Adaptive optical polishing system and method based on dynamic temperature control and multi-modal detection

Abstract

The application discloses a self-adaptive optical polishing system and method based on dynamic temperature control and multi-modal detection, and relates to the technical field of intelligent manufacturing and precision machining.The optical polishing system comprises a dynamic temperature control and multi-modal detection module, a flexible actuator, a closed-loop control and optimization module, and a cross-scale coordination and generalization module.The application has the following advantages: through the dynamic temperature control and multi-modal detection technology, the temperature field, surface topography and micro-crack signals of the polishing area are monitored in real time;the thermal deformation is quickly compensated by the flexible actuator;the problem of surface shape precision decline caused by temperature rise in the traditional process is effectively avoided;the multi-modal sensors work cooperatively to comprehensively capture multi-dimensional defects in the polishing process;the defect detection sensitivity and process stability are significantly improved;and the machining quality of high-precision optical elements is ensured.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent manufacturing and precision machining, in particular to an adaptive optical polishing system and method based on dynamic temperature control and multi-modal detection. BACKGROUND

[0002] With the wide application of high-precision optical elements in the fields of semiconductors, aerospace, etc., the traditional optical polishing technology faces many bottlenecks:

[0003] Firstly, local temperature rise in the polishing process easily causes material thermal deformation, resulting in a decrease in surface shape precision;

[0004] Secondly, single modal detection is difficult to capture multi-dimensional defects such as temperature field, surface roughness and micro-cracks in real time, and process parameter adjustment lags behind;

[0005] Thirdly, the existing actuator lacks flexibility and is difficult to adapt to complex curved surfaces and heterogeneous materials, relying on manual experience to adjust parameters, which is low in efficiency;

[0006] In view of the above problems, the prior art attempts to alleviate thermal deformation through fixed temperature control strategy or offline compensation method, but the dynamic response is insufficient, and it is impossible to cooperate with multi-sensor data for global optimization, at the same time, the control model based on rules has weak generalization ability, and it is difficult to adapt to the needs of new materials or process changes, the present application proposes an adaptive polishing system integrating dynamic temperature control, multi-modal detection and intelligent optimization, which realizes real-time multi-physical field monitoring by integrating infrared thermal imager, laser interferometer and acoustic emission sensor, dynamically compensates thermal deformation by combining shape memory alloy polishing head and micro-fluidic cooling system, realizes process parameter self-optimization by adopting hierarchical closed-loop control architecture (PID fast cycle + LSTM + PPO reinforcement learning slow cycle), and adapts to cross-material process migration through transfer learning framework, which significantly improves the polishing precision, efficiency and process generalization ability. SUMMARY

[0007] The present application aims to provide an adaptive optical polishing system and method based on dynamic temperature control and multi-modal detection.

[0008] To achieve the above-mentioned purpose, the present application provides the following technical scheme: an adaptive optical polishing system based on dynamic temperature control and multi-modal detection, the optical polishing system comprising:

[0009] a dynamic temperature control and multi-modal detection module integrating an infrared thermal imager, a laser interferometer and an acoustic emission sensor, for real-time acquisition of temperature field, surface topography and micro-crack signals of the polishing area;

[0010] a flexible actuator comprising a shape memory alloy polishing head and a multi-channel micro-fluidic cooling system, wherein:

[0011] The polishing head realizes local deformation compensation by current driving, and the deformation amount satisfies a relationship formula:

[0012] ;

[0013] In the formula, is a deformation coefficient, is a driving current.

[0014] The cooling system controls the flow and temperature of the cooling liquid by gradient, and the flow distribution satisfies an exponential decay formula:

[0015] ;

[0016] In the formula, is the radius from the center of the high-temperature area.

[0017] The closed-loop control and optimization module adopts a hierarchical control architecture, including a fast cycle layer (FPGA) for real-time regulation of temperature and pressure parameters, and a slow cycle layer (GPU acceleration) for running an LSTM+PPO reinforcement learning model for global process optimization.

[0018] The cross-scale collaboration and generalization module is based on a process parameter mapping table and a transfer learning framework, and is suitable for different materials and curved surface types.

[0019] As a further scheme of the present application, the multi-modal sensor satisfies the following conditions:

[0020] The infrared thermal imager has a spatial resolution of ≤0.1℃, a sampling frequency of 100Hz, and a positioning temperature difference of >5℃C, and a local hotspot with a diameter of >2mm;

[0021] The laser interferometer has a wavelength of 632.8nm, generates a three-dimensional height map with a resolution of 10nm every 5 seconds, and identifies the surface shape error PV value and the roughness Ra;

[0022] The acoustic emission sensor has a frequency band of 50-500KHz, and extracts a 200kHz frequency band energy burst signal through short-time Fourier transform (STFT).

[0023] As a further scheme of the present application, the temperature-pressure coupling model includes:

[0024] The physical driving layer: based on finite element simulation, a thermal expansion equation is established:

[0025] ;

[0026] In the formula, is a thermal expansion coefficient, is a temperature rise, is a characteristic dimension;

[0027] Data-driven layer: input real-time temperature and pressure sequence into LSTM network to predict thermal deformation trend in next 3 seconds;

[0028] Instruction generation layer: fuse physical and data prediction results to output high-temperature zone coolant flow instruction (120-200 ml / min) and temperature set value .

[0029] As a further scheme of the present application: the shape memory alloy polishing head is a nickel-titanium alloy (NiTiNOL) grid structure, and a micro-resistance heating sheet is embedded in a grid node.

[0030] As a further scheme of the present application: the multi-channel microfluidic cooling system comprises 16 channels of independently controlled piezoelectric micropumps and semiconductor refrigerating sheets, wherein:

[0031] The high-temperature zone coolant flow is 200 ml / min, and the temperature set value is 5℃;

[0032] The transition zone flow is distributed according to an exponential decay formula, and the temperature following strategy is .

[0033] As a further scheme of the present application: the hierarchical architecture of the closed-loop control module comprises:

[0034] Fast cycle layer (execution period 1 ms): control the coolant flow through a PID algorithm, and modulate the polishing head current through PWM;

[0035] Slow cycle layer (execution period 1 s): input temperature, pressure and defect data sequence in the past 60 seconds, and output pressure distribution, rotation speed and cooling gradient optimization parameters.

[0036] As a further scheme of the present application: the process parameter mapping table construction rule comprises:

[0037] Material attribute dimension: thermal expansion coefficient , Young's modulus , hardness ;

[0038] Process parameter dimension: coolant gradient mode, deformation compensation coefficient, allowable temperature rise ;

[0039] Real-time according to the workpiece material type (such as fused quartz, calcium fluoride) to load the corresponding temperature control strategy and pressure threshold value.

[0040] As a further scheme of the present application: the transfer learning framework comprises:

[0041] A pre-trained ResNet-34 backbone network extracts multi-modal data features;

[0042] The adaptive layer is a 256-node full connection layer, and is fine-tuned based on small sample data (for example, 10 groups of free-form surface machining data), and the loss function is:

[0043] ;

[0044] In the formula, is the surface roughness, is the machining efficiency.

[0045] The application also provides a self-adaptive optical polishing method based on dynamic temperature control and multi-modal detection, which comprises the following steps:

[0046] Data acquisition and preprocessing: synchronously trigger the infrared thermal imager (100 Hz), the laser interferometer (10 Hz), the pressure sensor (1 kHz) and the acoustic emission sensor (500 kHz), align the time sequence through the dynamic time warping (DTW) algorithm, and perform wavelet threshold denoising on the acoustic emission signal, and the temperature data is subjected to Kalman filtering;

[0047] Thermal-mechanical coupling analysis and instruction generation: based on finite element simulation and LSTM network to predict thermal deformation trend, output polishing head deformation compensation current and cooling liquid gradient parameters ;

[0048] Real-time defect detection and parameter correction: when the surface roughness Ra is greater than 0.5 nm or the acoustic emission energy exceeds 3 times the standard deviation of the baseline, trigger the PPO reinforcement learning model to correct the pressure (±10%), the rotating speed (±20%) and the cooling strategy within 50 ms;

[0049] Cross-cycle iterative optimization: store historical process data, train the LSTM model to output the optimal parameter combination, and adapt to new workpiece types through transfer learning.

[0050] As a further scheme of the application, the data preprocessing comprises:

[0051] The temperature data is normalized to the interval of 0-1, and the pressure is normalized according to the maximum range;

[0052] The laser interferometer data is interpolated and aligned to the high-frequency temperature data stream.

[0053] Compared with the prior art, the technical scheme has the advantages that:

[0054] 1、The present application realizes real-time monitoring of the temperature field, surface topography and micro-crack signals of the polishing area through dynamic temperature control and multi-modal detection technology, combined with the rapid compensation of thermal deformation of the flexible actuator, effectively avoids the problem of surface shape precision decline caused by temperature rise in traditional process, and the multi-modal sensor works cooperatively to comprehensively capture multi-dimensional defects in the polishing process, significantly improves the defect detection sensitivity and process stability, and ensures the machining quality of high-precision optical elements,

[0055] 2、The system adopts a hierarchical closed-loop intelligent control architecture, realizes real-time regulation and control of temperature and pressure parameters through a fast circulation layer, and globally optimizes process strategies based on a reinforcement learning model in a slow circulation layer, realizes self-adaptive adjustment and dynamic balance of process parameters, compared with the traditional fixed rule control model, the architecture can autonomously learn and optimize complex nonlinear process relationship, reduce the dependence on artificial experience, improve the polishing efficiency and the adaptability to different materials and curved surface types,

[0056] 3、The cross-material process adaptation capability based on the transfer learning framework enables the system to quickly transfer historical process data to new workpiece types through a pre-trained model, and realizes automatic loading of cross-material temperature control strategies and pressure thresholds combined with a process parameter mapping table, which breaks through the long parameter adjustment period and weak generalization capability of traditional process for new materials, significantly shortens the process development time, and at the same time, takes into account the machining precision and efficiency, meets the efficient production needs of multi-variety and small-batch optical elements. BRIEF DESCRIPTION OF DRAWINGS

[0057] Figure 1 is a self-adaptive optical polishing system architecture flow chart;

[0058] Figure 2 is an optical polishing method step flow chart. DETAILED DESCRIPTION

[0059] The specific embodiments of the present application will be further described below in conjunction with the drawings, and it should be noted that the description of these embodiments is used to help understand the present application, but does not constitute a limitation on the present application.

[0060] In addition, the technical features involved in each of the embodiments of the present application described below can be combined with each other as long as they do not conflict with each other.

[0061] Please refer to the accompanying Figure 1 , the present application is based on the adaptive optical polishing system of dynamic temperature control and multi-modal detection, the optical polishing system comprises:

[0062] The dynamic temperature control and multi-modal detection module integrates an infrared thermal imager, a laser interferometer and an acoustic emission sensor, and is used for real-time acquisition of the temperature field, surface topography and micro-crack signals of the polishing area;

[0063] The flexible actuator comprises a shape memory alloy polishing head and a multi-channel microfluidic cooling system, wherein:

[0064] The polishing head is locally deformed by current driving, and the deformation amount satisfies the relationship:

[0065] ;

[0066] In the formula, is the deformation coefficient, is the driving current.

[0067] The cooling system controls the flow and temperature of the cooling liquid by gradient, and the flow distribution satisfies the exponential decay formula:

[0068] ;

[0069] In the formula, is the radius from the center of the high-temperature area.

[0070] The closed-loop control and optimization module adopts a hierarchical control architecture, including a fast cycle layer (FPGA) for real-time regulation of temperature and pressure parameters, and a slow cycle layer (GPU accelerated) for running an LSTM+PPO reinforcement learning model for global process optimization.

[0071] The cross-scale collaboration and generalization module is based on a process parameter mapping table and a transfer learning framework, and is suitable for different materials and curved surface types.

[0072] In an embodiment of the present application, the multi-modal sensor satisfies the following conditions:

[0073] The infrared thermal imager has a spatial resolution of ≤0.1℃, a sampling frequency of 100Hz, and a positioning temperature difference of >5℃C, and a local hotspot diameter of >2mm;

[0074] The laser interferometer has a wavelength of 632.8nm, generates a three-dimensional height map with a resolution of 10nm every 5 seconds, and identifies the surface error PV value and roughness Ra;

[0075] The acoustic emission sensor has a frequency band of 50-500KHz, and extracts a 200kHz frequency band energy burst signal through short-time Fourier transform (STFT).

[0076] In an embodiment of the present application, the temperature-pressure coupling model includes:

[0077] The physical driving layer is based on a finite element simulation to establish a thermal expansion equation:

[0078] ;

[0079] In the formula, is the thermal expansion coefficient, is the temperature rise, characteristic dimension;

[0080] Data-driven layer: input real-time temperature and pressure sequence into LSTM network to predict thermal deformation trend in next 3 seconds;

[0081] Instruction generation layer: fuse physical and data prediction results to output high-temperature zone coolant flow instruction (120ml / min to 200ml / min) and temperature set value .

[0082] In an embodiment of the present application: the shape memory alloy polishing head is a nickel-titanium alloy (NiTiNOL) grid structure, and the grid nodes are embedded with micro-resistance heating sheets. The workpiece surface curvature is mapped to the node target displacement amount through inverse kinematics algorithm.

[0083] In an embodiment of the present application: the multi-channel microfluidic cooling system includes 16 channels of independently controlled piezoelectric micropumps and semiconductor refrigerating sheets, wherein:

[0084] The high-temperature zone coolant flow is 200ml / min, and the temperature is set to 5℃;

[0085] The transition zone flow is distributed according to the exponential decay formula, and the temperature following strategy is .

[0086] In an embodiment of the present application: the hierarchical architecture of the closed-loop control module includes:

[0087] Fast cycle layer (execution period 1ms): control the coolant flow through PID algorithm, and modulate the polishing head current through PWM;

[0088] Slow cycle layer (execution period 1s): input temperature, pressure and defect data sequence of the past 60 seconds, and output pressure distribution, rotation speed and cooling gradient optimization parameters.

[0089] In an embodiment of the present application: the process parameter mapping table construction rules include:

[0090] Material attribute dimension: thermal expansion coefficient , Young's modulus , hardness ;

[0091] Process parameter dimension: coolant gradient mode, deformation compensation coefficient, allowable temperature rise ;

[0092] Real-time according to the workpiece material type (such as fused quartz, calcium fluoride) to load the corresponding temperature control strategy and pressure threshold.

[0093] In an embodiment of the present application: the transfer learning framework includes:

[0094] Pre-training ResNet-34 backbone network extracts multi-modal data features;

[0095] The adaptive layer is a 256-node full connection layer, which is fine-tuned based on small sample data (such as 10 groups of free-form surface machining data), and the loss function is:

[0096] ;

[0097] In the formula, is the surface roughness, is the machining efficiency.

[0098] Please refer to the attached Figure 2 The application also provides a self-adaptive optical polishing method based on dynamic temperature control and multi-modal detection, which comprises the following steps:

[0099] Data acquisition and preprocessing: synchronously trigger infrared thermal imager (100Hz), laser interferometer (10Hz), pressure sensor (1kHz) and acoustic emission sensor (500kHz), align the time sequence through dynamic time warping (DTW) algorithm, and perform wavelet threshold denoising on the acoustic emission signal, and Kalman filtering is adopted for temperature data;

[0100] Thermal-mechanical coupling analysis and instruction generation: based on finite element simulation and LSTM network to predict thermal deformation trend, output polishing head deformation compensation current and cooling liquid gradient parameters ;

[0101] Real-time defect detection and parameter correction: when the surface roughness Ra is greater than 0.5nm or the acoustic emission energy exceeds 3 times the standard deviation of the baseline, trigger the PPO reinforcement learning model to correct the pressure (±10%), rotation speed (±20%) and cooling strategy within 50ms;

[0102] Cross-cycle iterative optimization: store historical process data, train LSTM model to output optimal parameter combination, and adapt to new workpiece types through transfer learning.

[0103] In an embodiment of the application: data preprocessing includes:

[0104] Temperature data is normalized to the range of 0-1, and pressure is standardized according to the maximum range;

[0105] Laser interferometer data is interpolated and aligned to high-frequency temperature data stream.

[0106] Example one,

[0107] Technical Background: In the high-precision polishing of fused silica aspherical lenses, the traditional constant pressure polishing process is difficult to solve the problem of material thermal deformation caused by local temperature rise. This invention achieves precise control of the polishing process through dynamic temperature control and multimodal detection technology.

[0108] Implementation steps:

[0109] 1. Multimodal data acquisition and synchronization:

[0110] The infrared thermal imager monitors the temperature field of the polished area in real time at a sampling frequency of 100Hz, achieving a spatial resolution of 0.08℃. It detected a local hot spot area with a diameter of 3.2mm, where the temperature at the center is higher than the surrounding area. .

[0111] The laser interferometer performs a full surface scan every 5 seconds, generating a three-dimensional height map (10nm resolution), and measures the surface shape error PV value. / 25( =632.8nm), surface roughness Ra=0.32nm.

[0112] The acoustic emission sensor captured high-frequency stress wave signals at a sampling rate of 500 kHz. Through short-time Fourier transform (STFT) analysis, it was found that the energy in the 200 kHz band suddenly increased to 3.2 times the baseline value at the 12th minute of processing, triggering a microcrack warning.

[0113] Data synchronization employs a dynamic time warping (DTW) algorithm to interpolate and align the low-frequency data (10Hz) from the laser interferometer to the temperature data stream (100Hz), eliminating timing discrepancies.

[0114] 2. Thermo-mechanical coupling modeling and instruction generation:

[0115] The physics-driven layer calculates the thermal expansion of fused silica based on finite element simulation:

[0116] ;

[0117] By combining LSTM network prediction of thermal deformation trend within the next 3 seconds, the deformation compensation amount of the polishing head is output. m corresponds to the driving current =3.5A.

[0118] Coolant gradient control command generation: In the high-temperature zone (radius r ≤ 5 mm), the coolant flow rate is increased to 200 ml / min, and the temperature is set to T_hotspot - 15℃ = 20℃. The flow rate in the transition zone is distributed according to the exponential decay formula, and the flow rate at radius r = 15 mm is reduced to... =200 =110ml / min.

[0119] 3. Closed-loop control and real-time optimization:

[0120] The fast cycle layer (FPGA) executes PID control with a 1ms cycle to control the temperature fluctuation in the high-temperature zone within 0.5℃;

[0121] When the acoustic emission energy exceeds 3 times the baseline standard deviation, the slow cycle layer (GPU) calls the PPO reinforcement learning model to reduce the polishing head speed from 1200 rpm to 1020 rpm (15% reduction) within 50ms, and preferentially allocates 80% cooling resources to the high-temperature zone.

[0122] Technical effects:

[0123] The post-processing fused quartz lens surface shape precision PV value is less than or equal to / 20( =632.8nm), and the surface roughness Ra is stabilized at 0.30-0.35nm, which is better than the traditional process Ra=0.5nm;

[0124] The micro-crack defect rate is reduced from 1.2% in the traditional process to 0.08%, the processing efficiency is improved by 22%, and the single piece processing time is shortened to 45 minutes.

[0125] Example two,

[0126] Technical background: For the low thermal conductivity of calcium fluoride free-form lens (curvature radius R=150mm, asphericity coefficient ), it is necessary to realize fast process adaptation across materials and complex surfaces.

[0127] Implementation steps:

[0128] 1. Cross-scale parameter mapping and transfer learning:

[0129] Call the process parameter mapping table to load the calcium fluoride material properties: thermal expansion coefficient , Young's modulus =80GPa, hardness =158.

[0130] According to the rule base, automatically match the cooling liquid flow to increase by 20%, the high-temperature zone flow is adjusted from 200ml / min to 240ml / min, and the maximum allowed temperature rise is set to 5℃.

[0131] The transfer learning framework loads the pre-trained fused quartz processing model (ResNet-34 backbone network), adds a 256-node fully connected layer, inputs 10 sets of calcium fluoride free-form surface processing data (including temperature field, pressure distribution and defect record), freezes the backbone network weights, and only fine-tunes the adaptive layer loss function to:

[0132] ​ ;

[0133] After 2 hours of training, the model output the optimal pressure distribution as 12% overpressure in the edge region and 8% underpressure in the center region.

[0134] 2. Dynamic deformation compensation and surface adaptation:

[0135] Inverse kinematics algorithm maps the curvature to the target displacement of the polishing head grid nodes, calculating the driving current of the corresponding nodes I = 2.8 A, triggering the nickel-titanium alloy austenitic phase transition, and realizing local deformation compensation.

[0136] The multi-channel microfluidic system distributes nozzles in concentric circles, with the cooling liquid temperature in the high-temperature zone (curvature mutation area) set to 5°C and the flow rate at 240 ml / min; the flow rate in the transition zone is distributed according to , and the flow rate at a radius of r = 40 mm is reduced to .

[0137] 3. Defect suppression and process verification:

[0138] The acoustic emission sensor detects energy fluctuations in the 250 kHz frequency band, determining sub-surface damage, triggering the PPO model to dynamically adjust the polishing head inclination angle by 2°, reducing local stress concentration;

[0139] The laser interferometer feedback surface roughness Ra = 0.45 nm, close to the fused quartz processing level, verifying the effectiveness of transfer learning.

[0140] Technical effects:

[0141] Freeform surface adaptation error ≤ 1.8 μm, surface shape consistency deviation < 0.3 μm. ;

[0142] The training period of the transfer learning model is shortened from the traditional 72 hours to 2 hours, the generalization ability is improved by 35%, and the compatible material types are expanded to more than 5.

[0143] Example Three,

[0144] Technical background: Applied to continuous processing of glass substrates with a curvature radius R = 50 mm, it needs to solve the problem of surface consistency degradation caused by dynamic temperature rise and stress mutation.

[0145] Implementation steps:

[0146] 1. Hierarchical closed-loop control and real-time response:

[0147] The fast cycle layer (FPGA) executes temperature control instructions at a period of 1 ms: adjusts 16-channel piezoelectric micropumps through PID algorithm, high-temperature zone cooling wave flow response delay ≤ 0.2 ms, temperature fluctuation control within +0.3°C.

[0148] Slow cycle layer (GPU) runs LSTM+PPO model every 1 second, inputs temperature, pressure and defect sequence of the past 60 seconds (total of 60000 data points), and outputs global optimization parameters: polishing head pressure distribution adjustment +8%, speed from 1500rpm to 1350rpm (decrease by 10%).

[0149] 2. Multimodal defect detection and suppression:

[0150] Acoustic emission sensor captures 300kHz high-frequency stress wave signals, with energy peak up to 4 times the baseline, and the system determines that the subsurface damage is expanding, triggering the following operations immediately:

[0151] Increase the flow of high-temperature zone cooling liquid to 200ml / min, and the temperature to 5℃℃;

[0152] The deformation compensation amount of the polishing head is increased to , compensating for thermal deformation error.

[0153] Laser interferometer monitors surface roughness Ra value in real time. When Ra increases from 0.4nm to 0.62nm, the reinforcement learning model generates parameter correction instructions within 50ms, reducing the polishing pressure from 8N to 7.2N (decrease by 10%).

[0154] 3. Gradient cooling and process stability guarantee:

[0155] Semiconductor refrigeration piece sets the temperature of the transition zone cooling liquid to 10℃℃, and the flow of the high-temperature zone (radius r≤10mm) is distributed according to , and the flow at radius r=30mm is ;

[0156] After continuous processing for 8 hours, the system automatically starts cross-cycle optimization, calls the historical process database (stores 120 parameter combinations), trains the LSTM model to predict the thermal deformation trend in the next 15 minutes, and adjusts the cooling strategy in advance.

[0157] Technical effects:

[0158] The surface consistency deviation during continuous processing is ≤5nm, with no batch overruns;

[0159] Dynamic response time ≤50ms, process stability improved by 42% compared to traditional methods, single-day production capacity increased to 300 pieces, and yield stabilized above 99.3%.

[0160] Although the present application is disclosed with reference to the preferred embodiments, it is not intended to limit the present application and any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the present application. Therefore, any modification, equivalent change and modification made to the above embodiments according to the technical essence of the present application without departing from the technical solutions of the present application, all fall within the protection scope defined by the claims of the present application.

Claims

1. An adaptive optical polishing system based on dynamic temperature control and multi-modal detection, characterized in that: The optical polishing system comprises: A dynamic temperature control and multi-modal detection module integrates an infrared thermal imager, a laser interferometer, and an acoustic emission sensor to collect temperature field, surface topography, and micro-crack signals of the polishing area in real time; A flexible actuator includes a shape memory alloy polishing head and a multi-channel micro-fluidic cooling system, wherein: The polishing head is driven by current to achieve local deformation compensation, and the deformation amount satisfies the relationship: ； wherein is a deformation coefficient, is a drive current; The cooling system controls the flow and temperature of the cooling liquid by gradient, and the flow distribution satisfies the exponential decay formula: ； wherein r is the radius from the center of the high temperature zone; A closed-loop control and optimization module adopts a hierarchical control architecture, including a fast cycle layer to adjust temperature and pressure parameters in real time, and a slow cycle layer to run an LSTM+PPO reinforcement learning model for global process optimization; A cross-scale collaboration and generalization module is based on a process parameter mapping table and a transfer learning framework to adapt to different materials and curved surface types. 2.The system of claim 1, wherein: The multi-modal sensor meets the following conditions: The infrared thermal imager has a spatial resolution of ≤0.1℃, a sampling frequency of 100Hz, and a positioning temperature difference of >5℃C with a local hotspot diameter of >2mm; The laser interferometer has a wavelength of 632.8nm, generates a three-dimensional height map with a resolution of 10nm every 5 seconds, and identifies the surface error PV value and roughness Ra; The acoustic emission sensor has a frequency band of 50-500KHz, and extracts the energy burst signal in the 200kHz frequency band through short-time Fourier transform. 3.The system of claim 1, wherein the system further comprises: a temperature control system configured to control the temperature of the polishing pad; and a multi-modal detection system configured to detect the polishing pad and the workpiece. The temperature-pressure coupling model includes: A physical driving layer: a thermal expansion equation is established based on finite element simulation: ； wherein is the coefficient of thermal expansion, is the temperature rise, is the characteristic dimension; A data-driven layer: an LSTM network is used to input real-time temperature and pressure sequences to predict the thermal deformation trend in the next 3 seconds; Instruction generation layer: fuse physical and data prediction results, output high temperature zone coolant flow instruction and temperature set value . 4.The system of claim 1, wherein the system further comprises: a temperature control system configured to control the temperature of the substrate. The shape memory alloy polishing head is a nickel-titanium alloy grid structure, and the grid nodes are embedded with micro-resistance heating sheets. The workpiece surface curvature is mapped to the node target displacement amount through inverse kinematics algorithm.

5. The adaptive optical polishing system based on dynamic temperature control and multi-modal detection of claim 1, wherein: The multi-channel micro-fluidic cooling system includes a 16-channel independently controlled piezoelectric micro-pump and a semiconductor refrigeration sheet, wherein: The cooling liquid flow in the high temperature zone is 200ml / min, and the temperature is set to 5℃; The flow in the transition zone is distributed according to an exponential decay formula, and the temperature following strategy is . 6.The system of claim 1, wherein: The hierarchical architecture of the closed-loop control module includes: Fast cycle layer: control the cooling liquid flow through PID algorithm, and modulate the polishing head current through PWM; Slow cycle layer: input temperature, pressure, and defect data sequences in the past 60 seconds, and output pressure distribution, speed, and cooling gradient optimization parameters. 7.The adaptive optical polishing system based on dynamic temperature control and multi-modal detection of claim 1, wherein: The process parameter mapping table construction rules include: Material attribute dimension: coefficient of thermal expansion , Young's modulus , hardness ; Process parameter dimensions: coolant gradient mode, deformation compensation coefficient, allowable temperature rise ; Real-time load corresponding temperature control strategy and pressure threshold according to workpiece material type. 8.The system of claim 1, wherein the system further comprises: a temperature controller configured to control the temperature of the substrate. The transfer learning framework includes: A pre-trained ResNet-34 backbone network extracts multi-modal data features; The adaptation layer is a 256-node fully connected layer, which is fine-tuned based on small sample data, and the loss function is: ； In the formula, Ra is the surface roughness, is the machining efficiency.

9. The method of any one of claims 1-8, wherein the method is a dynamic temperature control and multi-modal detection based adaptive optical polishing method. The optical polishing method comprises the following steps: Data acquisition and preprocessing: synchronize the infrared thermal imager, laser interferometer, pressure sensor (1kHz), and acoustic emission sensor, align the time sequence through dynamic time warping algorithm, and denoise the acoustic emission signal through wavelet threshold, and Kalman filter is used for temperature data; Thermal-mechanical coupling analysis and instruction generation: based on finite element simulation and LSTM network to predict thermal deformation trend, output polishing head deformation compensation current and coolant gradient parameters ; Real-time defect detection and parameter correction: when surface roughness Ra>0.5nm or acoustic emission energy exceeds 3 times the baseline standard deviation, trigger PPO reinforcement learning model to correct pressure, speed and cooling strategy within 50ms; Cross-cycle iterative optimization: store historical process data, train LSTM model to output optimal parameter combination, and adapt to new workpiece types through transfer learning.

10. The method of claim 9, wherein the method further comprises: The data preprocessing includes: Temperature data is normalized to the 0-1 interval, and pressure is normalized according to the maximum range; Laser interferometer data is interpolated and aligned to the high-frequency temperature data stream.

Images