Germane kettle residue recovery method and system combining machine learning and process simulation

Abstract

The embodiment of the invention provides a germane kettle residue recovery method and system combining machine learning and process simulation, and relates to the technical field of machine learning, firstly, key operation links and corresponding influence factors of a germane kettle residue recovery process are determined, and a process simulation structure is constructed to reproduce a substance conversion path; and then generating a recovery process simulation data set under different influence factor combinations, training a target machine learning model by using the recovery process simulation data set to establish a mapping relation and generate an optimized influence factor combination, finally applying the optimized influence factor combination to an actual recovery process, and executing an optimization process to obtain germane related recovery products. The method can improve the recovery rate and the product quality, reduce the cost and realize efficient utilization of resources.

Description

Technical Field

[0001] This application relates to the field of machine learning technology, and more specifically, to a method and system for recovering germanane reactor residue by combining machine learning and process simulation. Background Technology

[0002] Germanane, as an important semiconductor material precursor, has wide applications in semiconductor manufacturing, solar cells, and other fields. The production process of germanane generates a large amount of residue, which typically contains a certain amount of germanium and other valuable substances. Effective recovery of this germanane residue can not only improve resource utilization and reduce production costs, but also reduce environmental pollution.

[0003] Currently, existing methods for recovering germanane residue mainly rely on traditional chemical experiments and empirical operations. Researchers repeatedly conduct experiments, trying different operating parameters and conditions to find the optimal recovery process. However, these methods have many limitations. On the one hand, the experimental process requires a significant amount of time, manpower, and resources, resulting in high costs. On the other hand, due to limitations in experimental conditions, it is difficult to fully consider the complex interactions between numerous influencing factors in the recovery process, leading to unstable recovery results and difficulty in effectively improving recovery rates and product quality. Furthermore, traditional methods lack the ability to systematically simulate and predict the recovery process, making it impossible to accurately assess the recovery results under different operating conditions in advance, thus hindering the optimization of the recovery process. Summary of the Invention

[0004] In view of this, the purpose of this application is to provide a method and system for recovering germanane reactor residue by combining machine learning and process simulation.

[0005] According to a first aspect of this application, a method for recovering germanane reactor residue combining machine learning and process simulation is provided, the method comprising: Identify the key operational steps in the germane reactor residue recovery process and the corresponding influencing factors for each key operational step. A process simulation structure is constructed that includes each key operation step and corresponding influencing factors. The process simulation structure is used to reproduce the material conversion path in the germane residue recovery process. Based on the process simulation structure, a set of simulation data of the recycling process under different combinations of influencing factors is generated. The set of simulation data of the recycling process includes the operating status data and material conversion result data of each key operation link. A target machine learning model is trained using a dataset of simulated recycling processes to establish a mapping relationship between influencing factors and material transformation results, and an optimized combination of influencing factors is generated based on this mapping relationship. The optimized combination of influencing factors output by the target machine learning model is applied to each key operation step of the actual germane residue recovery process. The optimized germane residue recovery process is then executed to obtain germane-related recovered products.

[0006] According to a second aspect of this application, a germane slag recovery system combining machine learning and process simulation is provided. The germane slag recovery system combining machine learning and process simulation includes a processor and a readable storage medium storing a program that, when executed by the processor, implements the aforementioned germane slag recovery method combining machine learning and process simulation.

[0007] Based on any of the above aspects, by identifying the key operational steps and corresponding influencing factors in the germane residue recovery process, a process simulation structure capable of reproducing the material conversion pathway is constructed. Based on this simulation structure, a set of simulation data for the recovery process under different combinations of influencing factors is generated, covering the operational status and material conversion results of each key operational step. Using this data, a target machine learning model is trained, establishing a mapping relationship from influencing factors to material conversion results, enabling rapid and accurate prediction of recovery results under different combinations of influencing factors. The optimized combination of influencing factors generated based on this mapping relationship fully considers the complex interactions between various factors. Applying this optimized combination to the actual recovery process, the optimized process significantly improves the recovery rate and quality of germane-related recovery products, while reducing recovery costs and improving resource utilization efficiency, achieving an organic combination of machine learning and process simulation. Attached Figure Description

[0008] Figure 1 A schematic diagram of the process for the germanane reactor slag recovery method combining machine learning and process simulation provided in the embodiments of this application is shown. Figure 2 A schematic diagram of the component structure of the germanane reactor residue recovery system combining machine learning and process simulation provided in an embodiment of this application is shown. Detailed Implementation

[0009] Figure 1 A schematic diagram of the process for germane reactor slag recovery method combining machine learning and process simulation provided in this application embodiment is shown. The detailed steps are described below.

[0010] Step S110: Determine the key operational steps in the germane residue recovery process and the influencing factors corresponding to each key operational step.

[0011] In the field of germane residue recovery, the overall process typically encompasses multiple stages, including raw material pretreatment, chemical reaction, and separation and purification. To identify critical operational stages, it is essential to first conduct a material and energy flow analysis of the entire process to identify nodes that significantly impact germanium recovery rate, product purity, and process energy consumption. Taking a typical germane residue recovery process as an example, its flow includes unit operations such as residue crushing, acid leaching and dissolution, solid-liquid separation, extraction, back-extraction, germanium salt precipitation, and drying and calcination. Sensitivity analysis of historical operating data for each unit operation revealed that fluctuations in process parameters during acid leaching and dissolution, extraction, and back-extraction can cause germanium recovery rates to exceed preset thresholds. Therefore, these three stages were identified as critical operational stages.

[0012] For the acid leaching and dissolution stage, influencing factors can be divided into operating environment-related factors and material property-related factors. Operating environment-related factors include the temperature, pressure, stirring rate, acid solution concentration, and leaching time of the acid leaching reaction. Material property-related factors include the initial particle size distribution of the residue, the form of germanium in the residue (e.g., germanium oxide, germanium sulfide, etc.), the moisture content of the residue, and the composition of impurities in the residue (e.g., the content of iron, copper, silicon, etc.). For the extraction stage, operating environment-related factors include the extractant concentration, the volume ratio (total volume) of the organic phase and the aqueous phase, the extraction temperature, the mixing time, and the clarification time. Material property-related factors include the pH value of the feed solution, the germanium ion concentration in the feed solution, and the concentration of coexisting ions (e.g., chloride ions, sulfate ions, etc.). For the back-extraction stage, operating environment-related factors include the back-extractant concentration, back-extraction temperature, mixing time, and the ratio (total volume). Material property-related factors include the germanium concentration in the loaded organic phase and the aging degree of the organic phase (e.g., the number of times it has been used).

[0013] Step S120: Construct a process simulation structure that includes each key operation step and corresponding influencing factors. The process simulation structure is used to reproduce the material conversion path in the germane residue recovery process.

[0014] Step S121: Analyze the overall process route of the germane reactor residue recovery process, determine the execution order and material transfer relationship between each key operation step, and the material transfer relationship is used to represent the association form of the output material of the previous key operation step as the input material of the next key operation step.

[0015] In the overall process of germane residue recovery, the key operations are executed in the following sequence: acid leaching and dissolution – extraction – back-extraction. The output of the acid leaching and dissolution stage is an acidic solution containing germanium ions and other impurity ions. This solution, after solid-liquid separation, serves as the input for the extraction stage. In the extraction stage, the extractant contacts the acidic solution, transferring germanium ions from the aqueous phase to the organic phase. The output is a germanium-loaded organic phase, which serves as the input for the back-extraction stage. In the back-extraction stage, the back-extractant transfers germanium ions from the organic phase back to the aqueous phase, resulting in a high-concentration germanium solution for subsequent precipitation treatment. Specifically, the mass transfer parameters of the acid leaching and dissolution stage (such as flow rate, temperature, and concentration of each component) directly serve as the input parameters for the extraction stage; similarly, the output parameters of the extraction stage (such as the flow rate of the organic phase, germanium concentration, and temperature) directly serve as the input parameters for the back-extraction stage.

[0016] Step S122: For each key operation step, extract the factors affecting the material conversion efficiency of that key operation step. The factors include operating environment-related factors and material property-related factors. Operating environment-related factors are directly related to the execution conditions of the key operation step, and material property-related factors are related to the characteristics of the material itself participating in the key operation step.

[0017] For the acid leaching and dissolution process, temperature, among the environmental factors, directly affects the kinetic rate of the acid leaching reaction. Higher temperatures generally accelerate the reaction rate but may increase energy consumption and the risk of equipment corrosion. The stirring rate affects the mass transfer efficiency between the solid and liquid phases; too low a rate leads to excessively large local concentration gradients, while too high a rate may result in energy waste and equipment wear. Among the material properties, the initial particle size distribution of the slag affects the specific surface area of ​​the solid reactants; smaller particle sizes result in larger specific surface areas and greater reaction contact areas, but excessively fine particle sizes may lead to difficulties in subsequent solid-liquid separation. The occurrence form of germanium determines its solubility in acid solutions; different forms of germanium compounds require different acid concentrations and reaction conditions for effective dissolution.

[0018] Among the environmental factors related to the extraction process, the extractant concentration directly affects the extraction capacity and selectivity of the organic phase for germanium ions. Excessive concentration may lead to phase separation difficulties, while insufficient concentration reduces extraction efficiency. The choice of phase affects the mass transfer driving force and the amount of organic phase used in the extraction process; an excessively high phase will increase the amount of organic phase circulating, while an insufficient phase may result in incomplete extraction. Among the material properties, the pH value of the feed solution affects the form of germanium ions in the aqueous phase (e.g., whether they form complex ions), thus affecting their binding ability with the extractant. The concentration of coexisting ions may compete with the extractant for extraction, reducing the selectivity for germanium ions.

[0019] Among the environmental factors related to the back-extraction process, the concentration of the back-extractant determines the chemical driving force of the back-extraction reaction. Excessive concentration may lead to waste of the back-extractant and increased load on subsequent processing, while insufficient concentration results in inadequate back-extraction efficiency. The back-extraction temperature affects the miscibility of the organic and aqueous phases and the back-extraction reaction rate; it must be controlled within a suitable range to ensure effective back-extraction and minimize solvent loss. Among the material property factors, the concentration of germanium in the loaded organic phase affects the back-extraction equilibrium; excessively high concentrations may require multiple back-extractions to reach the target concentration. The aging of the organic phase can lead to a decrease in extractant activity, affecting back-extraction efficiency.

[0020] Step S123: Establish a dedicated process simulation unit for each key operation step. The process simulation unit contains the mathematical relationship of material transformation corresponding to the key operation step. The mathematical relationship of material transformation is formed by coupling the kinetic equation describing the chemical reaction rate, the mass transfer equation describing the diffusion of matter, and the heat transfer equation describing the heat exchange.

[0021] Step S1231: For a single key operation, consult relevant chemical principle literature and process practice data on germanane residue recovery to determine the type of chemical reaction and physical change process involved in the key operation.

[0022] Taking the acid leaching dissolution process as an example, according to the "Principles of Germanium Metallurgy" and related process reports, germanium in the reactor slag mainly exists in the forms of germanium dioxide and germanium sulfide. Germanium dioxide reacts with hydrogen ions in the acid solution to form soluble germanium salts, while germanium sulfide undergoes an oxidative dissolution reaction in the presence of oxidizing acids. The physical changes include the dispersion of reactor slag particles in the acid solution, the dissolution and diffusion of solid reactants, and the migration of reaction products into the bulk solution. Simultaneously, the acid leaching process involves heat exchange, including the heat effect of the reaction and the heat exchange between the equipment and the environment.

[0023] Step S1232: Extract the types of reactants, products, and reaction conditions required in this key operation step. The reaction conditions directly correspond to the operating environment-related factors among the influencing factors.

[0024] The reactants in the acid leaching and dissolution step include germanium dioxide and germanium sulfide in the residue, as well as hydrogen ions and oxidizing agents (such as hydrogen peroxide) in the acid solution (such as sulfuric acid or hydrochloric acid). The products include soluble germanium ions (such as germanium tetrachloride or germanium sulfate), water, and elemental sulfur or sulfate ions produced by the oxidation of sulfides. The required reaction conditions include a suitable temperature range (corresponding to temperature in the operating environment factors), acid concentration (corresponding to acid solution concentration), and reaction time (corresponding to acid leaching time). These requirements directly correspond to the operating environment factors extracted in step S122.

[0025] Step S1233: Determine the form of the reaction kinetic correlation based on the chemical reaction type. The reaction kinetic correlation reflects the relationship between reactant concentration, reaction temperature, reaction pressure and reaction rate.

[0026] The acid leaching reaction of germanium dioxide is a first-order irreversible reaction. Its kinetic correlation is that the reaction rate is directly proportional to the product of the germanium dioxide concentration and the hydrogen ion concentration, and is also corrected for by an Arrhenius term due to temperature. The oxidative dissolution reaction of germanium sulfide is a second-order reaction, and the reaction rate is directly proportional to the product of the germanium sulfide concentration and the oxidant concentration, also including a temperature-dependent term. The reaction pressure has a relatively small effect on the liquid-phase reaction and can be ignored or considered as a correction factor in the correlation.

[0027] Step S1234: Combine the physical change process to supplement the mass transfer correlation and heat transfer correlation. The mass transfer correlation reflects the diffusion and migration law of matter within the key operation link, and the heat transfer correlation reflects the heat exchange law within the key operation link.

[0028] The mass transfer correlation addresses the liquid film diffusion process on the surface of slag particles, described by Fick's law. The diffusion rate is proportional to the concentration gradient on the particle surface and the diffusion coefficient, which is affected by temperature and solution viscosity. For forced convection caused by stirring, the Sherwood number correlation is introduced, relating the mass transfer coefficient to the Reynolds number and Schmidt number to reflect the influence of flow state on mass transfer. The heat transfer correlation includes the calculation of reaction heat and equipment heat loss. The reaction heat is calculated by multiplying the enthalpy change of each reaction by the reaction rate. Equipment heat loss is calculated by the convective heat transfer coefficient, heat transfer area, and temperature difference between the inside and outside of the equipment, while also considering the portion of mechanical energy converted into heat energy during stirring.

[0029] Step S1235: Determine the undetermined parameters in the reaction kinetics correlation, mass transfer correlation, and heat transfer correlation based on process practice data. Couple the reaction kinetics correlation, mass transfer correlation, and heat transfer correlation to construct the core material transformation mathematical relationship for this key operation. The core material transformation mathematical relationship is used to comprehensively describe the combined effects of chemical and physical changes on material transformation.

[0030] By collecting historical operational data of the acid leaching process at a germane reactor residue recovery plant, including curves showing the change in germanium dissolution rate over time at different temperatures, acid concentrations, and stirring rates, the frequency factor and activation energy, among other indeterminate parameters, in the reaction kinetic correlation were fitted using the least squares method. For the diffusion coefficient and mass transfer coefficient in the mass transfer correlation, adjustments were made by referencing experimental data from similar systems and combining them with plant data. The convective heat transfer coefficient in the heat transfer correlation was calculated using equipment structural parameters (such as reactor diameter and impeller type) and operating parameters (such as stirring speed and fluid viscosity). These three correlations were coupled using the laws of mass conservation and energy conservation; for example, the amount of reactant consumed in the reaction is equal to the amount of reactant diffused to the reaction interface, and the heat released or absorbed by the reaction is equal to the heat exchanged through the heat transfer process, forming the core mathematical relationship of material transformation.

[0031] Step S1236: Add boundary condition constraints to the mathematical relationship of core material transformation. The boundary condition constraints are determined based on the equipment structure parameters and process operation limits of key operation links, so that the calculation results do not exceed the actual feasible range.

[0032] Boundary constraints include spatial boundaries and parameter boundaries. Regarding spatial boundaries, the geometry of the acid leaching reactor (e.g., diameter, height, and agitator position) defines the material flow region and reaction space, which is mathematically represented by boundary conditions in the diffusion equation (e.g., no-slip condition at the wall, zero gradient condition at the plane of symmetry). Regarding parameter boundaries, the upper temperature limit is constrained by the temperature resistance of the equipment material, while the lower limit is constrained by the reaction rate requirements; the upper acid concentration limit is constrained by the equipment's corrosion resistance and acid mist control requirements, while the lower limit is constrained by the dissolution efficiency; the upper stirring rate limit is constrained by the motor power and shaft strength, while the lower limit is constrained by the critical suspension speed. These boundary conditions are added to the core material transformation mathematical relationship in the form of inequalities to ensure that the calculated parameters are within the practically feasible range.

[0033] Step S1237: Establish the solution logic for the mathematical relationship of material transformation, and determine the solution order and iterative convergence conditions of the correlation formula.

[0034] The mathematical relationships of material transformation are solved using a combination of sequential solution and iterative coupling. First, the mass transfer correlation is solved to obtain the mass transfer rate from reactants to the reaction interface. This mass transfer rate is used as input to the reaction kinetic correlation to calculate the reaction rate, thus obtaining the change in substance concentration over time. Simultaneously, the heat of reaction is calculated based on the reaction rate, and the temperature distribution is calculated using the heat transfer correlation. The temperature distribution, in turn, affects the diffusion coefficient and reaction kinetic parameters, forming a coupled iteration. The convergence condition for the iteration is set as follows: the difference in germanium solubility calculated in two consecutive iterations is less than a preset minimum, and the residuals of each conservation equation (mass conservation, energy conservation) are less than the corresponding thresholds. The solution sequence proceeds according to the time step, with mass transfer calculation, reaction calculation, heat transfer calculation, and parameter update completed sequentially within each time step.

[0035] Step S1238: Integrate the mathematical relationships of core material transformation, boundary condition constraints, and solution logic into a unified calculation module, forming part of the process simulation unit specific to this key operation.

[0036] Employing object-oriented programming principles, the core material transformation mathematical relationships are encapsulated as a "reaction calculation class," encompassing methods for kinetic calculations, mass transfer calculations, and heat transfer calculations. Boundary condition constraints are stored as attribute parameters of the class, invoked in real-time during calculations to verify parameter rationality. The solution logic, as a "solver class," controls the calculation process, iterative progression, and convergence determination. Combining these class modules forms the core computational part of the process simulation unit. This core computational part can receive input from influencing factors and output material transformation results (such as germanium ion concentration, solid residue, and system temperature).

[0037] Step S1239: Perform logic testing on the core part of the process simulation unit, input the known values ​​of influencing factors, verify whether the output material conversion result data is consistent with the process practice data, adjust the parameters in the material conversion mathematical relationship according to the test results, and finally obtain a dedicated process simulation unit containing the material conversion mathematical relationship corresponding to the key operation link.

[0038] Three sets of known influencing factor values ​​(e.g., low temperature, low acid concentration; medium temperature, medium acid concentration; high temperature, high acid concentration) were selected and input into the core of the process simulation unit. After calculation, the output data on germanium solubility, reaction time, system temperature, and other material conversion results were compared with the actual process data collected concurrently at the plant. If the deviation between the simulated and actual solubility in any set of data exceeded the preset allowable error, the core material conversion mathematical relationship was re-examined, and deviations in the mass transfer coefficient or reaction kinetic parameters were checked. Corrections were made by fine-tuning the activation energy or diffusion coefficient values. The testing and correction process was repeated until the deviations between the simulation results and actual data under all test conditions were within the allowable range, thus completing the logical testing of the core of the process simulation unit.

[0039] Step S124: Based on the execution order and material transfer relationship between each key operation link, connect each dedicated process simulation unit in the corresponding order to form a preliminary process simulation structure. Configure an influencing factor input interface and a material conversion result output interface for each process simulation unit in the preliminary process simulation structure. The influencing factor input interface is used to receive the specific values ​​of the corresponding influencing factors, and the material conversion result output interface is used to output the corresponding operating status data and material conversion result data of the process simulation unit.

[0040] The dedicated process simulation units for the three key operational steps of acid leaching dissolution, extraction, and back-extraction are connected in the order of "acid leaching dissolution unit - extraction unit - back-extraction unit". Each process simulation unit is equipped with an independent input interface for influencing factors. For example, the input interface of the acid leaching unit receives specific values ​​of operating environment factors such as temperature, acid concentration, and stirring rate, as well as material property factors such as residue particle size and germanium occurrence form. The input interface of the extraction unit receives specific values ​​of operating environment factors such as extractant concentration, phase ratio, and temperature, as well as material property factors such as feed solution pH and germanium ion concentration. Regarding the output interface for material conversion results, the acid leaching unit outputs data such as the concentration of each component in the acid leaching solution (germanium ions, impurity ions), the amount of acid leaching residue, and the system temperature; the extraction unit outputs data such as the germanium concentration of the loaded organic phase, the composition of the raffinate, and the phase separation time; and the back-extraction unit outputs data such as the germanium concentration of the back-extraction solution and the composition of the regenerated organic phase. The interfaces adopt a standardized data format (such as JSON format), including data identifiers, values, units, timestamps, and other information.

[0041] Step S125: Define the material transfer interface specification between each process simulation unit. This material transfer interface specification specifies the material data format, data type and transmission protocol output by the material conversion result output interface of the previous process simulation unit, and matches the requirements of the material input port of the next process simulation unit.

[0042] The material transfer interface specification defines in detail all aspects of data exchange. Regarding data format, the material data output by the preceding unit must include a material flow identifier (e.g., "acid leaching solution flow"), total flow rate, mole fraction or mass fraction of each component, temperature, pressure, density, viscosity, and other physical property parameters. Each parameter must have a clearly defined data type (e.g., floating-point number, string) and precision requirements (e.g., retaining four decimal places). The transmission protocol uses TCP / IP-based Socket communication, defining a request-response mechanism. After receiving the material data from the preceding unit, the following unit must return a receipt confirmation message. If the data is incomplete or incorrectly formatted, an error code and reason must be returned. The interface specification also includes data validation rules, such as the sum of the mole fractions of each component should be 1, and temperature and pressure should be within reasonable ranges. The receiving unit must validate the received data before it can be used as input for calculations.

[0043] Step S126: Construct an overall control module for the process simulation structure. The overall control module sequentially calls and executes the calculation programs of each process simulation unit according to a preset sequence of key operation steps. A data acquisition unit is configured in the overall control module. The data acquisition unit reads the running status data of each process simulation unit in real time according to a preset sampling frequency, and stores the running status data in the database along with the corresponding influencing factor values ​​and timestamps.

[0044] The overall control module adopts a modular design, including sub-modules such as a scheduler, data acquisition unit, database interface, and human-machine interface. The scheduler generates a computational task queue based on the sequence of key operation steps (acid leaching—extraction—back-extraction) and the dependencies between each unit, sequentially calling the computational programs of each process simulation unit. For example, when starting the simulation, the scheduler first sends a start command and initial influencing factor values ​​to the acid leaching unit. After the acid leaching unit completes one computational step, it sends the output material data to the extraction unit through the mass transfer interface, then calls the extraction unit's computational program, and so on. The data acquisition unit is set with a preset sampling frequency (e.g., once per calculation step) and reads the operating status data of each unit in real time (e.g., the stirring motor current of the acid leaching unit, the interface height of the mixing and clarification tank of the extraction unit, and the pump outlet pressure of the back extraction unit). The data is packaged with the corresponding influencing factor values ​​(e.g., current acid concentration, extractant concentration) and timestamps (accurate to milliseconds) and stored in a specified data table of a relational database (e.g., MySQL) through a database interface. The data table contains the following fields: timestamp, unit identifier, influencing factor name, influencing factor value, operating status parameter name, and operating status parameter value.

[0045] Step S127: Retrieve the material balance verification operation to the process simulation structure. The material balance verification operation calculates the difference between the total mass of all input substances and the total mass of all output substances in the process.

[0046] The mass balance verification operation is performed in each calculation cycle of the process simulation structure (e.g., after each complete full-process simulation). First, the mass input and output data of all process simulation units are traversed, and the mass flow rates of each input and output stream are collected. For the acid leaching unit, the input stream includes slag and acid solution, and the output stream includes acid leaching liquor and acid leaching residue; for the extraction unit, the input stream includes acid leaching liquor and extractant, and the output stream includes loaded organic phase and raffinate; for the back-extraction unit, the input stream includes loaded organic phase and back-extractant, and the output stream includes back-extraction liquid and regenerated organic phase. The total mass of all input streams (the sum of the mass flow rates of each input stream multiplied by the time step) and the total mass of all output streams (the sum of the mass flow rates of each output stream multiplied by the time step) are calculated; the difference between these two values ​​is the mass balance deviation.

[0047] Step S128: Based on the verification results of the material balance verification operation, when there is a material non-conservation problem, execute the parameter and interface adjustment sub-step to adjust the material transformation mathematical relationship parameters and material transfer interface specifications of the relevant process simulation units; repeat the process simulation structure and material balance verification operation until the verification results of the material balance verification operation meet the preset tolerance range. At this time, it is determined that the process simulation structure is completed, forming the final process simulation structure containing each key operation link and corresponding influencing factors.

[0048] For example, step S1281: Analyze the verification results of the material balance verification operation, identify the process simulation unit connection nodes where there is a material non-conservation problem, and the connection nodes represent the material transfer interface between the previous process simulation unit and the next process simulation unit.

[0049] The verification results output by the mass balance check operation include the total mass input, total mass output, and deviation value. If the absolute value of the deviation value exceeds the preset tolerance range (e.g., 0.1% of the total mass), the mass transfer data at each connection node is queried from the database. For example, it checks whether the mass flow rate of the acid leaching solution output from the acid leaching unit to the extraction unit is consistent with the mass flow rate of the acid leaching solution received by the extraction unit. If they are inconsistent, the connection node (acid leaching-extraction interface) is a problem node for mass non-conservation; if they are consistent, the interface output from the extraction unit to the back-extraction unit is checked, and so on, to locate the problem connection node.

[0050] Step S1282: For connection nodes with material non-conservation problems, calculate the difference between the material input and the material output to determine the specific scale of material loss or redundancy.

[0051] For the identified problematic connection node (such as the interface between the extraction unit and the back-extraction unit), retrieve the output material data (mass flow rate of the loaded organic phase) of the preceding unit (extraction unit) and the input material data (mass flow rate of the received loaded organic phase) of the following unit (back-extraction unit), and calculate the difference between the two (output amount minus received amount). If the difference is positive, it indicates that there is material loss; if it is negative, it indicates that there is material redundancy. Based on the proportion of the absolute value of the difference to the total input amount, determine the scale level of loss or redundancy (e.g., slight, moderate, severe).

[0052] Step S1283: Analyze the reasons for the non-conservation of matter and determine whether it is due to unreasonable settings of mathematical relationship parameters for matter transformation in the previous process simulation unit or defects in the matter transfer interface specification.

[0053] The root cause analysis proceeded from two aspects: First, the mathematical parameters of the material transformation in the previous simulation unit were examined. For example, if the distribution coefficient in the material transformation mathematical relationship of the extraction unit was too low, the calculated concentration of germanium in the loaded organic phase would be lower than the actual value, thus causing a deviation in the calculated mass flow rate of the output stream. Second, the execution of the mass transfer interface specifications was checked, such as whether there were data packet losses, data format conversion errors (e.g., incorrect unit conversion, such as writing kg / h as g / h), or data overflow due to insufficient interface buffer capacity. By comparing the log files at both ends of the interface (the sending log of the previous unit and the receiving log of the next unit), any defects in the interface specifications were investigated.

[0054] Step S1284: If the reason is that the mathematical relationship parameters of material transformation are unreasonable, retrieve the core mathematical relationship of material transformation from the previous process simulation unit, adjust the reaction rate parameter in the reaction kinetic correlation based on the scale of material loss or redundancy, and simultaneously adjust the diffusion coefficient parameter in the mass transfer correlation and the heat transfer coefficient parameter in the heat transfer correlation to make the mathematical description of the material transformation process more in line with the actual law of conservation of matter.

[0055] When the cause is determined to be an unreasonable mathematical relationship parameter of the material transformation in the extraction unit (e.g., a low partition coefficient leading to material loss), retrieve the core mathematical relationship of the material transformation in the extraction unit and find the partition coefficient correlation describing the extraction equilibrium (e.g., an empirical formula related to extractant concentration and temperature). Based on the scale of material loss (e.g., loss accounting for 5% of the theoretical output), multiply the original partition coefficient by a correction factor (e.g., 1.05) to adjust the forward reaction rate constant in the reaction kinetic correlation, thereby increasing the calculated concentration of the loaded organic phase. Simultaneously, check if the diffusion coefficient in the mass transfer correlation is too low due to an incorrect temperature parameter input. If so, recalculate the diffusion coefficient based on the actual temperature (the diffusion coefficient is proportional to the square root of the temperature). If the thermal conductivity coefficient in the heat transfer correlation affects the temperature calculation and thus the partition coefficient, it also needs to be adjusted synchronously to ensure that the mathematical description of the material transformation process after parameter adjustments is closer to the actual material conservation situation.

[0056] Step S1285: If the cause is a defect in the material transfer interface specification, redefine the material transfer interface data format and transmission protocol of the connection node, optimize the buffer operation of the material transfer interface, and add a data verification field so that the receiving process simulation unit can verify whether the received material data is complete and accurate.

[0057] If the cause is a defect in the data transmission protocol of the extraction-re-extraction interface (such as the lack of a retransmission mechanism leading to data packet loss), redefine the transmission protocol: add a sequence number field and a checksum field to the original protocol. The sequence number is used to identify the order of data packets, and the checksum is obtained by hashing the data packet content. After receiving the data packet, the receiver first checks whether the sequence number is consecutive; if it is missing, it requests the sender to retransmit. Then, it calculates the checksum of the received data and compares it with the checksum attached by the sender. If they do not match, it indicates a data transmission error, and a retransmission is requested. Optimize the interface buffering operation by increasing the buffer capacity from the original setting to avoid data overflow during high traffic. Set a buffer data timeout handling mechanism; buffered data that has not been read by the receiver for more than a preset time (e.g., 5 seconds) is automatically marked as invalid and logged.

[0058] Step S1286: Retest the mathematical relationship parameters of material transformation and the material transfer interface specifications of the adjusted process simulation unit, start the process simulation structure operation, and perform the material balance verification operation again.

[0059] After completing parameter adjustments and interface specification optimization, the process simulation structure is initiated, running a complete simulation cycle (from acid leaching to back-extraction). During the simulation, mass transfer data at the problem connection nodes is closely monitored, and the data packet status (e.g., successful transmission, successful reception, number of retransmissions) is recorded through the data acquisition unit. After the simulation, a mass balance verification operation is performed to calculate the deviation between the total mass input and output.

[0060] Step S1287: If the material balance verification operation shows that the difference between the input and output of the material still exceeds the preset tolerance after retesting, repeat the cause analysis, parameter adjustment and interface optimization operations until the difference is reduced to within the preset tolerance.

[0061] If the deviation value still exceeds the preset tolerance (e.g., 0.1%) after retesting, return to step S1281, parse the verification results again, and relocate the problem node. It may be that a new connection node has a problem or the problem of the original node has not been completely resolved. Repeat the steps of material loss / redundancy calculation, cause analysis, parameter adjustment / interface optimization, and retesting until the material balance verification result meets the preset tolerance range.

[0062] Step S1288: After adjusting and verifying all connection nodes, start the process simulation structure to perform a full-process operation test. Monitor the material conservation of the entire process through the material balance verification operation. Based on the full-process operation test results, fine-tune the parameters of the process simulation units whose differences exceed the preset tolerance in the material balance verification until the full-process material balance verification results meet the tolerance requirements, and finally form the process simulation structure.

[0063] After all connection nodes have passed verification, a full-process operation test is conducted, simulating at least three process cycles to examine the material conservation during the dynamic process. The full-process material balance verification calculates the difference between the total material input (residue, acid, extractant, back-extraction agent, etc.) and the total material output (back-extraction liquor, raffinate, acid leaching residue, etc.) of the entire recovery process. If individual process simulation units (such as the back-extraction unit) exhibit persistent minor deviations during the full-process test, their material conversion mathematical parameters (such as the back-extraction efficiency correction coefficient) are fine-tuned. There is no need to retest the connection nodes; the adjustment effect is verified solely through the full-process verification. Finally, when the absolute value of the full-process material balance deviation is consistently below the preset tolerance range, the process simulation structure is considered complete.

[0064] Step S130: Generate a set of simulation data for the recycling process under different combinations of influencing factors based on the process simulation structure. The set of simulation data for the recycling process includes the operating status data and material conversion result data of each key operation link.

[0065] Step S131: For each key operation step, extract the value range of each influencing factor. The value range is determined based on the actual process feasibility and safe operation boundary of germane residue recovery.

[0066] Regarding the temperature-related factors in the acid leaching process, practical process feasibility requires that the temperature not be too low to ensure a sufficient reaction rate, while the safe operating boundary requires that the temperature not be too high to avoid excessive acid mist volatilization, accelerated equipment corrosion, and safety accidents. The temperature range is determined by consulting equipment manuals (e.g., for reactors made of 316L stainless steel, the upper limit of their long-term operating temperature) and process safety regulations (e.g., acid solutions are prone to decomposition above a certain temperature), combined with the stable operating temperature range in historical production. For the acid concentration-related factors, the lower limit of the range is determined by dissolution efficiency experimental data (below a certain concentration, germanium dissolution rate decreases significantly), while the upper limit is determined by the acid storage and transportation safety requirements (e.g., high-concentration acids are hazardous chemicals requiring special storage conditions) and the equipment's corrosion resistance (e.g., the acid resistance rating of pumps and valves).

[0067] Step S132: Using a uniform sampling method, extract multiple candidate values ​​from the value range of each influencing factor. According to the execution order of each key operation step, combine the candidate values ​​of different key operation steps to generate multiple influencing factor combination schemes. Each influencing factor combination scheme contains the candidate values ​​of influencing factors corresponding to all key operation steps.

[0068] Step S1321: Extract the execution order of each key operation step to form a key operation step sequence, and determine the connection relationship between the previous key operation step and the next key operation step in the key operation step sequence.

[0069] The key operational sequence is acid leaching-dissolving – extraction – back-extraction. The connection is as follows: the output stream (acid leaching solution) from the acid leaching-dissolving step serves as the input stream for the extraction step; the two are connected via a solid-liquid separation device. The operating parameters for the extraction step must be adapted to the properties of the acid leaching solution (e.g., pH value, germanium concentration). Conversely, the output stream (loaded organic phase) from the extraction step serves as the input stream for the back-extraction step; the two are connected via an organic phase pump and pipelines. The operating parameters for the back-extraction step must match the properties of the loaded organic phase (e.g., germanium concentration, organic phase composition).

[0070] Step S1322: For the first key operation in the key operation sequence, extract all its candidate values. Based on each candidate value of the first key operation, sequentially associate all candidate values ​​of the second key operation in the key operation sequence to form a combined segment containing the candidate values ​​of the first two key operation steps.

[0071] The first critical operation is acid leaching, which is influenced by factors including temperature, acid concentration, and stirring rate. Each factor is sampled uniformly to obtain multiple candidate values ​​(e.g., 5 values ​​for temperature, 4 values ​​for acid concentration, and 3 values ​​for stirring rate). Therefore, there are 5 × 4 × 3 = 60 possible combinations of candidate values ​​for the acid leaching step. Based on one of the candidate value combinations (temperature value A, acid concentration value B, stirring rate value C), we associate it with all the candidate value combinations of the second critical operation (extraction) (multiple values ​​for extractant concentration, phase ratio, and temperature, assuming a total of 40 combinations), forming 60 × 40 = 2400 combination fragments containing the candidate values ​​from the first two steps.

[0072] Step S1323: Perform a rationality analysis on the combined segments: Based on the process simulation unit of the previous key operation in the process simulation structure, calculate the composition and properties of the output material when using the current candidate value; determine whether the composition and properties of the output material fall within the range of input materials allowed by the process simulation unit of the next key operation, and eliminate combined segments whose output materials are not within the allowed range of input.

[0073] Taking a specific combination of segments (acid leaching candidate values ​​A, B, C, and extraction candidate values ​​D, E, F) as an example, the acid leaching candidate values ​​are input into the acid leaching process simulation unit. The unit then calculates the composition and properties of the output material (acid leaching solution), including germanium ion concentration, pH value, and impurity ion concentration. According to the design specifications of the extraction process simulation unit, the allowed input material pH value range is a certain interval, the lower limit of germanium ion concentration is a certain value, and the upper limit of impurity ion concentration is a certain value. The calculated pH value, germanium ion concentration, and impurity ion concentration of the acid leaching solution are compared with the allowed ranges. If the pH value exceeds the range, the germanium ion concentration is below the lower limit, or the impurity ion concentration is above the upper limit, the combination of segments is deemed unreasonable and discarded.

[0074] Step S1324: Associate the combined fragments that have passed the rationality analysis with all the candidate values ​​of the third key operation in the key operation sequence to form an extended combined fragment that includes the candidate values ​​of the first three key operation sequences.

[0075] After the rationality analysis, it is assumed that there are 1800 remaining combination fragments. Each combination fragment is associated with all candidate value combinations of the third key operation step (back-extraction) (multiple values ​​for back-extraction agent concentration, temperature, and ratio, assuming a total of 30 combinations), forming 1800 × 30 = 54000 extended combination fragments containing candidate values ​​of the first three steps.

[0076] Step S1325: Repeat the association and rationality analysis operation, and associate all candidate values ​​of each subsequent key operation step in the key operation step sequence with the current extended combination fragment in turn, and remove extended combination fragments that do not meet the material input requirements.

[0077] A feasibility analysis was performed on 54,000 extended combination fragments: Extraction candidate values ​​from the combination fragments were input into the extraction process simulation unit. The composition and properties of the output material (supported organic phase) (e.g., germanium concentration, organic phase moisture content) were calculated and compared with the input material range allowed by the back-extraction process simulation unit (e.g., lower limit of germanium concentration, upper limit of moisture content). Extended combination fragments that did not meet the requirements were eliminated. Since there are only three key operational steps in the current sequence, this step yields a combination scheme containing candidate values ​​for all steps.

[0078] Step S1326: After all candidate values ​​for key operation steps are associated, multiple complete combination schemes containing candidate values ​​for all key operation steps are formed. The complete combination schemes are logically consistent and checked to see if there are any conflicting settings among the candidate values ​​of each key operation step. The conflicting settings include conflicting operating environment requirements or material property compatibility conflicts. Complete combination schemes with conflicting settings are eliminated, and logically consistent complete combination schemes are retained. The retained complete combination schemes are numbered and their information is organized. The candidate values ​​of influencing factors corresponding to each key operation step in each scheme are determined, and finally multiple combination schemes of influencing factors are generated.

[0079] The complete combination scheme includes all candidate values ​​of influencing factors in the three stages of acid leaching, extraction, and back-extraction. Logical consistency comparison includes: conflicting operating environment requirements, such as candidate values ​​for acid leaching requiring high temperatures (close to equipment limits) while candidate values ​​for extraction require low temperatures (close to ambient temperature), potentially leading to excessive energy consumption for cooling the acid leaching solution, constituting an unreasonable conflict; and conflicting material property compatibility, such as candidate values ​​for acid leaching resulting in high acidity (low pH) of the acid leaching solution, while candidate values ​​for extraction use an extractant sensitive to acid (easily decomposed at low pH), constituting a compatibility conflict. The complete combination scheme is screened through manual review and preset rules (such as establishing a conflict detection matrix). After eliminating conflicting schemes, the remaining schemes are numbered sequentially (e.g., Scheme 1, Scheme 2… Scheme N), and the names of the influencing factors and corresponding candidate values ​​for each stage in each scheme are compiled to form a list of influencing factor combination schemes.

[0080] Step S133: Input each combination of influencing factors into the influencing factor input interface of the process simulation structure, start the process simulation structure to run, so that each process simulation unit performs simulation operation according to the corresponding candidate values ​​of influencing factors. During the operation of the process simulation structure, the operation status data output by each process simulation unit is collected in real time through the status monitoring function of the overall control module. The operation status data reflects the execution status of key operation links under the corresponding candidate values ​​of influencing factors.

[0081] The combination of influencing factors for Scheme 1 (including acid leaching temperature T1, acid concentration C1, stirring rate R1; extractant concentration E1, phase ratio V1, temperature T2; back-extraction agent concentration C2, temperature T3, phase ratio V2) is input through the human-machine interface of the process simulation structure. The overall control module assigns the candidate values ​​of each influencing factor to the influencing factor input interface of the corresponding process simulation unit. After the simulation starts, the acid leaching unit performs simulation calculations according to T1, C1, and R1; the extraction unit performs calculations according to E1, V1, and T2; and the back-extraction unit performs calculations according to C2, T3, and V2. The status monitoring function of the overall control module collects real-time operating status data through the communication interface with each unit, such as the temperature distribution in the reactor of the acid leaching unit, the stirring shaft torque, and the inlet and outlet flow rates; the stirring power in the mixing chamber of the extraction unit, the interface height in the clarification chamber, and the organic phase circulation flow rate; and the pump inlet and outlet pressure, heat exchanger heat load, and stirring paddle speed of the back-extraction unit.

[0082] Step S134: Synchronously collect the material conversion result data output by each process simulation unit through the material conversion result output interface. The material conversion result data reflects the material conversion effect of key operation links under the corresponding candidate values ​​of influencing factors.

[0083] While collecting operational status data, material conversion result data is simultaneously collected through the material conversion result output interface. The acid leaching unit output data includes: germanium ion concentration in the acid leaching solution, concentration of major impurity ions (iron, copper), residual germanium content in the acid leaching residue, acid leaching reaction time, and acid consumption. The extraction unit output data includes: germanium concentration in the loaded organic phase, germanium concentration in the raffinate (raffinate rate), extractant consumption, and phase separation time. The back-extraction unit output data includes: germanium concentration in the back-extraction solution, germanium concentration in the regenerated organic phase (back-extraction rate), back-extraction agent consumption, and pH value of the back-extraction solution. All of the above data is stored in association with the influencing factor combination scheme number and simulation timestamp through the data acquisition unit.

[0084] Step S135: Link and bind the operational status data and material conversion result data of all key operation links corresponding to the same combination of influencing factors to form a single set of simulation data records. Summarize all the single sets of simulation data records to form an initial set of recovery process simulation data. Perform format unification processing on the data in the initial set of recovery process simulation data, convert each operational status data and material conversion result data into a predefined standardized data format, and obtain the final set of recovery process simulation data.

[0085] After the simulation of the same combination of influencing factors (such as Scheme 1), operational status data and material conversion result data for the three stages of acid leaching, extraction, and back-extraction are generated. These data are linked and bound according to the structure "Scheme Number - Stage Name - Data Type - Data Value," for example, Scheme 1 - Acid Leaching - Operational Status - Stirring Shaft Torque - Value A, Scheme 1 - Extraction - Material Conversion Result - Residual Rate - Value B, forming a single set of simulation data records. The single sets of records for all combinations of influencing factors are summarized to obtain the initial recovery process simulation data set. Format standardization includes: unifying the unit of all concentration data to moles per liter, the unit of time to minutes, and the unit of flow rate to cubic meters per hour; converting text-based data (such as equipment status descriptions "normal / abnormal") to numerical data (1 / 0); marking missing data (e.g., using specific symbols); and standardizing data precision (retaining three decimal places), ultimately forming a standardized recovery process simulation data set.

[0086] Step S140: Use the simulation dataset of the recycling process to train the target machine learning model to establish a mapping relationship from influencing factors to material transformation results data, and generate an optimized combination of influencing factors based on the mapping relationship.

[0087] Step S141: Split the single set of simulated data records in the simulation data set of the recycling process, take the candidate values ​​of the influencing factors in each single set of simulated data records as input data, and take the corresponding material conversion result data as output labels to form multiple sets of training samples. Divide the multiple sets of training samples into training sample subset, validation sample subset and test sample subset according to a preset ratio. The training sample subset is used for parameter learning of the target machine learning model, the validation sample subset is used for parameter adjustment of the target machine learning model during the training process, and the test sample subset is used for the final performance evaluation of the target machine learning model.

[0088] Single sets of simulated data records are extracted from the dataset of the recycling process simulation, and input data and output labels are separated. The input data consists of all candidate values ​​of influencing factors in the record, arranged in the order of the stages: [acid leaching temperature, acid concentration, acid leaching stirring rate, extractant concentration, extraction ratio, extraction temperature, back-extraction agent concentration, back-extraction temperature, back-extraction ratio]. The output labels are the corresponding material conversion results, with key indicators selected: [acid leaching germanium dissolution rate, extraction germanium extraction rate, back-extraction germanium back-extraction rate, total germanium recovery rate]. Each training sample consists of an input data vector and an output label vector. All training samples are divided into a training sample subset (70%), a validation sample subset (20%), and a test sample subset (10%) according to a preset ratio of 7:2:1. Stratified sampling is used in the partitioning process to ensure that the output label distribution of each subset is consistent with the original dataset, avoiding sample bias.

[0089] Step S142: Construct the network structure of the target machine learning model. The network structure includes an input layer, a feature mapping layer, a feature fusion layer, a prediction output layer, and an optimization recommendation layer. The input layer is used to receive candidate values ​​of influencing factors. The feature mapping layer is used to extract features and transform dimensions of the input data. The feature fusion layer is used to integrate extracted features from different dimensions. The prediction output layer is used to output predicted values ​​of material transformation result data. The optimization recommendation layer is used to output optimized combinations of influencing factors based on the prediction results.

[0090] The target machine learning model employs a deep neural network structure. The input layer contains the same number of neurons as the influencing factors (e.g., 9 neurons for 9 influencing factors), with each neuron receiving a standardized candidate value for the influencing factor (e.g., normalizing the temperature range to the [0,1] interval). The feature mapping layer contains three parallel feature extraction sub-networks (sub-network 1, sub-network 2, and sub-network 3), each extracting features for different types of influencing factors: Sub-network 1 handles temperature-related factors (acid leaching temperature, extraction temperature, and back-extraction temperature), containing two fully connected layers and a ReLU activation function; Sub-network 2 handles concentration-related factors (acid concentration, extractant concentration, and back-extractant concentration), containing two convolutional layers (1D convolutions for extracting local features) and a pooling layer; Sub-network 3 handles rate / proportion-related factors (stirring rate and ratio), containing a Long Short-Term Memory (LSTM) layer (simulating the influence of dynamic processes). The feature fusion layer receives the output features from the three sub-networks, calculates the weights of the features from each sub-network using an attention mechanism, and then concatenates them to obtain a weighted feature vector. The prediction output layer consists of two fully connected layers. The first fully connected layer maps the integrated feature vector to a high-dimensional space, and the second fully connected layer outputs a predicted value equal to the number of output labels (4 neurons, corresponding to 4 key indicators). The optimization recommendation layer takes the result of the prediction output layer as input and combines it with a preset optimization objective (such as maximizing the total germanium recovery rate and minimizing energy consumption) to search for the optimal combination of influencing factors using a genetic algorithm.

[0091] Step S143: Configure multiple parallel feature extraction channels for the feature mapping layer. Each feature extraction channel performs exclusive feature extraction for the input data of a certain type of influencing factor, so that the features of different types of influencing factors are fully captured. Set an adaptive fusion weight generation operation for the feature fusion layer. This adaptive fusion weight generation operation dynamically calculates and updates the fusion weights of the output features of different feature extraction channels based on the gradient information of backpropagation during training.

[0092] The parallel feature extraction channels of the feature mapping layer correspond one-to-one with the types of influencing factors. Taking the temperature-related influencing factor channel as an example, the input consists of three temperature values ​​(acid leaching, extraction, and back-extraction temperatures). The first fully connected layer maps the input dimension from 3 to a higher dimension (e.g., 3-16), introducing a nonlinear transformation through the ReLU activation function to extract the nonlinear influence of temperature on reaction kinetics. The second fully connected layer further maps the dimension from 16 to 8, extracting more abstract temperature features. The concentration-related channel takes three concentration values ​​as input. The first 1D convolutional layer uses multiple convolutional kernels (e.g., eight kernels of size 2) to perform convolution operations on the input sequence, extracting the interaction features between concentrations. The pooling layer uses max pooling to retain key features and reduce dimensionality. The second convolutional layer further extracts higher-order concentration features. The rate / proportion-related channel takes two rate / proportion values ​​as input. The LSTM layer learns the dynamic influence features of stirring rate and time through a gating mechanism (although the input is a static value, the LSTM can simulate its cumulative effect on the dynamic process). In the adaptive fusion weight generation operation of the feature fusion layer, the output features of each feature extraction channel first pass through a fully connected layer to obtain a channel importance score. The score value is then normalized using the softmax function to obtain the fusion weight. During training, when the error of the prediction output layer is backpropagated, the gradient of the channel importance score is calculated. The weights of the fully connected layer are updated using the gradient descent algorithm, thereby dynamically adjusting the fusion weights of each channel feature, making the model pay more attention to the feature channels that contribute more to the prediction result.

[0093] Step S144: Input the input data of the training sample subset into the input layer of the target machine learning model. Extract features through multiple feature extraction channels of the feature mapping layer to obtain the specific features corresponding to each influencing factor. Input each specific feature into the feature fusion layer. Generate fusion weights through adaptive fusion weight generation operation. Perform weighted fusion processing on the specific features to obtain a comprehensive feature vector. Input the comprehensive feature vector into the prediction output layer. Calculate the predicted value of the material conversion result data through the mapping function of the prediction output layer. Calculate the difference between the predicted value of the material conversion result data and the output label corresponding to the training sample subset to obtain the training error.

[0094] The input data of a subset of training samples (such as the values ​​of 9 influencing factors for a sample) is input to the input layer, standardized, and then distributed to three parallel channels in the feature mapping layer. Temperature-related channels extract temperature-specific feature vectors (8 dimensions), concentration-related channels extract concentration-specific feature vectors (10 dimensions), and rate / proportion-related channels extract rate-specific feature vectors (6 dimensions). The feature fusion layer processes these three feature vectors: the importance score of the temperature channel is softmaxed to obtain weight W1, the concentration channel to W2, and the rate channel to W3 (W1+W2+W3=1). These weighted vectors are then fused into a comprehensive feature vector = W1 × temperature feature + W2 × concentration feature + W3 × rate feature (vector concatenation). This comprehensive feature vector is input to the prediction output layer. The first fully connected layer (32 hidden neurons, ReLU activation function) maps it to a 32-dimensional vector. The second fully connected layer (4 neurons, linear activation function) outputs four predicted values ​​for material conversion (such as predicted germanium dissolution rate from acid leaching, predicted total germanium recovery rate, etc.). The mean squared error (MSE) between the predicted value and the corresponding output label (actual value in the simulated data) of the training sample subset is calculated as the training error.

[0095] Step S145: Based on the training error, adjust the network parameters of the target machine learning model through backpropagation, including the weight parameters of the feature extraction channel, the fusion weight parameters of the feature fusion layer, and the mapping function parameters of the prediction output layer.

[0096] The network parameters are adjusted using backpropagation (BP) and the Adam optimizer. The training error (MSE) is backpropagated from the prediction output layer, calculating the gradient of each layer's parameters with respect to the error. For the weights and biases of the fully connected layers in the prediction output layer, the gradient is calculated using the chain rule, and the parameters are updated using gradient descent (parameter = parameter - learning rate × gradient). The fusion weights of the feature fusion layer (fully connected layer weights for channel importance scoring) are updated based on their gradient contribution to the comprehensive feature vector, increasing the weights of channels with a greater impact on the error. The parameters of the feature extraction channels (sub-networks) (e.g., the fully connected layer weights of sub-network 1, the convolutional kernel parameters of sub-network 2, and the LSTM unit weights of sub-network 3) are also updated using gradient backpropagation; for example, the gradient of the convolutional kernel parameters is calculated using the backpropagation formula for convolution operations. An adaptive adjustment strategy is used for the learning rate. The initial learning rate is set to a certain value, and as the training epochs increase, the learning rate is halved when the validation error no longer decreases. Regularization measures include adding L2 regularization terms (weight decay) to each fully connected layer to prevent overfitting.

[0097] Step S146: Repeatedly input the training sample subset into the target machine learning model for iterative training. In each iteration, the prediction accuracy of the target machine learning model is verified using the validation sample subset. During the iterative training process, the prediction accuracy of the validation sample subset is continuously monitored. When the improvement of the prediction accuracy of the validation sample subset is lower than the preset threshold in N consecutive iterations, the training of the target machine learning model is stopped.

[0098] Set the maximum number of iterations (e.g., 1000) and an early stopping mechanism (N=20). In each epoch, input a subset of training samples into the model in batches (batchsize=32) to update parameters during forward and backward propagation. After each iteration, input a subset of validation samples into the model and calculate the validation error (MSE) and prediction accuracy metrics (e.g., coefficient of determination R²). Continuously monitor the validation accuracy. If the improvement in validation accuracy is less than a preset threshold (e.g., 0.01%) for 20 consecutive iterations, the model is considered converged, and training is stopped. During training, record the training error and validation error for each iteration and plot the learning curve to analyze whether the model is overfitting (e.g., if the training error continuously decreases while the validation error increases, it indicates overfitting). If overfitting occurs, adjust the regularization coefficient or add a dropout layer (add a dropout layer with a dropout rate of 0.3 between the feature mapping layer and the fusion layer).

[0099] Step S147: Use a subset of test samples to perform performance testing on the trained target machine learning model, calculate the root mean square error or mean absolute error between the model's prediction results on the subset of test samples and the true labels, and adjust and optimize the filtering logic of the recommendation layer based on the test results of the subset of test samples; the optimization recommendation layer uses the germanium element recovery rate or target product yield in the predicted value of the material conversion result data as the optimization objective, selects the combination of influencing factors that maximizes the optimization objective from the training samples, and obtains the target machine learning model after fine-tuning.

[0100] Input the subset of test samples into the trained target machine learning model to obtain the predicted output value for each sample. Calculate the root mean square error (RMSE) and mean absolute error (MAE) between the predicted value and the true label (the material conversion result data of the test sample) to evaluate the model's generalization ability. If both RMSE and MAE are lower than the preset acceptable values ​​(set according to process requirements), the model performance is qualified; otherwise, return to step S142 to adjust the network structure (such as increasing the number of neurons or adjusting the subnet type) or step S141 to re-divide the sample subset. Adjust and optimize the filtering logic of the recommendation layer based on the test results. For example, if the test finds that the model's prediction accuracy for the back-extraction rate is low, reduce the weight of the back-extraction rate in the optimization objective, or add constraints (the back-extraction rate should not be lower than a certain threshold). The optimization recommendation layer aims to maximize the total germanium recovery rate. A genetic algorithm is used to search within the range of influencing factor values. The initial population consists of combinations of influencing factors from the training samples, and the fitness function is the model's predicted total germanium recovery rate. Iterative evolution is achieved through selection (roulette wheel selection), crossover (single-point crossover), and mutation (randomly perturbing a factor's value). After a preset number of generations (e.g., 50 generations), the individual with the highest fitness is output as the optimized influencing factor combination. The parameters of the optimization recommendation layer (such as the crossover rate and mutation rate of the genetic algorithm) are fine-tuned to ensure that the output optimized combinations fall within the range of each factor's value and are logically consistent, ultimately yielding the trained target machine learning model.

[0101] Step S150: Apply the optimized combination of influencing factors output by the target machine learning model to each key operation in the actual germane residue recovery process, execute the optimized germane residue recovery process, and obtain germane-related recovery products.

[0102] Step S151: Receive the optimized combination of influencing factors output by the target machine learning model, analyze the specific values ​​of the influencing factors corresponding to each key operation step in the optimized combination of influencing factors, and for each key operation step, associate the analyzed specific values ​​of the influencing factors with the actual control parameters of the key operation step to determine the adjustment direction and adjustment target of the actual control parameters.

[0103] The optimized influencing factor combination output by the target machine learning model is transmitted in structured data form (such as JSON format), including the scheme identifier, the name of the influencing factor for each key operation step, and the corresponding optimized value, for example: {"Scheme ID":"Opt1","Acid Immersion":{"Temperature":T_opt,"Acid Concentration":C_opt,"Stirring Rate":R_opt},"Extraction":{"Extractant Concentration":E_opt,"Comparison":V_opt,"Temperature":T2_opt},"Back Extraction":{"Back Extractant Concentration":C2_opt,"Temperature":T3_opt,"Comparison":V3_opt}}. This data is parsed to extract the specific values ​​of the influencing factors for each key operation step. The optimized temperature value T_opt for the acid immersion step is associated with the actual temperature control parameter of the acid immersion reactor (such as the setpoint of the temperature control system). The current actual control parameter is T_current, and the adjustment direction is T_opt - T_current (positive values ​​indicate the need for heating, negative values ​​indicate the need for cooling). The adjustment goal is to control the actual temperature within the range of T_opt ± allowable fluctuation. Similarly, the optimal acid concentration C_opt is associated with the control parameters of the acid solution preparation system (such as the flow rate setting of the acid pump), and the target concentration is achieved by adjusting the mixing ratio of acid and water; the stirring rate R_opt is associated with the frequency setting of the frequency converter of the stirring motor to determine the frequency adjustment target.

[0104] Step S152: According to the execution order of each key operation step, adjust the parameters of the control equipment of the key operation steps in the actual germane residue recovery process in sequence, so that the control equipment of each key operation step operates according to the specific values ​​of the corresponding influencing factors. During the parameter adjustment process, monitor the actual operating status of each key operation step in real time and collect the actual operating status data. The actual operating status data and the operating status data in the process simulation structure adopt the same data collection dimension.

[0105] Step S1521: Obtain a list of control equipment corresponding to each key operation in the actual germane residue recovery process, and determine the function and parameter adjustment range of each control equipment.

[0106] The control equipment list for the acid leaching process includes: a reactor temperature controller (function: regulates the reactor temperature via steam heating or cold water cooling; adjustment range: room temperature to a specific temperature value), an acid solution metering pump (function: delivers concentrated acid to the reactor; adjustment range: 0 to maximum flow rate), a stirring motor frequency converter (function: controls the stirring paddle speed; adjustment range: 0 to rated speed), and a pH meter (function: monitors the pH value of the acid leaching solution, used to indirectly verify acid concentration). The control equipment for the extraction process includes: an extractant tank outlet pump (adjusts the extractant flow rate and controls the concentration), a mixing and clarification tank interface controller (adjusts the phase ratio), and an extraction temperature heat exchanger temperature control valve (adjusts the temperature). The control equipment for the back-extraction process includes: a back-extractant metering pump, a back-extraction stirrer frequency converter, and an organic / aqueous phase flow control valve. The parameter adjustment range for each piece of equipment is determined by consulting the equipment nameplate and operation manual; for example, the adjustment range for the acid solution metering pump is 0-1000 liters / hour.

[0107] Step S1522: Map the specific values ​​of the influencing factors in each key operation link of the optimized influencing factor combination to the adjustable parameters of the control equipment, and establish a correspondence table between influencing factors and equipment parameters.

[0108] Establish a corresponding relationship table, for example: Acid leaching temperature T_opt—temperature controller setpoint; Acid concentration C_opt—ratio of acid metering pump flow rate Q1 to water metering pump flow rate Q2 (C_opt=Q1×concentrated acid concentration / (Q1+Q2)); Stirring rate R_opt—frequency converter frequency f (R_opt=k×f, k is the speed-frequency conversion coefficient). Extractant concentration E_opt—ratio of extractant pump flow rate Qe to diluent pump flow rate Qs (E_opt=Qe / (Qe+Qs)); Phase ratio V_opt—ratio of aqueous phase flow rate Qw to organic phase flow rate Qo (V_opt=Qw / Qo); Extraction temperature T2_opt—heat exchanger temperature control valve opening (opening is positively correlated with temperature). Store the above correspondence in the control system database in tabular form for querying during parameter adjustment.

[0109] Step S1523: Construct a parameter adjustment execution plan according to the execution sequence of key operation steps, and determine the adjustment time window and adjustment sequence of the control equipment for each key operation step.

[0110] The key operational steps are executed in the following sequence: acid leaching – extraction – back-extraction. Parameter adjustments follow the "upstream priority" principle to avoid downstream adjustments interfering with upstream operations. The parameter adjustment window for the acid leaching step is set 30 minutes before the new batch of slag is fed, ensuring stable adjustment before feeding. Adjustments for the extraction step begin 10 minutes after the acid leaching step is completed and running stably (waiting for the acid leaching solution to stabilize). Adjustments for the back-extraction step begin 15 minutes after the extraction step is completed and running stably (waiting for the loaded organic phase to stabilize). The equipment adjustment sequence within each step is as follows: For the acid leaching step, first adjust the acid concentration (acid solution needs to be prepared in advance), then adjust the stirring rate, and finally adjust the temperature (heating / cooling takes a long time); for the extraction step, first adjust the phase ratio (by adjusting the aqueous / organic phase flow rate), then adjust the extractant concentration, and finally adjust the temperature. Adjustment time windows must avoid unstable periods of process operation (such as feeding and discharging).

[0111] Step S1524: Within the adjustment time window of the first critical operation step, find the target value of the equipment parameter corresponding to the specific value of the influencing factor of the critical operation step according to the corresponding relationship table.

[0112] The first key operational step is acid leaching. Within its adjustment time window (e.g., t=0), the control system retrieves the corresponding relationship table from the database. Based on the optimized acid leaching temperature value T_opt, it looks up the target setpoint value Ts of the temperature controller (Ts=T_opt). Based on the optimized acid concentration value C_opt, it calculates the target values ​​of the acid metering pump flow rate Q1 and the water metering pump flow rate Q2 (based on the current total liquid flow requirement, Q1 and Q2 are solved using the formula C_opt=(Q1×C_concentrated acid) / (Q1+Q2)). Based on the optimized stirring rate value R_opt, it calculates the target value f_target of the frequency converter (f_target=R_opt / k). These target values ​​of the equipment parameters are then sent to the corresponding control equipment controllers.

[0113] Step S1525: Operate the control equipment for this key operation step, and gradually adjust the equipment parameters from the current value to the target value. During the parameter adjustment process, collect the equipment operating parameters in real time through the feedback monitoring function of the control equipment, and check whether the parameter adjustment is accurate and in place.

[0114] Acid leaching temperature control: The temperature controller receives the setpoint Ts and adjusts the steam valve opening using a PID control algorithm. The current actual temperature is fed back in real time by an in-vessel temperature sensor. The control output approximates the actual temperature from T_current to Ts. The adjustment process uses a stepped heating / cooling method, with each adjustment not exceeding the maximum allowable step size (e.g., 2℃ / minute) to avoid drastic temperature fluctuations. Equipment operating parameters such as steam valve opening and heating power are collected in real time to determine if the response is as expected. Acid concentration control: The acid and water metering pumps receive flow target values ​​Q1 and Q2, respectively. The pump speed is changed by adjusting the frequency converter. The flow meter provides real-time feedback of the actual flow rate, which is compared with the target value. Fine-tuning is performed when the deviation exceeds the allowable range (e.g., ±2%). Stirring rate control: The frequency converter receives the frequency target value f_target and gradually adjusts the output frequency. The motor speed sensor provides feedback of the actual speed, which is compared with R_opt to ensure proper adjustment (deviation < ±1%).

[0115] Step S1526: After the control equipment parameters of the first critical operation are adjusted to the target value and run stably, the adjustment time window for the second critical operation begins.

[0116] After the parameters of each control device in the acid leaching process are adjusted to the target values, the operating status data (temperature, concentration, stirring rate) are continuously monitored. If the values ​​remain within the target value ± allowable fluctuation range within a preset stabilization time (e.g., 15 minutes), the acid leaching process is considered to have completed adjustment and is operating stably. At this point, the control system triggers the start signal for the adjustment time window of the second critical operation (extraction), and enters the parameter adjustment process for the extraction process.

[0117] Step S1527: Repeat the operation process of parameter search, equipment operation, and feedback monitoring, and adjust the parameters of the control equipment in the second key operation step to the corresponding target values.

[0118] The parameter adjustment process for the extraction stage is similar to that for acid leaching: Based on the corresponding relationship table, find the aqueous phase flow rate Qw_target and organic phase flow rate Qo_target corresponding to V_opt, and control the flow rate by adjusting the opening of the raffinate outlet valve and the loaded organic phase outlet valve; find the extractant pump flow rate Qe_target and diluent pump flow rate Qs_target corresponding to the extractant concentration E_opt, and adjust the speed of the corresponding pumps; find the target value for the heat exchanger temperature control valve opening corresponding to the extraction temperature T2_opt, and adjust the valve opening. During the adjustment process, the actual parameters are fed back in real time through flow meters, concentration meters, and temperature sensors to verify the adjustment accuracy. Once stability is ensured, proceed to the next stage.

[0119] Step S1528: According to the parameter adjustment execution plan, continuously adjust the parameters of the control equipment for each subsequent key operation link, so that each control equipment operates according to the specific values ​​of the corresponding influencing factors. After the parameter adjustment of the control equipment in all key operation links is completed, conduct an overall inspection of the operating parameters of all control equipment to check for any parameter deviations or omissions in adjustment, so that each key operation link of the actual germane reactor residue recovery process can operate according to the optimized combination of influencing factors.

[0120] After completing the adjustment of the extraction stage, the parameter adjustment of the back-extraction stage is initiated as planned, following the same process. Once all key operational adjustments are completed, the control system performs a comprehensive inspection: it traverses each control device, reads the current setpoint and actual feedback values, compares them with the target values ​​in the optimized combination of influencing factors, and generates an inspection report. If a device parameter deviation exceeds a threshold (e.g., a 3°C deviation in back-extraction temperature), the parameter adjustment process for that device is re-executed; if any adjustments are missed (e.g., a corresponding device parameter is not set for a certain influencing factor), supplementary adjustments are made. After the inspection passes, the actual germane reactor residue recovery process enters the optimized operating condition phase.

[0121] Step S153: Calculate the difference between the actual operating status data and the operating status data of the corresponding optimized influencing factor combination in the simulation data set of the recycling process. When the absolute value of the difference exceeds the preset allowable deviation, generate a parameter fine-tuning instruction based on the difference to correct the control equipment parameters of the key operation links. Repeat the data acquisition, difference calculation and parameter correction operations until the absolute value of the difference is lower than the preset allowable deviation.

[0122] After the actual process is run according to the optimized combination of influencing factors, the data acquisition system collects the actual operating status data of each key operation link at a preset frequency (e.g., once every 5 minutes). This includes data such as the actual temperature T_actual of the acid leaching reactor, the torque M_actual of the stirring shaft; the actual phase ratio V_actual of the extraction mixing chamber; and the actual temperature T3_actual of the back-extraction. Simulated operating status data (T_sim, M_sim, V_sim, T3_sim, etc.) corresponding to the optimized combination of influencing factors are retrieved from the simulation data set of the recovery process. The difference between each operating status parameter is calculated: Δ=T_actual-T_sim (temperature), Δ=M_actual-M_sim (torque), etc. Preset allowable deviations are set according to process requirements (e.g., allowable temperature deviation ±1℃, allowable torque deviation ±5%). If the |Δ| of a certain parameter exceeds the allowable deviation (e.g., T_actual - T_sim = 2.5℃ > 1℃), a parameter fine-tuning instruction is generated: For temperature deviation, the correction amount ΔT_corr = Δ × correction coefficient (determined according to the proportional term coefficient of PID control) is calculated and sent to the temperature controller to adjust the setpoint (Ts_new = Ts_old + ΔT_corr); for relative deviation, correction is achieved by fine-tuning the flow rate of the aqueous or organic phase. The process of data acquisition (new actual operating data), difference calculation (new Δ), and parameter correction is repeated until the |Δ| of all operating status parameters is lower than the preset allowable deviation.

[0123] Step S154: After the parameters of all key operation links have been adjusted and the actual operating status has stabilized, start the continuous operation of the optimized germane residue recovery process. During the continuous operation, collect the actual material conversion result data of each key operation link at the set time interval. The actual material conversion result data and the material conversion result data in the process simulation structure adopt the same data characterization standard.

[0124] After all key operational parameters were adjusted and verified for deviation, the actual germane residue recovery process was started continuously: residue was fed at a set rate, acid solution was added continuously, and each separation unit operated synchronously. During continuous operation, the data acquisition system collected actual material conversion results data at set time intervals (e.g., once per hour): in the acid leaching stage, the germanium ion concentration of the acid leaching solution was collected (detected by atomic absorption spectrometry), and the residual germanium content of the acid leaching residue was analyzed by X-ray fluorescence spectrometry; in the extraction stage, the germanium concentration of the raffinate was collected (detected by spectrophotometry), and the germanium concentration of the loaded organic phase was detected after centrifugation; in the back-extraction stage, the germanium concentration of the back-extraction solution was collected (detected by ICP-MS), and the germanium concentration of the regenerated organic phase was collected. The data characterization standards were consistent with the process simulation structure: the concentration unit was moles per liter, the detection method was performed according to national standards, the data retained three significant figures, and outliers (exceeding 3 times the standard deviation) were marked and stored separately.

[0125] Step S155: Input the continuously collected actual material conversion result data into the prediction output layer of the target machine learning model, calculate the error between the predicted value of the material conversion result data and the actual material conversion result data; continuously execute the optimized germane residue recovery process until the preset amount of germane residue is completed, collect the material output at the end of the process, and obtain germane-related recovery products after separation and purification.

[0126] Continuously collected actual material conversion result data (such as germanium concentration in the back-extraction solution, C_actual) is input into the prediction output layer of the target machine learning model. The model outputs the corresponding predicted material conversion result value, C_pred, based on the current influencing factor values ​​(optimized combination). The calculation error is E = C_actual - C_pred. If the absolute value of E is within the preset range (e.g., ±5%), it indicates that the model prediction is accurate and the process is running stably. If it exceeds the range, the cause is analyzed (e.g., changes in raw material properties, equipment aging), and if necessary, the process is returned to step S140 to retrain the model. The process continues to run until the cumulative amount of treated residue reaches the preset value (e.g., batch processing volume or daily processing volume). The final output of the process is a back-extraction solution (high-concentration germanium solution), which undergoes separation and purification: ammonia is added to adjust the pH to neutral, generating germanium hydroxide precipitate; the precipitate is filtered, washed, and dried under an inert atmosphere; the dried germanium hydroxide is calcined and decomposed in a high-temperature furnace to obtain germanium dioxide; germanium dioxide reacts with a reducing agent (such as hydrogen) at a specific temperature and pressure to generate germanane gas; the germanane gas is condensed and purified by distillation to obtain high-purity germanane-related recovery products (such as germanane with a purity of 99.999%).

[0127] Throughout the implementation process, the data collected includes some process parameters and material property data, but does not involve any sensitive personal privacy data. For equipment operation logs and material composition data generated during the process, access control technologies (such as setting user permissions and differentiating between administrator, operator, and maintenance personnel permissions) and data encryption technologies (such as database encryption and SSL encryption for transmitted data) are used to protect privacy and prevent leakage, ensuring that the data is used only for process optimization and quality control and is not disclosed to external parties.

[0128] Furthermore, Figure 2 A schematic diagram of the hardware structure of a germanane reactor slag recovery system 100 combining machine learning and process simulation for implementing the methods provided in the embodiments of this application is shown. Figure 2As shown, the germane residue recovery system 100, which combines machine learning and process simulation, may include at least one processor 102 (the processor 102 may be, but is not limited to, a microprocessor MCU or a programmable logic device FPGA, etc.), a memory 104 for storing data, a transmission device 106 for communication functions, and a controller 108. Those skilled in the art will understand that... Figure 2 The structure shown is for illustrative purposes only and does not limit the structure of the germane residue recovery system 100 that combines machine learning and process simulation. For example, the germane residue recovery system 100 that combines machine learning and process simulation may also include more than Figure 2 The more or fewer components shown, or having the same Figure 2 The different configurations shown.

[0129] The memory 104 can be used to store software programs and modules for application software, such as the program instructions corresponding to the method embodiments described above in this application. The processor 102 executes various functional applications and data processing by running the software programs and modules stored in the memory 104, thereby realizing the aforementioned method for germanane reactor slag recovery that combines machine learning and process simulation. The transmission device 106 is used to acquire or send data via a network.

[0130] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

Claims

1. A method for recovering germanane reactor residue combining machine learning and process simulation, characterized in that, The method includes: Identify the key operational steps in the germane reactor residue recovery process and the corresponding influencing factors for each key operational step. A process simulation structure is constructed that includes each key operation step and corresponding influencing factors. The process simulation structure is used to reproduce the material conversion path in the germane residue recovery process. Based on the process simulation structure, a set of simulation data of the recycling process under different combinations of influencing factors is generated. The set of simulation data of the recycling process includes the operating status data and material conversion result data of each key operation link. A target machine learning model is trained using a dataset of simulated recycling processes to establish a mapping relationship between influencing factors and material transformation results, and an optimized combination of influencing factors is generated based on this mapping relationship. The optimized combination of influencing factors output by the target machine learning model is applied to each key operation step of the actual germane residue recovery process. The optimized germane residue recovery process is then executed to obtain germane-related recovered products.

2. The germanane reactor slag recovery method combining machine learning and process simulation according to claim 1, characterized in that, The construction of the process simulation structure, which includes each key operational step and its corresponding influencing factors, includes: The overall process route of germane reactor slag recovery process is analyzed to determine the execution sequence and material transfer relationship between each key operation. The material transfer relationship is used to represent the association form in which the output material of the previous key operation is used as the input material of the next key operation. For each key operation step, the factors affecting the material conversion efficiency of that key operation step are extracted. These factors include operating environment-related factors and material property-related factors. Operating environment-related factors are directly related to the execution conditions of the key operation step, while material property-related factors are related to the characteristics of the materials involved in the key operation step. A dedicated process simulation unit is established for each key operation step. The process simulation unit contains the mathematical relationship of material transformation corresponding to the key operation step. The mathematical relationship of material transformation is formed by coupling the kinetic equation describing the chemical reaction rate, the mass transfer equation describing the diffusion of matter, and the heat transfer equation describing the heat exchange. Based on the execution order and material transfer relationship between each key operation, each dedicated process simulation unit is connected in the corresponding order to form a preliminary process simulation structure. Each process simulation unit in the preliminary process simulation structure is configured with an input interface for influencing factors and an output interface for material conversion results. The input interface for influencing factors is used to receive the specific values ​​of the corresponding influencing factors, and the output interface for material conversion results is used to output the corresponding operating status data and material conversion result data of the process simulation unit. Define the material transfer interface specification between each process simulation unit. This material transfer interface specification specifies the material data format, data type and transmission protocol output by the material conversion result output interface of the previous process simulation unit, and matches it with the requirements of the material input port of the next process simulation unit. An overall control module for constructing a process simulation structure is provided. The overall control module sequentially calls and executes the calculation programs of each process simulation unit according to a preset sequence of key operation steps. A data acquisition unit is configured in the overall control module. The data acquisition unit reads the running status data of each process simulation unit in real time according to a preset sampling frequency, and stores the running status data in the database along with the corresponding influencing factor values ​​and timestamps. The material balance verification operation is retrieved to the process simulation structure. The material balance verification operation calculates the difference between the total mass of all input substances and the total mass of all output substances in the process. Based on the verification results of the material balance check operation, when there is a material non-conservation problem, the parameter and interface adjustment sub-step is executed to adjust the material transformation mathematical relationship parameters and material transfer interface specifications of the relevant process simulation units; the process simulation structure and material balance check operation are run repeatedly until the verification results of the material balance check operation meet the preset tolerance range. At this time, the process simulation structure is determined to be completed, forming the final process simulation structure containing each key operation link and corresponding influencing factors.

3. The germanane reactor slag recovery method combining machine learning and process simulation according to claim 1, characterized in that, The dataset of simulated recovery processes under different combinations of influencing factors, generated based on the process simulation structure, includes: For each key operational step, the influencing factors are identified, and the value range of each influencing factor is extracted. The value range is determined based on the actual process feasibility and safe operation boundaries of germane residue recovery. Multiple candidate values ​​are extracted from the range of values ​​of each influencing factor using a uniform sampling method. The candidate values ​​of different key operation steps are combined according to the execution order of each key operation step to generate multiple influencing factor combination schemes. Each influencing factor combination scheme contains the candidate values ​​of influencing factors corresponding to all key operation steps. Each combination of influencing factors is input into the influencing factor input interface of the process simulation structure, and the process simulation structure is started to run. Each process simulation unit performs simulation operations according to the corresponding candidate values ​​of influencing factors. During the operation of the process simulation structure, the running status data output by each process simulation unit is collected in real time through the status monitoring function of the overall control module. The running status data reflects the execution status of key operation links under the corresponding candidate values ​​of influencing factors. The material conversion result data output by each process simulation unit through the material conversion result output interface is collected synchronously. The material conversion result data reflects the material conversion effect of key operation links under the corresponding candidate values ​​of influencing factors. The operational status data and material conversion result data of all key operational links corresponding to the same combination of influencing factors are linked and bound to form a single set of simulation data records. All single sets of simulation data records are summarized to form an initial set of recovery process simulation data. The data in the initial set of recovery process simulation data are processed to unify the format, and the operational status data and material conversion result data are converted into a predefined standardized data format to obtain the final set of recovery process simulation data.

4. The germanane reactor slag recovery method combining machine learning and process simulation according to claim 1, characterized in that, The method of training a target machine learning model using a dataset of simulated recycling processes to establish a mapping relationship from influencing factors to material transformation results, and generating an optimized combination of influencing factors based on this mapping relationship, includes: The simulation data records in the simulation dataset of the recycling process are split into multiple sets. The candidate values ​​of the influencing factors in each set of simulation data records are used as input data, and the corresponding material conversion result data are used as output labels to form multiple sets of training samples. The multiple sets of training samples are divided into training sample subset, validation sample subset and test sample subset according to a preset ratio. The training sample subset is used for parameter learning of the target machine learning model, the validation sample subset is used for parameter adjustment of the target machine learning model during the training process, and the test sample subset is used for the final performance evaluation of the target machine learning model. A network structure for constructing a target machine learning model is provided, comprising an input layer, a feature mapping layer, a feature fusion layer, a prediction output layer, and an optimization recommendation layer. The input layer receives candidate values ​​of influencing factors, the feature mapping layer extracts features and transforms dimensions of the input data, the feature fusion layer integrates extracted features from different dimensions, the prediction output layer outputs predicted values ​​of material transformation results, and the optimization recommendation layer outputs optimized combinations of influencing factors based on the prediction results. Multiple parallel feature extraction channels are configured for the feature mapping layer. Each feature extraction channel performs exclusive feature extraction for the input data of a certain type of influencing factor, so that the features of different types of influencing factors are fully captured. An adaptive fusion weight generation operation is set for the feature fusion layer. This adaptive fusion weight generation operation dynamically calculates and updates the fusion weights of the output features of different feature extraction channels based on the gradient information of backpropagation during training. The input data of the training sample subset is input into the input layer of the target machine learning model. Features are extracted through multiple feature extraction channels of the feature mapping layer to obtain the specific features corresponding to each influencing factor. The specific features are input into the feature fusion layer, and fusion weights are generated through adaptive fusion weight generation. The specific features are weighted and fused to obtain a comprehensive feature vector. The comprehensive feature vector is input into the prediction output layer. The predicted value of the material conversion result data is calculated through the mapping function of the prediction output layer. The difference between the predicted value of the material conversion result data and the output label corresponding to the training sample subset is calculated to obtain the training error. Based on the training error, the network parameters of the target machine learning model are adjusted through backpropagation, including the weight parameters of the feature extraction channel, the fusion weight parameters of the feature fusion layer, and the mapping function parameters of the prediction output layer. The training sample subset is repeatedly input into the target machine learning model for iterative training. In each iteration, the prediction accuracy of the target machine learning model is verified using the validation sample subset. During the iterative training process, the prediction accuracy of the validation sample subset is continuously monitored. When the improvement of the prediction accuracy of the validation sample subset is less than a preset threshold in N consecutive iterations, the training of the target machine learning model is stopped. The trained target machine learning model is tested using a subset of test samples. The root mean square error or mean absolute error between the prediction results of the target machine learning model on the subset of test samples and the true labels is calculated. Based on the test results of the subset of test samples, the selection logic of the optimization recommendation layer is adjusted and optimized. The optimization recommendation layer uses the germanium element recovery rate or target product yield in the predicted value of the material conversion result data as the optimization objective. The combination of influencing factors that maximizes the optimization objective is selected from the training samples. After fine-tuning, the target machine learning model is obtained.

5. The germanane reactor slag recovery method combining machine learning and process simulation according to claim 1, characterized in that, The optimized combination of influencing factors output by the target machine learning model is applied to each key operational step of the actual germane residue recovery process. The optimized germane residue recovery process is then executed to obtain germane-related recovered products, including: Receive the optimized combination of influencing factors output by the target machine learning model, analyze the specific values ​​of the influencing factors corresponding to each key operation step in the optimized combination of influencing factors, and for each key operation step, associate the analyzed specific values ​​of the influencing factors with the actual control parameters of the key operation step to determine the adjustment direction and adjustment target of the actual control parameters; According to the execution sequence of each key operation step, the parameters of the control equipment of the key operation steps in the actual germane reactor slag recovery process are adjusted in turn, so that the control equipment of each key operation step operates according to the specific values ​​of the corresponding influencing factors. During the parameter adjustment process, the actual operating status of each key operation step is monitored in real time, and the actual operating status data is collected. The actual operating status data and the operating status data in the process simulation structure adopt the same data collection dimension. The difference between the actual operating status data and the operating status data corresponding to the optimized combination of influencing factors in the simulation data set of the recycling process is calculated. When the absolute value of the difference exceeds the preset allowable deviation, a parameter fine-tuning instruction is generated based on the difference to correct the control equipment parameters of the key operation links. The data acquisition, difference calculation and parameter correction operations are repeated until the absolute value of the difference is lower than the preset allowable deviation. After all the parameters of the key operation links have been adjusted and the actual operating status has stabilized, the optimized germane residue recovery process is started to run continuously. During the continuous operation, the actual material conversion result data of each key operation link is continuously collected at the set time interval. The actual material conversion result data and the material conversion result data in the process simulation structure adopt the same data characterization standard. The continuously collected actual material conversion result data is input into the prediction output layer of the target machine learning model to calculate the error between the predicted value of the material conversion result data and the actual material conversion result data; the optimized germane residue recovery process is continuously executed until the preset amount of germane residue is processed, and the material output at the end of the germane residue recovery process is collected. After separation and purification, the related germane recovery products are obtained.

6. The germanane reactor slag recovery method combining machine learning and process simulation according to claim 2, characterized in that, The establishment of a dedicated process simulation unit for each key operational step includes: For each key operational step, we consulted relevant literature on the chemical principles and process data on germane residue recovery to determine the types of chemical reactions and physical changes involved in that key operational step. Extract the types of reactants, products, and reaction conditions required in this key operational step. The reaction conditions directly correspond to the operational environment-related factors among the influencing factors. The form of the reaction kinetic correlation is determined based on the type of chemical reaction, and the reaction kinetic correlation reflects the relationship between reactant concentration, reaction temperature, reaction pressure and reaction rate. In conjunction with the physical change process, supplement the mass transfer correlation and heat transfer correlation. The mass transfer correlation reflects the diffusion and migration law of matter within the key operation link, and the heat transfer correlation reflects the heat exchange law within the key operation link. Based on process practice data, the undetermined parameters in the reaction kinetics correlation, mass transfer correlation, and heat transfer correlation are determined. The reaction kinetics correlation, mass transfer correlation, and heat transfer correlation are coupled to construct the core material transformation mathematical relationship of this key operation. The core material transformation mathematical relationship is used to comprehensively describe the combined effects of chemical and physical changes on material transformation. Boundary constraints are added to the mathematical relationships of core material transformation. These boundary constraints are determined based on the equipment structure parameters and process operation limits of key operational links, so that the calculation results do not exceed the actual feasible range. Establish the solution logic for the mathematical relationships of material transformation, and determine the solution order and iterative convergence conditions of the correlation formula; The mathematical relationships, boundary condition constraints, and solution logic of the core material transformation are integrated into a unified calculation module, forming part of the process simulation unit dedicated to this key operation. The core part of the process simulation unit is logically tested. Known influencing factor values ​​are input, and the output material conversion result data is verified to be consistent with the process practice data. The parameters in the material conversion mathematical relationship are adjusted according to the test results, and finally a dedicated process simulation unit containing the material conversion mathematical relationship corresponding to the key operation link is obtained.

7. The germanane reactor slag recovery method combining machine learning and process simulation according to claim 3, characterized in that, The process involves combining candidate values ​​for different key operational steps according to their execution order to generate multiple combinations of influencing factors, including: Extract the execution order of each key operation step to form a key operation step sequence, and determine the connection relationship between the previous key operation step and the next key operation step in the key operation step sequence. For the first critical operation step in the sequence of critical operation steps, extract all its candidate values. Based on each candidate value of the first critical operation step, sequentially associate all candidate values ​​of the second critical operation step in the sequence of critical operation steps to form a combined fragment containing the candidate values ​​of the first two critical operation steps. Perform a rationality analysis on the combined segments: Based on the process simulation unit of the previous key operation in the process simulation structure, calculate the composition and properties of the output material when using the current candidate value; determine whether the composition and properties of the output material fall within the range of input materials allowed by the process simulation unit of the next key operation, and eliminate combined segments whose output materials are not within the allowed range of input. The combined fragments obtained through rationality analysis are associated with all candidate values ​​of the third key operation in the key operation sequence to form an extended combined fragment containing the candidate values ​​of the first three key operation steps. Repeatedly perform the association and rationality analysis operations, and sequentially associate all candidate values ​​of each subsequent key operation step in the key operation step sequence with the current extended combination fragment, and eliminate extended combination fragments that do not meet the material input requirements; Once the candidate values ​​for all key operational steps are associated, multiple complete combination schemes containing the candidate values ​​for all key operational steps are formed. The complete combination schemes are then compared for logical consistency to check whether there are any conflicting settings among the candidate values ​​for each key operational step. These conflicts include inconsistencies in operational environment requirements or material property compatibility. Complete combination schemes with conflicts are eliminated, and logically consistent complete combination schemes are retained. The retained complete combination schemes are numbered and their information is organized to determine the candidate values ​​of influencing factors for each key operational step in each scheme. Finally, multiple combination schemes of influencing factors are generated.

8. The germanane reactor slag recovery method combining machine learning and process simulation according to claim 5, characterized in that, The process involves adjusting the parameters of the control equipment for each critical operation step in the actual germane reactor slag recovery process according to the sequential execution order of each key operation step. This ensures that the control equipment for each critical operation step operates according to the specific values ​​of the corresponding influencing factors. This includes: Obtain a list of control equipment corresponding to each key operation in the actual germane residue recovery process, and determine the function and parameter adjustment range of each control device. The specific values ​​of the influencing factors in each key operational step of the optimized influencing factor combination are mapped to the adjustable parameters of the control equipment, and a correspondence table between influencing factors and equipment parameters is established. Based on the sequential execution order of key operational steps, construct a parameter adjustment execution plan and determine the adjustment time window and adjustment sequence of the control equipment for each key operational step; Within the adjustment time window of the first critical operation step, the target values ​​of the equipment parameters corresponding to the specific values ​​of the influencing factors of the critical operation step are found according to the corresponding relationship table. The control equipment for this critical operation step gradually adjusts the equipment parameters from the current value to the target value. During the parameter adjustment process, the control equipment's feedback monitoring function collects the equipment's operating parameters in real time to verify whether the parameter adjustment is accurate and in place. Once the control equipment parameters for the first critical operation are adjusted to the target value and the system is running stably, the adjustment time window for the second critical operation begins. Repeat the parameter search, equipment operation, and feedback monitoring process to adjust the parameters of the control equipment to the corresponding target values ​​in the second key operation step. According to the parameter adjustment execution plan, the parameters of the control equipment in each subsequent key operation link are continuously adjusted to ensure that each control equipment operates according to the specific values ​​of the corresponding influencing factors. After the parameter adjustment of the control equipment in all key operation links is completed, the operating parameters of all control equipment are inspected to check for any parameter deviations or omissions in adjustment, so that each key operation link in the actual germane reactor residue recovery process can operate according to the optimized combination of influencing factors.

9. A computer program product, characterized in that, The computer program product includes machine-executable instructions stored in a computer-readable storage medium, the processor reads the machine-executable instructions from the computer-readable storage medium, and the processor executes the machine-executable instructions to perform the germanane residue recovery method combining machine learning and process simulation as described in any one of claims 1 to 8.

10. A germanane reactor slag recovery system combining machine learning and process simulation, characterized in that, The device includes a processor and a readable storage medium storing a program that, when executed by the processor, implements the germanane reactor residue recovery method combining machine learning and process simulation as described in any one of claims 1-9.

Images