A method and system for optimizing process parameters for preparing high-purity lanthanum fluoride based on machine learning

Abstract

The application discloses a kind of high-purity lanthanum fluoride preparation process parameter optimization method and system based on machine learning, and relates to the technical field of preparation process parameter optimization. It includes: at least two in-situ fiber-optic spectroscopy probes are arranged in the preparation process of lanthanum fluoride by arrangement method, original spectral data is collected, and original spectral data set is obtained;Data preprocessing is carried out based on original spectral data set.The application realizes the global dynamic synergic regulation and control of multiple process parameters by real-time acquisition of multi-dimensional feature data of precipitation and calcination process with double in-situ fiber-optic spectroscopy probes, combined with deep deterministic policy gradient reinforcement learning model, breaks the limitation of traditional single-process optimization, completely solves the problem that traditional method cannot consider multiple processes and multiple index optimization, and intelligently binds the impurity characteristic parameters of lanthanum source and the time sequence of production data, so that the agent can identify the influence of impurity fluctuations such as calcium and carbonate on production in real time, and automatically adjust process parameters.

Description

Technical Field

[0001] This invention relates to the field of process parameter optimization technology, specifically to a method and system for optimizing process parameters in the preparation of high-purity lanthanum fluoride based on machine learning. Background Technology

[0002] Lanthanum fluoride, also known as lanthanum trifluoride, is a fluoride of the rare earth element lanthanum. It is a widely used inorganic functional material, a white crystalline powder at room temperature, almost insoluble in water and sparingly soluble in dilute acids. Common preparation methods include precipitation, where a soluble lanthanum salt reacts with a fluoride to form a sparingly soluble lanthanum fluoride precipitate, which is then filtered, washed, and calcined at high temperature to obtain the pure product. Alternatively, it can be prepared by directly reacting lanthanum oxide with hydrogen fluoride gas at high temperatures. High-purity lanthanum fluoride typically refers to products with a purity ≥99.99% and strictly controlled impurity content. These are often used in high-end applications with extremely high material performance requirements. Optimizing the preparation process parameters involves systematically adjusting and matching the controllable conditions throughout the entire preparation process to meet these high performance requirements, ultimately achieving the goal of meeting product performance standards and optimizing production efficiency / cost. Therefore, the preparation process requires highly precise controllable parameters.

[0003] The patent publication number CN120595750A discloses a method and system for optimizing parameters in organosilicon preparation. This method utilizes a thermally integrated DCS system for organosilicon distillation to acquire thermal balance parameters and load information. The load information represents the material information in the organosilicon distillation system during the current time period. Based on the thermal balance parameters, it determines a first type of parameter for outer-layer optimization and a second type of parameter for inner-layer optimization during distillation system control optimization. Based on the load information, it determines the optimization range of the first type of parameter. In other words, the range of the first type of parameter is limited according to different load information. Based on the optimization range and the second type of parameter, a parallel outer-layer optimization strategy is used to obtain optimization results for achieving inner-layer optimization. Based on the inner-layer optimization results, the optimization result with the lowest energy consumption is selected as the target optimization result for inner-layer optimization. The second type of parameter corresponding to the target optimization result is sent to the DCS system for real-time optimal control of the organosilicon distillation system. In this way, two types of parameters can be optimized hierarchically, breaking down the complex multi-dimensional problem and improving optimization efficiency.

[0004] When preparing lanthanum fluoride, the above-mentioned and similar technical solutions have a strong nonlinear coupling relationship between the core process parameters in the entire process of liquid-phase precipitation and gas-phase roasting. Existing process parameter optimization methods can only perform static fitting for single processes and single indicators, and cannot adapt to the coupling relationship between core parameters, nor can they dynamically respond to the fluctuation of trace impurities in different batches of lanthanum source. As a result, the synchronous compliance rate of core performance indicators during mass production is insufficient, and it cannot meet the mass production requirements of deep ultraviolet fluoride optical fibers and scintillation crystals. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for optimizing process parameters in the preparation of high-purity lanthanum fluoride based on machine learning, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for optimizing process parameters in the preparation of high-purity lanthanum fluoride based on machine learning, comprising:

[0007] At least two in-situ fiber optic spectral probes were arranged in the lanthanum fluoride preparation process to collect raw spectral data and obtain the raw spectral dataset.

[0008] Data preprocessing is performed on the original spectral dataset to obtain a processed dataset. Based on the processed dataset, the preset characteristic peaks are quantitatively transformed to obtain a set of characteristic parameters.

[0009] Obtain process parameters synchronized with the timestamp of the feature parameter set, and simultaneously obtain the raw material impurity feature parameters of the target batch of lanthanum source to obtain a benchmark dataset;

[0010] A deep deterministic policy gradient reinforcement learning agent framework is built based on the benchmark dataset. The action space of the framework is the adjustment range of process parameters synchronized with the timestamp of the feature parameter set, which is the action space term. The reward function is constructed with the core performance compliance rate of high-purity lanthanum fluoride as the highest weight, combined with production energy consumption and production cycle weighting, which is the learning model term.

[0011] A digital twin model for lanthanum fluoride preparation driven by hybrid mechanism data was constructed. Based on the learning model terms, interactive pre-training was performed to obtain a general regulatory agent. At the same time, actual production data of the target production line was collected, and the general regulatory agent was fine-tuned by small sample transfer to obtain a dedicated regulatory agent adapted to the target production line.

[0012] A sampling threshold is set, which is a fixed time value. The original spectral dataset is collected based on the sampling threshold and input into a dedicated control agent. The agent outputs process parameter adjustment instructions to obtain an output instruction set. Parameter optimization is performed based on the output instruction set, thereby realizing the adaptive adjustment of continuous process parameters and closed-loop dynamic optimization.

[0013] Furthermore, the arrangement method includes:

[0014] A first in-situ fiber optic spectroscopy probe was deployed on the side wall of the liquid-phase precipitation reactor for lanthanum fluoride preparation to collect liquid-phase reaction spectral data of the precipitation process.

[0015] A second in-situ fiber optic spectral probe is deployed in the gas outlet pipe of the gas phase roasting furnace to collect the spectral data of the gas phase products in the roasting process and obtain the original spectral dataset.

[0016] Furthermore, the method for obtaining the processed dataset includes:

[0017] The real-time collected air background spectrum is obtained to obtain background removal items. The background removal items are then removed based on the original spectral dataset to obtain the first processed data.

[0018] Based on the first processed data, a filter window value is set, and the first processed dataset is smoothed and denoised to obtain the second processed data.

[0019] Based on the second processed data, the spectral baseline correction is completed using the adaptive iterative reweighted penalized least squares method, resulting in the processed dataset.

[0020] Furthermore, the method for obtaining the feature parameter set includes:

[0021] A calibration curve is pre-constructed, and a gradient value is set. The gradient values ​​are fluoride ion standard solutions with different gradient concentrations, hexagonal lanthanum fluoride standard suspensions with different crystal phase ratios, and mixed standard gases with different concentrations.

[0022] Based on the original spectral dataset, the spectra of each standard sample were collected, and a calibration curve was fitted with the characteristic peak absorbance as the independent variable and the standard concentration and crystal phase ratio as the dependent variables.

[0023] Extract the absorbance of the corresponding preset feature peaks in the processed dataset and substitute them into the pre-constructed calibration curve to obtain the feature parameter set.

[0024] Furthermore, the benchmark dataset includes:

[0025] The set of characteristic parameters includes the fluoride ion concentration in the precipitation system, the hexagonal lanthanum fluoride crystal phase conversion rate, the total amount of gaseous oxygen in the calcination system, and the residual HF concentration in the calcination tail gas.

[0026] The process dataset includes the lanthanum source feeding rate, reaction system pH value, HF inlet flow rate during calcination, calcination heating rate, precipitation aging time, and calcination holding temperature.

[0027] The impurity dataset includes the calcium impurity content, carbonate content, and free water content of the target batch of lanthanum source.

[0028] Furthermore, the action space term includes:

[0029] Set a first adjustment range value, which is the lanthanum source feeding rate adjustment range; set a second adjustment range value, which is the HF inlet flow rate adjustment range; set a third adjustment range value, which is the calcination heating rate adjustment range; set a fourth adjustment range value, which is the reaction system pH adjustment range; set a fifth adjustment range value, which is the precipitation aging time adjustment range; and set a sixth adjustment range value, which is the calcination holding temperature adjustment range.

[0030] Furthermore, the method for obtaining the learning model terms includes:

[0031] Define the network structure of the deep deterministic policy gradient reinforcement learning agent framework, including the number of fully connected network layers and the number of input dimensions, with the number of input dimensions corresponding to the total dimensions of the target dataset.

[0032] Define the output dimension, which corresponds to the process parameters synchronized with the feature parameter set timestamp. Define the activation functions for the hidden layer and the output layer to obtain the activation function settings.

[0033] Furthermore, the method for acquiring the general-purpose regulatory agent includes:

[0034] By integrating the enthalpy of lanthanum fluoride reaction, fluorine vacancy formation energy, and impurity diffusion barrier parameters obtained from first-principles calculations, and acquiring corresponding data on historical production process performance, a hybrid driving model is constructed using a regression model. Simultaneously, interactive pre-training is performed based on the learning model terms to obtain a general regulatory agent.

[0035] Furthermore, the method for obtaining the output instruction set includes:

[0036] Process parameter adjustment commands are sent to the production line PLC control system for automatic execution. After each batch of production is completed, the status, actions, rewards, and next status data of that batch are stored in the experience playback pool. Each time a fixed batch of data is stored, incremental model training is triggered.

[0037] A machine learning-based process parameter optimization system for the preparation of high-purity lanthanum fluoride uses the aforementioned machine learning-based process parameter optimization method for the preparation of high-purity lanthanum fluoride, comprising:

[0038] Data acquisition and processing module: At least two in-situ fiber optic spectral probes are arranged in the lanthanum fluoride preparation process to acquire raw spectral data and obtain raw spectral datasets. Data preprocessing is performed based on the raw spectral datasets to obtain processed datasets. Based on the processed datasets, preset characteristic peaks are quantitatively converted to obtain a set of characteristic parameters.

[0039] Benchmarking data acquisition module: Acquires process parameters synchronized with the feature parameter set timestamp, and simultaneously acquires the raw material impurity feature parameters of the target batch of lanthanum source to obtain the benchmarking dataset;

[0040] The learning framework construction module: Based on the benchmark dataset, a deep deterministic policy gradient reinforcement learning agent framework is built. The action space of the framework is the adjustment range of process parameters synchronized with the timestamp of the feature parameter set, which is the action space term. The reward function is constructed with the core performance achievement rate of high-purity lanthanum fluoride as the highest weight, combined with production energy consumption and production cycle weighting, which is the learning model term. A mechanism data hybrid-driven digital twin model of lanthanum fluoride preparation is constructed. Based on the learning model term, interactive pre-training is performed to obtain a general control agent. At the same time, actual production data of the target production line is collected, and small sample transfer fine-tuning is performed on the general control agent to obtain a dedicated control agent adapted to the target production line.

[0041] Command output optimization module: Set the acquisition threshold, which is a fixed time value. Based on the acquisition threshold, acquire the original spectral dataset and input it into the dedicated control agent. Output process parameter adjustment commands to obtain the output command set. Parameter optimization is performed based on the output command set.

[0042] Compared with the prior art, the beneficial effects of the present invention are:

[0043] This machine learning-based method and system for optimizing process parameters in the preparation of high-purity lanthanum fluoride uses dual in-situ fiber optic spectral probes to collect multi-dimensional feature data from precipitation and calcination processes in real time. Combined with a deep deterministic strategy gradient reinforcement learning model, it breaks through the limitations of traditional single-process optimization and achieves global dynamic collaborative control of multiple process parameters. This completely solves the problem that traditional methods cannot simultaneously optimize multiple processes and indicators. By binding the characteristic parameters of lanthanum source impurities with the production data time series, the intelligent agent can identify the impact of fluctuations in impurities such as calcium and carbonate on production in real time and automatically adjust process parameters. This effectively avoids the risk of product scrap caused by raw material impurities, and significantly improves the consistency of product performance during batch production.

[0044] Meanwhile, by weighting the core performance achievement rate, energy consumption and production cycle through the reward function, product quality is improved while production energy consumption is reduced and the production cycle is shortened. Furthermore, based on the mechanism-data hybrid driven digital twin model, only ≤5 batches of production line data are needed to complete the fine-tuning of the exclusive intelligent agent, which reduces the adaptation cost compared to traditional methods. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of the overall process of the present invention;

[0046] Figure 2 This is a schematic diagram of the data acquisition and data processing flow of the present invention;

[0047] Figure 3This is a schematic diagram illustrating the process of acquiring the benchmark dataset and constructing the intelligent agent framework in this invention;

[0048] Figure 4 This is a schematic diagram of the closed-loop optimization and incremental training process of the present invention;

[0049] Figure 5 This is a schematic diagram of the feature parameter set acquisition process of the present invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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.

[0051] Lanthanum fluoride is an important rare-earth fluoride with excellent optical properties, such as high transmittance, low dispersion, and low refractive index, making it promising for applications in deep ultraviolet (DUV) optical fibers and scintillation crystals. DUV optical fibers are core materials for next-generation fiber optic communication and laser technology, and lanthanum fluoride, as a key component, directly determines the performance and lifespan of the fiber. However, the preparation process of lanthanum fluoride is complex, involving multiple steps such as liquid-phase precipitation, filtration, washing, drying, and gas-phase calcination. Liquid-phase precipitation controls the size, morphology, and crystallinity of lanthanum fluoride nanoparticles, while gas-phase calcination affects the purity, defect density, and optical properties of the material. Due to the strong nonlinear coupling between the process parameters of these steps, such as precipitation temperature, reactant concentration, calcination temperature, atmosphere composition, and heating rate, existing process optimization methods mainly focus on static fitting of single steps and single indicators. For example, for the liquid-phase precipitation process, researchers adjust particle size and morphology by controlling the precipitation temperature or adding surfactants. For the gas-phase calcination process, they reduce impurity content and improve crystallinity by changing the calcination temperature and atmosphere. However, these methods neglect the coupling relationship between process parameters and cannot achieve overall performance optimization. Furthermore, different batches of lanthanum sources often exhibit fluctuations in trace impurities, which can significantly affect the crystallization and optical properties of lanthanum fluoride. Existing optimization methods struggle to dynamically respond to these fluctuations, resulting in poor product performance consistency during mass production. The technical solution provided in this application, however, utilizes dual in-situ fiber optic spectral probes to collect multi-dimensional feature data from the precipitation and calcination processes in real time. Combined with a deep deterministic strategy gradient reinforcement learning model, it overcomes the limitations of traditional single-process optimization, achieving global dynamic collaborative control of multiple process parameters. This completely solves the problem of traditional methods being unable to simultaneously address multiple processes and indicators. By binding lanthanum source impurity feature parameters to production data timelines, the intelligent agent can identify the impact of fluctuations in impurities such as calcium and carbonate on production in real time, automatically adjusting process parameters. This effectively avoids the risk of product scrapping due to raw material impurities, significantly improving product performance consistency during mass production. By weighting core performance achievement rate, energy consumption, and production cycle using a reward function, product quality is improved while reducing production energy consumption and shortening the production cycle. Furthermore, based on a mechanism-data hybrid-driven digital twin model, only ≤5 batches of production line data are needed to complete the dedicated intelligent agent fine-tuning, reducing adaptation costs compared to traditional methods. Figure 1 As shown, it includes steps S100-S700.

[0052] Step S100: Arrange at least two in-situ fiber optic spectral probes in the lanthanum fluoride preparation process using the arrangement method to collect raw spectral data and obtain the raw spectral dataset.

[0053] It is important to note that, such as Figure 2As shown, the arrangement method includes: deploying a first in-situ fiber optic spectral probe on the side wall of the liquid-phase precipitation reactor for lanthanum fluoride preparation to collect the liquid-phase reaction spectral data of the precipitation process; deploying a second in-situ fiber optic spectral probe on the gas outlet pipe of the gas-phase roasting furnace to collect the gas-phase product spectral data of the roasting process, thereby obtaining the original spectral dataset.

[0054] Specifically, the in-situ fiber optic spectral probes all employ sapphire-windowed fluoride fiber Fourier transform infrared probes resistant to hydrofluoric acid corrosion, with a spectral acquisition wavenumber range of 400 cm⁻¹. -1 -4500cm -1 The sampling interval is ≤1s. The first in-situ fiber optic spectroscopy probe is immersed 2-5cm below the surface of the reaction liquid to avoid the window being blocked by sediment particles at the bottom of the vessel. The installation point is sealed with a PTFE flange to prevent HF gas leakage. The pipeline for the second in-situ fiber optic spectroscopy probe installation section needs to be heated to 120℃ to prevent HF condensation from corroding the probe window. A 200-mesh PTFE filter element is installed at the front end to block the calcination dust from contaminating the optical path. The FTIR probe was chosen because it can simultaneously cover the characteristic absorption peaks of multiple components in the liquid and gas phases. The sapphire window and PTFE seal are compatible with the highly corrosive HF environment of lanthanum fluoride preparation. The heating and filtration design ensures long-term operational stability.

[0055] In the specific implementation process, on a mass production line of a 100L precipitation reactor and a 5m³ vacuum calcining furnace, the first in-situ fiber optic spectroscopy probe was installed on the side wall of the reactor 30cm from the bottom, and the second in-situ fiber optic spectroscopy probe was installed on the heat-insulated pipe 50cm from the furnace opening at the gas outlet of the calcining furnace. After 30 days of continuous operation, there was no window corrosion or signal attenuation, and the sampling delay was stable within 1.2s.

[0056] Step S200: Perform data preprocessing based on the original spectral dataset to obtain the processed dataset.

[0057] It should be noted that the method for obtaining the processed dataset includes: acquiring the real-time collected air background spectrum, obtaining background removal items, removing background spectra from the background removal items based on the original spectral dataset, and obtaining the first processed data; based on the first processed data, setting a filter window value, and smoothing and denoising the first processed dataset to obtain the second processed data; based on the second processed data, using an adaptive iterative reweighted penalized least squares method to complete spectral baseline correction, and obtaining the processed dataset.

[0058] Specifically, the background removal item is the clean air spectrum collected every 2 hours. The difference between the spectrum and the original spectrum is directly calculated to eliminate the interference of characteristic peaks of CO2 and H2O in the environment. The filter window value is set to 9-point Savitzky-Golay filter, and the polynomial fitting order is 2. The principle is to remove random noise from in-situ acquisition while preserving the peak shape of characteristic peaks. The distortion rate is more than 60% lower than that of ordinary mean filtering. The number of iterations of the adaptive iterative reweighted penalized least squares method is set to 100, and the penalty factor is set to 1e6. The baseline shift caused by temperature drift and probe contamination can be automatically fitted without the need for manual selection of baseline points. The three-step preprocessing eliminates environmental interference, random noise, and baseline drift in turn, ensuring the accuracy of subsequent quantification. The spectral signal-to-noise ratio after preprocessing is ≥40dB.

[0059] Step S300: Quantitatively transform the preset feature peaks based on the processed dataset to obtain a set of feature parameters.

[0060] It is important to note that, such as Figure 5 As shown, the method for obtaining the feature parameter set includes: pre-constructing a calibration curve, setting and assigning gradient values, and assigning gradient values ​​to fluoride ion standard solutions with different gradient concentrations, hexagonal lanthanum fluoride standard suspensions with different crystal phase ratios, and mixed standard gases with different gradient concentrations; based on the original spectral dataset, collecting the spectra of each standard sample, fitting the calibration curve with the absorbance of the characteristic peak as the independent variable and the standard concentration and crystal phase ratio as the dependent variables; extracting the absorbance of the corresponding preset characteristic peak in the processed dataset, and substituting it into the pre-constructed calibration curve to obtain the feature parameter set.

[0061] Specifically, the quantitative conversion is based on the Lambert-Beer law, with a set distribution gradient of 5-8. Standard solutions of fluoride ions at 5-8 gradient concentrations, hexagonal lanthanum fluoride standard suspensions with varying crystal phase proportions, and mixed standard gases of CO2, H2O, and HF at gradient concentrations are prepared. The fluoride ion standard solutions are prepared at 0.1, 0.5, 1, 2, 5, and 10 mol / L, for a total of 6 gradients. The hexagonal lanthanum fluoride standard suspensions are prepared with crystal phase proportions of 10%, 30%, 50%, 70%, and 90%. There are 6 gradients in total, 0% and 100%, and 6 sets of gradients prepared with mixed standard gases, including CO2 (1-50ppm), H2O (1-100ppm), and HF (0.1-5%). The pre-set characteristic peak correspondences are as follows: 1350cm⁻¹ is the fluoride ion stretching vibration peak, 420cm⁻¹ / 760cm⁻¹ is the hexagonal LaF3 lattice vibration peak (peak area ratio is linearly related to crystal phase conversion rate), 2350cm⁻¹ is the CO2 antisymmetric stretching peak, and 3650cm⁻¹ is the peak of the hexagonal phase LaF3 crystal (peak area ratio is linearly related to crystal phase conversion rate). -1 The peaks represent the H2O stretching vibrations, and the peak areas and corresponding total gaseous oxygen content are 4000 cm⁻¹. -1The selected characteristic peaks are all specific absorption peaks of the corresponding substances, and there is no overlap of characteristic peaks of other reaction products. The linear correlation coefficient R² of the calibration curve is ≥0.995, and the quantitative error is ≤3%.

[0062] Step S400: Obtain process parameters synchronized with the feature parameter set timestamp, and simultaneously obtain the raw material impurity feature parameters of the target batch of lanthanum source to obtain the benchmark dataset.

[0063] It is important to note that, such as Figure 3 As shown, the benchmark dataset includes: the feature parameter set, which includes the fluoride ion concentration in the precipitation system, the hexagonal lanthanum fluoride crystal phase conversion rate, the total gas phase oxygen in the calcination system, and the residual HF concentration in the calcination tail gas; the process dataset, which includes the lanthanum source feeding rate, the pH value of the reaction system, the HF inlet flow rate in calcination, the calcination heating rate, the precipitation aging time, and the calcination holding temperature; and the impurity dataset, which includes the calcium impurity content, carbonate content, and free water content of the target batch of lanthanum source.

[0064] Specifically, timestamp synchronization is achieved using the NTP network time protocol. The two spectral probes, the production line PLC control system, and the raw material detection system are all connected to the same industrial Ethernet switch, with a time deviation of ≤100ms. This ensures that the timing of the feature parameters and process parameters is completely aligned. The feature parameters of the raw material impurities are obtained by ICP-OES, Karl Fischer titration, and total carbon analyzer when each batch of lanthanum source enters the factory. These parameters are pre-entered into the system and bound to the batch ID. Timing alignment is the core prerequisite for ensuring the effectiveness of the input data of the reinforcement learning model and avoiding model training bias caused by time misalignment.

[0065] In the specific implementation process, a batch of lanthanum source was tested upon entering the factory and found to contain 1.2 ppm calcium impurities, 28 ppm carbonate ions, and 42 ppm free water. The process parameters corresponding to the timestamp 14:02 were: feeding rate 5 L / min, pH 4.2, HF flow rate 10 L / min, heating rate 10℃ / h, aging time 4h, and calcination temperature 750℃. The corresponding characteristic parameters were: fluoride ion concentration 2.3 mol / L, crystal phase conversion rate 72%, total gas phase oxygen 12 ppm, and residual HF concentration 1.2%. The alignment deviation of the 13-dimensional data was only 32 ms.

[0066] Step S500: Based on the benchmark dataset, build a deep deterministic policy gradient reinforcement learning agent framework. The action space of the framework is the process parameter adjustment range synchronized with the feature parameter set timestamp, thus obtaining the action space term. The reward function is constructed with the core performance compliance rate of high-purity lanthanum fluoride as the highest weight, combined with production energy consumption and production cycle weighting, thus obtaining the learning model term.

[0067] It is important to note that, such as Figure 3As shown, the action space items include: setting a first adjustment range value, the first adjustment range value being the lanthanum source feeding rate adjustment range; setting a second adjustment range value, the second adjustment range value being the HF inlet flow rate adjustment range; setting a third adjustment range value, the third adjustment range value being the calcination heating rate adjustment range; setting a fourth adjustment range value, the fourth adjustment range value being the reaction system pH adjustment range; setting a fifth adjustment range value, the fifth adjustment range value being the precipitation aging time adjustment range; and setting a sixth adjustment range value, the sixth adjustment range value being the calcination holding temperature adjustment range.

[0068] Specifically, the settings of the six adjustment range values ​​have all been verified for process safety. The first adjustment range is ±20% of the lanthanum source feeding rate to avoid precipitation and agglomeration due to excessive feeding and reduced production efficiency due to excessive feeding. The second adjustment range is ±30% of the HF inlet flow rate to avoid incomplete fluorination due to insufficient HF and corrosion and waste of raw materials due to excessive HF. The third adjustment range is ±15℃ / h of the calcination heating rate to avoid grain cracking due to excessive heating and prolonged production cycle due to excessively slow heating. The fourth adjustment range is ±0.5 of the pH of the reaction system to avoid corrosion of the reactor body due to excessively low pH and generation of lanthanum hydroxide impurities due to excessively high pH. The fifth adjustment range is ±2h of the precipitation aging time to avoid excessively short aging time resulting in small grains and excessively long aging time resulting in grain agglomeration. The sixth adjustment range is ±50℃ of the calcination holding temperature to avoid incomplete fluorination due to excessively low temperature and grain sintering due to excessively high temperature.

[0069] It is important to note that the methods for obtaining the learning model terms include: setting the network structure of the deep deterministic policy gradient reinforcement learning agent framework, including the number of fully connected network layers and the number of input dimensions, with the number of input dimensions corresponding to the total dimensions of the target dataset; setting the output dimensions, which correspond to the process parameters synchronized with the feature parameter set timestamps; setting the activation functions of the hidden layers and the output layers, thus obtaining the activation function settings.

[0070] Specifically, the network structure is as follows: The Actor network is a 4-layer fully connected network with an input dimension of 13, corresponding to 4-dimensional feature parameters, 6-dimensional process parameters, and 3-dimensional impurity parameters. It has two hidden layers, each with 64 neurons and the ReLU activation function. The output dimension is 6, corresponding to 6 adjustable process parameters and the Tanh activation function. Mapping the output to the [-1, 1] interval corresponds to the adjustment range of each parameter. The Critic network is a 3-layer fully connected network with an input dimension of 19, consisting of 13-dimensional states and 6-dimensional actions. It has 128 hidden layers with the ReLU activation function and an output dimension of 1, corresponding to the action value Q. The specific weights of the reward function are w1=0.6, w2=0.2, and w3=0.2. Training uses a capacity of 1×10⁻⁶. 6The priority experience replay pool is set with a priority coefficient of 0.7, the target network update delay step size is 3, the loss function adopts mean square error loss, the action space with process constraints avoids the agent outputting invalid and dangerous actions, the DDPG network adapts to the adjustment requirements of continuous process parameters, and the priority experience replay improves training efficiency by more than 30%.

[0071] Step S600: Construct a digital twin model for lanthanum fluoride preparation driven by hybrid mechanism data, perform interactive pre-training based on the learning model terms to obtain a general regulatory agent, and simultaneously collect actual production data from the target production line to perform small-sample transfer fine-tuning on the general regulatory agent to obtain a dedicated regulatory agent adapted to the target production line.

[0072] It is important to note that, such as Figure 3 As shown, the method for obtaining the general regulatory agent includes: integrating the enthalpy of lanthanum fluoride reaction, fluorine vacancy formation energy, and impurity diffusion barrier parameters obtained from first-principles calculations, acquiring corresponding data on historical production process performance, constructing a hybrid driving model using a regression model, and simultaneously performing interactive pre-training based on the learning model terms to obtain the general regulatory agent.

[0073] Specifically, the mechanistic parameters were calculated using VASP first-principles calculation software, with PBE as the functional. The calculated enthalpy of fluorination was -124 kJ / mol, the fluorine vacancy formation energy was 2.1 eV, and the calcium impurity diffusion barrier was 1.8 eV. These mechanistic parameters were incorporated as features into the regression model to avoid predictions that violate physical laws from a purely data-driven model. Historical production data consisted of 1200 sets of mass production process-performance data from leading domestic fluoride companies. The regression model used LightGBM, and the constructed digital twin model achieved a simulation accuracy of ≥92%. During pre-training, the agent interacted with the digital twin model, and each round of output action yielded corresponding performance and reward values. After 1e5 rounds of training, the agent... When the average reward value of the body is stable at around 0.62 and the fluctuation is ≤3% for 1000 consecutive rounds, the pre-training is considered to have converged and a general regulatory agent has been obtained. When fine-tuning with small samples, the first 3 general feature extraction layers of the Actor network are frozen, and only the last output layer of the Actor and the last output layer of the Critic are fine-tuned. The fine-tuning learning rate is set to 1 / 10 of the pre-training learning rate. The actual data of the target production line used for fine-tuning is ≤5 batches, and the number of iterations is ≤1000 rounds to converge and obtain a specific regulatory agent. The mechanism-data hybrid driven twin model takes into account both physical rationality and fitting accuracy. The transfer fine-tuning solves the pain points of insufficient self-owned data and high modeling cost of small and medium production lines. The adaptation cost is reduced by more than 70% compared with traditional methods.

[0074] Step S700: Set the acquisition threshold, which is a fixed time value. Based on the acquisition threshold, acquire the original spectral dataset and input it into the dedicated control agent. Output the process parameter adjustment command to obtain the output command set. Optimize the parameters based on the output command set, thereby realizing the adaptive adjustment of continuous process parameters and closed-loop dynamic optimization.

[0075] It is important to note that, such as Figure 4 As shown, the method for obtaining the output instruction set includes: the process parameter adjustment instruction is sent to the production line PLC control system for automatic execution; after each batch of production is completed, the status, action, reward, and next status data of the batch are stored in the experience playback pool; and each time a fixed batch of data is stored, the model incremental training is triggered once.

[0076] Specifically, the acquisition threshold is set to 30 seconds, and the raw spectrum is acquired every 30 seconds. After preprocessing and quantification, the data is input into a dedicated control agent. The output control commands are directly sent to the production line PLC control system via the OPC UA protocol for automatic execution without manual intervention. The incremental training rule for the model is as follows: after each batch of production is completed, the four-tuple of the batch's state, action, reward, and next state is stored in the experience replay pool. Incremental training is triggered every 10 batches of data stored. During training, only the entire Critic network and the last output layer of the Actor network are updated. The training time for a single session is ≤5 minutes, which does not affect normal production. Fixed-interval acquisition ensures the real-time nature of the control. Incremental training adapts to long-term variables such as equipment aging and operating condition drift. The accuracy of the model decreases by ≤3% over long-term operation.

[0077] In the specific implementation process, when a batch was produced for 2 hours, the total oxygen content in the gas phase was detected by spectroscopy to have risen to 22 ppm, while the threshold was ≤15 ppm. The intelligent agent automatically judged that the lanthanum source carbonate content in this batch was too high and automatically output adjustment instructions: aging time +1 hour, calcination HF flow rate +20%, calcination temperature +30℃. The final product oxygen content was 4.1 ppm, which met the standard. If the traditional fixed parameter process were used, the oxygen content of this batch would have reached 12 ppm and it would have been scrapped directly, avoiding a loss of about 12,000 yuan. Batch operation statistics show that this method increased the simultaneous compliance rate of the three core indicators of ultra-high purity lanthanum fluoride from 24% to 86%, reduced production energy consumption by 13%, and shortened the production cycle by 8%.

[0078] A machine learning-based system for optimizing process parameters in the preparation of high-purity lanthanum fluoride utilizes the aforementioned machine learning-based optimization method. The system includes: a data acquisition and processing module: deploying at least two in-situ fiber optic spectral probes in the lanthanum fluoride preparation process to acquire raw spectral data, obtaining a raw spectral dataset; performing data preprocessing on the raw spectral dataset to obtain a processed dataset; and quantitatively converting preset characteristic peaks based on the processed dataset to obtain a set of characteristic parameters; a benchmark data acquisition module: acquiring process parameters synchronized with the timestamp of the characteristic parameter set, and simultaneously acquiring the raw material impurity characteristic parameters of the target batch of lanthanum source to obtain a benchmark dataset; and a learning framework construction module: building a deep deterministic policy gradient reinforcement learning agent framework based on the benchmark dataset. The action space is the range of process parameter adjustments synchronized with the timestamps of the feature parameter set, resulting in an action space term. The reward function is constructed with the core performance compliance rate of high-purity lanthanum fluoride as the highest weight, combined with production energy consumption and production cycle weighting, resulting in a learning model term. A mechanistic data-driven digital twin model for lanthanum fluoride preparation is constructed. Interactive pre-training is performed based on the learning model term to obtain a general control agent. Simultaneously, actual production data from the target production line is collected, and the general control agent is fine-tuned through small-sample transfer to obtain a dedicated control agent adapted to the target production line. The instruction output optimization module sets a collection threshold, which is a fixed time value. Based on the collection threshold, the original spectral dataset is collected and input into the dedicated control agent, outputting process parameter adjustment instructions to obtain an output instruction set. Parameter optimization is performed based on the output instruction set.

[0079] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended embodiments and their equivalents.

Claims

1. A method for optimizing process parameters in the preparation of high-purity lanthanum fluoride based on machine learning, comprising: At least two in-situ fiber optic spectral probes were arranged in the lanthanum fluoride preparation process to collect raw spectral data and obtain the raw spectral dataset. Its characteristic is that it further includes: Data preprocessing is performed on the original spectral dataset to obtain a processed dataset. Based on the processed dataset, the preset characteristic peaks are quantitatively transformed to obtain a set of characteristic parameters. Obtain process parameters synchronized with the timestamp of the feature parameter set, and simultaneously obtain the raw material impurity feature parameters of the target batch of lanthanum source to obtain a benchmark dataset; A deep deterministic policy gradient reinforcement learning agent framework is built based on the benchmark dataset. The action space of the framework is the adjustment range of process parameters synchronized with the timestamp of the feature parameter set, which is the action space term. The reward function is constructed with the core performance compliance rate of high-purity lanthanum fluoride as the highest weight, combined with production energy consumption and production cycle weighting, which is the learning model term. A digital twin model for lanthanum fluoride preparation driven by hybrid mechanism data was constructed. Based on the learning model terms, interactive pre-training was performed to obtain a general regulatory agent. At the same time, actual production data of the target production line was collected, and the general regulatory agent was fine-tuned by small sample transfer to obtain a dedicated regulatory agent adapted to the target production line. A sampling threshold is set, which is a fixed time value. The original spectral dataset is collected based on the sampling threshold and input into a dedicated control agent. The agent outputs process parameter adjustment instructions to obtain an output instruction set. Parameter optimization is performed based on the output instruction set, thereby realizing the adaptive adjustment of continuous process parameters and closed-loop dynamic optimization.

2. The method for optimizing process parameters for preparing high-purity lanthanum fluoride based on machine learning according to claim 1, characterized in that: The arrangement method includes: A first in-situ fiber optic spectroscopy probe was deployed on the side wall of the liquid-phase precipitation reactor for lanthanum fluoride preparation to collect liquid-phase reaction spectral data of the precipitation process. A second in-situ fiber optic spectral probe is deployed in the gas outlet pipe of the gas phase roasting furnace to collect the spectral data of the gas phase products in the roasting process and obtain the original spectral dataset.

3. The method for optimizing process parameters for preparing high-purity lanthanum fluoride based on machine learning according to claim 1, characterized in that: The method for obtaining the processing dataset includes: The real-time collected air background spectrum is obtained to obtain background removal items. The background removal items are then removed based on the original spectral dataset to obtain the first processed data. Based on the first processed data, a filter window value is set, and the first processed dataset is smoothed and denoised to obtain the second processed data. Based on the second processed data, the spectral baseline correction is completed using the adaptive iterative reweighted penalized least squares method, resulting in the processed dataset.

4. The method for optimizing process parameters for preparing high-purity lanthanum fluoride based on machine learning according to claim 1, characterized in that: The method for obtaining the feature parameter set includes: A calibration curve is pre-constructed, and a gradient value is set. The gradient values ​​are fluoride ion standard solutions with different gradient concentrations, hexagonal lanthanum fluoride standard suspensions with different crystal phase ratios, and mixed standard gases with different concentrations. Based on the original spectral dataset, the spectra of each standard sample were collected, and a calibration curve was fitted with the characteristic peak absorbance as the independent variable and the standard concentration and crystal phase ratio as the dependent variables. Extract the absorbance of the corresponding preset feature peaks in the processed dataset and substitute them into the pre-constructed calibration curve to obtain the feature parameter set.

5. The method for optimizing process parameters for preparing high-purity lanthanum fluoride based on machine learning according to claim 1, characterized in that: The benchmark dataset includes: The set of characteristic parameters includes the fluoride ion concentration in the precipitation system, the hexagonal lanthanum fluoride crystal phase conversion rate, the total amount of gaseous oxygen in the calcination system, and the residual HF concentration in the calcination tail gas. The process dataset includes the lanthanum source feeding rate, reaction system pH value, HF inlet flow rate during calcination, calcination heating rate, precipitation aging time, and calcination holding temperature. The impurity dataset includes the calcium impurity content, carbonate content, and free water content of the target batch of lanthanum source.

6. The method for optimizing process parameters for preparing high-purity lanthanum fluoride based on machine learning according to claim 5, characterized in that: The action space items include: Set a first adjustment range value, which is the lanthanum source feeding rate adjustment range; set a second adjustment range value, which is the HF inlet flow rate adjustment range; set a third adjustment range value, which is the calcination heating rate adjustment range; set a fourth adjustment range value, which is the reaction system pH adjustment range; set a fifth adjustment range value, which is the precipitation aging time adjustment range; and set a sixth adjustment range value, which is the calcination holding temperature adjustment range.

7. The method for optimizing process parameters for preparing high-purity lanthanum fluoride based on machine learning according to claim 1, characterized in that: The methods for obtaining the learning model terms include: Define the network structure of the deep deterministic policy gradient reinforcement learning agent framework, including the number of fully connected network layers and the number of input dimensions, with the number of input dimensions corresponding to the total dimensions of the target dataset. Define the output dimension, which corresponds to the process parameters synchronized with the feature parameter set timestamp. Define the activation functions for the hidden layer and the output layer to obtain the activation function settings.

8. The method for optimizing process parameters for preparing high-purity lanthanum fluoride based on machine learning according to claim 1, characterized in that: The method for acquiring the general regulatory agent includes: By integrating the enthalpy of lanthanum fluoride reaction, fluorine vacancy formation energy, and impurity diffusion barrier parameters obtained from first-principles calculations, and acquiring corresponding data on historical production process performance, a hybrid driving model is constructed using a regression model. Simultaneously, interactive pre-training is performed based on the learning model terms to obtain a general regulatory agent.

9. The method for optimizing process parameters for preparing high-purity lanthanum fluoride based on machine learning according to claim 1, characterized in that: The method for obtaining the output instruction set includes: Process parameter adjustment commands are sent to the production line PLC control system for automatic execution. After each batch of production is completed, the status, actions, rewards, and next status data of that batch are stored in the experience playback pool. Each time a fixed batch of data is stored, incremental model training is triggered.

10. A machine learning-based system for optimizing process parameters in the preparation of high-purity lanthanum fluoride, characterized in that: The method for optimizing process parameters for preparing high-purity lanthanum fluoride based on machine learning, as described in any one of claims 1-9, includes: Data acquisition and processing module: At least two in-situ fiber optic spectral probes are arranged in the lanthanum fluoride preparation process to acquire raw spectral data and obtain raw spectral datasets. Data preprocessing is performed based on the raw spectral datasets to obtain processed datasets. Based on the processed datasets, preset characteristic peaks are quantitatively converted to obtain a set of characteristic parameters. Benchmarking data acquisition module: Acquires process parameters synchronized with the feature parameter set timestamp, and simultaneously acquires the raw material impurity feature parameters of the target batch of lanthanum source to obtain the benchmarking dataset; The learning framework construction module: Based on the benchmark dataset, a deep deterministic policy gradient reinforcement learning agent framework is built. The action space of the framework is the adjustment range of process parameters synchronized with the timestamp of the feature parameter set, which is the action space term. The reward function is constructed with the core performance achievement rate of high-purity lanthanum fluoride as the highest weight, combined with production energy consumption and production cycle weighting, which is the learning model term. A mechanism data hybrid-driven digital twin model of lanthanum fluoride preparation is constructed. Based on the learning model term, interactive pre-training is performed to obtain a general control agent. At the same time, actual production data of the target production line is collected, and small sample transfer fine-tuning is performed on the general control agent to obtain a dedicated control agent adapted to the target production line. Command output optimization module: Set the acquisition threshold, which is a fixed time value. Based on the acquisition threshold, acquire the original spectral dataset and input it into the dedicated control agent. Output process parameter adjustment commands to obtain the output command set. Parameter optimization is performed based on the output command set.

Images