Laser processing control method, system, device and medium based on neural network

Abstract

The application belongs to the technical field of laser processing, and particularly relates to a laser processing control method based on a neural network, and specifically comprises the following steps: a target definition and input stage; a physical model and database stage: establishing a basic model library; establishing a mapping database; an AI core engine stage: training a model to learn a complex nonlinear relationship between 'laser parameters-material response-final processing result'; based on a prediction model and a target set by a user, an optimal laser parameter combination is searched by using an optimization algorithm; a laser parameter automatic adjustment and execution stage; a real-time monitoring and feedback stage; and an iterative learning and system improvement stage. The application also discloses a control system, an electronic device and a computer storage medium. The application realizes sub-micron precision control and can make thermal damage approach to zero, shortens a development cycle, realizes online real-time regulation and control, reduces energy consumption and improves material utilization, has an interpretable decision mechanism and cross-material generalization capability.

Description

Technical Field

[0001] This invention belongs to the field of laser processing technology, specifically relating to a laser processing control method, system, equipment, and medium based on neural networks. Background Technology

[0002] Femtosecond laser ingot cutting is a technique that utilizes ultrashort pulse lasers (pulse widths in the femtosecond range, i.e., 10^66 seconds). −15 This technology (in seconds) is used to precisely process high-hardness crystalline ingot materials. Its core principle is to induce a nonlinear absorption effect inside the material through the high peak energy of an extremely short pulse, thereby achieving precise breaking of molecular bonds instead of thermal melting and thus avoiding thermal damage.

[0003] Problems with existing technology: Machining accuracy out-of-control problem Inaccurate energy deposition: Traditional laser processing relies on fixed parameters, which are difficult to adapt to differences in material microstructure (such as lattice defects and doping concentration fluctuations), leading to expansion of the heat-affected zone or uncontrollable microcracks; Lack of cross-scale tomography prediction: There is a lack of quantitative correlation models between laser parameters (such as spatiotemporal shaping), electron dynamics response (electron temperature / density distribution) and macroscopic results (cut morphology), and process development relies on trial and error; Efficiency and cost bottlenecks Long parameter tuning cycle: A single process optimization requires hundreds of experiments, especially in new materials (such as GaN and silicon carbide) which can take several months; Lack of dynamic disturbance compensation: Real-time disturbances such as environmental temperature and humidity fluctuations and optical system drift lack adaptive correction mechanisms, resulting in yield fluctuations of over ±15%. Summary of the Invention

[0004] The purpose of this invention is to provide a laser processing control method, system, device, and medium based on neural networks, which can achieve submicron-level precision control and reduce thermal damage to near zero, shorten the development cycle, enable online real-time control, reduce energy consumption, and improve material utilization. It also has an interpretable decision-making mechanism and cross-material generalization capability.

[0005] The specific technical solution adopted by this invention is as follows: The laser processing control method based on neural networks specifically includes the following steps: Target definition and input phase; Physical Model and Database Phase: Establish a basic model library; establish a mapping database of "laser parameters-electron dynamics-processing results": construct an associated database based on historical experimental data and model simulation results; AI Core Engine Stage: Machine Learning / Deep Learning Model: Training the model to learn the complex nonlinear relationship between "laser parameters - material response - final processing result"; Optimization Algorithm: Based on the prediction model and the user-defined goals, the optimization algorithm searches for the optimal combination of laser parameters; Automatic laser parameter adjustment and execution stage: Control command generation: The optimal parameter combination output by the AI ​​optimization engine is converted into specific control commands; Hardware control: Commands are sent to the laser control system, pulse shaper, and spatial light modulator. Real-time monitoring and feedback phase: Process monitoring: Real-time or in-situ monitoring using diagnostic technology; Result evaluation: Obtaining key indicators of actual processing results; Feedback loop: Feeding the monitored process signals back to the AI ​​engine, which compares this information with the expected target. Iterative learning and system improvement phase: Each processing attempt and its data are used to update the AI ​​model and physical model database; the system continuously learns the optimal parameter strategy under new materials and new targets, and achieves automatic parameter adjustment.

[0006] The process of establishing the basic model library is as follows: Step 1: Model selection and adaptation; Step 2: Parameter calibration and verification; Step 3: Multiphysics Coupled Simulation Process Input the initial laser pulse shape, calculate the electron excitation density and energy deposition profile through the electron dynamics model, solve the electron temperature and lattice temperature through the dual-temperature model, determine the phase transition by the molecular dynamics model, analyze the free electron growth, shielding and reverse bremsstrahlung radiation loss through the plasma model, and finally output the spatially resolved energy deposition map, temperature evolution curve, phase transition region volume, and initial surface morphology. Step 4: Model reduction and acceleration.

[0007] The process of establishing the "laser parameters-electron dynamics-processing results" mapping database is as follows: Step 1: Multimodal data generation mechanism to build a closed-loop data production pipeline; Step 2: Dimensional unification and correlation modeling to solve the cross-scale discontinuity problem between laser parameters, microscopic processes and macroscopic results; Step 3: Knowledge graph construction and reasoning support, using graph structures to express complex causal chains; Step 4: Dynamic updates and version management.

[0008] The machine learning / deep learning model operates as follows: Step 1: Data Acquisition and Preprocessing; Step 2: Feature Engineering and Label Generation: Combining offline experiments and simulations, establish the mapping relationship between "process parameters - morphology of internal modified zones - fracture behavior" and generate training labels; Step 3: Model training and validation: The model is trained under multiple material and thickness conditions using a cross-validation strategy, and interpretability analysis is performed using SHAP values ​​or Grad-CAM. Step 4: Online reasoning and feedback: Deployed in edge computing units, it predicts the quality risks of the current processing point in real time and outputs early warning signals or adjustment suggestions.

[0009] The control command generation process is as follows: Step 1: Parameter Reception and Parsing The AI ​​engine outputs a structured parameter package, and the instruction generation module parses the parameters and matches them with a preset hardware driver template. Step 2: Multi-device instruction collaborative generation The instruction generation logic of the spatial light modulator is to call the GPU-accelerated algorithm to generate a phase hologram and output a 512×512 grayscale bitmap; the instruction generation logic of the pulse shaper is to compile the time series into the RF drive signal of the AOM; the instruction generation logic of the laser power supply is to calculate the energy-voltage mapping curve and generate the DAC control waveform. Step 3: Cross-system timing synchronization Construct timing chains with nanosecond-level precision; timing jitter control <100ns to ensure precise matching of spatial-temporal parameters; Step 4: Virtual Execution Verification The digital twin platform simulates the effect of commands: it simulates the propagation path of light beams through ray tracing; it calls the physical model to predict the material response; if the prediction result deviates from the target by more than 5%, it automatically triggers the parameter rollback mechanism. Step 5: Instruction encapsulation and issuance Packed into device-specific binary instruction frames, they are broadcast to each execution unit via real-time Ethernet with a latency of <1ms.

[0010] The process monitoring operation is as follows: Step 1: Data Acquisition and Feature Extraction Input: Real-time multimodal data stream of the laser-affected area; Output: Quantitative indicators such as plasma density, molten pool size, and surface roughness Ra value; Step 2: State Diagnosis and Decision Generation Rules engine: Preset thresholds trigger alarms; AI prediction model: Predicts the final processing quality under current parameters based on historical data, and triggers adjustments if the prediction deviation is greater than 10%; Step 3: Closed-loop control execution Command issued: Adjust laser energy: Adjust laser power supply voltage via DAC module; Correct optical path: Spatial light modulator generates new phase diagram in real time to optimize beam spatial distribution; Dynamic response timing: Data acquisition to capture transient phenomena; edge computing to complete feature analysis and diagnosis; command execution to adjust laser parameters; Step 4: Knowledge Accumulation and Model Iteration Data from each processing cycle is automatically stored in the mapping database; incremental training is triggered weekly to update AI model weights and optimize control strategies.

[0011] The process of learning the optimal parameters is as follows: Step 1: Cold Start of New Materials / New Tasks Input the parameters of the new material, the system queries the knowledge graph, loads the migration strategy template when similar materials exist, starts the digital twin simulation when there is no matching record, and generates the safety parameter boundary. The results of both are then used to conduct small sample exploration experiments, and finally output the evaluation of the initial performance. Step 2: Online Incremental Learning Loop Step 1: Data Awareness; Step 2: Policy Optimization: RL agents select actions based on state; the meta-model updates the policy network weights every 10 minutes; Step 3: Knowledge Accumulation: Successful policies are stored in the case library, and failed cases trigger causal analysis; the confidence weights of the material-parameter mapping database are updated; Step 4: System Self-Check: The model's generalization ability is periodically verified; outdated policies are eliminated. Step 3: Target Co-evolution When users add constraints: Pareto front search: using the NSGA-III algorithm to solve for the optimal set of three-dimensional solutions for cutting quality / energy consumption / speed; constraint satisfaction surrogate model: training a graph neural network to predict the feasibility of parameter combinations; human-computer collaborative decision-making: visually displaying the distribution of optimal solutions and supporting interactive adjustment of weights by users.

[0012] A neural network-based laser processing control system includes: The user interaction and target input system includes core components such as a processing parameter setting module and a file reading module, which are used to receive user instructions, parse processing files, and set constraints. The intelligent decision-making central system's core components include an industrial control host, specifically an AI engine and a physical model library. The physical model library is used for predicting material responses using models such as electronic dynamics / molecular dynamics; the mapping database is used to store historical parameter-result relationships; and the AI ​​engine is used for machine learning model prediction results and optimization algorithms to search for the best parameter combinations. The laser parameter execution layer system consists of a laser source and a modulation optical path. Energy / pulse width control is used to adjust the pulse energy of the laser. Spatiotemporal shaping generates Bessel light / double-peak beams through a spatial light modulator and regulates the time sequence through a pulse shaper. The core components of the precision motion control system include a high-precision three-axis displacement platform and a controller. The real-time monitoring system's core components include a CCD camera, a spectrometer, and a sensor array. In-situ imaging uses a CCD to monitor surface morphology / plasma emission; spectral analysis uses LIBS to detect plasma composition and infer material state; and anomaly warning uses temperature / vibration sensors to capture thermal disturbances. The closed-loop learning mechanism system comprises a data platform and incremental learning algorithms as its core components. These are used to feed processing results back to the physical model library and AI engine, update model parameters online, and build a process knowledge graph to guide new tasks.

[0013] An electronic device includes a processor, a storage medium, and a computer program, the computer program being stored in the storage medium, and when executed by the processor, the computer program implements a neural network-based laser processing control method.

[0014] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a neural network-based laser processing control method.

[0015] The technical effects achieved by this invention are as follows: 1. Significant improvement in processing quality Submicron-level precision control: Through pulsed spatiotemporal shaping (such as Bessel beams) and real-time feedback, the sidewall roughness of the microchannel is reduced to <0.1μm, and the aspect ratio is increased to 20:1; Thermal damage approaches zero: Electron dynamics model (EDC) optimizes the non-thermal ablation path, reducing the thickness of the heat-affected zone from 10μm to <1μm, and the ingot fracture strength retention rate is >98%.

[0016] 2. Breakthrough improvement in efficiency Process development cycle shortened by 90%: The AI ​​engine (Bayesian optimization + reinforcement learning) converges to the optimal parameters in less than 50 iterations, which is 10 times faster than traditional methods; Online real-time control: Model predictive control (MPC) adjusts laser parameters every millisecond, the disturbance compensation response time is less than 100μs, and the processing yield is stable at over 99.5%.

[0017] 3. Resource consumption optimization Energy consumption reduced by 40%: The "laser parameter-processing result" mapping database supports precise energy matching, reducing single-point cutting energy from 2μJ to 1.2μJ; Material utilization improved: Crack propagation prediction accuracy reached 95%, and wafer cutting fragmentation rate decreased from 0.8% to 0.05%.

[0018] 4. System intelligent upgrade Explainable decision-making mechanism: Knowledge graph-based causal chains (e.g., pulse delay Δt → peak electron temperature → melt pool depth) support reverse reasoning of processes; Cross-material generalization capability: Small-sample transfer learning enables zero-sample parameter transfer from silicon to sapphire to gallium nitride, with a switching time of less than 1 hour. Attached Figure Description

[0019] Figure 1 This is a flowchart of the laser processing control method provided in an embodiment of the present invention; Figure 2 This is a flowchart of the intelligent control closed loop provided in an embodiment of the present invention; Figure 3 This is a flowchart illustrating the execution instructions provided by an embodiment of the present invention, which utilizes sensor feedback. Detailed Implementation

[0020] To make the objectives and advantages of this invention clearer, the invention will be specifically described below with reference to embodiments. It should be understood that the following text is merely used to describe one or more specific embodiments of the invention and does not strictly limit the scope of protection specifically claimed by the invention.

[0021] By actively controlling the shape (time and space) of femtosecond laser pulses, the electronic dynamics (EDC) inside the material can be modulated, thereby optimizing the micromachining results; by automatically adjusting the laser parameters through AI, a closed-loop control method and system combining physical models, real-time monitoring, and AI algorithms can be constructed. The laser processing control method based on neural networks specifically includes the following steps: I. Target Definition and Input The specific steps are as follows: (a) Setting processing targets: Specify processing requirements (e.g., specify the morphology required for crystal processing, microchannels with a specific aspect ratio, substrates with high SERS enhancement factor, nanowires with specific conductivity, taper-free micropores, high etching rate, etc.). (ii) Material property input: Input the known physical properties of the material to be processed (e.g., band structure, thermal conductivity, plasma frequency, etc.); (iii) Initial laser parameters: Input the initial laser parameter range (center wavelength, pulse width range, energy range, etc.).

[0022] II. Physical Model and Database: (a) Establishing a basic model library: Its core physical models include: (1) Electronic dynamics model: including non-equilibrium quantum dynamics simulation, carrier excitation and relaxation rate equation modeling, multi-band Keldysh theory for ionization rate prediction and non-perturbation processing method under strong optical field-matter coupling. This model is used to predict the electronic excitation and temperature change inside the material under different laser pulse shapes (energy distribution, time delay sub-pulse); (2) Molecular dynamics model: Classical molecular dynamics combines EAM or REBO potential functions and includes a reaction force field for phase transition and ablation processes. It also has phase transition criterion identification algorithms (based on local atomic coordination number, bond angle distribution, etc.) and thermodynamic path sampling techniques (such as Metadynamics). This model is used to predict phase transition (melting, ablation, resolidification) processes. (3) Plasma model: including two-fluid or multi-fluid models under the fluid dynamics framework, particle simulation methods, self-consistent electromagnetic field solvers, critical density surface tracking and shielding effect modeling. This model is used to predict possible ionization and plasma formation processes. (4) Dual-temperature model / improved model: including nonlocal thermal conduction correction, temperature-dependent electron-phonon coupling coefficient G(T), interface thermal resistance modeling in multilayer heterostructures, and implicit numerical solution scheme under time scale separation. This model is used to predict and describe the electron-lattice energy transfer process. The specific technical requirements are analyzed as follows: The specific operation process is as follows: Step 1: Model Selection and Adaptation Based on the target material (such as semiconductor ingots like silicon, sapphire, and silicon carbide), select the appropriate theoretical level: for wide bandgap materials, use the improved Keldysh+TTM combination; for metals or highly conductive materials, prioritize full PIC simulation; for composite structures (such as multilayer films), introduce cross-scale coupling interfaces; establish a "material-model matching rule base" to automatically recommend the optimal model combination.

[0023] Step 2: Parameter Calibration and Verification Using known experimental data (such as pump-probe spectra, transient reflectance variation curves), the key parameter, electron specific heat capacity C, is retrieved. e Electronic thermal conductivity k e Electron-phonon coupling constant G and ionization threshold intensity I th Parameter fitting is performed using Bayesian inference or least squares optimization methods to ensure that the model prediction accuracy error is less than 10%.

[0024] Step 3: Multiphysics Coupled Simulation Process Input initial laser pulse shape (time-controlled) → Electron dynamics model: calculate electron excitation density and energy deposition profile → Dual-temperature model: solve for electron temperature Te(t) and lattice temperature Tl(t) → Molecular dynamics model: determine phase transition (melting / vaporization / recrystallization) → Plasma model: analyze free electron growth, shielding and reverse bremsstrahlung radiation loss → Output: spatially resolved energy deposition map, temperature evolution curve, phase transition region volume, and initial surface morphology.

[0025] Step 4: Model order reduction and acceleration To meet the real-time call requirements of the subsequent AI engine, the high-fidelity model undergoes the following processes: intrinsic solution space compression, surrogate model training, and output of a lightweight "black box" model, which can complete a forward prediction within milliseconds.

[0026] (ii) Establish a mapping database of "laser parameters-electron dynamics-processing results": Based on historical experimental data (such as the examples listed in the document) and model simulation results, construct an associated database; for example: specific spatiotemporal pulse shape-predicted electron temperature / density distribution-expected microchannel morphology / etching rate / SERS enhancement factor / nanowire properties, etc. The specific technical requirements are analyzed as follows: (1) Data acquisition and fusion technology: including automated experimental data acquisition system (high-speed camera, spectrometer, AFM / SEM image analysis), numerical simulation batch generation toolchain (Python + LAMMPS / PIC script), and multi-source heterogeneous data cleaning and normalization processing; (2) Feature engineering and representation learning: including laser pulse spatiotemporal coding (Zernike moments, wavelet packet decomposition, pulse fingerprint vector); extraction of quantitative indicators of processing results: edge roughness Ra, aspect ratio AR, SERS enhancement factor EF, nanowire orientation order S; and automatic extraction of latent variable representation using VAE or CNN. (3) Database architecture design: Use graph database to store causal relationships, such as: [laser pulse] → [peak electron temperature] → [melting depth] → [crack length]; use vector database to support semantic retrieval of similar cases; use distributed SQL / NoSQL hybrid architecture to ensure scalability; (4) Uncertainty modeling and confidence assessment: Bayesian neural network (BNN) is used to output the prediction interval; Monte Carlo Dropout is used to estimate the model confidence; and a "knowledge integrity index" is introduced to evaluate the credibility level of each record. The specific operation process is as follows: Step 1: Multimodal data generation mechanism Construct a closed-loop data production pipeline: First, standard laser parameter set → physical model simulation, historical experimental log → data preprocessing; then, generate virtual samples through model simulation and extract real samples from historical data; subsequently, store the virtual samples and real samples in a unified format in the laser-EDC-result mapping library. The test covers typical operating conditions, including different material types, thicknesses, doping concentrations, ambient atmospheres (vacuum / inert gas), and scanning speeds. It also introduces a perturbation test set, incorporating noise, drift, and abnormal events (such as focus shift) to enhance robustness and generalization capabilities.

[0027] Step 2: Dimensional unification and correlation modeling to solve the cross-scale discontinuity problem between laser parameters, microscopic processes and macroscopic results; Step 2.1: Input Encoding Temporal shaping: chirp, burst mode, sub-pulse delay Δt; Spatial shaping: vortex beam, Bessel beam, multifocal array; Encoded as a high-dimensional tensor P∈R T×X×Y ; Step 2.2: Intermediate State Characterization Maximum electron extraction temperature T e max Rise time τ rise Energy deposition gradient ∇E; Define "non-thermal damage factor" η=T e / T l As a criterion for cold ablation; Step 2.3: Output Quantization Geometric morphology: microgroove aspect ratio, bottom flatness, sidewall cone angle; Functional properties: Raman enhancement factor (SERS), carrier mobility improvement, optical transmittance variation; Defect indicators: microcrack density, recast layer thickness, stress concentration factor.

[0028] Step 3: Knowledge Graph Construction and Reasoning Support. This uses a graph structure to express complex causal chains. This structure supports: forward queries: "Which pulses can produce non-thermal ablation?"; and backward reasoning: "How to adjust parameters to enhance SERS by >10?". 5 "?"; Analogical recommendation: "What are some better results under similar conditions?"

[0029] Step 4: Dynamic Updates and Version Management Design an incremental learning mechanism: automatically trigger incremental index updates in the database after a new experiment or simulation is completed; use a version control system (Git-LFS + DVC) to track each change to ensure traceability; set an "active data window": prioritize retaining the best 5% of samples from the most recent N iterations and discard outdated and inefficient data.

[0030] The aforementioned collaborative mechanism and systemic significance of establishing the basic model library and mapping database are as follows: At the positive-driving collaborative level, the basic model provides the database with a large number of high-quality virtual samples, compensating for the high cost and long cycle of experiments; at the feedback-closed-loop collaborative level, the real experimental results in the database feed back into the model parameter correction, forming a spiral upward progression of "simulation-experiment-optimization"; at the AI ​​readiness support collaborative level, the mapping database becomes the "prior knowledge pool" of the AI ​​core engine, significantly reducing the exploration space of reinforcement learning; and at the cross-scale bridging collaborative level, the model explains microscopic mechanisms, while the database connects macroscopic performance, jointly supporting "explainable AI decision-making". Without an accurate physical model, AI is prone to falling into the "black box fitting" trap; without large-scale mapping data, the model cannot generalize to new scenarios, and both are indispensable.

[0031] III. AI Core Engine: (a) Machine learning / deep learning model: The model is trained to learn the complex nonlinear relationship between "laser parameters - material response (electron dynamics prediction) - final processing result"; input: processing target, material properties, current / candidate laser parameters (including complex spatiotemporal shaping parameters); output: predicted processing result (morphology, size, efficiency, quality index) and uncertainty estimate; The specific technical requirements are analyzed as follows: (1) Deep Neural Networks (DNN) and Convolutional Neural Networks (CNN): used to process high-dimensional sensor data (such as high-speed imaging, spectral signals, acoustic emission signals) and extract spatial-temporal features; (2) Temporal modeling technology (LSTM / Transformer): captures the dynamic evolution process between laser pulse sequence and material response, suitable for online monitoring; (3) Physics-Informed Neural Networks (PINN): Embed physical laws such as heat conduction equations and nonlinear optical equations into the model structure to improve generalization ability and interpretability; (4) Few-shot learning and transfer learning: to address the problem of scarce labeled data when switching between different materials (such as silicon, glass, GaN); (5) Multimodal fusion model: Integrates multi-source sensing information such as optical microscopic images, infrared thermograms, and vibration signals to construct a comprehensive state representation; The specific operation process is as follows: Step 1: Data Acquisition and Preprocessing: Laser parameters (pulse width, repetition frequency, power, scanning speed) and material response signals are acquired through high-speed CCD, spectrometer, AE sensor, etc., and then denoised, normalized and time-frequency transformed. Step 2: Feature Engineering and Tag Generation: Combining offline experiments and simulations (such as FDTD and finite element analysis), establish the mapping relationship between "process parameters → morphology of internal modified zone → fracture behavior" and generate training tags; Step 3: Model training and validation: The model is trained under multiple material and thickness conditions using a cross-validation strategy, and interpretability analysis is performed using SHAP values ​​or Grad-CAM. Step 4: Online reasoning and feedback: Deployed on edge computing units, it predicts the quality risks of the current processing point (such as the probability of microcrack propagation) in real time and outputs early warning signals or adjustment suggestions.

[0032] (ii) Optimization algorithm: Based on the prediction model and the user-defined target (and possible constraints, such as maximum energy and minimum processing time), the optimal combination of laser parameters (center wavelength, pulse width, total energy, time-shaping sequence, spatial light modulator phase diagram, etc.) is searched using optimization algorithms (such as Bayesian optimization, evolutionary algorithm, reinforcement learning). The specific technical requirements are analyzed as follows: (1) Bayesian optimization: suitable for global optimization under low sampling budget, especially suitable for expensive experimental scenarios (high cost of each trial cut); (2) Reinforcement Learning (RL): The MDP framework is constructed with “environment = laser system + material”, “action = parameter adjustment”, and “reward = cutting quality − energy consumption” to achieve real-time adaptive control; (3) Multi-objective evolutionary algorithm (NSGA-II / MOEA / D): balances multiple conflicting objectives such as cutting speed, edge quality, and energy consumption, and generates Pareto front for decision-making; (4) Model Predictive Control (MPC): Combining a machine learning model as an internal predictor, the laser path and parameters for the next few steps are continuously optimized to suppress disturbances; (5) Online learning and adaptive correction mechanism: The surrogate model is continuously updated using newly collected data to prevent model drift; The specific operation process is as follows: Step 1: Initialize the search space: Set the initial parameter range based on experience or simulation (e.g., pulse energy 0.5–2 μJ, scan spacing 1–5 μm). Step 2: Iterative Exploration and Evaluation: Bayesian optimization uses a Gaussian process to model the uncertainty of the objective function and selects the next point with the largest information gain for experimentation; the reinforcement learning agent decides whether to increase the pulse density or adjust the focus position based on the current state (such as surface topography deviation); Step 3: Closed-loop execution and dynamic adjustment: MPC calls the prediction model every millisecond to calculate the optimal parameter sequence for the next 10 scan points; when a sudden change in material thickness is detected, the online fine-tuning mechanism is triggered to quickly converge to the new optimal solution; Step 4: Knowledge Accumulation and Transfer: Store the successfully optimized path in the process knowledge graph to support the rapid launch of new materials.

[0033] The two modules mentioned above do not operate in isolation, but are embedded in an intelligent control closed loop of "perception-decision-execution-feedback". Please refer to the appendix for details. Figure 2 ; This architecture achieves a leap from "passive control" to "active cognition," and has the following advantages: strong anti-interference: it can automatically compensate for batch differences in materials and changes in environmental temperature and humidity; shortened process development cycle: the process debugging that traditionally requires hundreds of trials can be converged by AI in less than 50 trials; edge intelligent deployment: it can run in real time on FPGA or industrial control computer through model compression (knowledge distillation, quantization).

[0034] IV. Automatic Adjustment and Execution of Laser Parameters: (a) Control command generation: The optimal parameter combination output by the AI ​​optimization engine is converted into specific control commands; The specific technical requirements are analyzed as follows: (1) Parameter mapping algorithm: Spatiotemporal shaping coding: Converting the pulse shape (such as Bessel beam, vortex phase) output by AI into the phase diagram of the spatial light modulator (SLM) requires Zernike polynomial decomposition or wavelet packet transform technology; Time series generation: Converting parameters such as pulse width and sub-pulse delay into the driving signal of the acousto-optic modulator (AOM) using chirped pulse coding technology. (2) Constraint Satisfaction Engine: Introducing the Mixed Integer Programming (MIP) algorithm to adjust parameters while satisfying hardware limits (such as maximum laser power and scanning galvanometer acceleration); (3) Safety boundary constraint technology Real-time threshold monitoring: Integrating a database of material damage thresholds (e.g., the ablation threshold of silicon, 2.0 J / cm). 2 Automatically limits AI output parameters from exceeding safe limits; Redundancy check mechanism: By pre-simulating the execution results of instructions in Monte Carlo simulation, abnormal combinations that may damage optical components (such as ultra-high energy + minimum spot size) are intercepted. The specific operation process is as follows: Step 1: Parameter Reception and Parsing The AI ​​engine outputs a structured parameter package (JSON format), and the instruction generation module parses the parameters and matches them with a preset hardware driver template. Step 2: Multi-device instruction collaborative generation The instruction generation logic of the spatial light modulator is to call the GPU-accelerated algorithm to generate a phase hologram, and output a 512×512 grayscale bitmap (8-bit depth); the instruction generation logic of the pulse shaper is to compile the time series into an AOM RF drive signal (time resolution 50fs, voltage accuracy ±0.1V); the instruction generation logic of the laser power supply is to calculate the energy-voltage mapping curve and generate a DAC control waveform (e.g., 1.8μJ → output voltage 124.7V ±0.5%). Step 3: Cross-system timing synchronization Construct a timing chain with nanosecond-level precision (based on the PTP protocol); timing jitter control <100ns to ensure precise matching of spatial-temporal parameters; Step 4: Virtual Execution Verification The digital twin platform simulates the effect of commands: it simulates the propagation path of light beams through ray tracing; it calls the physical model to predict the material response (such as peak electronic temperature and melting depth); if the prediction result deviates from the target by more than 5%, it automatically triggers the parameter rollback mechanism. Step 5: Instruction encapsulation and issuance Packed into device-specific binary instruction frames, they are broadcast to each execution unit via real-time Ethernet with a latency of <1ms.

[0035] (ii) Hardware control: Commands are sent to the laser control system (adjusting energy and pulse width), pulse shaper (adjusting time shape), and spatial light modulator (adjusting spatial shape to generate double peaks, Bessel beams, etc.). The specific technical requirements are analyzed as follows: (1) Laser dynamic control technology High-precision power / pulse width control: Microsecond-level adjustment of laser energy is achieved using an acousto-optic modulator (AOM) or an electro-optic modulator (EOM), with power stability required to reach ±1%; Wavelength tuning mechanism: The center wavelength (typically 400nm–2000nm) can be adjusted by using a tunable laser crystal (such as Ti: Sapphire) or an optical parametric oscillator (OPO) to adapt to the absorption characteristics of different materials; (2) Pulse time shaping technology Fourier domain pulse shaper: It is composed of a liquid crystal spatial light modulator (LC-SLM) or a digital micromirror device (DMD) and generates complex pulse sequences (such as chirped pulses and multi-pulse trains) by programming the phase diagram. Real-time delay control: The sub-pulse interval is precisely adjusted (accuracy <10fs) using fiber optic delay lines or piezoelectric displacement stages to achieve active control of electron dynamics; (3) Beam spatial shaping technology Phase-type spatial light modulator (SLM): By loading Zernike polynomials or holograms, non-Gaussian distributions such as Bessel beams and vortex beams can be generated, and the focal size can be compressed to 70% of the diffraction limit; Multi-focal parallel processing: By using diffractive optical elements (DOE) to split the beam, more than 10 independent focal points can be controlled simultaneously to improve processing efficiency; (4) Multi-device collaborative control architecture Hardware synchronization interface: Adopting an FPGA + real-time Ethernet (EtherCAT) architecture, ensuring that the command synchronization deviation of the laser, galvanometer, and modulator is <100ns; Safety interlock mechanism: Integrated overcurrent / overtemperature sensor and emergency stop circuit, which can cut off laser output within 1ms in case of abnormality; The specific operation process is as follows: Step 1: Instruction parsing and distribution AI engine output parameter set → controller parses into device commands: Laser: Adjusts the pump current to the target energy value; Pulse shaper: Loads the corresponding phase mask; Spatial light modulator: Generates a Bessel beam to calculate a hologram (CGH); Step 2: Dynamic Compensation and Calibration Commands are executed using sensor feedback; please refer to the appendix for details. Figure 3 ; Step 3: Joint Execution of Spatiotemporal Parameters Timing control: The pulse shaper completes the loading of a new waveform within 50μs → the laser emits in sequence (e.g., BurstMode@500kHz); Spatial control: SLM refresh rate > 100Hz → galvanometer scans at a speed of 5m / s in coordination with focus movement; Step 4: Closed-loop verification and fault tolerance Real-time quality monitoring: The processing depth is detected by a confocal microscope probe, and parameter re-optimization is triggered when the deviation is >5%. Emergency response to faults: If the plasma sensor detects abnormal sputtering, immediately switch to safe pulse mode (energy reduced to 10%).

[0036] V. Real-time monitoring and feedback: (a) Process monitoring: Real-time or in-situ monitoring using the advanced diagnostic technologies mentioned in the document: Laser-induced breakdown spectroscopy (LIBS): Analyzes plasma composition to indirectly reflect the material removal status; Fast imaging CCD: Observes changes in surface morphology or plasma luminescence during the processing; The specific technical requirements are analyzed as follows: (1) Multimodal sensing technology High-speed imaging and spectral sensing: High-speed CCD (>100,000 fps) is used to capture material surface deformation and plasma plume dynamics; integrated Raman spectrometer is used to analyze material phase transitions (such as amorphization and lattice vibrations) in real time. Acoustic emission (AE) and thermal imaging sensing: AE sensors detect stress wave signals of microcrack propagation (frequency range 20kHz-1MHz); infrared thermal imagers monitor the temperature field distribution in the heat-affected zone (HAZ) with an accuracy of ±0.5℃. (2) Edge computing and real-time data processing Based on an FPGA-based streaming architecture, it enables data preprocessing (denoising and feature extraction) with microsecond-level latency; lightweight AI model deployment (such as CNN accelerated by TensorRT); and real-time identification of processing anomalies (such as uneven ablation and cracks). (3) Feedback control mechanism Adaptive PID controller: dynamically adjusts laser power and scanning speed based on monitoring data; Reinforcement learning (RL) agent: optimizes parameters online using processing quality scores (such as roughness and crack density) as the reward function; The specific operation process is as follows: Step 1: Data Acquisition and Feature Extraction Input: Real-time multimodal data stream of the laser-affected area (image, spectrum, acoustic signal, temperature); Processing flow: High-speed imaging, Raman spectroscopy, acoustic emission, thermal imaging → Feature extraction → Edge computing node → Anomaly detection model → Generation of status report; Outputs include quantitative indicators such as plasma density, molten pool size, and surface roughness Ra value. Step 2: State Diagnosis and Decision Generation Rule engine: Preset thresholds trigger alarms (e.g., forced power reduction when temperature > material melting point); AI prediction model: Predicts the final processing quality under current parameters based on historical data, and triggers adjustments if the prediction deviation > 10%; Step 3: Closed-loop control execution Command issued: Adjust laser energy: Adjust laser power supply voltage via DAC module; Correct optical path: Spatial light modulator (SLM) generates new phase map in real time to optimize beam spatial distribution; Dynamic response timing: Data acquisition to capture transient phenomena; edge computing to complete feature analysis and diagnosis; command execution to adjust laser parameters; Step 4: Knowledge Accumulation and Model Iteration Data from each processing cycle (parameters - monitoring results - final quality) is automatically stored in the mapping database; incremental training is triggered weekly to update AI model weights and optimize control strategies.

[0037] (ii) Result evaluation (optional): After processing, key indicators of the actual processing results are obtained by means of microscopic imaging (SEM, AFM), electrical testing, optical testing (such as SERS measurement).

[0038] (iii) Feedback loop: The monitored process signals (or final measurement results) are fed back to the AI ​​engine. The AI ​​engine compares this information with the expected target: if the result meets the expectation, the parameters are valid and can be stored in the database; if there is a deviation, the AI ​​engine adjusts its prediction model or re-optimizes the parameters based on the new information and makes the next processing attempt (adaptive control). The specific technical requirements are analyzed as follows: (1) Edge computing and real-time analysis engine Streaming data processing architecture: Apache Kafka / Kinesis enables millisecond-level pipelined processing of sensor data; FPGA-accelerated parallel computing modules support real-time feature extraction (such as plasma area and temperature gradient). Online quality prediction model: Lightweight LSTM network predicts the probability of processing defects (e.g., sidewall cone angle deviation > 5%); Physics-based anomaly detection (dual-temperature model simulation results vs. measured temperature curves); (2) Dynamic decision-making and instruction generation Reinforcement learning control agent: Optimizes laser parameters online using processing quality (e.g., surface roughness Ra) as the reward function; Safety constraint module: Prevents energy exceeding limits or sudden changes in focus position; Digital twin-assisted decision-making: Real-time invocation of reduced-order physics models (such as simplified molecular dynamics) to preview the consequences of parameter adjustments; (3) Control execution and hardware interface Low-latency communication protocol: EtherCAT bus (period ≤ 100μs) transmits adjustment commands; OPC UA protocol integrates spatial light modulator (SLM) phase map updates; Adaptive actuator: a piezoelectric ceramic driven focusing lens (response time 0.2ms); an acousto-optic modulator (AOM) enables fine-tuning of pulse energy (accuracy ±0.5%). The specific operation process is as follows: Step 1: Real-time Status Analysis and Quality Assessment Feature extraction: Spectral analysis: ablation depth was calculated using PLS regression (error < 5 μm); Image processing: Plasma region was segmented using U-Net, and plume symmetry index was calculated; Anomaly diagnosis: Compare with standard process window: If the width of the heat-affected zone is greater than the set threshold, trigger a level 3 alarm; Causal reasoning engine: Correlate energy fluctuations with the probability of crack formation (Bayesian network). Step 2: Dynamic parameter adjustment and execution Decision logic: For example, when the plasma asymmetry index is >0.7, adjust the SLM phase diagram to generate a Bessel beam; when the thermal gradient is >10^6 K / m, reduce the pulse energy by 20% and increase the pulse interval by 50 fs; at the same time, maintain the current parameters and update the reward function. Command issuance: Control commands are prioritized and scheduled by the real-time operating system (such as RT-Linux); SLM loads a new phase map (refresh rate 1kHz), AOM calibrates the energy; Phase 3: Closed-loop verification and knowledge accumulation Effect verification: Next scan point monitoring: verify the roughness improvement rate (e.g., from Ra 1.2μm→0.8μm); Post-processing in-situ inspection: verify the consistency of microgroove depth using confocal microscopy; Data closed loop: Successful cases are stored in the graph database: "Parameter Adjustment Strategy" and "Quality Improvement Indicators" are linked; Failed cases trigger physical model recalibration: update the electron-phonon coupling coefficient G(T).

[0039] VI. Iterative Learning and System Improvement: (i) Every successful (or even unsuccessful) processing attempt and its data will be used to update the AI ​​model and physical model database; The specific technical requirements are analyzed as follows: (1) Incremental learning algorithm: The AI ​​prediction model is updated in real time by using online learning models (such as Online Random Forest and incremental SVM); the new data features are mapped to the existing model by combining transfer learning to avoid repeated training; new data is quickly integrated on the basis of retaining historical knowledge to adapt to material batch differences or process changes; (2) Bayesian optimization of physical model parameters: Based on Markov chain Monte Carlo (MCMC) sampling or variational inference, the physical model parameters are corrected by inversion using new processed data; the parameter uncertainty is quantified by Gaussian process regression (GPR) to guide subsequent experimental design; (3) Multi-source data fusion and knowledge distillation: Graph neural networks (GNN) are used to integrate experimental data, simulation results and sensor stream data (such as thermal imaging and spectral signals) to construct a unified feature space; knowledge distillation is used to compress complex models into lightweight proxy models to adapt to edge device deployment; (4) Dynamic indexing and version control of the database: Vector databases (such as Pinecone) update embedded vectors in real time and support similar case retrieval; Git-LFS + DVC toolchain manages data versions and records the impact of each update on model performance; The specific operation process is as follows: Step 1: Data Collection and Quality Assessment Real-time acquisition of processing data: laser parameters (energy, pulse width), material response (eTe, phase transition region), and result indicators (crack density, surface roughness Ra); filtering of noisy data through anomaly detection models (such as Isolation Forest) and calculation of "data confidence score"; Step 2: Incremental Model Update AI model update: After new data input, online backpropagation is triggered to update the neural network weights, and elastic weight solidification (EWC) is used to prevent catastrophic forgetting; Example: If new data shows that a certain pulse shape causes crack propagation, the model reduces the recommendation priority of this type of parameter combination; Physical model calibration: The processing results are compared with the predictions of the dual-temperature model, and the parameters of the G(T) function are adjusted through Bayesian optimization. The error is fed back to the molecular dynamics model. Step 3: Dynamic Expansion of the Knowledge Graph Analyze causal relationships in new data (e.g., "Bessel beam → electron density gradient ↑ → sidewall cone angle ↓"); update the associated edges and node attributes in the graph database (Neo4j), support semantic queries, and introduce causal inference models (e.g., Do-Calculus) to distinguish between correlation and causality, avoiding misleading associations; Step 4: Closed-loop verification and version release Test the updated model in a digital twin environment to simulate extreme conditions (such as material impurities and focus shift); compare the performance of the new and old models through A / B testing, and deploy only if the following conditions are met: prediction error reduction >5% and stability of key indicators (such as heat-affected zone thickness) is improved; release the new model version to edge devices, and archive the old version for disaster recovery.

[0040] (ii) The system continuously learns the optimal parameter strategy under new materials and new targets, improves prediction accuracy and optimization efficiency, and achieves increasingly intelligent automatic parameter adjustment.

[0041] The specific technical requirements are analyzed as follows: (1) Meta-Learning: adopts the paradigm of "learning how to learn" to train models to quickly adapt to new tasks. For example, MAML (Model-Agnostic Meta-Learning) is used to initialize a set of model parameters that are sensitive to changes in the task, and convergence can be achieved with only a small amount of new data for fine-tuning; a prototype space of new material features is constructed through Prototypical Networks, and an initial parameter strategy is generated through similarity matching. (2) Transfer learning and federated learning Cross-material knowledge transfer: Pre-trained models learn physical laws (such as thermal conduction and phase transition threshold) on silicon / sapphire data and transfer them to new materials through feature decoupling technology; Distributed collaborative evolution: Multiple devices share encrypted model gradients through a federated learning framework to build an industry-level knowledge pool without leaking the original data (meeting industrial confidentiality requirements). (3) Reinforcement learning adaptive optimization Contextual RL: Material properties (band gap, heat capacity) and target indicators (crack tolerance, shear rate) are used as state variables; Safety exploration mechanism: Constrains the search range of parameters to avoid material explosion caused by excessive energy. Improved reward function: R = α⋅cutting quality + β⋅efficiency − γ⋅energy loss − δ⋅defect penalty term; (4) Digital twin-driven simulation and pre-play High-fidelity virtual environment: Integrating molecular dynamics (MD) and two-temperature model (TTM) for real-time simulation of nonlinear interactions between lasers and new materials; Active sampling strategy: Based on Bayesian optimization uncertainty sampling, priority is given to simulating high-risk parameter regions (such as near ablation threshold pulses). (5) Causal inference and interpretability analysis Structural Causal Model (SCM): Analyzes the causal chain of "pulse shape → electron temperature → phase transition type → cutting quality" to avoid misleading correlations; Dynamic knowledge graph update: New experimental conclusions are entered into the database in the form of triplets (e.g., (novel silicon carbide, pulse width tolerance, <100fs)). The specific operation process is as follows: Step 1: Cold Start of New Materials / New Tasks Input the parameters of the new material, the system queries the knowledge graph, loads the migration strategy template when similar materials exist, starts the digital twin simulation when there is no matching record, and generates the safety parameter boundary. The results of both are then used to conduct small sample exploration experiments, and finally output the evaluation of the initial performance. Step 2: Online Incremental Learning Loop Step 1: Data Awareness; Step 2: Strategy Optimization: The RL agent selects actions based on the state (material response + target offset) (e.g., increasing the number of spatial modulation rings); the meta-model updates the policy network weights every 10 minutes; Step 3: Knowledge Accumulation: Successful strategies are stored in the case library, and failed cases trigger causal analysis; the confidence weights of the material-parameter mapping database are updated; Step 4: System Self-Check: The model's generalization ability is periodically verified (e.g., through adversarial example testing); outdated strategies are phased out (e.g., solutions with stability <90% are automatically deactivated); Step 3: Target Co-evolution When a user adds a new constraint (such as "reduce energy consumption by 30%)): Pareto Front Search: Solving for the optimal three-dimensional solution set of cutting quality / energy consumption / speed using the NSGA-III algorithm; Constraints satisfying surrogate model: feasibility of training a graph neural network (GNN) to predict parameter combinations; Human-machine collaborative decision-making: Visualizes the distribution of optimal solutions and supports interactive adjustment of weights by users.

[0042] A neural network-based laser processing control system includes: The user interaction and target input system has core components including a processing parameter setting module and a file reading module, which are used to receive user instructions (such as cutting patterns and quality requirements), parse processing files (CAD / G code), and set constraints (energy limit and time limit). The intelligent decision-making central system consists of an industrial control host (AI engine + physical model library). The physical model library is used to predict material responses using models such as electronic dynamics / molecular dynamics; the mapping database is used to store historical parameter-result relationships; and the AI ​​engine is used to predict results from machine learning models and optimize algorithms to search for the best parameter combinations. The laser parameter execution layer system consists of core components including a laser source and a modulation optical path. Energy / pulse width control is used to adjust the pulse energy of the laser (μJ level). Spatiotemporal shaping is achieved by generating Bessel light / double-peak beams through a spatial light modulator and controlling the time sequence through a pulse shaper. The core components of the precision motion control system include a high-precision three-axis displacement platform and a controller. The real-time monitoring system's core components include a CCD camera, a spectrometer, and a sensor array. In-situ imaging uses a CCD to monitor surface morphology / plasma emission; spectral analysis uses LIBS to detect plasma composition and infer material state; and anomaly warning uses temperature / vibration sensors to capture thermal disturbances. The closed-loop learning mechanism system comprises a data platform and incremental learning algorithms. Its core components include data feedback (processing results are fed back to the physical model library and AI engine), dynamic optimization (model parameters are updated online, such as the electron-phonon coupling coefficient), and knowledge accumulation (constructing a process knowledge graph to guide new tasks).

[0043] An electronic device includes a processor, a storage medium, and a computer program, wherein the computer program is stored in the storage medium and, when executed by the processor, implements a neural network-based laser processing control method.

[0044] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a neural network-based laser processing control method.

[0045] The above description is merely a preferred embodiment of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention. Structures, devices, and operating methods not specifically described or explained in this invention are implemented according to conventional methods in the art unless otherwise specified or limited.

Claims

1. A laser processing control method based on neural networks, characterized in that, Specifically, the following steps are included: Target definition and input phase; Physical Model and Database Phase: Establish a basic model library; establish a mapping database of "laser parameters-electron dynamics-processing results": construct an associated database based on historical experimental data and model simulation results; The process of establishing the "laser parameters-electron dynamics-processing results" mapping database is as follows: A closed-loop data production pipeline is constructed based on a multimodal data generation mechanism: standard laser parameters are collected for physical model simulation, and data preprocessing is performed based on historical experimental logs; virtual samples are generated through model simulation, and real samples are extracted from historical data; Virtual samples and real samples are entered into the laser-EDC-result mapping library in a unified format. Dimensional unification and correlation modeling solve the cross-scale discontinuity problem between laser parameters, microscopic processes and macroscopic results; Input encoding: Time shaping is performed through chirping, pulse trains, and sub-pulse delays; spatial shaping is performed through vortex beams, Bessel beams, and multifocal arrays; encoding is done as a high-dimensional tensor P∈R. T×X×Y ; Intermediate state characterization: extracting the maximum electron temperature, rise time, and energy deposition gradient; defining a "non-thermal damage factor" as a criterion for cold ablation; Outputs the geometry of the microgroove, including aspect ratio, bottom flatness, and sidewall cone angle; outputs functional properties such as Raman enhancement factor, carrier mobility improvement, and optical transmittance change; and outputs defect indices such as microcrack density, recast layer thickness, and stress concentration factor. Knowledge graph construction and reasoning support utilizes graph structures to express complex causal chains; Dynamic updates and version management: Design an incremental learning mechanism to automatically trigger incremental index updates in the database upon completion of a new experiment or simulation; a version control system tracks each change to ensure traceability; set an "active data window" to prioritize retaining the best 5% of samples from the most recent N iterations and discard outdated and inefficient data; AI Core Engine Stage: Machine Learning / Deep Learning Model: Training the model to learn the complex nonlinear relationship between "laser parameters - material response - final processing result"; Optimization Algorithm: Based on the prediction model and the user-defined target, the optimization algorithm searches for the optimal combination of laser parameters; Automatic laser parameter adjustment and execution stage: Control command generation: The optimal parameter combination output by the AI ​​optimization engine is converted into specific control commands; Hardware control: Commands are sent to the laser control system, pulse shaper, and spatial light modulator. Real-time monitoring and feedback phase: Process monitoring: Real-time or in-situ monitoring using diagnostic technology; Result evaluation: Obtaining key indicators of actual processing results; Feedback loop: Feeding the monitored process signals back to the AI ​​engine, which compares this information with the expected target. Iterative learning and system improvement phase: Each processing attempt and its data are used to update the AI ​​model and physical model database; the system continuously learns the optimal parameter strategy under new materials and new targets, and achieves automatic parameter adjustment; The machine learning / deep learning model operates as follows: Laser parameters and material response signals are acquired by high-speed CCD, spectrometer, AE sensor, etc., and then denoised, normalized and time-frequency transformed. By combining offline experiments and simulations, a mapping relationship between "process parameters - morphology of internal modified zone - fracture behavior" is established, and training labels are generated. A cross-validation strategy was adopted to train the model under multiple material and thickness conditions, and interpretability analysis was performed using SHAP values ​​or Grad-CAM. Deployed in edge computing units, it can predict the quality risks of the current processing point in real time and output early warning signals or adjustment suggestions; The search process for the optimal combination of laser parameters is as follows: Set the initial parameter range based on experience or simulation; Bayesian optimization uses a Gaussian process to model the uncertainty of the objective function and selects the next point with the maximum information gain for the experiment; the reinforcement learning agent decides whether to increase the impulse density or adjust the focus position based on the current state. MPC calls the prediction model every millisecond to calculate the optimal parameter sequence for the next 10 scan points; when a sudden change in material thickness is detected, an online fine-tuning mechanism is triggered to quickly converge to the new optimal solution; Successfully optimized paths will be stored in the process knowledge graph to support the rapid launch of subsequent new materials.

2. The laser processing control method based on neural networks according to claim 1, characterized in that: The process of establishing the basic model library is as follows: Select the appropriate theoretical level based on the target material; establish a "material-model matching rule base" to automatically recommend the optimal model combination; Use known experimental data to invert key parameters and complete parameter fitting; Input the initial laser pulse shape to obtain the output spatially resolved energy deposition map, temperature evolution curve, phase transition region volume, and initial surface morphology. The high-fidelity model undergoes intrinsic solution space compression, surrogate model training, and output of a lightweight "black box" model.

3. The laser processing control method based on neural networks according to claim 1, characterized in that: The process of establishing the "laser parameters-electron dynamics-processing results" mapping database is as follows: By utilizing a multimodal data generation mechanism, a closed-loop data production pipeline can be constructed. Dimensional unification and correlation modeling solve the cross-scale discontinuity problem between laser parameters, microscopic processes and macroscopic results; It uses graph structures to express complex causal chains, supporting forward queries, backward reasoning, and analogical recommendations; Dynamic updates and version management.

4. The laser processing control method based on neural networks according to claim 1, characterized in that: The control command generation process is as follows: The AI ​​engine outputs a structured parameter package, and the instruction generation module parses the parameters and matches them with a preset hardware driver template. The instruction generation logic of the spatial light modulator is to call the GPU-accelerated algorithm to generate a phase hologram, and the output is a 512×512 grayscale bitmap. The instruction generation logic of the pulse shaper is to compile the time series into the RF drive signal of the AOM; the instruction generation logic of the laser power supply is to calculate the energy-voltage mapping curve and generate the DAC control waveform. Construct timing chains with nanosecond-level precision; timing jitter control <100ns to ensure precise matching of spatial-temporal parameters; The digital twin platform simulates the effect of commands: it simulates the propagation path of light beams through ray tracing; it calls the physical model to predict the material response; if the prediction result deviates from the target by more than 5%, it automatically triggers the parameter rollback mechanism. Packed into device-specific binary instruction frames, they are broadcast to each execution unit via real-time Ethernet with a latency of <1ms.

5. The laser processing control method based on neural networks according to claim 1, characterized in that: The process monitoring operation is as follows: Input the real-time multimodal data stream of the laser-acting region to obtain quantitative indicators such as plasma density, molten pool size, and surface roughness Ra value; An alarm is triggered based on a preset threshold; the final processing quality under the current parameters is predicted based on historical data, and adjustments are triggered if the prediction deviation is greater than 10%. The laser energy is adjusted by issuing commands, and the laser power supply voltage is adjusted through the DAC module; the optical path is corrected, and a new phase diagram is generated in real time through the spatial light modulator. Data acquisition, capturing transient phenomena, completing feature analysis and diagnosis, and adjusting laser parameters; Data from each processing cycle is automatically stored in the mapping database; AI model weights are updated, and control strategies are optimized.

6. The laser processing control method based on neural networks according to claim 1, characterized in that: The operational process of the optimal parameter strategy for learning new materials and new objectives is as follows: Input the parameters of the new material, the system queries the knowledge graph, loads the migration strategy template when similar materials exist, starts the digital twin simulation when there is no matching record, and generates the safety parameter boundary. The results of both are then used to conduct small sample exploration experiments, and finally output the evaluation of the initial performance. The online incremental learning closed loop consists of four steps: First, data awareness; Second, policy optimization: the RL agent selects actions based on state; the meta-model updates the policy network weights every 10 minutes; Third, knowledge accumulation: successful policies are stored in the case library, and failed cases trigger causal analysis; the confidence weights of the material-parameter mapping database are updated; Fourth, system self-checking: the model's generalization ability is periodically verified; outdated policies are eliminated. When users add constraints: the NSGA-III algorithm is used to solve for the optimal three-dimensional solution set of cutting quality / energy consumption / speed; the feasibility of parameter combinations for training a graph neural network to predict the surrogate model is satisfied by the constraints; the optimal solution distribution is visualized through human-computer collaborative decision-making, and users can interactively adjust the weights.

7. A neural network-based laser processing control system, used to execute the neural network-based laser processing control method according to claims 1-6, characterized in that, include: The user interaction and target input system includes core components such as a processing parameter setting module and a file reading module, which are used to receive user instructions, parse processing files, and set constraints. The intelligent decision-making central system's core components include an industrial control host, specifically an AI engine and a physical model library. The physical model library is used for predicting material responses using models such as electronic dynamics / molecular dynamics; the mapping database is used to store historical parameter-result relationships; and the AI ​​engine is used for machine learning model prediction results and optimization algorithms to search for the best parameter combinations. The laser parameter execution layer system has core components including a laser source and a modulation optical path, where energy / pulse width control is used to adjust the pulse energy of the laser. Spatiotemporal shaping involves generating Bessel light / double-peak beams using a spatial light modulator and controlling the time series using a pulse shaper. The core components of the precision motion control system include a high-precision three-axis displacement platform and a controller. The real-time monitoring system's core components include a CCD camera, a spectrometer, and a sensor array. In-situ imaging uses a CCD to monitor surface morphology / plasma luminescence; spectral analysis uses LIBS to detect plasma composition and infer material state; and anomaly warning uses temperature / vibration sensors to capture thermal disturbances. The closed-loop learning mechanism system comprises a data platform and incremental learning algorithms as its core components. These are used to feed processing results back to the physical model library and AI engine, update model parameters online, and build a process knowledge graph to guide new tasks.

8. An electronic device comprising a processor, a storage medium, and a computer program, wherein the computer program is stored in the storage medium, characterized in that, When the computer program is executed by the processor, it implements the neural network-based laser processing control method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the neural network-based laser processing control method according to any one of claims 1 to 6.

Images