit via a DN100 stainless steel pipe, which contains an online mixer for further homogenizing the slurry. The crystal growth unit includes a horizontal tank 4 with a circular cross-section. The tank is made of 316L stainless steel, and its inner surface is polished to Ra0.4 micrometers to reduce particle adhesion.

[0031] The core design feature of the horizontal tank 4 is that its interior is divided into multiple series compartments by several highly staggered overflow baffles 5. This design forces the slurry to advance in an upward-downward piston flow, effectively avoiding the back-mixing phenomenon present in traditional continuous stirred tanks. The number of overflow baffles 5 is 8-12, preferably 10, and each baffle is made of 316L stainless steel. The staggered design of the baffles is a key technical point: odd-numbered baffles extend upward from the bottom of the tank to a distance of 100-200 mm from the liquid surface, and even-numbered baffles extend downward from the top of the tank to a distance of 100-200 mm from the bottom. This staggered arrangement forces the slurry to complete a full upward-downward cycle in each compartment before entering the next compartment.

[0032] Each tandem compartment is equipped with a low-shear agitator. The sole purpose of these agitators is to maintain the suspension of crystal particles and prevent violent collisions. The agitators employ a blade design, with a rotational speed controlled at 20-60 rpm and a blade-to-compartment gap of 15-30 mm. This low agitation intensity ensures that the crystal particles remain well suspended, preventing sedimentation and avoiding crystal breakage or deformation due to excessive shear force. The residence time in the crystal growth unit is 15-40 minutes, preferably 25-30 minutes. During this period, the crystal nuclei gradually grow, the supersaturation gradually decreases, and the growth rate dominates.

[0033] To maintain a stable growth environment, multiple distributed feed ports are installed at the top of the horizontal tank 4 along the slurry flow direction. These feed ports are connected to precision metering pumps to accurately add low-concentration reactant solutions to each series compartment. Programmed control maintains the supersaturation in each compartment at the optimal level for crystal growth. The number of feed ports is typically 4-8, evenly distributed above each compartment. The flow rate of each feed port can be independently adjusted, ranging from 0.1-2 liters per hour.

[0034] The crystal form control unit, as the third-stage reactor, primarily functions to perform final crystal form stabilization and surface modification treatment on the grown crystals, ensuring a stable crystal structure, smooth surface, and good dispersibility in the product. The inlet of this unit is fluidly connected to the outlet of the crystal growth unit via a DN80 stainless steel pipe. The crystal form control unit includes a vertical continuous stirred tank 6, which adopts a cylindrical design and is equipped with multiple layers of agitators, including a lower-layer propeller agitator for axial circulation, a middle-layer turbine agitator for radial mixing, and an upper-layer paddle agitator for surface renewal.

[0035] The vertical continuous stirred tank 6 is equipped with a heating and cooling jacket 7 on its outer wall. This jacket adopts a spiral coil design, and the heating and cooling jacket 7 precisely controls the aging temperature through a circulating medium. The circulating medium can be heat transfer oil, steam, or cooling water, selected according to process requirements. The temperature control accuracy can reach ±1 degree Celsius, and the aging temperature is generally controlled within the range of 40-80 degrees Celsius. At this temperature, calcium carbonate crystals can undergo slow lattice rearrangement and surface reconstruction, achieving crystal stabilization. The aging time is 20-60 minutes, precisely controlled through continuous feeding and discharging of the stirred tank.

[0036] To achieve precise process monitoring and automated control, the device also includes a sensor network and an actuator network. The sensor network forms the perception foundation of the entire intelligent control system and includes various types of sensors deployed at key locations in the nucleation unit, crystal growth unit, and crystal form control unit. Temperature sensors, using Pt100 platinum resistance thermometers, have a measurement accuracy of ±0.1 degrees Celsius and are deployed at the inlet, outlet, and key internal locations of each reaction unit, totaling 12-16 measurement points. pH sensors, using composite glass electrodes, have a measurement range of 6-12 pH and an accuracy of ±0.02 pH, and are mainly deployed at the inlet of each unit and inside the nucleation unit, totaling 4-6 measurement points. Conductivity sensors, using four electrodes, have a measurement range of 0-200 mS / cm and an accuracy of ±1%, used to monitor changes in ion concentration in the slurry, and are deployed at the inlet and outlet of each unit, totaling 6-8 measurement points. Flow sensors, using electromagnetic sensors, have a measurement accuracy of ±0.5% and are deployed on the main material pipelines, including lime slurry feed, CO2 intake, and various feed additions, totaling 8-12 measurement points.

[0037] An online image analyzer is deployed at the outlet of the crystal growth unit or in the crystal form control unit. This online image analyzer uses a high-resolution CCD camera, equipped with an LED light source and a flow cell, providing a continuous real-time video stream for monitoring dynamic changes in crystal morphology. The image acquisition frequency is 10-30 frames / second, with a resolution of no less than 1024×768 pixels. Image processing algorithms can analyze key quality parameters such as crystal particle size distribution, shape factor, and agglomeration degree in real time. The flow cell is designed with transparent quartz glass, and the internal flow channel cross-section is rectangular, 10 mm wide and 2 mm high, ensuring that crystal particles can pass through the imaging area in a single layer.

[0038] The actuator network comprises various actuators used for process control. All material handling utilizes low-shear screw pumps, characterized by minimal flow pulsation and slight shearing action, preventing mechanical damage to crystal particles. Each screw pump is equipped with a frequency converter, achieving precise flow regulation with an accuracy of ±2% by changing the motor speed. Piping for industrial media such as CO2, steam, and cooling water is equipped with high-precision electric regulating valves. These valves employ straight-through or angle valve designs, exhibiting equal percentage flow characteristics, a regulation accuracy of ±1%, and a response time of less than 30 seconds. Temperature control is achieved by adjusting the flow rate and temperature of the circulating medium in the heating / cooling jacket 7, utilizing a combination of three-way regulating valves and proportional regulating valves.

[0039] The working process of the device is as follows: First, the prepared lime slurry is pumped from the raw material tank to the crystal nucleation induction unit at a constant flow rate via a screw pump. The concentration of the lime slurry is generally 8-15%, and the temperature is controlled at 20-30 degrees Celsius. The lime slurry enters the vertical cylindrical tank 1 at high speed through the tangential inlet, forming a stable rotating vortex inside the tank. Simultaneously, CO2 gas is forcefully drawn in through the inlet pipe 3 under the negative pressure generated by the Venturi injector 2, and undergoes intense micro-mixing with the lime slurry in the throat and diffuser section. During this process, CO2 rapidly dissolves and reacts with Ca(OH)2 to form CaCO3. Due to the uniformity of mixing and the speed of reaction, a large number of tiny, uniform calcium carbonate crystal nuclei are formed within a residence time of 5-30 seconds.

[0040] The slurry containing numerous crystal nuclei then enters the crystal growth unit, flowing sequentially through each series-connected compartment within the horizontal tank 4 according to a designed piston flow path. Within each compartment, the crystal nuclei grow slowly in a relatively mild environment, with low-shear agitators maintaining particle suspension. Precisely controlled feeding via distributed feed ports ensures that the supersaturation in each compartment is maintained at an appropriate level, guaranteeing continuous crystal growth while preventing new nucleation. After a growth process of 15-40 minutes, the crystal particle size reaches the expected range, and the shape becomes nearly complete.

[0041] Finally, the grown crystals are placed in the vertical continuous stirred tank 6 of the crystal form control unit for aging. At a temperature of 40-80 degrees Celsius, through precise temperature control by the heating and cooling jacket 7, the crystals undergo slow lattice rearrangement and surface reconstruction, eliminating internal stress and stabilizing the crystal structure. After aging for 20-60 minutes, a food-grade calcium carbonate product with narrow particle size distribution, uniform crystal form, and good dispersibility is finally obtained.

[0042] The three-stage segmented reactor design achieves physical separation and independent optimization of the nucleation, growth, and aging processes, fundamentally solving the problem of mutual interference between processes in traditional methods. The product particle size distribution coefficient is reduced from 35-50% in traditional processes to 15-25%. The application of the Venturi injector 2 enables microscopic mixing of lime slurry and CO2, resulting in a more uniform nucleation process and a reduction of the standard deviation of crystal nucleus size by more than 60%. The multi-compartment piston flow design effectively suppresses backmixing, ensuring a stable and consistent crystal growth environment and preventing particle size broadening. Compared to batch processes, continuous production improves production efficiency, reduces energy consumption, and lowers labor costs. A comprehensive sensor network lays a solid foundation for subsequent intelligent control, and the application of online image analyzers enables real-time monitoring of product quality. The modular design facilitates factory prefabrication and rapid on-site installation, shortening the construction cycle and reducing maintenance costs.

[0043] Example 2: like Figure 9 As shown, this embodiment, based on the modular continuous production device provided in Embodiment 1, further provides a hierarchical predictive twin intelligent control method. This method achieves comprehensive predictive control of the continuous calcium carbonate production process from the macroscopic physicochemical field to the microscopic crystal morphology by constructing a dual predictive architecture of a reaction field twin model and a crystal evolution twin model. It fundamentally changes the passive mode of "correcting deviations after they occur" in traditional chemical process control, and transforms it into a feedforward intelligent control of "predicting deviations and actively guiding them".

[0044] This hierarchical predictive twin intelligent control method, based on the sensor and actuator networks deployed in Example 1, utilizes four core steps—multidimensional data synchronous acquisition, reaction field modeling, crystal evolution prediction, and collaborative predictive control—to construct a complete closed-loop control system from data perception to intelligent decision-making. The entire control system employs a distributed computing architecture, deploying edge computing nodes on-site for real-time data processing and rapid response control, and a high-performance computing cluster in the cloud for complex model training and long-term optimization.

[0045] The multi-dimensional data synchronous acquisition step serves as the sensing foundation of the entire intelligent control system. Through the sensor network deployed in Example 1, it synchronously acquires process parameter data characterizing the macroscopic state of the production process—temperature, pH, conductivity, and flow rate—in a high-precision and high-frequency manner, as well as real-time video streams from an online image analyzer characterizing the crystal morphology of the product's microscopic state. The acquisition frequency of the process parameter data is set to 1-10Hz to ensure the capture of rapid changes during the calcium carbonate reaction. Specifically, temperature sensors acquire temperature data from 12-16 measuring points within each reaction unit, with an accuracy of ±0.1 degrees Celsius; pH sensors acquire pH data from 4-6 key locations, with an accuracy of ±0.02 pH; conductivity sensors acquire ion concentration change data from 6-8 measuring points, with an accuracy of ±1%; and flow sensors acquire flow data from 8-12 pipelines, with an accuracy of ±0.5%. This sensor data is transmitted in real-time to the edge computing node via a fieldbus system, forming a multi-dimensional process state vector.

[0046] Simultaneously, the online image analyzer continuously acquires real-time video streams of crystal morphology at a frequency of 10-30 frames per second, with each frame having a resolution of no less than 1024×768 pixels. The image data is transmitted to the image processing server via high-speed Ethernet. To ensure data synchronization, all sensors and image devices employ a unified timestamp system, achieving millisecond-level time synchronization accuracy. The acquired multidimensional data undergoes preprocessing, including data cleaning, noise filtering, and missing value imputation, to form a standardized data format for subsequent use in the twin model.

[0047] The reaction field modeling step achieves accurate prediction of the invisible physicochemical fields inside the reactor by constructing a reaction field twin model. This step inputs the collected process parameter data into the reaction field twin model, which, based on the calcium carbonate reaction mechanism, divides the entire reaction system into three coupled physical subdomains: a nucleation field, a growth field, and an aging field. Each subdomain corresponds to the three-stage reactor in Example 1. The reaction field twin model employs a field separation attention mechanism to model the spatiotemporal evolution of calcium ion enrichment, carbonate diffusion, and crystal nucleation, achieving differentiated prediction of the physical fields in different regions inside the reactor through localized attention windows.

[0048] The core algorithm of the reaction field Siamese model is based on an improved Transformer architecture, where the mathematical expression for the field separation attention mechanism is: in, Let these represent the query matrix, key matrix, and value matrix of the i-th field, respectively. Let be the dimension of the key vector. These correspond to the three fields of nucleation, growth, and aging, respectively. The physical quantities in each field include, but are not limited to, temperature distribution. Concentration distribution Flow velocity distribution etc., among which t represents the spatial coordinates, and t represents the time coordinates.

[0049] The localized attention window mechanism improves computational efficiency by limiting the spatial scope of attention computation. Its window function is defined as: in, The coordinates of the window center The window radius is typically set to 10%-20% of the reactor's characteristic dimensions. Through this localization mechanism, the model can efficiently handle the evolution of physical fields in three-dimensional space and predict the state changes of each field within the next 5-10 minutes.

[0050] The crystal evolution prediction step achieves accurate prediction of the future evolution of the product's crystal microstructure by constructing a crystal evolution twin model. This step inputs real-time video streams of crystal morphology from an online image analyzer into the crystal evolution twin model. This model constructs a dual-path prediction architecture based on the principle of crystal growth kinetics. One path predicts the overall morphological evolution trajectory of the crystal group, while the other path identifies and tracks the growth hotspots of individual crystals. Through a growth kinetics sensing mechanism, it automatically identifies active crystal regions undergoing lattice rearrangement, surface adsorption, and agglomeration.

[0051] The first path of the dual-path prediction architecture employs a sequence prediction model based on a recurrent neural network to predict the overall statistical properties of the crystal population, including macroscopic morphological parameters such as grain size distribution, shape factor distribution, and aggregation degree. The loss function for this path is defined as: ; Where N is the number of time steps, Let these represent the particle size distribution, shape factor, and aggregation degree at the nth time step, respectively, with superscripts indicating the particle size distribution, shape factor, and aggregation degree. and These represent the predicted value and the actual value, respectively. The weighting coefficient is usually set to 1. .

[0052] The second path of the dual-path prediction architecture employs a convolutional neural network-based target detection and tracking algorithm to identify and track the growth behavior of individual crystals. This path first extracts the crystal contour from video frames using a semantic segmentation algorithm, then tracks changes in the crystal boundaries using an optical flow algorithm. The growth dynamics perception mechanism identifies active growth regions by analyzing the displacement velocity and direction of the crystal boundaries. The criteria for determining active regions are as follows: in, Let be the velocity vector at the crystal boundary at position (x, y). This is the speed threshold, typically set to 0.1-0.5 pixels per frame. This represents the velocity divergence; a positive value indicates expansive growth.

[0053] The collaborative predictive control step is a crucial step in translating the predictions of the dual twin model into specific control actions. In this step, the model predictive controller receives the future states of the three subdomains predicted by the reaction field twin model and the crystal evolution trajectory predicted by the crystal evolution twin model. It constructs a field-crystal collaborative optimization objective function, adjusts the mass transfer rate between the fields through the inner controller to maintain an ideal concentration gradient distribution, and corrects the reaction driving force of each field based on the deviation of the crystal evolution trajectory through the outer controller. This achieves predictive control throughout the entire process, from shaping the reaction environment to guiding the product form.

[0054] The collaborative optimization objective function comprehensively considers three aspects: process stability, product quality, and energy consumption optimization. Its mathematical expression is: in, To predict the length of the time domain, Let the field prediction error be at the k-th time step. For crystal morphology prediction error, To control the increment, To control the input, These are the weighting coefficients. The inner controller obtains the optimal control sequence by solving this optimization problem. The control variables include the feed flow rate, temperature setpoint, and stirring speed of each reactor stage.

[0055] The outer layer controller corrects the setpoint of the inner layer controller based on the deviation information of the crystal evolution trajectory. The correction algorithm is as follows: in, For the corrected control input, For the inner controller output, To correct the gain matrix, This is due to deviations in the crystal evolution trajectory.

[0056] In the crystal evolution prediction step, the local supersaturation gradient distribution predicted in the reaction field modeling step is further used as the driving force input of the crystal evolution twin model. Through field-crystal coupling prediction, integrated modeling of the macroscopic mass transfer environment and microscopic crystal growth kinetics is achieved. The calculation of local supersaturation is based on solubility theory and the mass transfer equation, and its expression is: in, This is localized oversaturation. These represent the local concentrations of calcium ions and carbonate ions, respectively. This is the temperature-dependent solubility product constant. The supersaturation gradient is obtained through spatial differentiation: This gradient information is added as an additional input channel to the crystal evolution twin model, enabling the model to more accurately predict the growth behavior of crystals under different local environments.

[0057] To optimize computational resource allocation, the system employs a growth activity adaptive scheduler. When the proportion of active regions in the crystal exceeds a preset threshold, the prediction focus is automatically switched to optimize resource allocation. The formula for calculating the proportion of active regions is: in, The percentage of active area. For the crystal region mask. When When the scheduler reduces the allocation of computational resources for group path prediction, it improves the accuracy of individual path tracking; when If the threshold is not met, then the adjustment is reversed. It is usually set to 0.3-0.5.

[0058] The entire control system has an execution cycle of 100-500 milliseconds, ensuring timely response to process changes. The control system is also equipped with fault diagnosis and safety protection functions. When a sensor malfunction or abnormal model prediction is detected, it automatically switches to backup sensor or traditional PID control mode to ensure the safe and stable production process.

[0059] Through collaborative prediction using dual twin models, comprehensive monitoring and forecasting of the calcium carbonate production process are achieved, with prediction accuracy improved by 30-50% compared to traditional methods. Secondly, the feedforward control strategy proactively adjusts before quality deviations occur, improving product quality stability and reducing the defect rate to below 1%. Thirdly, field separation modeling technology allows for independent optimization of different reaction regions, improving overall reaction efficiency. Fourthly, the growth activity adaptive scheduler effectively reduces computational complexity, lowering computational costs while maintaining prediction accuracy. Fifthly, the field-crystal coupling prediction mechanism achieves seamless connection between macroscopic and microscopic scales, providing a new technical path for multi-scale modeling of chemical processes.

[0060] Example 3: like Figure 9As shown, this embodiment, based on the modular continuous production device provided in Embodiment 1 and the hierarchical predictive twin intelligent control method provided in Embodiment 2, further provides advanced control functions with refined modeling and adaptive optimization. Through the spatial token partitioning mechanism characterized by chemical process, the sequence modeling method guided by growth kinetics, the field-driven feedforward control strategy, and the collaborative diagnosis and adaptive optimization technology, it achieves the ultimate refined control and intelligent optimization of the continuous production process of food-grade calcium carbonate, reaching the international advanced level of intelligent manufacturing.

[0061] The core of this embodiment lies in the in-depth optimization of the dual twin model and the comprehensive upgrade of the control strategy in Embodiment 2. The reaction field twin model adopts a spatial token partitioning mechanism characterized by chemical processes. Based on the characteristics of the calcium carbonate reaction, the reactor space is divided into sub-regions with different flow characteristics, including a turbulent mixing zone, a laminar mass transfer zone, and a wall boundary zone. Each sub-region uses an independent physical quantity encoding channel. Through a cross-regional attention mechanism, the mass and energy transfer between different regions is modeled, achieving refined prediction of the complex flow field of the chemical reactor. This spatial token partitioning mechanism breaks through the limitations of traditional uniform grid partitioning and can adaptively divide the space according to the actual fluid dynamics characteristics and mass and heat transfer laws.

[0062] Spatial tokenization for chemical process characterization is primarily based on the spatial distribution of dimensionless numbers such as Reynolds number and Schmidt number to identify different flow and mass transfer regions. The turbulent mixing region is mainly located near the outlet of Venturi ejector 2 and at the feed inlets of each reactor stage. This region is characterized by a Reynolds number (Re) > 4000, high mixing intensity, and high mass transfer rate. The laminar mass transfer region is mainly located within the series compartments of the crystal growth unit. This region is characterized by a Re < 2300, stable flow, and mass transfer mainly relying on molecular diffusion. The wall boundary region refers to the area within a certain distance from the reactor wall. This region is significantly affected by wall effects, exhibiting large temperature and concentration gradients.

[0063] The mathematical expression for space token partitioning is: in, Let Re(x,y,z) be the token type corresponding to the spatial location (x,y,z), and let Re(x,y,z) be the local Reynolds number. It is a local velocity vector. Let be the norm of the velocity gradient tensor. The threshold for determining turbulence is typically set to a value of [value to be filled in]. The distance to the nearest wall. For boundary layer thickness, These represent four token types: turbulent, laminar, boundary, and transition.

[0064] Each sub-region uses an independent physical quantity encoding channel for feature extraction, and the encoding process is implemented through a multilayer perceptron network. in, This is the region encoding vector for the l-th layer. These are the weight matrix and the bias vector, respectively. The activation function is typically ReLU or GELU. Different regions have different network parameters in their encoding channels, enabling the model to learn the unique physicochemical properties of each region.

[0065] Cross-regional attention mechanisms are used to model the transfer of matter and energy between different regions. The core algorithm is based on graph attention networks, treating different regions in space as nodes in a graph. The adjacency relationships between regions are determined by physical connectivity and transfer strength. The formula for calculating the cross-regional attention weights is: in, Let be the attention weight of region i to region j. Let i and j be the feature vectors of regions i and j, respectively. For a shared linear transformation matrix, Let | be the attention parameter vector, where | denotes the vector concatenation operation. Let be the set of neighboring regions of region i. The updated region feature vector is: ; in, The value transformation matrix, This is the updated feature vector for region i.

[0066] Based on Example 2, the crystal evolution twin model further adopts a growth kinetics-guided sequence modeling method. It constructs a stage-aware video prediction model based on the multi-stage evolution mechanism of calcium carbonate crystal nucleation-growth-agglomeration. The model identifies the location of new crystal nuclei through a motion detector during the nucleation stage, monitors crystal boundary expansion through a deformation tracker during the growth stage, and warns of crystal collisions and aggregation through a contact detector during the agglomeration stage, thereby achieving accurate prediction of the entire crystal life cycle.

[0067] Growth kinetics-guided sequence modeling first trains three specialized detectors simultaneously using a multi-task learning framework. The nucleation phase motion detector, based on optical flow analysis and background subtraction techniques, identifies newly appearing moving objects in video frames. Its detection algorithm is as follows: in, Let be the nucleation probability of position (x, y) at time t. The optical flow velocity vector, The minimum motion speed threshold is typically 0.5-1.0 pixels per frame. For background difference intensity, The intensity threshold, The size of the detected object. This is the upper limit for the core size, typically 5-10 pixels.

[0068] The growth-stage deformation tracker employs a deep learning-based target tracking algorithm, combined with the physical constraints of crystal growth, to monitor the expansion process of the crystal boundary. The loss function for deformation tracking is designed as follows: in, This is the contour matching loss, used to ensure tracking accuracy; The loss is due to physical constraint, based on the thermodynamic and kinetic laws of crystal growth; To compensate for temporal consistency loss, we ensure tracking continuity between adjacent frames. The weighting coefficient is usually set to 1. .

[0069] Physical constraint loss The specific expression is: ; in Let be the area of ​​the i-th crystal. This is a function of the crystal growth rate. Let represent the supersaturation, temperature, and pH value surrounding the i-th crystal, respectively. The growth rate function is based on classical crystal growth theory: in, This is a temperature- and pH-dependent growth rate constant. The supersaturation driving force is n, which is the growth index, typically ranging from 1.5 to 2.5.

[0070] Agglomeration-stage contact detectors provide early warning of potential aggregation events by analyzing changes in the distance and contact area between crystals. The contact detection algorithm is based on the minimum bounding rectangle and convex hull algorithms. in, Let be the minimum distance between crystals i and j at time t. These are the contour point sets of two crystals, respectively. When And relative motion speed At that time, it was determined to be a potential family reunion event, among which This is a distance threshold, typically 1-3 pixels. This is the unit normal vector in the contact direction.

[0071] The collaborative predictive control step further adopts a field-driven feedforward control strategy. When the crystal evolution twin model predicts that the crystal evolution trajectory deviates from the target morphology, the model predictive controller calculates the optimal inter-field driving force redistribution scheme based on the mass transfer constraints provided by the reaction field twin model. By adjusting the supersaturation peak of the nucleation field, the concentration gradient distribution of the growth field, and the temperature stability of the aging field, the active correction of the crystal evolution path is achieved.

[0072] The core of the field-driven feedforward control strategy is to construct a multi-objective optimization problem, in which the objective function comprehensively considers crystal quality deviation, process stability, and energy consumption optimization. ; in, Let be the objective function of the field-driven control. This is the crystal quality deviation vector. This is the field state deviation vector. To control the increment, This is the weight matrix. Let be the field state vector. For the system matrix, These are process noise and measurement noise, respectively.

[0073] The redistribution of driving forces between fields is achieved by solving the above optimization problem to obtain the optimal control input for each field. The peak supersaturation control of the nucleation field is achieved by adjusting the lime slurry flow rate and the CO2 inlet flow rate. in, The target supersaturation of the nucleation field, For reference oversaturation, To control the gain, This refers to particle size deviation.

[0074] Concentration gradient distribution in the growth field is controlled by adjusting the feed flow rate of each compartment: in, The target concentration gradient in the growth field. As a reference concentration gradient, For the control gain matrix, This is the shape deviation vector.

[0075] Temperature stability control in the aging field is achieved by adjusting the flow rate and temperature of the circulating medium in the heating and cooling jacket 7. in, The target temperature for the aging field, For reference temperature, To control the gain, This is due to crystal form deviation.

[0076] This embodiment also includes a field-crystal co-diagnosis and adaptive optimization step. By continuously comparing the prediction results of the reaction field twin model and the crystal evolution twin model with the actual data collected by the sensor network, when the reaction field prediction is accurate but the crystal evolution prediction is inaccurate, it is determined to be a deviation of the crystal growth kinetic model. When the reaction field prediction is inaccurate, it is determined to be a deviation of the mass transfer or heat transfer model. Based on the deviation type, the coupling parameters between the fields and the crystal growth kinetic parameters are calibrated in a targeted manner through an online learning algorithm.

[0077] The collaborative diagnostic mechanism employs multivariate statistical analysis to determine the type of deviation by calculating the statistical distance between predicted and measured values. Diagnostic indicators include root mean square error. Mean Absolute Percentage Error (MAPE) and Correlation Coefficient : ; ; ; in, These are the predicted value and the measured value, respectively. is the mean of the measured values, and N is the sample size.

[0078] The rules for determining the type of deviation are as follows: in, For the error threshold, The correlation coefficient threshold is usually set to 1. The adaptive optimization algorithm, based on online learning and parameter estimation techniques, employs a combination of recursive least squares and Kalman filters for parameter updates. For inter-field coupling parameters, the update algorithm is as follows: in, The estimated value of the coupling parameters at step k. Let P(k) be the regression vector and P(k) be the covariance matrix. These are the observed values.

[0079] For crystal growth kinetic parameters, an extended Kalman filter is used for updating: in, Here are the updated crystal parameter estimates, and z(k) is the observation vector. Let H(k) be the observation function, H(k) be the Jacobian matrix, R(k) be the observation noise covariance matrix, and K(k) be the Kalman gain.

[0080] This advanced intelligent optimization control system significantly improves the control precision and intelligence level of continuous production of food-grade calcium carbonate. Through a spatial token partitioning mechanism characterized by chemical process features, the prediction accuracy of the reaction field is improved by 25-35%, accurately capturing local features in complex flow fields. The growth kinetics-guided sequence modeling method achieves a second-level temporal resolution for crystal evolution prediction, improving prediction accuracy by 40-50%. The field-driven feedforward control strategy shortens the control response time to less than 100 milliseconds, reducing product quality fluctuations by 60-70%. Collaborative diagnostics and adaptive optimization functions enable the system to possess self-learning and self-evolution capabilities, achieving an online calibration accuracy of over 95% for model parameters and significantly improving long-term system stability. This system provides a complete technical solution for the digital transformation and intelligent manufacturing of chemical processes, possessing significant engineering application value and broad industrialization prospects.

[0081] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A continuous production apparatus for food-grade calcium carbonate, characterized in that, The device comprises, in sequence according to the slurry flow order: The crystal nucleus induction unit includes a vertical cylindrical tank (1), on which a Venturi injector (2) is installed. The Venturi injector (2) is connected to an air inlet pipe (3) for inputting carbon dioxide gas. The Venturi injector (2) uses the instantaneous negative pressure generated by the high-speed flowing lime milk in the throat contraction section to forcefully draw in carbon dioxide gas and tear it into bubbles in the diffusion section, thereby realizing the micro-mixing of reactants. The crystal growth unit, whose inlet is in fluid communication with the outlet of the crystal nucleus induction unit, includes a horizontal tank (4). The interior of the horizontal tank (4) is divided into multiple series compartments by multiple highly staggered overflow baffles (5), which force the slurry to move forward in a piston flow manner of upward and downward flow. The series compartments are equipped with low-shear stirring paddles to maintain the suspension of crystal particles and avoid violent collisions. The crystal form control unit, whose inlet is fluidly connected to the outlet of the crystal growth unit, includes a vertical continuous stirring tank (6). The outer wall of the vertical continuous stirring tank (6) is provided with a heating and cooling jacket (7). The aging temperature is controlled by circulating medium to achieve crystal form stabilization and surface modification.

2. The apparatus according to claim 1, characterized in that, The device also includes a sensor network and an actuator network. The sensor network includes temperature sensors, pH sensors, conductivity sensors, and flow sensors deployed at key locations in the crystal nucleation unit, crystal growth unit, and crystal form control unit, as well as an online image analyzer deployed at the outlet of the crystal growth unit or the crystal form control unit. The online image analyzer provides a continuous real-time video stream for monitoring dynamic changes in crystal morphology.

3. The apparatus according to claim 1, characterized in that, In the crystal nucleus induction unit, lime slurry enters the vertical cylindrical tank (1) through the tangential inlet to form a rotating vortex with a residence time of 5-30 seconds. The throat contraction ratio and diffusion angle of the Venturi injector (2) are optimized to achieve micro-mixing of gas and liquid.

4. The apparatus according to claim 1, characterized in that, The residence time of the crystal growth unit is 15-40 minutes, the length of the horizontal tank (4) is 3-5 meters, the number of overflow baffles (5) is 8-12, and multiple distributed feed ports are provided on the top of the horizontal tank (4) along the slurry flow direction for adding low concentrations of reactants to the series compartments to maintain a stable supersaturation.

5. A smart control method for the device according to claim 1, characterized in that, Includes the following steps: Multidimensional data synchronous acquisition steps: Synchronously acquire process parameter data, such as temperature, pH value, conductivity, and flow rate, which characterize the macroscopic state of the production process, as well as real-time video streams of crystal morphology from the online image analyzer, which characterize the microscopic state of the product, through a sensor network; Reaction field modeling steps: Input the process parameter data into the reaction field twin model. The reaction field twin model divides the calcium carbonate reaction system into three mutually coupled physical subdomains: nucleation field, growth field, and aging field. The field separation attention mechanism is used to model the spatiotemporal evolution process of calcium ion enrichment, carbonate diffusion, and crystal nucleation. The localized attention window is used to achieve differentiated prediction of the physical field in different regions inside the reactor. Crystal evolution prediction steps: The real-time video stream of the crystal morphology is input into the crystal evolution twin model. The crystal evolution twin model is constructed based on the principle of crystal growth dynamics to build a dual-path prediction architecture. One path predicts the overall morphological evolution trajectory of the crystal group, and the other path identifies and tracks the growth hotspots of individual crystals. The active crystal regions that are undergoing lattice rearrangement, surface adsorption, and agglomeration are automatically identified through the growth dynamics perception mechanism. Collaborative predictive control steps: The model predictive controller receives the future states of the three subdomains predicted by the reaction field twin model and the crystal evolution trajectory predicted by the crystal evolution twin model, constructs a field-crystal collaborative optimization objective function, adjusts the mass transfer rate between each field through the inner controller to maintain the ideal concentration gradient distribution, and corrects the reaction driving force of each field according to the deviation of the crystal evolution trajectory through the outer controller, so as to realize the whole-process predictive control from the shaping of the reaction environment to the guidance of product form.

6. The method according to claim 5, characterized in that, In the crystal evolution prediction step, the local supersaturation gradient distribution predicted in the reaction field modeling step is also used as the driving force input of the crystal evolution twin model. The integrated modeling of macroscopic mass transfer environment and microscopic crystal growth dynamics is realized through field-crystal coupling prediction. When the proportion of active crystal region exceeds a preset threshold, the prediction focus is automatically switched to optimize the allocation of computing resources.

7. The method according to claim 5, characterized in that, The reaction field twin model adopts a spatial token partitioning mechanism characterized by chemical processes. Based on the characteristics of calcium carbonate reaction, the reactor space is divided into sub-regions with different flow characteristics, such as turbulent mixing zone, laminar mass transfer zone, and wall boundary zone. Each sub-region adopts an independent physical quantity encoding channel. Through a cross-region attention mechanism, the material and energy transfer between different regions is modeled to achieve a refined prediction of the complex flow field of the chemical reactor.

8. The method according to claim 5, characterized in that, The crystal evolution twin model adopts a growth kinetics-guided sequence modeling method. Based on the multi-stage evolution mechanism of calcium carbonate crystal nucleation-growth-agglomeration, a stage-aware video prediction model is constructed. The model identifies the location of new crystal nuclei through a motion detector during the nucleation stage, monitors crystal boundary expansion through a deformation tracker during the growth stage, and warns of crystal collisions and aggregation through a contact detector during the agglomeration stage, thereby achieving accurate prediction of the entire crystal life cycle.

9. The method according to claim 5, characterized in that, The collaborative predictive control step adopts a field-driven feedforward control strategy. When the crystal evolution twin model predicts that the crystal evolution trajectory deviates from the target morphology, the model predictive controller calculates the optimal inter-field driving force redistribution scheme based on the mass transfer constraints provided by the reaction field twin model. It actively corrects the crystal evolution path by adjusting the supersaturation peak of the nucleation field, the concentration gradient distribution of the growth field, and the temperature stability of the aging field.

10. The method according to claim 5, characterized in that, The method further includes a field-crystal co-diagnosis and adaptive optimization step. By continuously comparing the prediction results of the reaction field twin model and the crystal evolution twin model with the actual data collected by the sensor network, when the reaction field prediction is accurate but the crystal evolution prediction is inaccurate, it is determined to be a deviation of the crystal growth kinetic model. When the reaction field prediction is inaccurate, it is determined to be a deviation of the mass transfer or heat transfer model. Based on the deviation type, the coupling parameters between the fields and the crystal growth kinetic parameters are calibrated in a targeted manner through an online learning algorithm.

Images