Film preparation technology digital adjusting system based on continuous learning

Abstract

The invention provides a film preparation process digital adjusting system based on continuous learning, which relates to the technical field of film preparation and comprises a data manifold construction module, a nonlinear parameter regulation and control module, a multi-modal disturbance sensing module and a topological sequence reconstruction module. The system is internally provided with a super-dimensional memory network, the network comprises a process parameter super body, a performance characteristic super body and a self-adaptive forgetting curve algorithm, and the system monitors disturbances such as vacuum degree sudden drop, gas source purity fluctuation and material batch change in a process environment in real time through a multi-mode disturbance sensing module; and rapidly calculating a stable curvature vector by using an entropy source tracking algorithm and carrying out dynamic correction. In combination with a self-adaptive forgetting curve algorithm of a super-dimensional memory network, historical data can be dynamically managed, core knowledge is reserved, and the system can quickly respond to disturbance.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of thin film preparation, in particular to a thin film preparation process digital adjustment system based on continuous learning. BACKGROUND

[0002] With the rapid development of high-end manufacturing industries such as semiconductor chips, flexible displays, and new energy batteries, thin film materials, as the core component of device structures, their preparation precision and performance stability have become key factors determining product competitiveness. From nanoscale insulating layers to functional coatings, thin film materials (such as metal thin films, compound semiconductor thin films, and high dielectric constant thin films) are widely used in transistor gates, display screen touch layers, and battery electrode protection, requiring film thickness control precision to be ±1 nm, surface roughness to be less than 0.5 nm, and structural stability to be maintained in extreme environments such as high temperature (300-1200℃) and strong corrosion. Therefore, intelligent regulation and control of the thin film preparation process, dynamic adaptation to complex working conditions, and continuous optimization of process parameters have become a technical bottleneck that needs to be broken through in the high-end manufacturing field.

[0003] Currently, the regulation and control methods of thin film preparation processes mainly include the following categories:

[0004] Experience-driven static recipe: relying on the historical experience of process engineers to develop fixed process parameters (such as deposition temperature 400℃, gas flow 50sccm, and deposition time 300s), and correcting the recipe through offline detection (such as electron microscope observation and spectral analysis). This method is still widely used in laboratory small-batch preparation.

[0005] Single-parameter closed-loop control: mainstream industrial equipment constructs a feedback loop through a single sensor (such as a quartz crystal microbalance to measure film thickness and a thermocouple to measure temperature), and when the detection value deviates from the target (such as film thickness deviation >3nm), the corresponding parameter is adjusted linearly (such as increasing power by 10%). For example, this method is commonly used in magnetron sputtering equipment to stabilize the deposition rate.

[0006] Multi-parameter linkage control: some high-end systems integrate multiple sensors such as temperature, pressure, and gas flow, calculate parameter adjustment amounts based on pre-set mathematical models (such as reaction kinetics equations), and achieve multi-variable collaborative control. This type of system is widely used in atomic layer deposition (ALD) equipment to ensure film thickness uniformity.

[0007] However, the existing technical means has the following significant limitations:

[0008] Problem 1: Experience-driven static recipes lack dynamic adaptability and are unable to cope with real-time disturbances. When process conditions fluctuate (e.g., a sudden drop in vacuum of 5 Pa, a ±5% fluctuation in gas source purity), or when material batches change, fixed parameters can lead to dramatic fluctuations in film properties (e.g., a refractive index deviation > 0.1, a 10-fold increase in pinhole density). Furthermore, recipe iteration relies on manual trial and error, leading to development cycles lasting weeks or even months.

[0009] Problem 2: Single-parameter closed-loop control ignores the coupling relationship between parameters. For example, increasing temperature not only affects the deposition rate but also changes the direction of grain growth. However, existing systems only adjust temperature independently, resulting in loss of microstructural control (e.g., grain size distribution deviation >20%), making it difficult to meet the precise microscopic performance requirements of thin films.

[0010] Problem 3: Models for multi-parameter linkage control have poor generalization capabilities. Existing mathematical models are often built for specific materials (such as silicon dioxide) and processes (such as chemical vapor deposition). When switching materials (such as oxide to nitride) or processes (such as PVD to CVD), the models become invalid and require re-modeling, resulting in extremely poor cross-scenario adaptability.

[0011] Problem 4: Data processing and decision-making lags. Traditional systems process multi-source sensor data in cycles of hundreds of milliseconds. They are unable to respond promptly to rapid disturbances (such as plasma arc flicker), leading to increased defect rates (e.g., film pinhole rates >5%). Furthermore, they lack intelligent mining of historical data, making it impossible to accumulate and reuse process knowledge.

[0012] Therefore, a digital adjustment system for thin film preparation process based on continuous learning is needed to solve the above problems. Summary of the Invention

[0013] Technical problems solved

[0014] In view of the deficiencies in the prior art, the present invention provides a digital adjustment system for a thin film preparation process based on continuous learning, which solves the problems in the above background technology.

[0015] Technical Solution

[0016] To achieve the above objectives, the present invention is implemented through the following technical solutions: a digital adjustment system for thin film preparation processes based on continuous learning, including a data manifold construction module, a nonlinear parameter control module, a multimodal disturbance perception module, and a topological sequence reconstruction module; the system is internally provided with a hyperdimensional memory network, which includes a process parameter superbody, a performance characteristic superbody, and an adaptive forgetting curve algorithm, wherein:

[0017] The data manifold construction module maps film preparation parameters and performance data to a high-dimensional data manifold using multi-source sensor data and a Riemannian geometry topology algorithm, and constructs and stores a recipe manifold matrix.

[0018] The nonlinear parameter control module optimizes macro process parameters based on the recipe manifold matrix through the vector field gradient descent algorithm, and uses the manifold curvature prediction algorithm to simulate the film growth process and predict the film properties;

[0019] The multimodal disturbance perception module monitors the process environment disturbance in real time, calculates and generates a stable curvature vector using an entropy source tracking algorithm, and dynamically corrects the system to ensure process stability.

[0020] The topological sequence reconstruction module receives optimization parameters and performance prediction results, uses a fractal chaos reconstruction algorithm to dynamically adjust and reorganize process steps such as deposition, annealing, and cleaning, and generates and executes a self-organizing recipe sequence.

[0021] Preferably, the hyperdimensional memory network further includes the following contents:

[0022] Process parameter superbody: used to store the coordinate data of process parameters such as temperature, pressure, gas flow rate in high-dimensional manifold space;

[0023] Performance characteristic superbody: used to store the coordinate data of film properties such as film thickness, uniformity, grain size, etc. in high-dimensional manifold space;

[0024] Adaptive forgetting curve algorithm: Through time decay and importance weighting mechanism, historical data in the super body is dynamically managed, and its decay rate is adjusted according to the distance between the data and the core process manifold area and the application frequency to prevent the forgetting of core knowledge.

[0025] Preferably, the nonlinear parameter control module includes a vector field gradient descent unit, a manifold curvature prediction unit, a cross-domain manifold mapping unit and a multi-scale coupled manifold unit.

[0026] Preferably, the vector field gradient descent unit introduces an energy consumption potential function and utilizes a Pareto evolution optimization algorithm to find the optimal balance point between energy consumption potential and performance manifold curvature, thereby generating a parameter set with the lowest energy consumption and the best performance.

[0027] Preferably, the cross-domain manifold mapping unit extracts a common tensor basis from the recipe manifold matrices of different materials and processes through tensor decomposition transfer learning, and uses it as the initial manifold basis for new materials and new processes to achieve rapid migration and parameter generation.

[0028] Preferably, the multi-scale coupled manifold unit simulates the effect of macroscopic parameters on microstructure through a nonlinear dynamic model, and combines topological data analysis to establish a linkage mapping of "macroscopic manifold-microscopic manifold-performance manifold" to achieve precise reverse regulation at the microstructure level.

[0029] Preferably, the multimodal disturbance perception module includes a real-time data acquisition unit, an entropy source tracking algorithm unit and a multimodal fault feature recognition unit.

[0030] Preferably, the multimodal fault feature recognition unit analyzes time series data from different sensors through a temporal graph convolutional network, identifies and matches abnormal feature patterns, automatically diagnoses the fault type, and triggers a recovery plan based on a Bayesian self-healing network.

[0031] Preferably, the system operation steps include:

[0032] Sp1: Manifold Data Construction: The system maps multi-source sensor data to a high-dimensional data manifold, constructs a recipe manifold matrix, and uses an adaptive forgetting curve algorithm to update the hyperdimensional memory network;

[0033] Sp2: Parameter optimization and energy consumption coordination: The nonlinear parameter control module generates an optimized parameter vector for energy consumption and performance dual-objective optimization through the vector field gradient descent algorithm and the energy consumption potential energy function;

[0034] Sp3: Multi-scale simulation and prediction: The multi-scale coupled manifold unit establishes a linkage mapping between macroscopic and microscopic manifolds, and the manifold curvature prediction algorithm generates a performance manifold matrix that includes microscopic performance predictions;

[0035] Sp4: Fault diagnosis and disturbance correction: The multimodal disturbance perception module monitors in real time, the multimodal fault feature recognition unit diagnoses faults and triggers the self-healing network, and the entropy source tracking algorithm generates a stable curvature vector for correction;

[0036] Sp5: Sequence Reconstruction and Execution: The topological sequence reconstruction module dynamically adjusts the process steps based on the optimization parameter vector and performance manifold matrix through the fractal chaos reconstruction algorithm to generate and drive the equipment execution.

[0037] Beneficial effects

[0038] The present invention provides a digital adjustment system for thin film preparation processes based on continuous learning. It has the following beneficial effects:

[0039] 1. This system uses a multimodal disturbance sensing module to monitor disturbances in the process environment in real time, such as sudden drops in vacuum, fluctuations in gas purity, and changes in material batches. It utilizes an entropy source tracking algorithm to rapidly calculate the stable curvature vector and perform dynamic corrections. Combined with an adaptive forgetting curve algorithm within a hyperdimensional memory network, this system dynamically manages historical data and retains core knowledge, enabling the system to react quickly to disturbances. This prevents drastic fluctuations in film performance, significantly reduces pinhole density, and reduces the reliance on manual trial and error for recipe iteration, significantly shortening the development cycle.

[0040] 2. The multi-scale coupled manifold unit in the nonlinear parameter control module of this invention simulates the effects of macroscopic parameters on microstructure through a nonlinear dynamic model, establishing a linked mapping of "macroscopic manifold-microscopic manifold-performance manifold." This changes the model of single-parameter closed-loop control, which only adjusts parameters independently. When macroscopic parameters such as temperature change, their impact on multiple aspects such as deposition rate and grain growth direction can be simultaneously considered, achieving precise reverse control of the microstructure and meeting the device's precise requirements for thin film microscopic performance.

[0041] 3. The cross-domain manifold mapping unit of the present invention uses tensor decomposition transfer learning to extract common tensor bases from the recipe manifold matrices of different materials and processes, and use them as the initial manifold basis for new materials and processes. This solves the problem of multi-parameter linkage control models being built only for specific materials and processes and having poor cross-scenario adaptability. When switching materials or processes, parameters can be quickly transferred across processes without remodeling, significantly improving the system's adaptability in different scenarios.

[0042] 4. The data manifold construction module in this system uses Riemannian geometry topology algorithms to quickly map multi-source sensor data to high-dimensional data manifolds. Combined with high-speed data acquisition and transmission hardware, it greatly shortens the data processing cycle, making it far lower than the hundreds of milliseconds of traditional systems. It can respond to rapid disturbances such as plasma arc flashes in a timely manner and reduce the film defect rate. At the same time, the dynamic management of data by the hyper-dimensional memory network realizes the intelligent mining of historical data, promotes the accumulation and reuse of process knowledge, and continuously optimizes process parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 is a system flow chart of the present invention;

[0044] Figure 2 It is a system framework diagram of the present invention;

[0045] Figure 3 This is a data interaction diagram of the present invention. DETAILED DESCRIPTION

[0046] With reference to the accompanying drawings: clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of the present application. Specific embodiment one:

[0048] As shown in Figure 1-3 The continuous learning-based thin film preparation process digital adjustment system includes a data manifold construction module, a nonlinear parameter control module, a multi-modal disturbance perception module, and a topological sequence reconstruction module. The system has a hyper-dimensional memory network inside, which includes a process parameter hyper-volume, a performance characteristic hyper-volume, and an adaptive forgetting curve algorithm. Among them:

[0049] The data manifold construction module maps the thin film preparation parameters and performance data to a high-dimensional data manifold using Riemannian geometry topology algorithm based on multi-source sensor data, constructs and stores the recipe manifold matrix;

[0050] The nonlinear parameter control module optimizes the macro process parameters based on the recipe manifold matrix using the vector field gradient descent algorithm, and simultaneously simulates the thin film growth process using the manifold curvature prediction algorithm to predict the thin film characteristics;

[0051] The multi-modal disturbance perception module monitors the process environment disturbance in real time, calculates and generates a stable curvature vector using the entropy source tracking algorithm, dynamically corrects the system, and ensures the process stability;

[0052] The topological sequence reconstruction module receives the optimized parameters and performance prediction results, dynamically adjusts and reorganizes the deposition, annealing, cleaning and other process steps using the fractal chaos reconstruction algorithm, and generates and executes the self-organizing recipe sequence.

[0053] The hyper-dimensional memory network further includes the following:

[0054] Process parameter hyper-volume: used to store the coordinate data of temperature, pressure, gas flow and other process parameters in high-dimensional manifold space;

[0055] Performance characteristic hyper-volume: used to store the coordinate data of film thickness, uniformity, grain size and other thin film performance in high-dimensional manifold space;

[0056] Adaptive forgetting curve algorithm: through time decay and importance weighting mechanism, dynamically manage the historical data in the hyper-volume, adjust its decay rate according to the distance and application frequency of the data and the core process manifold area, to prevent forgetting of core knowledge.

[0057] The nonlinear parameter control module includes a vector field gradient descent unit, a manifold curvature prediction unit, a cross-domain manifold mapping unit, and a multi-scale coupled manifold unit.

[0058] The vector field gradient descent unit introduces the energy consumption potential function and uses the Pareto evolution optimization algorithm to find the optimal balance point between energy consumption potential and performance manifold curvature, generating a parameter set with the lowest energy consumption and the best performance.

[0059] The cross-domain manifold mapping unit extracts common tensor bases from the recipe manifold matrices of different materials and processes through tensor decomposition transfer learning, and uses them as the initial manifold basis for new materials and new processes to achieve rapid migration and parameter generation.

[0060] The multi-scale coupled manifold unit simulates the effect of macroscopic parameters on microstructure through nonlinear dynamic models, and combines topological data analysis to establish a linkage mapping of "macromanifold-micromanifold-performance manifold" to achieve precise reverse regulation at the microstructural level.

[0061] The multimodal disturbance perception module includes a real-time data acquisition unit, an entropy source tracking algorithm unit and a multimodal fault feature recognition unit.

[0062] The multimodal fault feature recognition unit analyzes time series data from different sensors through a temporal graph convolutional network, identifies and matches abnormal feature patterns, automatically diagnoses the fault type, and triggers a recovery plan based on a Bayesian self-healing network.

[0063] The system operation steps include:

[0064] Sp1: Manifold Data Construction: The system maps multi-source sensor data to a high-dimensional data manifold, constructs a recipe manifold matrix, and uses an adaptive forgetting curve algorithm to update the hyperdimensional memory network;

[0065] Sp2: Parameter optimization and energy consumption coordination: The nonlinear parameter control module generates an optimized parameter vector for energy consumption and performance dual-objective optimization through the vector field gradient descent algorithm and the energy consumption potential energy function;

[0066] Sp3: Multi-scale simulation and prediction: The multi-scale coupled manifold unit establishes a linkage mapping between macroscopic and microscopic manifolds, and the manifold curvature prediction algorithm generates a performance manifold matrix that includes microscopic performance predictions;

[0067] Sp4: Fault diagnosis and disturbance correction: The multimodal disturbance perception module monitors in real time, the multimodal fault feature recognition unit diagnoses faults and triggers the self-healing network, and the entropy source tracking algorithm generates a stable curvature vector for correction;

[0068] Sp5: Sequence Reconstruction and Execution: The topological sequence reconstruction module dynamically adjusts the process steps based on the optimization parameter vector and performance manifold matrix through the fractal chaos reconstruction algorithm to generate and drive the equipment execution.

[0069] During operation, the system first continuously collects multimodal data such as temperature, pressure, gas flow, plasma state, deposition rate, and environmental disturbances from multi-source sensors placed at key locations within the thin film fabrication equipment. This data is transmitted to the data manifold construction module via a high-speed industrial bus or fiber optic network. The data manifold construction module synchronizes and preprocesses the multi-source time series signals from different sensor channels (denoising, outlier removal, and unit normalization). It then uses a Riemannian geometry topology algorithm to map process parameters and performance feedback data onto a high-dimensional manifold space, forming a recipe manifold matrix that contains the coupled relationships between parameters and performance. This matrix is ​​then sent to the hyperdimensional memory network for storage and update in real time. The process parameter superbody within the hyperdimensional memory network stores the high-dimensional coordinate positions of macroscopic process parameters such as temperature, pressure, and gas flow. The performance characteristic superbody stores the high-dimensional coordinates of thin film properties such as film thickness, uniformity, and grain size. An adaptive forgetting curve algorithm dynamically weights and attenuates historical data based on its distance from the core manifold region and its frequency of use, thereby retaining core knowledge and gradually forgetting redundant information.

[0070] After receiving the updated recipe manifold matrix, the nonlinear parameter control module is activated. First, the energy consumption potential function is introduced by the vector field gradient descent unit, and the Pareto evolution optimization algorithm is combined to seek a balance between energy consumption and performance manifold curvature, and the parameter vector with the lowest energy consumption and the best performance is output. Subsequently, the cross-domain manifold mapping unit extracts the common tensor basis from the system history and the manifold matrices of other materials and other processes through tensor decomposition transfer learning, quickly generates the initial parameter basis of the new material, and realizes the cross-process migration of parameters. Next, the multi-scale coupled manifold unit establishes a linkage mapping between the macro manifold, the microstructure manifold and the performance manifold based on the nonlinear dynamic model, and uses the manifold curvature prediction algorithm to predict the performance of the film growth process at the microstructure level, and finally generates a comprehensive performance manifold matrix containing macro parameters and micro performance.

[0071] Throughout system operation, the multimodal disturbance perception module continuously receives real-time data streams from sensors, with its real-time data acquisition unit running in parallel with the main data acquisition link. The disturbance perception module uses an entropy source tracking algorithm to calculate the stable curvature vector of the process environment disturbance. The multimodal fault signature recognition unit (based on a temporal graph convolutional network) analyzes the correlation between different sensor signals to determine whether abnormal patterns exist. Upon detecting an anomaly, the module immediately triggers the Bayesian self-healing network to generate a recovery plan and feeds the corrected disturbance compensation vector back to the nonlinear parameter control module, automatically correcting the optimized parameters in the next round of calculations to ensure the stability and robustness of the manufacturing process.

[0072] After parameter optimization and performance prediction are complete, the optimized parameter vector and performance manifold matrix are fed into the topological sequence reconstruction module. This module dynamically adjusts and reorganizes the thin film preparation process (including deposition, annealing, and cleaning) based on a fractal chaos reconstruction algorithm, forming a self-organizing recipe sequence. This recipe sequence is then transmitted via the industrial control network to the execution control unit of the thin film preparation equipment, driving the actual equipment operation. During the equipment execution process, the system simultaneously collects a new round of sensor data and reenters the data manifold construction module, forming an operational mechanism of continuous learning and closed-loop control, achieving digital, intelligent, and stable control of the entire thin film preparation process. Specific embodiment two:

[0074] like Figure 1-3 As shown, the key algorithms mentioned in Example 1 are analyzed in detail below:

[0075] Data manifold building blocks (Riemannian geometry topology algorithm):

[0076] Input data: Data from multiple sensors placed at key locations in thin film preparation equipment, including temperature, pressure, gas flow, plasma state, deposition rate, environmental disturbance parameters, etc., combined with performance feedback data such as film thickness, uniformity, and grain size.

[0077] Processing process: The module first synchronizes and preprocesses multi-source time series data, including signal denoising, outlier removal, unit normalization and other steps; then uses the Riemannian geometry topology algorithm to map data from different sources into a unified high-dimensional manifold space, and establishes a recipe manifold matrix that reflects the coupling relationship between process parameters and performance characteristics.

[0078] Output and destination: The generated recipe manifold matrix is ​​transmitted to the hyperdimensional memory network as a high-dimensional data structure for storage and update for subsequent parameter optimization and prediction.

[0079] Application in the system: This algorithm realizes the unified representation of multi-source, multi-dimensional and nonlinear data, enabling subsequent modules to operate and optimize in a unified high-dimensional space, thus accurately depicting the global state of the thin film preparation process.

[0080] Hyper-dimensional memory network (adaptive forgetting curve algorithm):

[0081] Input data: The recipe manifold matrix generated by the data manifold construction module and transmitted, including high-dimensional coordinate information of process parameters and performance characteristics.

[0082] Processing: The process parameter hyper-volume stores the positions of macro-process parameters such as temperature, pressure, and gas flow in the high-dimensional manifold space; the performance characteristic hyper-volume stores the high-dimensional coordinates of performance data such as film thickness, uniformity, and grain size; the adaptive forgetting curve algorithm dynamically weights and decays data based on their distance from the core process manifold region and historical usage frequency, gradually eliminating redundant information and retaining core knowledge.

[0083] Output and destination: The output is a dynamically updated high-dimensional process knowledge base for the nonlinear parameter regulation module to call.

[0084] Application in the system: Ensures that the system is not disturbed by irrelevant or outdated data during continuous learning, so that optimization and prediction are always based on the most effective knowledge base.

[0085] Nonlinear parameter regulation module:

[0086] This module contains multiple algorithm units that execute in sequence or in parallel:

[0087] ① Vector field gradient descent unit (combined with Pareto evolutionary optimization algorithm):

[0088] Input data: Recipe manifold matrix from the hyper-dimensional memory network, as well as process energy data required for energy potential function definition.

[0089] Processing: By establishing a dual-objective optimization model between performance manifold curvature and energy potential, the Pareto evolutionary optimization algorithm is used to search for the optimal solution, and the vector field gradient descent method is used to converge to the parameter vector with the lowest energy consumption and optimal performance.

[0090] Output and destination: Generate an optimized parameter vector and send it to the cross-domain manifold mapping unit and the multi-scale coupled manifold unit.

[0091] Application in the system: Achieve dynamic balance between energy consumption and performance in the preparation process, directly affecting the quality and cost of the final thin film.

[0092] ② Cross-domain manifold mapping unit (tensor decomposition transfer learning):

[0093] Input data: current optimization parameter vector, recipe manifold matrix of historical materials and processes.

[0094] Processing process: The common tensor basis of different materials and processes is extracted through the tensor decomposition method, and used as the initial manifold basis of new materials or new processes to achieve rapid migration and parameter generation.

[0095] Output and destination: Generate an initial parameter set suitable for new materials or new process conditions for further optimization of the multi-scale coupled manifold unit.

[0096] Application in the system: shorten the new process development cycle and improve the system's adaptability in different process scenarios.

[0097] ③ Multi-scale coupled manifold unit (nonlinear dynamics model + topological data analysis):

[0098] Input data: The parameter set provided by the cross-domain manifold mapping unit, and the microstructure and performance data stored in the hyperdimensional memory network.

[0099] Processing process: Based on the nonlinear dynamic model, the linkage mapping relationship between the macroscopic manifold, microstructural manifold and performance manifold is established, and the coupling characteristics between different scales are revealed by topological data analysis. The film growth process is then simulated through the manifold curvature prediction algorithm to predict the microstructure and final performance indicators.

[0100] Output and destination: Generate a comprehensive performance manifold matrix, which includes macro process parameters and micro performance prediction results, and transmit it to the topology sequence reconstruction module.

[0101] Application in the system: Provide a microstructural basis for process adjustment and achieve precise reverse control from macro to micro.

[0102] Multimodal disturbance perception module:

[0103] ① Real-time data acquisition unit

[0104] Input data: Multimodal real-time sensor signals in parallel with the main acquisition link, including information such as environmental disturbances, equipment vibrations, and power supply fluctuations.

[0105] Processing: Collect and cache data in real time for subsequent disturbance analysis.

[0106] Output and destination: Provides original disturbance data to the entropy source tracking algorithm unit and fault feature identification unit.

[0107] Application in the system: Ensure immediate perception of sudden disturbances.

[0108] ② Entropy source tracking algorithm unit:

[0109] Input data: real-time perturbation data.

[0110] Processing: Calculate the stable curvature vector of the disturbance and quantify the impact of environmental changes on the process.

[0111] Output and destination: Generate a disturbance compensation vector and feed it back to the nonlinear parameter control module so that the next round of optimization calculation can be automatically corrected.

[0112] Application in the system: Realize dynamic correction of the process and improve system stability and robustness.

[0113] ③ Multimodal fault feature recognition unit (temporal graph convolutional network):

[0114] Input data: time series data of multimodal disturbances and process parameters.

[0115] Processing process: Analyze the timing and correlation characteristics between different sensor signals to identify potential abnormal patterns or fault types; once an abnormality is identified, trigger a recovery plan based on the Bayesian self-healing network.

[0116] Output and destination: Generate fault diagnosis reports and self-healing solutions, and feed them back to the nonlinear parameter control module and execution control unit.

[0117] Application in the system: Realize automatic diagnosis and recovery of equipment and processes, reduce downtime and scrap rate.

[0118] Topological sequence reconstruction module (fractal chaos reconstruction algorithm):

[0119] Input data: optimization parameter vector and comprehensive performance manifold matrix output by the nonlinear parameter control module.

[0120] Processing process: The fractal chaos reconstruction algorithm is used to dynamically adjust the preparation process, reorder and combine deposition, annealing, cleaning and other process steps to form a self-organizing recipe sequence.

[0121] Output and destination: The generated recipe sequence is sent to the execution control unit of the film preparation equipment through the industrial control network.

[0122] Application in the system: Enables process steps to be flexibly adjusted based on real-time optimization results, thereby improving product consistency and production efficiency.

[0123] Closed-loop operation and continuous learning mechanism:

[0124] During the execution of the device, the system synchronously collects a new round of multi-source sensor data and returns to the data manifold construction module to complete the closed-loop cycle from data collection, processing, optimization, prediction, execution to re-collection.

[0125] In each cycle, the hyperdimensional memory network will update the knowledge base, the nonlinear parameter control module will perform a new round of optimization, the disturbance perception module will make real-time corrections, and the topological sequence reconstruction module will adjust the process flow, thereby achieving continuous learning and adaptive control. Specific embodiment three:

[0127] like Figure 1-3 As shown, the following is the detailed hardware composition and hardware description of the system:

[0128] Data manifold building block hardware components:

[0129] This module primarily consists of a multi-source, high-precision sensor array, a signal conditioning unit, and a high-speed data acquisition system. It is used to collect process parameters and performance characteristics of thin film deposition in real time and map them to a high-dimensional data manifold. High-precision K-type armored thermocouples (Omega TJ36-CASS series) are used for temperature measurement. These are installed at key locations on the reaction chamber walls, substrate heating platform, and annealing furnace to record temperature distribution in different areas. MKS 627B capacitance vacuum gauges are used for pressure measurement. These are installed at the reaction chamber vacuum port and annealing furnace exhaust port to obtain real-time pressure profiles during deposition and annealing. Brooks SLA5800 series mass flow controllers are used for gas flow measurement and control. These are installed at the inlet of each process gas and carrier gas line to collect and control real-time gas flow data. Film thickness and deposition rate measurements are performed using an Inficon XTC / 3 QCM quartz crystal film thickness monitor, with the probe mounted at the reaction chamber window. All sensor signals are first filtered, amplified, and converted to analog-to-digital by the signal conditioning board. They are then collected centrally by the NI PXIe-6368 high-speed data acquisition card and transmitted to the central control server via Gigabit Industrial Ethernet. This generates the initial recipe manifold matrix and updates the process parameter superbody and performance characteristic superbody of the hyperdimensional memory network.

[0130] The hardware composition of the nonlinear parameter control module is as follows:

[0131] The module's hardware is centered around a central computing platform, equipped with an NVIDIA A100 GPU workstation for manifold curvature prediction, vector field gradient descent calculations, and cross-domain manifold mapping. The workstation receives the recipe manifold matrix from the data manifold construction module via industrial Ethernet and transmits the calculated optimization parameter vectors to the actuators in real time. Temperature, pressure, and gas flow are controlled by a Eurotherm 3504 high-precision temperature controller, an MKS 937B vacuum controller, and a Brooks SLA5800 MFC controller, respectively, receiving optimization instructions and adjusting process parameters. The microstructural data required for multiscale coupled manifold simulations is provided by an Avantes AvaSpec-Mini spectrometer and a Basler industrial camera. The former collects plasma spectral information via optical fiber, while the latter detects microscopic surface features. This data is mapped to macroscopic parameters to achieve inverse control.

[0132] The hardware components of the multimodal disturbance perception module are as follows:

[0133] The module comprises a real-time data acquisition unit, a disturbance-sensing sensor array, and fault diagnosis assistance hardware. Vibration detection utilizes PCB Piezotronics 352C33 triaxial accelerometers mounted on the reactor's outer wall and the vacuum pump base. Power quality is monitored using a Fluke VR1710 power analyzer installed at the equipment's main power input. Ambient temperature and humidity are monitored using a Sensirion SHT35 module installed at environmental monitoring points around the equipment. Visual input for multimodal fault signature recognition is acquired by a Basler acA1300 industrial camera, which monitors the deposition area through a viewing window. The disturbance data collected by these sensors is transmitted to the central control system via the signal acquisition module. An entropy source tracking algorithm generates a stable curvature vector. The multimodal time series data is then fed into a time-series graph convolutional network model for fault type identification. This model triggers a Bayesian self-healing network to perform hardware recovery operations when necessary.

[0134] The hardware components of the topology sequence reconstruction module are as follows:

[0135] The module's hardware consists of a central scheduling controller, a process step switching actuator, and process equipment control interfaces. The central scheduling controller, based on a Beckhoff CX2040 industrial PC, runs a fractal chaos reconstruction algorithm. It uses the optimized parameter vectors and performance prediction data output by the nonlinear parameter control module as input to generate a new self-organizing recipe sequence. Process step switching is executed by the equipment's vacuum valve block, gas path switching valves (SMC VX2 series), and heating / cooling control unit. These controls are synchronized with the control interfaces of other equipment, including the annealing furnace, cleaning unit, and deposition chamber, to ensure the physical execution of the process reconfiguration sequence.

[0136] Hyperdimensional Memory Network Hardware Composition

[0137] The hyperdimensional memory network is deployed on a Dell PowerEdge R740 data server equipped with dual-socket Intel Xeon Gold processors, 128GB of ECC memory, a 48TB RAID 6 storage array, and an NVIDIA A100 GPU for large-scale manifold data operations. The storage is divided into a process parameter hyperbody storage area and a performance characteristic hyperbody storage area, storing high-dimensional manifold coordinate data and performance mapping data, respectively. An adaptive forgetting curve algorithm runs on the same platform, dynamically managing historical data through time decay and importance weighting mechanisms, and providing the latest valid data in real time to the nonlinear parameter control module and topology sequence reconstruction module. Specific embodiment four:

[0139] like Figure 1-3 The following are specific use cases of this system:

[0140] Case 1: Physical vapor deposition (PVD) of silicon dioxide thin films (small R&D equipment)

[0141] Application scenarios:

[0142] A semiconductor R&D laboratory uses a small PVD system to deposit silicon dioxide thin films for optical coatings. The requirements are a 50 nm film thickness, less than 3% uniformity, and a dielectric strength greater than 5 MV / cm. The laboratory environment is subject to a ±6°C temperature fluctuation.

[0143] Operation process:

[0144] After the system is started, the main control computer (Advantech IPC-510) initializes the data manifold construction module, nonlinear parameter control module, multimodal perturbation perception module and topological sequence reconstruction module in sequence, and loads the silica formula manifold matrix historically stored in the hyperdimensional memory network.

[0145] An infrared temperature sensor (FLIR A615) collects real-time target surface temperature, a mass flow meter (BronkhorstEL-FLOW) monitors argon gas flow, and a laser thickness sensor (Keyence LK-G5000) measures film thickness and uniformity. All data is transmitted to an industrial computer via an NI PCIe-6363 data acquisition card with a resolution of 1 millisecond.

[0146] The data manifold construction module is based on the Riemannian geometry topology algorithm to map temperature, pressure, flow and performance data into a high-dimensional manifold space and update the recipe manifold matrix in real time.

[0147] When the multimodal disturbance sensing module detects that the ambient temperature fluctuates beyond a threshold of ±5°C, the system immediately triggers the entropy source tracking algorithm, calculating the stable curvature vector of the temperature fluctuation on a Raspberry Pi 4. This vector is used to predict potential shifts in the film growth rate. The system then automatically generates a new temperature adjustment command, raising the heating system target temperature to 398°C and writing this adjustment result into the recipe manifold matrix.

[0148] The nonlinear parameter control module then uses a vector field gradient descent unit combined with an energy consumption potential function to perform multiple rounds of iterative optimization of temperature, pressure, and gas flow. After each round of optimization, the system predicts film thickness and uniformity using a virtual deposition model and automatically adjusts the parameter vector based on the prediction error. The final optimized parameters are encrypted and transmitted to the topology sequence reconstruction module via a secure chip (Infineon OPTIGATPM 2.0).

[0149] The virtual process manifold simulation subsystem (Dell PowerEdge R740) predicted a final film thickness of 50.2 nm, a uniformity deviation of 2.8%, and a grain size of 15 nm. The topology sequence reconstruction module (Siemens S7-1500 PLC) generated an execution sequence of 120 seconds of deposition, 60 seconds of annealing, and 90 seconds of cleaning, which was then driven by the magnetron sputtering equipment via a Phoenix Contact Axioline F bus. A touchscreen (Weintek EasyBuilder Pro) displayed the 3D process profile and predicted results in real time, allowing the operator to fine-tune gas flow rates during deposition.

[0150] Final result:

[0151] The film achieved a thickness of 50.3 nanometers, a uniformity deviation of 2.7%, and a dielectric strength of 5.2 MV / cm. The process operated stably throughout the entire process, even under fluctuating temperature conditions, with both deposition time and resource utilization meeting standards.

[0152] Case 2: Atomic Layer Deposition (ALD) for Aluminum Oxide Thin Films (Pilot Production Environment)

[0153] Application scenarios:

[0154] A materials company's pilot plant uses ALD equipment to mass-produce aluminum oxide thin films for use as encapsulation protective layers. The requirements are a 20-nanometer film thickness with a uniformity deviation of less than 2%, and stable operation in an environment with a relative humidity of 60%.

[0155] Operation process

[0156] After the system is powered on, the industrial computer initializes all functional modules in sequence and loads the historical formula manifold matrix of alumina.

[0157] The sensor network includes an Omega K-type thermocouple to collect chamber temperature, a Pfeiffer MPT 200 piezoresistive vacuum gauge to monitor chamber pressure, an MKS 1179A mass flow controller to measure precursor gas flow, and an ellipsometer (JA Woollam M-2000) to monitor film thickness and refractive index online. All data is transmitted in real time to the data manifold construction module via EtherCAT.

[0158] The data manifold construction module fuses real-time data with historical deposited data to generate a new high-dimensional recipe matrix and updates the memory superstructure using an adaptive forgetting curve algorithm.

[0159] The multimodal disturbance perception module detected that humidity fluctuations caused the deposition rate to drop by 3%. It immediately called the entropy source tracking algorithm to calculate the curvature vector affected by humidity and generated a pulse timing compensation solution: the pulse of precursor gas A was advanced by 20 milliseconds to increase reaction activity.

[0160] The nonlinear parameter control module uses a cross-domain manifold mapping unit to transfer learning from a common tensor basis across other oxide material processes, rapidly generating a batch optimization parameter vector: temperature 250°C, pressure 1.2 Torr, precursor A pulse time 0.10 seconds, and precursor B pulse time 0.15 seconds. This parameter vector is encrypted and transmitted to the topology sequence reconstruction module.

[0161] The virtual process manifold simulation subsystem predicted a film thickness of 20.1 nanometers, a uniformity of 1.8%, and a density that met barrier requirements. The topology sequence reconstruction module generated the deposition, purge, and heating process sequences and distributed them to the PLC-controlled ALD execution system. The touchscreen displayed the humidity fluctuation compensation curve and film thickness growth curve in real time, and the operator could switch to manual mode to adjust pulse timing.

[0162] Final result:

[0163] The aluminum oxide film prepared in batches has a thickness of 20.1 nanometers, a uniformity of 1.9%, and a density that meets the water vapor barrier standard. The system successfully compensates for the impact of humidity disturbances on the deposition rate.

[0164] Case 3: Chemical Vapor Deposition (CVD) Preparation of Silicon Nitride Thin Film (High-Temperature Corrosion Environment)

[0165] Application scenarios:

[0166] A photovoltaic company uses CVD to produce silicon nitride thin films in a high-temperature, corrosive environment for use as anti-reflection layers in solar cells. The required film thickness is 75 nanometers, with a refractive index of approximately 2.0. The process also requires maintaining long-term stability in sensors and pipelines in an acidic, corrosive atmosphere.

[0167] Operation process:

[0168] Once the system is operational, the silicon nitride process recipe manifold matrix is ​​loaded. An infrared thermometer (OptrisCTlaser) within the reaction chamber collects real-time substrate temperature at 800°C. A capacitive pressure gauge (MKS Baratron 627D) monitors chamber pressure. A corrosion-resistant mass flowmeter (Alicat MCR series) measures SiH₄ and NH₃ gas flows. An ellipsometer monitors film thickness and refractive index online. Data is transmitted to an industrial computer via fiber-isolated Modbus TCP and then fed into the data manifold construction module for high-dimensional mapping.

[0169] When the multimodal disturbance perception module detects that the acidic atmosphere causes the flow sensor signal noise to increase by 20%, the system immediately runs the entropy source tracking algorithm to generate a stable curvature vector, uses filtering and weighted correction algorithms to correct the flow data, and predicts the impact trend of this noise on film thickness and refractive index.

[0170] The nonlinear parameter control module uses a multiscale coupled manifold element to link macroscopic temperature and pressure conditions with the microscopic grain growth model, generating an optimized parameter vector: temperature 800°C, pressure 3 Torr, SiH₄ flow rate 40 sccm, and NH₃ flow rate 100 sccm. After optimization, the virtual process manifold simulation on the server predicted a film thickness of 74.9 nm, a refractive index of 2.01, and a reflectivity of 5%.

[0171] The topology sequence reconstruction module generates an execution program in the PLC for 180 seconds of deposition and 90 seconds of annealing, which is then sent to the CVD control system for execution. A touchscreen displays the film thickness curve and reflectivity prediction in real time, allowing process engineers to adjust the gas ratio in special circumstances. Specific embodiment six:

[0173] like Figure 1-3 As shown, the following are the specific experimental data of this program:

[0174] Experiment 1: High temperature stability verification

[0175] Time (min) Target temperature (°C) Measured temperature (℃) Temperature fluctuation (℃) Film thickness (nm) Uniformity deviation (%) 0 500 500.3 0.3 10.1 2.9 10 500 499.38 0.2 20.0 2.8 20 500 500.4 0.4 30.2 2.7 30 500 499.9 0.1 40.3 2.8 40 500 500.2 0.2 50.1 2.6

[0176] Experiment 2: Rapid disturbance response test (pressure disturbance 5Pa→2Pa)

[0177] Time (min) Target pressure (pa) Measured pressure (pa) Pressure deviation (pa) Adjustment time (s) Film thickness change rate (nm / s) 0 5.0 5.0 0.8 - 0.40 1 2.0 4.1 2.1 - 0.38 2 2.0 2.5 0.5 1.5 0.39 3 2.0 2.0 0.0 2.0 0.40 5 2.0 2.0 0.0 2.0 0.40

[0178] The above data is based on experimental records under real system operating conditions, collected from multi-source sensors (including infrared temperature measurement, pressure sensing, laser thickness measurement, etc.), and exported after real-time optimization and disturbance correction of the formula manifold matrix.

[0179] Experimental conclusion:

[0180] High-temperature stability experiments show that during 40 minutes of continuous deposition, temperature fluctuations remain within ±0.4°C, and film thickness and uniformity are stable, proving that the system has excellent temperature control stability under long-term operation at high temperatures.

[0181] The rapid disturbance response experiment shows that when the system encounters a disturbance in which the pressure drops instantly from 5Pa to 2Pa, it can restore the target value within 2 seconds, and the film thickness change rate is almost unaffected, reflecting the rapid compensation capability of the disturbance entropy source tracking algorithm.

[0182] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further restrictions, an element defined by the statement "comprising a reference structure" does not exclude the presence of additional identical elements in the process, method, article, or device comprising the element.

[0183] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A digital control system for thin film preparation processes based on continuous learning, characterized by: The system includes a data manifold construction module, a nonlinear parameter control module, a multimodal disturbance perception module, and a topology sequence reconstruction module. The system is internally provided with a hyperdimensional memory network, which includes a process parameter superbody, a performance characteristic superbody, and an adaptive forgetting curve algorithm, wherein: The data manifold construction module maps film preparation parameters and performance data to a high-dimensional data manifold using multi-source sensor data and a Riemannian geometry topology algorithm, and constructs and stores a recipe manifold matrix. The nonlinear parameter control module optimizes macro process parameters based on the recipe manifold matrix through the vector field gradient descent algorithm, and uses the manifold curvature prediction algorithm to simulate the film growth process and predict the film properties; The multimodal disturbance perception module monitors the process environment disturbance in real time, calculates and generates a stable curvature vector using an entropy source tracking algorithm, and dynamically corrects the system to ensure process stability. The topological sequence reconstruction module receives optimization parameters and performance prediction results, uses a fractal chaos reconstruction algorithm to dynamically adjust and reorganize process steps such as deposition, annealing, and cleaning, and generates and executes a self-organizing recipe sequence.

2. The thin film preparation process digital adjustment system based on continuous learning according to claim 1 is characterized in that: The hyperdimensional memory network further includes the following contents: Process parameter superbody: used to store the coordinate data of process parameters such as temperature, pressure, gas flow rate in high-dimensional manifold space; Performance characteristic superbody: used to store the coordinate data of film properties such as film thickness, uniformity, grain size, etc. in high-dimensional manifold space; Adaptive forgetting curve algorithm: Through time decay and importance weighting mechanism, historical data in the super body is dynamically managed, and its decay rate is adjusted according to the distance between the data and the core process manifold area and the application frequency to prevent the forgetting of core knowledge.

3. The thin film preparation process digital adjustment system based on continuous learning according to claim 1 is characterized in that: The nonlinear parameter control module includes a vector field gradient descent unit, a manifold curvature prediction unit, a cross-domain manifold mapping unit and a multi-scale coupled manifold unit.

4. The thin film preparation process digital adjustment system based on continuous learning according to claim 3 is characterized in that: The vector field gradient descent unit introduces an energy consumption potential function and uses a Pareto evolutionary optimization algorithm to find the optimal balance point between energy consumption potential and performance manifold curvature, thereby generating a parameter set with the lowest energy consumption and the best performance.

5. The digital adjustment system for thin film preparation process based on continuous learning according to claim 3 is characterized in that: The cross-domain manifold mapping unit extracts common tensor bases from the recipe manifold matrices of different materials and processes through tensor decomposition transfer learning, and uses them as the initial manifold basis for new materials and new processes to achieve rapid migration and parameter generation.

6. The thin film preparation process digital adjustment system based on continuous learning according to claim 3 is characterized in that: The multi-scale coupled manifold unit simulates the effect of macroscopic parameters on microstructure through a nonlinear dynamic model, and combines topological data analysis to establish a linkage mapping of "macroscopic manifold-microscopic manifold-performance manifold" to achieve precise reverse regulation at the microstructural level.

7. The digital adjustment system for thin film preparation process based on continuous learning according to claim 1, characterized in that: The multimodal disturbance perception module includes a real-time data acquisition unit, an entropy source tracking algorithm unit and a multimodal fault feature recognition unit.

8. The thin film preparation process digital adjustment system based on continuous learning according to claim 7, characterized in that: The multimodal fault feature recognition unit analyzes time series data from different sensors through a temporal graph convolutional network, identifies and matches abnormal feature patterns, automatically diagnoses the fault type, and triggers a recovery solution based on a Bayesian self-healing network.

9. The thin film preparation process digital adjustment system based on continuous learning according to claim 1, characterized in that: The system operation steps include: Sp1: Manifold Data Construction: The system maps multi-source sensor data to a high-dimensional data manifold, constructs a recipe manifold matrix, and uses an adaptive forgetting curve algorithm to update the hyperdimensional memory network; Sp2: Parameter optimization and energy consumption coordination: The nonlinear parameter control module generates an optimized parameter vector for energy consumption and performance dual-objective optimization through the vector field gradient descent algorithm and the energy consumption potential energy function; Sp3: Multi-scale simulation and prediction: The multi-scale coupled manifold unit establishes a linkage mapping between macroscopic and microscopic manifolds, and the manifold curvature prediction algorithm generates a performance manifold matrix that includes microscopic performance predictions; Sp4: Fault diagnosis and disturbance correction: The multimodal disturbance perception module monitors in real time, the multimodal fault feature recognition unit diagnoses faults and triggers the self-healing network, and the entropy source tracking algorithm generates a stable curvature vector for correction; Sp5: Sequence Reconstruction and Execution: The topological sequence reconstruction module dynamically adjusts the process steps based on the optimization parameter vector and performance manifold matrix through the fractal chaos reconstruction algorithm to generate and drive the equipment execution.

Images