A large-diameter substrate spin coating process adaptive optimization and collaborative decision system

Abstract

The application discloses a kind of large aperture substrate spin coating process self-adaptive optimization and collaborative decision system, including spin coater, multimode sensor array, online state-parameter estimation module, predictive controller, actuator, event-triggered adaptive optimization module, fast safety constraint optimization module, edge AI control device, double-layer intelligent decision module and natural language conversion module.The application can accurately capture and predict film thickness distribution dynamics through global multimodal real-time monitoring and high-precision state estimation based on digital twinning, and then carry out feedforward-feedback composite regulation through model predictive control.The application is particularly aimed at the stubborn edge thick peak problem in spin coating process, and through the coordinated action of zoned temperature control and zoned airflow on the edge of the substrate, the solvent evaporation rate and photoresist leveling behavior in this area are adjusted, achieving a significant effect of reducing the edge thick peak by ≥60%.The application improves computational efficiency, is safe and reliable, and reduces energy consumption.

Description

Technical Field

[0001] This invention belongs to the field of semiconductor manufacturing technology, specifically relating to an adaptive optimization and collaborative decision-making system for large-diameter substrate spin coating process based on multimodal monitoring and digital twins. It is applicable to real-time monitoring, adaptive control and collaborative optimization of photoresist spin coating process on substrates with diameters of 300-1200 mm. Background Technology

[0002] Spin coating, a crucial step in photoresist coating during semiconductor manufacturing, directly determines the precision and reliability of subsequent processes such as development and etching. With the continuous miniaturization of process structures and the increasing complexity of device structures, stringent requirements have been placed on the uniformity of photoresist film thickness, interface flatness, and defect control. However, as substrate sizes expand (diameters from 300mm to 1200mm), traditional spin coating processes face unprecedented technical challenges. Spin coating on large-diameter substrates exhibits significant spatial non-uniformity and temporal dynamics. During spin coating, the differences in centrifugal force distribution between the substrate edge and center, solvent evaporation rate gradients, and uneven heat conduction all contribute to difficulties in ensuring film thickness uniformity, particularly leading to noticeable "edge thickness peaks" at the substrate edges. Furthermore, as substrate size increases, the temperature, airflow, and solvent concentration fields within the process chamber become more complex, causing the photoresist leveling process, solvent evaporation kinetics, and curing behavior to exhibit strong non-linear characteristics, further exacerbating the non-uniformity of film thickness distribution and resulting in a significantly increased defect rate.

[0003] Current spin coating control strategies widely used in industry are mostly based on fixed process parameters or simple single-variable feedback control mechanisms, lacking real-time multimodal monitoring and adaptive adjustment capabilities. These strategies exhibit significant technical limitations when coating large-diameter substrates, leading to insufficient process stability. For example, spin coating control methods using single-point film thickness sensors cannot comprehensively monitor film thickness distribution and changes in physical properties; using temperature control to improve uniformity also fails to integrate multi-sensor data and advanced estimation algorithms. Furthermore, existing control methods generally lack a deep understanding of the coupling effects between process parameters and effective decoupling methods. Complex nonlinear relationships exist between key process parameters such as rotation speed, temperature, pressure, and airflow. Simple single-variable control easily leads to mutual interference between parameters, even causing process instability. This parameter coupling effect is particularly pronounced in the processing of large-diameter substrates, and current technologies have not yet established effective multi-parameter collaborative control mechanisms.

[0004] Furthermore, traditional real-time optimization control methods (such as model predictive control) typically involve continuous data acquisition and calculation at fixed intervals. This results in a significant data communication burden and computational delay in high-speed, high-cycle production lines, causing control system "jitter," and a large amount of computational resources are consumed on minor, insignificant changes in process conditions. Simultaneously, the spin coating process involves volatile organic compound emissions and potential pollution risks, requiring strict adherence to safety constraints during optimization. However, existing safety optimization methods often involve excessively large amounts of computational data and slow convergence speeds, failing to meet the second- or even millisecond-level response requirements of production lines. Moreover, current technologies often employ single-layer optimization methods, such as optimizing only process parameters or equipment structure, lacking collaborative decision-making mechanisms, leading to a narrow manufacturable window and high trial-and-error costs. Changes in nozzle diameter (e.g., ±200 mm) necessitate recalibrating the entire system, resulting in low efficiency. Current systems also lack intelligent interfaces, failing to translate optimization results into natural language recommendations, which is detrimental to production line personnel's operation and knowledge accumulation.

[0005] Therefore, there is an urgent need for an integrated solution that can estimate film thickness in real time, adaptively optimize process parameters, reduce computational load, ensure safety constraints, and achieve coordinated optimization of equipment and process. Summary of the Invention

[0006] To address the aforementioned problems in existing technologies, this invention designs an adaptive optimization and collaborative decision-making system for spin coating processes on large-diameter substrates. This system solves the technical problems of poor uniformity of spin coating thickness and significant edge thickness peaks on large-diameter substrates, thereby achieving efficient, safe, and adaptive control of the spin coating process.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows: an adaptive optimization and collaborative decision-making system for spin coating process of large-diameter substrates, comprising a spin coater, a multimodal sensor array, an online state-parameter estimation module, a predictive controller, an actuator, an event-triggered adaptive optimization module, a fast safety constraint optimization module, an edge AI control device, a dual-layer intelligent decision-making module, and a natural language conversion module;

[0008] The multimodal sensing array includes a multi-wavelength laser interferometric thickness measurement channel, a miniature infrared thermal imager, a solvent partial pressure sensor, a vibration accelerometer, and a torque sensor, which are used to monitor the film thickness, temperature, solvent partial pressure, vibration parameters, and spindle torque in real time during the spin coating process.

[0009] The online state-parameter estimation module estimates key parameters of film thickness state online based on a multiphysics model and a nonlinear filtering algorithm.

[0010] The predictive controller uses a model predictive control algorithm, i.e., MPC algorithm, to adjust the output process parameters based on the output process parameter control instructions from the online state-parameter estimation module.

[0011] The actuator is used to adjust the rotation speed, temperature, partial pressure and airflow parameters of the spin coater;

[0012] The event-triggered adaptive optimization module sets an event trigger threshold and performs parameter optimization and distribution only when the threshold is triggered.

[0013] The fast safety constraint optimization module adopts a safe Bayesian optimization framework to optimize parameters under safety constraints.

[0014] The edge AI control device enables the local execution of a reduced-order digital twin model and a fast solver by configuring an edge AI controller on the device side.

[0015] The dual-layer intelligent decision-making module includes an upper-layer decision-making module, a lower-layer decision-making module, and a coupling interface module; the upper-layer decision-making module is used to optimize equipment configuration, the lower-layer decision-making module is used to optimize process trajectory, and the coupling interface module realizes collaboration between the upper and lower layers.

[0016] The natural language conversion module transforms the optimized scheme into a natural language process control recommendation scheme through an intelligent translation center, and constructs a knowledge graph.

[0017] Furthermore, the spin coater includes a lower dust cover, an upper dust cover, an edge AI controller, a servo motor, a control center, a spin coating atmosphere cover, a spin coating tray, a transmission mechanism, a frame, liquid distribution holes, a ring support, a multi-wavelength laser interferometric thickness measurement channel, a miniature infrared thermal imager, a solvent partial pressure sensor, a solvent vapor injection and exhaust port, a spin coater spindle, a vibration accelerometer, rolling bearings, and a torque sensor.

[0018] The transmission mechanism transmits the torque of the servo motor to the glue-spreading tray.

[0019] The servo motor and transmission mechanism are mounted on the frame;

[0020] The lower dust cover is located at the top of the frame and below the glue coating tray;

[0021] The upper dust cover is located above the lower dust cover and the top coat tray. The upper dust cover is rotatably connected to one side of the lower dust cover by a pin. The upper dust cover is opened when the substrate is placed and the spin coating atmosphere is set up, and is fastened when the spin coating is in place and when the substrate is stationary.

[0022] The edge AI controller is mounted on the side of the upper dust cover.

[0023] The spin coating atmosphere hood is installed on the upper part of the lower dust cover via a slide groove. The top of the spin coating atmosphere hood is located below the upper dust cover, covering the spin coating tray and the substrate. When placing the substrate, the spin coating atmosphere hood is removed via the slide groove, the substrate is placed on the spin coating tray, and then the spin coating atmosphere hood is fixed by the slide groove.

[0024] The servo motor is connected to the control center via a data cable;

[0025] The liquid distribution hole is located at the center of the upper end of the spin coating atmosphere cover and is used to drop the solution onto the substrate;

[0026] The annular support is located on the inner wall of the top cover of the spin coating atmosphere cover, with a radius of 100~300 mm, and is embedded in the center of the substrate rotation.

[0027] The multi-wavelength laser interferometric thickness measurement channels are evenly distributed along the circumference of the annular bracket and are embedded in the mounting holes on the annular bracket.

[0028] The miniature infrared thermal imager is embedded in the top cover of the spin coating atmosphere hood at a radius of 2 / 3 from the center.

[0029] One solvent partial pressure sensor is installed at the center of the top of the spin coating atmosphere hood and one at the side wall near the exhaust port.

[0030] The edge AI controller is connected to the multi-wavelength laser interferometric thickness measurement channel, miniature infrared thermal imager, and solvent partial pressure sensor on the spin coating atmosphere hood via data cables, and is also connected to an external control center via data cables.

[0031] The solvent vapor injection and exhaust ports are 8 to 12 cylindrical channels evenly distributed circumferentially along the side wall of the spin coating atmosphere hood.

[0032] The vibration accelerometer is located between the spin coater tray and the rolling bearing, and is arranged around the spin coater spinner shaft inside the frame.

[0033] The torque sensor is located below the rolling bearing and is arranged around the spindle of the spin coater inside the frame;

[0034] The vibration accelerometer and torque sensor are integrated in the spindle drive unit of the spindle coater, and together they form the spindle dynamics sensor.

[0035] The multi-wavelength laser interferometric thickness measurement channel has 8 to 12 measurement channels, the accuracy of the miniature infrared thermal imager is ±0.2℃, and the measurement range of the solvent partial pressure sensor is 0-80% saturated vapor pressure.

[0036] Furthermore, the online state-parameter estimation module uses the partial differential equations coupled with thin film lubrication, evaporation, and the Marangoni effect as a priori models, and employs unscented Kalman filtering or particle filtering for state estimation. The state update time is ≤10 ms. The output parameters of the state estimation include film thickness h(r,θ,t), chamber temperature T(r,θ,t), mass fraction c(r,θ,t), and interfacial tension γ(r,θ,t).

[0037] Furthermore, the predictive controller is a constrained model predictive control (MPC), with constraints including a temperature gradient |dT / dr| ≤ 0.02℃ / mm and a partial pressure step ≤ 10% / 0.1 s; and it uses Lagrange multiplication for safe reinforcement learning (Safe-RL), sets parameter control thresholds, and optimizes the process parameter control path using particle swarm optimization.

[0038] Furthermore, the event triggering thresholds of the event-triggered adaptive optimization module include the threshold for film thickness error |Δh|, the threshold for ripple growth rate d(Astripe) / dt, and the threshold for evaporation flux mutation |ΔJev|, and the event triggering thresholds are dynamically adjusted based on historical data.

[0039] The threshold for the film thickness error |Δh| is set to trigger when |Δh|>0.5%·h, where h is the target film thickness;

[0040] The threshold value of the ripple growth rate d(Astripe) / dt is preset according to the substrate diameter and process requirements, and is triggered when d(Astripe) / dt > the threshold value;

[0041] The threshold for the sudden change in evaporation flux, |ΔJev|, is set to be triggered when |ΔJev|>10%.

[0042] Furthermore, the fast safety constraint optimization module uses a Gaussian process of multi-objective safety constraints as a surrogate model, and converges to the optimal parameter configuration within ≤30 experimental points by fusing simulation data with multi-fidelity kernels and small sample experimental data, and outputs a confidence interval; the safety constraints include VOC emission upper limit, pollution threshold and stripe contrast upper limit.

[0043] Furthermore, the edge AI control device adopts FPGA or Edge-GPU hardware, with a single optimization calculation latency of ≤20 ms. The optimization variables include the linkage speed curve ω(t), zone temperature control Ti(t), zone pressure pi(t), and zone airflow ui(t). It also supports collaboration with the host computer model predictive control or digital twin, and automatically reverts to a safe configuration in case of anomalies.

[0044] Furthermore, the upper-level decision module optimizes the configuration of the spin coater, which includes the number of partitions, nozzle layout, pressure-dividing annular cavity topology, and edge air gap geometry. The achievable manufacturing window is maximized through topology optimization and a multi-objective particle swarm genetic algorithm.

[0045] The lower-level decision module optimizes the trajectory of the spin coating process, which includes the rotation speed curve ω(t), zone control temperature Ti(t), zone pressure pi(t), and zone airflow ui(t). The thickness gradient and defect risk are minimized through model predictive control and a safe Bayesian optimization framework.

[0046] Furthermore, the coupling interface module uses a unified cost function and sensitivity interface to couple the upper-level decision-making module with the lower-level decision-making module, automatically generates A / B verification schemes and performs closed-loop iterations, and outputs a joint optimization scheme for equipment configuration and spin coating process trajectory; the unified cost function simultaneously considers film uniformity, energy consumption and edge thickness peak index.

[0047] Furthermore, the natural language conversion module interfaces with the Manufacturing Execution System (MES) and the Semiconductor Equipment Communication Standard (SECS-GEM) to support natural language queries and audit trails; the knowledge graph stores the "defect-cause-solution-effect" relationship.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] 1. High-precision film thickness control: This invention, through real-time monitoring of multimodal phenomena across the entire domain and high-precision state estimation based on digital twins, can accurately capture and predict the dynamics of film thickness distribution. This is then addressed through model predictive control (MPC) for feedforward-feedback composite regulation. This approach stably controls the film thickness uniformity within the effective aperture at a high level of ≤ ±1.2% (3σ), significantly improving film quality and fundamentally ensuring linewidth consistency in subsequent photolithography processes.

[0050] 2. Significant reduction in edge thickness peaks: Since this invention is specifically designed to address the persistent problem of edge thickness peaks in spin coating processes, it achieves a significant reduction in edge thickness peaks of ≥60% by using zoned temperature control and zoned airflow to work synergistically at the substrate edge and adjusting the solvent evaporation rate in this area to be parallel to the photoresist flow.

[0051] 3. Improved computational efficiency: This invention utilizes an event-triggered mechanism to perform optimizations only at critical moments, reducing parameter update frequency by ≥50% and decreasing network load and computational latency. The edge AI controller runs the optimization algorithm locally, with a single computation latency of ≤20 ms, adapting to high-cycle production lines and improving system response speed.

[0052] 4. Safety and Reliability: This invention introduces constrained model predictive control and combines it with safe Bayesian optimization to set safety thresholds (such as temperature gradient, partial pressure step, and VOC emission limit) for the dynamic adjustment of all process parameters. This ensures that the optimization process is always carried out within the safety envelope of the equipment and process, effectively avoiding process failure or equipment damage caused by parameter out-of-bounds, and ensuring the continuity and reliability of large-scale production.

[0053] 5. Collaborative Optimization and Rapid Reconfiguration: This invention utilizes a dual-layer intelligent decision-making system for equipment-process collaborative optimization to achieve thickness uniformity ≤±1.2% and reduce defect rate by over 20% on large-diameter substrates. When switching diameters by ±200 mm, only the lower-level process trajectory needs to be re-optimized, rapidly responding to production demands, improving efficiency by 25%, and reducing trial-and-error costs by 35%.

[0054] 6. Intelligent Operation and Knowledge Accumulation: The natural language conversion module of this invention transforms optimized solutions into intuitive natural language recommendations, facilitating operation by production line personnel; the knowledge graph supports defect diagnosis and solution recommendations, improving audit tracking capabilities and knowledge reusability.

[0055] 7. Energy consumption optimization: This invention optimizes energy consumption through a unified cost function, reducing overall energy consumption by 10% and improving production economy. Attached Figure Description

[0056] Figure 1 This is a logic diagram of the present invention.

[0057] Figure 2 This is a schematic diagram of the entire spin coating machine of the present invention;

[0058] Figure 3 for Figure 1 Schematic diagram of the structure of the vortex coating atmosphere cover and the location of the multimodal sensor array;

[0059] Figure 4 This is a schematic diagram showing the positions of each sensor in the multimodal sensor array for spindle dynamics, viewed in half-section.

[0060] In the diagram: 1-Lower dust cover; 2-Upper dust cover; 3-Edge AI controller; 4-Servo motor; 5-Control center; 6-Spin coating atmosphere hood; 7-Spreading tray; 8-Transmission mechanism; 9-Frame; 10-Dispensing hole; 11-Ring support; 12-Multi-wavelength laser interferometry thickness measurement channel; 13-Miniature infrared thermal imager; 14-Solvent partial pressure sensor; 15-Solvent vapor injection and exhaust port; 16-Spinning machine spindle; 17-Vibration accelerometer; 18-Rolling bearing; 19-Torque sensor. Detailed Implementation

[0061] The present invention will be described in detail below with reference to the embodiments. In a preferred embodiment of the present invention, the system hardware platform is based on a large spin coater suitable for 300mm silicon wafers and is upgraded.

[0062] like Figure 2-4As shown, the spin coater includes a lower dust cover, an upper dust cover 2, an edge AI controller 3, a servo motor 4, a control center 5, a spin coating atmosphere cover 6, a spin coating tray 7, a transmission mechanism 8, a frame 9, a liquid distribution hole 10, a ring support 11, a multi-wavelength laser interferometric thickness measurement channel 12, a miniature infrared thermal imager 13, a solvent partial pressure sensor 14, a solvent vapor injection and exhaust port 15, a spin coater spindle 16, a vibration accelerometer 17, a rolling bearing 18, and a torque sensor 19.

[0063] like Figure 1 As shown, the spin coater is driven by a large servo motor 4; the transmission mechanism 8 transmits the torque of the servo motor 4 to the spin coating tray 7; the servo motor 4 and the transmission mechanism 8 are mounted on the frame 9; the lower dust cover 1 is located above the frame 9 and below the spin coating tray 7; the upper dust cover 2 is located above the lower dust cover 1 and the spin coating tray 7; the spin coating atmosphere cover 6 is located between the spin coating tray 7 and the upper dust cover 2; and the edge AI controller 3 is mounted on the side of the upper dust cover 2.

[0064] Furthermore, the spin coating atmosphere cover 6 is installed above the lower dust cover 1 via a sliding groove, making it easy to remove for substrate placement and sensor installation;

[0065] Furthermore, the upper dust cover 2 and the lower dust cover 1 are connected by a pin or other means, so that the upper dust cover 2 can rotate along the pin to expose the glue spreading tray 7, which facilitates the placement of the substrate and the application of the liquid.

[0066] Furthermore, the servo motor 4 is connected to the control center 5 via a data cable;

[0067] Furthermore, the edge AI controller 3 adopts the AI ​​controller of Xilinx FPGA, runs a reduced-order CFD digital twin and SQP solver, and is connected to the multi-wavelength laser interferometric thickness measurement channel 12, miniature infrared thermal imager 13, solvent partial pressure sensor 14, vibration accelerometer 17, torque sensor 19 and control center 5 on the spin coating atmosphere hood 6 via data lines.

[0068] like Figure 3 As shown, the liquid distribution hole 10 is located at the center of the upper end of the spin coating atmosphere hood 6 and is used to drop the solution onto the substrate; the annular support 11 is located on the inner wall of the top cover of the spin coating atmosphere hood 6 with a radius of 150 mm and is embedded with the center of rotation of the substrate as the center; the eight independent multi-wavelength laser interferometric thickness measurement channels 12 are evenly distributed along the circumference of the annular support 11; the miniature infrared thermal imager 13 is embedded at 2 / 3 radius from the center of the top cover of the spin coating atmosphere hood 6; the solvent partial pressure sensor 14 is installed at the center of the top of the spin coating atmosphere hood 6 and one on the side wall near the exhaust port 15; the solvent vapor injection and exhaust port 15 are eight cylindrical channels evenly distributed along the circumference of the side wall of the spin coating atmosphere hood 6.

[0069] Furthermore, the measurement spot diameter of the multi-wavelength laser interferometric thickness measurement channel 12 is ≤1mm, and the data acquisition frequency is 1kHz, ensuring full coverage and real-time scanning of the substrate surface.

[0070] Furthermore, the miniature infrared thermal imager 13 has an accuracy of ±0.2℃, faces the four quadrant regions of the substrate surface respectively, and has a spatial resolution better than 0.5mm. It is used to monitor the two-dimensional temperature field T(r,θ,t) distribution on the substrate surface during the coating process, forming an infrared temperature measurement system.

[0071] Furthermore, the solvent partial pressure sensor 14 is based on a quartz microbalance (QCM) to monitor the spatial distribution and dynamic changes of solvent vapor concentration in the cavity in real time, with a measurement range covering 0% to 80% of the ambient saturated vapor pressure.

[0072] Furthermore, the solvent vapor injection and exhaust port 15 has a diameter of 20-100mm and is used to regulate the air pressure in the spin coating chamber. The solvent vapor injection and exhaust are controlled by a micro piezoelectric valve and a mass flow controller, which precisely regulates the partition solvent pressure pi(t) and the purified airflow ui(t).

[0073] like Figure 4 As shown, the vibration accelerometer 17 is located between the spin coater tray 7 and the rolling bearing 18, and is arranged around the spin coater main shaft 16 inside the frame 9; the torque sensor 19 is located below the rolling bearing 18 and is arranged around the spin coater main shaft 16 inside the frame 9.

[0074] Furthermore, the vibration accelerometer 17 and the torque sensor 19 are integrated in the spindle drive unit of the spindle spindle 16 of the transmission unit, together forming a spindle dynamics sensor.

[0075] Furthermore, the vibration accelerometer 17 has a frequency response range of 0-5kHz and is used to monitor the mechanical stability of the transmission mechanism 8.

[0076] Furthermore, the torque sensor 19 has high-speed response characteristics and is used to monitor load changes in the transmission mechanism 8.

[0077] like Figure 1 As shown, this invention utilizes the improved spin coater hardware platform described above to achieve real-time estimation of film thickness, adaptive optimization of process parameters, reduction of computational load, assurance of safety constraints, and integration of equipment and process co-optimization.

[0078] To achieve digital twin model construction and state estimation, this invention constructs a multiphysics coupled model. Based on thin-film lubrication theory, convection-diffusion equations, and thermodynamic equations, a set of coupled partial differential equations (PDEs) describing photoresist flow, solvent evaporation, and the Marangoni effect induced by temperature / concentration gradients is established as the core prior model for the digital twin. This model is solved spatially and temporally using the finite volume method on an industrial control computer. To balance computational efficiency and estimation accuracy, this embodiment prioritizes the unscented Kalman filter (UKF) algorithm. This algorithm uses the aforementioned PDEs as the state transition model and real-time readings from a multimodal sensor array as observations, running on an embedded industrial control computer equipped with an Intel i7 processor and a real-time Linux system. The algorithm execution cycle is strictly controlled within 10 ms, outputting real-time two-dimensional distribution estimates of the film thickness h(r,θ,t), temperature T(r,θ,t), solvent mass fraction c(r,θ,t), and interfacial tension γ(r,θ,t) across the entire field of view. This estimation result accurately reflects the real-time rheological and curing process of the photoresist.

[0079] An event-triggered and safety optimization algorithm is embedded within the system, with the target film thickness h* set to 2.0 μm. Accordingly, the film thickness error trigger threshold is set to |Δh|>0.01 μm (0.5% × 2.0 μm). The ripple growth rate threshold is set to 0.05 μm / s, and the evaporation flux mutation threshold is set to 10%. After the spin coating process begins, the edge AI controller 3 receives data from the multimodal sensor array in real time. Based on safety process conditions, such as the absolute value of the radial temperature gradient |dT / dr| not exceeding 0.02℃ / mm to prevent film cracking due to thermal stress, and the partial pressure adjustment rate not exceeding 10% saturated vapor pressure / 0.1 seconds to avoid condensation risks, the film thickness error trigger condition is met when the real-time film thickness deviates from the target value by 0.012 μm, initiating safety Bayesian optimization. The optimizer calls a locally stored multi-fidelity Gaussian process model trained with CFD simulation data and 50 sets of historical small-sample experimental data. Under safety constraints of VOC emissions <50 ppm, pollution <1 ppb, and stripe contrast <5%, ω(t), Ti(t), pi(t), and ui(t) were jointly optimized. After 22 iterations, the algorithm converged, finding a set of optimal parameters and predicting that it could achieve a film thickness uniformity of ≤±1.1% (3σ) with all safety indicators meeting the requirements. The edge controller completed the calculation within 15 ms and distributed the new set of process parameters to each actuator. If an anomaly was detected, it automatically reverted to the safe configuration.

[0080] During parameter optimization, the system employs a two-layer decision-making and collaborative optimization approach. The upper-layer decision-making module uses a multi-objective particle swarm optimization (MPC) genetic algorithm to optimize the number of zones (selected as 12 zones), nozzle layout (uniform distribution), and pressure chamber topology (annular), calculating the manufacturable window. The lower-layer decision-making module uses model predictive control (MPC) to dynamically adjust the rotational speed curve ω(t) (linearly increasing from 500 rpm to 2000 rpm), and combines this with Safe Bayesian optimization (Safe-BO) to adjust the zone temperature control Ti(t) (range 20–25℃), zone pressure pi(t) (0.1–0.5 MPa), and zone airflow ui(t) (5–10 L / min) to minimize the thickness gradient. The coupling interface module calculates the weighted sum of uniformity, energy consumption, and edge thickness peaks using a unified cost function, generating an A / B verification scheme for iterative optimization. The natural language processing module translates the optimization plan into natural language, such as: "It is recommended to increase the engine speed from 500 rpm to 2000 rpm within 10 seconds and adjust the temperature of partition 3 to 22℃." The recommended plan is then sent through the MES / SECS-GEM interface to update the knowledge graph.

[0081] Finally, the execution and closed-loop control process is carried out. The optimization instructions of the predictive controller are sent to each actuator through a high-speed fieldbus (such as EtherCAT). The spindle servo motor receives the ω(t) curve to achieve smooth and accurate speed tracking. The zone temperature control system adjusts the current of the Peltier element installed on the chamber wall to accurately control the temperature Ti(t) of each zone. The micro piezoelectric valve and the mass flow controller work together to control the injection and exhaust of solvent vapor, and accurately regulate the zone solvent partial pressure pi(t) and the purification airflow ui(t). The whole system forms a high-frequency closed loop, dynamically adjusting the process parameters according to the real-time estimated film thickness state to achieve adaptive optimization of "sensing-decision-execution".

[0082] Performance Verification: In this embodiment, the film thickness uniformity (3σ) across the entire substrate (excluding the 3mm edge exclusion area) reached ±0.8%, significantly better than the ±3% level achieved by traditional control methods. Through precise temperature control and airflow management in the edge region, the height and width of the edge thickness peaks were effectively suppressed, with a reduction of up to 65%. Throughout the dynamic coating process, the system exhibited excellent anti-interference capabilities and repeatability, resulting in a significant improvement in yield.

[0083] This invention is not limited to the embodiments described above. Those skilled in the art can make various modifications without departing from the principles of this invention, such as adjusting and optimizing the algorithm or interface. Any equivalent concept or change within the scope of the technology disclosed in this invention is included within the protection scope of this invention.

Claims

1. An adaptive optimization and collaborative decision-making system for spin coating processes on large-diameter substrates, characterized in that: It includes a spin coater, a multimodal sensor array, an online state-parameter estimation module, a predictive controller, an actuator, an event-triggered adaptive optimization module, a fast safety constraint optimization module, an edge AI control device, a two-layer intelligent decision-making module, and a natural language conversion module; The multimodal sensing array includes a multi-wavelength laser interferometric thickness measurement channel (12), a miniature infrared thermal imager (13), a solvent partial pressure sensor (14), a vibration accelerometer (17), and a torque sensor (19), which are used to monitor the film thickness, temperature, solvent partial pressure, vibration parameters, and spindle torque in real time during the spin coating process. The online state-parameter estimation module estimates key parameters of film thickness state online based on a multiphysics model and a nonlinear filtering algorithm. The predictive controller uses a model predictive control algorithm, i.e., the MPC algorithm, to adjust the process parameters based on the output of the online state-parameter estimation module. The actuator is used to adjust the rotation speed, temperature, partial pressure and airflow parameters of the spin coater; The event-triggered adaptive optimization module sets an event trigger threshold and performs parameter optimization and distribution only when the threshold is triggered. The fast safety constraint optimization module adopts a safe Bayesian optimization framework to optimize parameters under safety constraints. The edge AI control device, by configuring an edge AI controller (3) on the device side, enables local operation of a reduced-order digital twin model and a fast solver; The dual-layer intelligent decision-making module includes an upper-layer decision-making module, a lower-layer decision-making module, and a coupling interface module; the upper-layer decision-making module is used to optimize equipment configuration, the lower-layer decision-making module is used to optimize process trajectory, and the coupling interface module realizes collaboration between the upper and lower layers. The natural language conversion module transforms the optimized scheme into a natural language process control recommendation scheme through an intelligent translation center, and constructs a knowledge graph.

2. The adaptive optimization and collaborative decision-making system for spin coating process of large-diameter substrates according to claim 1, characterized in that: The spin coater includes a lower dust cover (1), an upper dust cover (2), an edge AI controller (3), a servo motor (4), a control center (5), a spin coating atmosphere cover (6), a spin coating tray (7), a transmission mechanism (8), a frame (9), a liquid distribution hole (10), a ring support (11), a multi-wavelength laser interferometric thickness measurement channel (12), a miniature infrared thermal imager (13), a solvent partial pressure sensor (14), a solvent vapor injection and exhaust port (15), a spin coater spindle (16), a vibration accelerometer (17), a rolling bearing (18), and a torque sensor (19). The transmission mechanism (8) transmits the torque of the servo motor (4) to the glue-coating tray (7). The servo motor (4) and transmission mechanism (8) are mounted on the frame (9); The lower dust cover (1) is located on the upper part of the frame (9) and below the glue-coating tray (7); The upper dust cover (2) is located above the lower dust cover (1) and the coating tray (7). The upper dust cover (2) is rotatably connected to one side of the lower dust cover (1) by a pin. The upper dust cover (2) is opened when the substrate is placed and the spin coating atmosphere cover (6) is set up, and is fastened when the spin coating is performed and the substrate is left to stand still. The edge AI controller (3) is mounted on the side of the upper dust cover (2); The spin coating atmosphere hood (6) is installed on the upper part of the lower dust cover (1) through a sliding groove. The top cover of the spin coating atmosphere hood (6) is located below the upper dust cover (2) and covers the spin coating tray (7) and the substrate. When placing the substrate, the spin coating atmosphere hood (6) is removed through the sliding groove, the substrate is placed on the spin coating tray (7), and then the spin coating atmosphere hood (6) is fixed through the sliding groove. The servo motor (4) is connected to the control center (5) via a data cable; The liquid distribution hole (10) is located at the center of the upper end of the spin coating atmosphere cover (6) and is used to drop the solution onto the substrate; The annular bracket (11) is located on the inner wall of the top cover of the spin coating atmosphere cover (6), with a radius of 100~300 mm, and is embedded in the substrate rotation center as the center. The multi-wavelength laser interferometric thickness measurement channel (12) is evenly distributed along the circumference of the annular bracket (11) and is embedded in the mounting hole on the annular bracket (11); The miniature infrared thermal imager (13) is embedded in the top cover of the spin coating atmosphere hood (6) at a radius of 2 / 3 from the center. One solvent partial pressure sensor (14) is installed at the top center of the spin coating atmosphere hood (6) and one at the side wall near the exhaust port; The edge AI controller (3) is connected to the multi-wavelength laser interferometric thickness measurement channel (12), miniature infrared thermal imager (13), and solvent partial pressure sensor (14) on the spin coating atmosphere hood (6) via data cables, and is also connected to the external control center (5) via data cables. The solvent vapor injection and exhaust port (15) consists of 8 to 12 cylindrical channels evenly distributed along the circumference of the side wall of the spin coating atmosphere hood (6); The vibration accelerometer (17) is located between the spin coater tray (7) and the rolling bearing (18), and is arranged around the spin coater spinner shaft (16) inside the frame (9); The torque sensor (19) is located below the rolling bearing (18) and is arranged around the spindle (16) of the spin coater inside the frame (9); The vibration accelerometer (17) and torque sensor (19) are integrated in the spindle (16) drive unit of the spindle coater, and together form the spindle dynamic sensor; The multi-wavelength laser interferometric thickness measurement channel (12) has 8 to 12 measurement channels, the accuracy of the miniature infrared thermal imager (13) is ±0.2℃, and the measurement range of the solvent partial pressure sensor (14) is 0-80% saturated vapor pressure.

3. The adaptive optimization and collaborative decision-making system for spin coating process of large-diameter substrates according to claim 1, characterized in that: The online state-parameter estimation module uses the partial differential equations coupled with thin film lubrication, evaporation and Marangoni effect as the prior model, and uses unscented Kalman filtering or particle filtering for state estimation. The state update time is ≤10 ms. The output parameters of the state estimation include film thickness h(r,θ,t), chamber temperature T(r,θ,t), mass fraction c(r,θ,t) and interfacial tension γ(r,θ,t).

4. The adaptive optimization and collaborative decision-making system for spin coating process of large-diameter substrates according to claim 1, characterized in that: The predictive controller is a constrained model predictive control (MPC), with constraints including temperature gradient |dT / dr|≤0.02℃ / mm and partial pressure step ≤10% / 0.1 s; The process parameter control path is optimized by using the particle swarm optimization algorithm after setting parameter control thresholds through the Lagrange method to perform safe reinforcement learning (Safe-RL).

5. The adaptive optimization and collaborative decision-making system for spin coating process of large-diameter substrates according to claim 1, characterized in that: The event triggering thresholds of the event-triggered adaptive optimization module include the threshold for film thickness error |Δh|, the threshold for ripple growth rate d(Astripe) / dt, and the threshold for evaporation flux mutation |ΔJev|, and the event triggering thresholds are dynamically adjusted based on historical data; The threshold for the film thickness error |Δh| is set to trigger when |Δh|>0.5%·h, where h is the target film thickness; The threshold value of the ripple growth rate d(Astripe) / dt is preset according to the substrate diameter and process requirements, and is triggered when d(Astripe) / dt > the threshold value; The threshold for the sudden change in evaporation flux, |ΔJev|, is set to be triggered when |ΔJev|>10%.

6. The adaptive optimization and collaborative decision-making system for spin coating process of large-diameter substrates according to claim 1, characterized in that: The rapid safety constraint optimization module uses a Gaussian process with multi-objective safety constraints as a surrogate model. By fusing simulation data with multi-fidelity kernels and small sample experimental data, it converges to the optimal parameter configuration within ≤30 experimental points and outputs a confidence interval. The safety constraints include VOC emission limits, pollution thresholds, and stripe contrast limits.

7. The adaptive optimization and collaborative decision-making system for spin coating process of large-diameter substrates according to claim 1, characterized in that: The edge AI control device uses FPGA or Edge-GPU hardware, with a single optimization calculation latency of ≤20 ms. The optimization variables include the linkage speed curve ω(t), zone temperature control Ti(t), zone pressure pi(t), and zone airflow ui(t). It supports collaboration with host computer model predictive control or digital twins and automatically reverts to a safe configuration in case of anomalies.

8. The adaptive optimization and collaborative decision-making system for spin coating process of large-diameter substrates according to claim 1, characterized in that: The upper-level decision module optimizes the configuration of the spin coater, which includes the number of partitions, nozzle layout, pressure-dividing annular cavity topology, and edge air gap geometry. The achievable manufacturing window is maximized through topology optimization and a multi-objective particle swarm genetic algorithm. The lower-level decision module optimizes the trajectory of the spin coating process, which includes the rotation speed curve ω(t), zone control temperature Ti(t), zone pressure pi(t), and zone airflow ui(t). The thickness gradient and defect risk are minimized through model predictive control and a safe Bayesian optimization framework.

9. The adaptive optimization and collaborative decision-making system for spin coating process of large-diameter substrates according to claim 1, characterized in that: The coupling interface module uses a unified cost function and sensitivity interface to couple the upper-level decision-making module with the lower-level decision-making module, automatically generates A / B verification schemes and performs closed-loop iterations, and outputs a joint optimization scheme for equipment configuration and spin coating process trajectory; the unified cost function simultaneously considers film uniformity, energy consumption and edge thickness peak index.

10. The adaptive optimization and collaborative decision-making system for spin coating process of large-diameter substrates according to claim 1, characterized in that: The natural language conversion module interfaces with the Manufacturing Execution System (MES) and the Semiconductor Equipment Communication Standard (SECS-GEM), respectively, supporting natural language queries and audit trails; the knowledge graph stores the "defect-cause-solution-effect" relationship.

Images