A multi-segment dynamic collaborative heating control method and system for a single-crystal silicon thermal field

Abstract

The present invention relates to a method and system for dynamic collaborative heating control of a multi-segment single-crystal thermal field, comprising the following steps: constructing a temperature coupling model for the multi-segment thermal field, dividing the thermal field into n axial heating segments, and establishing a dynamic relationship between power input and temperature gradient; designing a deep reinforcement learning agent, whose state space integrates the temperatures of each heating segment, the crystal pulling speed, the crucible lifting speed, and the temperature gradient, and whose action space outputs the power adjustment amount for each heating segment; defining a multi-objective reward function to jointly optimize temperature gradient tracking, total power change, and the crucible bottom safety threshold; offline training the agent based on the thermal field model to optimize the policy network; online deploying the agent to generate power adjustment instructions in real time; dynamically correcting the control strategy and performing online fine-tuning by setting trigger adjustments. The present invention solves the deficiencies of traditional methods in terms of temperature gradient control accuracy, energy consumption economy, and adaptability to abnormal conditions, and significantly improves the quality and efficiency of single-crystal growth.

Description

Technical Field

[0001] The present invention relates to the technical field of single-crystal silicon thermal fields, and in particular to a multi-segment dynamic collaborative heating control method and system for a single-crystal silicon thermal field. Background Art

[0002] During the growth process of single-crystal silicon, the control of the temperature gradient in the thermal field directly determines the crystal quality. The traditional technologies mainly adopt the following control schemes:

[0003] 1. PID control: The heating power is adjusted through the proportional-integral-derivative algorithm, but there are the following problems:

[0004] Difficult decoupling of multi-variable coupling: Each heating section in the thermal field affects each other. Fixed PID parameters lead to serious overshoot under dynamic working conditions, and the temperature gradient fluctuation range reaches ±10°C / cm;

[0005] Defect of lag response: The heat conduction delay effect is significant (typical delay 3 - 10 seconds), and the PID feedback control cannot predict the temperature change trend.

[0006] 2. Fuzzy control: It depends on expert experience to set the rule base, and the adaptability is insufficient:

[0007] Strong dependence on experience: Different furnace types and silicon material purities require re-tuning of parameters, and the deployment period is long (> 2 weeks);

[0008] Lack of multi-objective optimization: It is difficult to simultaneously take into account the temperature gradient, energy consumption and equipment safety threshold.

[0009] 3. Static partition heating: The thermal field is divided into fixed heating sections, and the power of each section is manually set:

[0010] Poor coordination: The power adjustment of each section is isolated from each other, resulting in uneven energy distribution in the thermal field (local temperature difference > 50°C);

[0011] High energy consumption: To compensate for the thermal field distortion, the total power often needs to be increased by an excess (15% - 20%).

[0012] In summary, the existing single-crystal silicon thermal field has the following problems: insufficient dynamic collaborative ability: unable to respond in real time to the disturbances of the crystal lifting speed and the melt level change on the thermal field; single control target: only focusing on the temperature gradient or energy consumption, ignoring the comprehensive optimization of multiple constraint conditions; rigid handling of abnormal working conditions: relying on manual intervention, resulting in a high frequency of shutdowns (average 3 - 5 times per furnace) and a high risk of crystal growth interruption. Summary of the Invention

[0013] To this end, the technical problem to be solved by the present invention is to overcome the problems existing in the single-crystal thermal field in the prior art, and provide a multi-segment dynamic collaborative heating control method and system for a single-crystal thermal field, which solves the deficiencies of the traditional method in terms of temperature gradient control accuracy, energy consumption economy, and adaptability to abnormal working conditions, and can improve the quality and efficiency of single-crystal growth.

[0014] To solve the above technical problems, the present invention provides a multi-segment dynamic collaborative heating control method for a single-crystal thermal field, including the following steps:

[0015] Construct a multi-segment temperature coupling model for the thermal field, divide the thermal field into n axial heating segments, and establish a dynamic relationship between the power input of each heating segment and the temperature gradient of crystal growth;

[0016] Design a deep reinforcement learning agent, whose state space includes: the real-time temperature values T1~T of each heating segment n , the crystal pulling speed v, the crucible lifting speed s, and the measured value G of the longitudinal temperature gradient of the thermal field; the action space is the power adjustment amount ΔP1~ΔP of each heating segment n ;

[0017] Define the reward function:

[0018] ;

[0019] where α, β, γ are weight coefficients related to G, ΔP, and T respectively, G t : the target temperature gradient, ΔP is the total power change amount, T c : the temperature at the bottom of the crucible, T th : the safety temperature threshold;

[0020] Based on the multi-segment temperature coupling model of the thermal field and the reward function, simulate the whole process of crystal growth, and optimize the parameter of the agent's policy network;

[0021] Connect the trained agent to the thermal field, read the sensor data in real time, and output the power adjustment instructions for each segment;

[0022] Dynamically correct the control strategy, set the trigger condition, automatically trigger the strategy update, and correct the control parameters.

[0023] In an embodiment of the present invention, the construction of the multi-segment temperature coupling model of the thermal field includes:

[0024] Based on the finite element analysis method, establish a dynamic relationship between the power input of each heating segment and the temperature gradient, satisfying:

[0025] G(t)=f(P1(t - τ1), P2(t - τ2),..., Pn(t - τ n )) + ε(t);

[0026] Among them, τ1 to τ n are the heat conduction delay times of each heating section, and ε(t) is the environmental noise disturbance term.

[0027] In an embodiment of the present invention, a plurality of independently temperature-controlled heaters are arranged along the axial direction of the thermal field, including:

[0028] The upper-section annular heater covers the area 50 - 100 mm above the solid-liquid interface and isolates the top water-cooled area;

[0029] The middle-section multi-stage gradient heater includes at least 3 concentric heating rings that can be independently temperature-controlled;

[0030] The lower-section bottom auxiliary heater isolates the bottom area.

[0031] In an embodiment of the present invention, an infrared temperature measurement array is arranged at the crystal growth interface, the crucible bottom, and multiple axial positions, including:

[0032] Laser pyrometer: Detect the temperature of each heating section;

[0033] Contact temperature measurement module: Used to measure the temperature of the crucible bottom;

[0034] Distributed optical fiber temperature sensor, used to collect axial temperature gradient data:

[0035] In an embodiment of the present invention, the crystal pulling speed v and the crucible lifting speed s satisfy a dynamic synchronization relationship:

[0036] s(t)=v(t)×(1 + k×dD(t) / dt)

[0037] Among them, k = 0.1 - 0.3 is the diameter compensation coefficient, D(t) is the real-time crystal diameter, and dD(t) / dt is the crystal diameter change rate.

[0038] In an embodiment of the present invention, the weight coefficient in the reward function is dynamically adjusted according to the crystal growth stage:

[0039] Seeding stage: Increase the α value, and give priority to stabilizing the temperature gradient;

[0040] Shoulder release stage: Increase the β value to suppress power mutation;

[0041] Constant diameter stage: Balance the coefficient and synchronously optimize the temperature gradient and energy consumption.

[0042] In an embodiment of the present invention, based on the multi-section temperature coupling model of the thermal field and the reward function, the whole process of crystal growth is simulated, including at least:

[0043] The dynamic response process of the thermal field, including: the evolution of the multi-section temperature field and the formation of the longitudinal temperature gradient;

[0044] The crystal morphology evolution process includes: the growth of the seed crystal in the seeding stage, the diameter expansion in the shoulder release stage, and the steady-state growth in the constant diameter stage;

[0045] The thermo-mechanical-electrical coupling process includes: the interaction between the pulling speed v and the thermal field, the crucible lifting speed s and the melt level control, and the power regulation and energy consumption accumulation;

[0046] The abnormal working conditions and fault tolerance process include: equipment failure simulation, process disturbance simulation, and safety boundary test;

[0047] The long-term evolution and equipment life include: heater aging simulation, quartz crucible corrosion.

[0048] In an embodiment of the present invention, the dynamic correction control strategy includes:

[0049] First-level response: when the diameter deviation is > 0.5 mm and ≤ 1 mm, adjust the crystal pulling speed, and preferentially adjust the power of the top heating section;

[0050] Second-level response: when the diameter deviation is > 1 mm and the melt level deviation ≤ 1 mm, adjust the crystal pulling speed, and preferentially adjust the power of the middle heating section;

[0051] Third-level response: when the melt level line deviation is > 1 mm and the bottom crucible temperature fluctuation is > 15 °C, pause the pulling action, the crucible lifting action, the crystal rotation and crucible rotation actions, set the top and middle heating powers to zero, and start the constant temperature protection mode of the bottom heating section.

[0052] In an embodiment of the present invention, a PID closed-loop compensation module is set, and the power adjustment amount output by the reinforcement learning is used as a feedforward signal, which is superimposed on the PID feedback control amount and then output to the temperature control heater.

[0053] In an embodiment of the present invention, after each crystal growth is completed, the intelligent agent performs the following operations:

[0054] Drive the temperature control heaters of each heating section to step up the load from 10% to 95% of the rated power, and record the temperature rise curve;

[0055] Compare with the historical data. If the temperature rise rate deviation of a certain section is > 15%, mark the temperature control heater of this section as the state to be repaired.

[0056] To solve the above technical problems, the present invention also provides a multi-section dynamic cooperative heating system for the single crystal thermal field, including:

[0057] Thermal field coupling modeling module: including: a storage unit for presetting a thermal field multi-segment temperature coupling model and defining the dynamic relationship between the power input and temperature gradient of n axial heating segments; a processing unit: calculating the power-temperature gradient mapping relationship based on the finite element analysis equation;

[0058] Agent control module: used to receive the temperatures T1 to T of each heating segment n , crystal pulling speed v, crucible lifting speed s, and the measured value G of the longitudinal temperature gradient of the thermal field, and output the power adjustment amounts ΔP1 to ΔP of each heating segment n ;

[0059] Reward calculation module: preset the target temperature gradient G and the safety temperature threshold T th , and perform the reward function operation:

[0060] ;

[0061] Policy optimization module: based on the thermal field coupling model, simulate the whole process of crystal growth, and optimize the parameters of the agent policy network;

[0062] Real-time control module: collect temperature, speed, and gradient data in real time, deploy the trained policy network, generate power adjustment instructions, and send the adjustment signals of ΔP1 to ΔP to the temperature control heater n ;

[0063] Policy update module: used to automatically trigger policy update according to the preset trigger conditions to correct the control parameters.

[0064] The above technical solutions of the present invention have the following advantages compared with the prior art:

[0065] For the multi-segment dynamic collaborative heating control method of single crystal thermal field of the present invention, compared with the traditional partition PID control, the present invention realizes the collaborative regulation of the power of multiple segments of the thermal field through dynamic coupling modeling and multi-objective reinforcement learning control. A deep reinforcement learning agent is designed to synchronously sense the temperatures, pulling speeds, crucible positions and real-time temperature gradients of all segments in its state space, and directly output the power adjustment amounts of each heating segment in the action space. The collaborative rules for increasing and decreasing the power of multiple segments are automatically learned through the neural network policy, breaking through the limitations of manual experience in parameter tuning. A temperature gradient tracking term and a power change penalty term are set in the reward function to drive the agent to avoid drastic power adjustments while reducing gradient fluctuations; an online policy fine-tuning module is introduced, and trigger conditions are set to trigger the reweighting of the local reward function to achieve dynamic compensation for the thermal field; through the synergistic effect of the above technical features, the inherent defects of the traditional method in dynamic collaboration, multi-objective optimization and adaptive control are theoretically solved. Brief Description of the Drawings

[0066] To make the content of the present invention easier to be clearly understood, the following further details the present invention according to specific embodiments of the present invention in conjunction with the accompanying drawings, where:

[0067] Figure 1 is the step flow chart of the multi-segment dynamic collaborative heating control method for the single crystal thermal field of the present invention;

[0068] Figure 2 is the framework diagram of the multi-segment dynamic collaborative heating control system for the single crystal thermal field of the present invention. Specific Embodiments

[0069] The following further illustrates the present invention in conjunction with the accompanying drawings and specific embodiments, so that those skilled in the art can better understand the present invention and be able to implement it, but the exemplified embodiments do not limit the present invention. Embodiment 1

[0070] Referring to Figure 1 as shown, the present invention discloses a multi-segment dynamic collaborative heating control method for a single crystal thermal field, including the following steps:

[0071] S10. Construct a multi-segment temperature coupling model for the thermal field, divide the thermal field into n axial heating segments, and establish the dynamic relationship between the power input of each heating segment and the temperature gradient of crystal growth.

[0072] It should be noted that during the single crystal growth process, the requirements for the temperature gradient near the solid-liquid interface (upper segment), the main body of the crystal (middle segment), and the bottom of the crucible (lower segment) are different. In this embodiment, the thermal field is divided into axially independent temperature control zones, and the temperature of each segment can be controlled. However, in the single crystal thermal field, the power inputs of each heating segment will affect each other through heat conduction, radiation, etc., forming a complex multi-variable coupling system. It is difficult for traditional methods to decouple this non-linear relationship, resulting in low temperature gradient control accuracy. Therefore, the present application establishes a multi-segment temperature coupling model for the thermal field, which can represent the dynamic relationship between the power input of each heating segment and the temperature gradient of crystal growth.

[0073] Specifically, in this embodiment, based on the finite element analysis method, the dynamic relationship between the power input of each heating segment and the temperature gradient is established, satisfying:

[0074] G(t)=f(P1(t - τ1), P2(t - τ2),..., Pn(t - τ n )) + ε(t);

[0075] where τ1~τ n are the heat conduction delay times of each heating segment, and ε(t) is the environmental noise disturbance term.

[0076] The thermal field is discretized into finite elements, and the transfer relationship between the power of each section and the temperature gradient is described by a matrix equation to clarify the weight coefficients among variables and achieve the mathematical decoupling of the coupled system;

[0077] It is found through experiments that there is a significant delay in the heat transfer in the thermal field. The measured delay time is 3 - 10 s. Due to the inability of traditional feedback control to predict the delay effect, overshoot phenomena occur, and the temperature gradient fluctuation reaches ±10 °C / cm. In the dynamic equation of this embodiment, delay terms τ1 - τ are explicitly introduced n , and the lag effect is offset through a pre-compensation mechanism to make the control instruction take effect in advance;

[0078] Moreover, in the thermal field, there are also influences of environmental noises such as melt convection and cooling water flow fluctuation, which will damage the thermal field stability. It is difficult for the empirical model to distinguish noises from real thermal field changes. To solve this problem, in the dynamic equation of this embodiment, the Kalman filter is also combined to perform online estimation and suppression on the noise term ε(t) to enhance the anti-interference ability of the model.

[0079] When actually dividing the axial heating sections, it is necessary to divide according to the size of the thermal field. At least three sections are set along the axis of the thermal field, and independent temperature control heaters are correspondingly set for each section. In this embodiment, it includes:

[0080] The upper-section annular heater covers the area 50 - 100 mm above the solid-liquid interface and isolates the top water-cooled area. On the one hand, it can quickly adjust the temperature gradient of the solid-liquid interface, and on the other hand, it can isolate the top water-cooled area to reduce heat loss;

[0081] The middle-section multi-stage gradient heater includes at least 3 concentric heating rings that can be independently temperature-controlled. The multi-stage gradient heater in the middle section plays a crucial role in adjusting the temperature gradient. The more heaters are set in this section, the more stable the temperature gradient adjustment effect is, and it can maintain the longitudinal temperature uniformity;

[0082] The lower-section bottom auxiliary heater isolates the bottom area. On the one hand, it can stably heat the melt to stabilize the melt convection, and on the other hand, it can also prevent the bottom of the crucible from being supercooled and reduce heat loss.

[0083] S20. Design a deep reinforcement learning agent, whose state space includes: the real-time temperature values T1 - T of each heating section n , the crystal pulling speed v, the crucible lifting speed s, and the measured value G of the longitudinal temperature gradient of the thermal field; the action space is the power adjustment amount ΔP1 - ΔP of each heating section n .

[0084] In reinforcement learning, the state space is the set of all possible states that an agent can perceive of the environment, while the action space is the set of all possible actions that the agent can execute. The parameters in the state space need to comprehensively describe the state of the current environment so that the agent can make decisions, and the action space is the action selected by the agent based on the current state.

[0085] In this embodiment, the state space introduces the real-time temperature values T1 to T of each heating section n , which reflects the instantaneous state of the axial energy distribution of the thermal field and is the basis for realizing independent zone temperature control, and can suppress local temperature mutations.

[0086] The state space introduces the crystal pulling speed v and the crucible lifting speed s: mechanical control quantities used to characterize the crystal growth rate. During the crystal pulling process, as the silicon crystal is slowly pulled up, its lifting speed is v, and the melt level will gradually drop. Subsequently, the crucible will rise upward. The role of the crucible lifting speed s is to synchronously lift the crucible position to compensate for the drop in the melt level and avoid the offset of the solid-liquid interface position. At the same time, the rising of the crucible will change the relative position between the heater and the melt, directly affecting the radiative heat transfer path of the thermal field. The crucible lifting speed s is strongly correlated with the crystal growth stage. The agent needs to identify the current growth stage through the crucible lifting speed s, so as to adapt to the control strategies of different stages;

[0087] Specifically, the crystal pulling speed v and the crucible lifting speed s need to satisfy the dynamic mass conservation equation:

[0088] s(t) = v(t) × (1 + k × dD(t) / dt)

[0089] where k = 0.1 to 0.3 is the diameter compensation coefficient, D(t) is the real-time crystal diameter, and dD(t) / dt is the crystal diameter change rate; if s ≠ v × (1 + k × dD / dt), the melt level will continuously rise or fall, resulting in a change in the thermal field boundary conditions (such as the offset of the solid-liquid interface position), leading to G temperature gradient runaway. For example, when the crystal diameter increases (dD / dt > 0), it is necessary to increase the crucible lifting speed s to supplement the melt consumption and avoid a sudden increase in the temperature of the lower section caused by the drop in the liquid level;

[0090] Moreover, the pulling speed v directly affects the crystal growth rate, and the change in the growth rate will change the heat flux demand at the solid-liquid interface. It is necessary to synchronously adjust the power adjustment amounts ΔP1 to ΔP of each heating section n to maintain the target temperature gradient G. Incorporating the pulling speed v into the state space of the agent, the agent can learn the mapping relationship of pulling speed v → ΔP1 to ΔP n and can control the power adjustment amounts ΔP1 to ΔP of each heating section according to the actual pulling speed v. n .

[0091] Introduce the measured value G of the longitudinal temperature gradient of the thermal field into the state space, which is a core index for quantifying the thermodynamic stability of the solid-liquid interface. Directly measure G and use it as the state input to improve the real-time performance of control.

[0092] In this embodiment, the action space is defined as the power adjustment amounts ΔP1 to ΔP of each heating section n For multiple heating sections arranged correspondingly, different power adjustment requirements are realized. For example, the upper section ΔP1 needs to quickly respond to the change of the heat flow at the solid-liquid interface and suppress the diameter fluctuation. The middle sections ΔP2 to ΔP n-1 Need to maintain the longitudinal temperature uniformity and reduce the temperature fluctuation. The lower section ΔP n Needs to control the temperature at the bottom of the crucible to prevent the bottom of the crucible from being overcooled / overheated.

[0093] Specifically, corresponding to the division of the heating sections, in order to realize the temperature detection requirements of each heating section, an infrared temperature measurement array is set at the crystal growth interface, the bottom of the crucible and multiple axial positions, including:

[0094] Laser thermometer: Detect the temperature of each heating section and can non-contact measure the surface temperature of each heating section;

[0095] Contact temperature measurement module: Used to measure the temperature at the bottom of the crucible;

[0096] Distributed optical fiber temperature sensor, used to collect axial temperature gradient data and can realize continuous measurement of the temperature gradient.

[0097] S30. Define the reward function:

[0098] ;

[0099] where α, β, γ are weight coefficients related to G, ΔP, and T respectively. G t : Target temperature gradient, ΔP is the total power change amount, T c : Temperature at the bottom of the crucible, T th : Safety temperature threshold.

[0100] It should be noted that the growth of single crystal silicon needs to simultaneously meet the objectives such as temperature gradient accuracy, energy consumption economy, and equipment safety. Traditional PID or fuzzy control is difficult to balance these mutually restrictive factors. Therefore, in this embodiment, introducing a reward function plays the role of a strategy optimization guiding target in deep reinforcement learning. Through the reward function, the complex single crystal silicon thermal field control problem can be transformed into a multi-objective optimization task. In the reward function, multiple objectives are integrated into a single scalar signal through weighted summation to drive the intelligent agent to learn the optimal strategy.

[0101] Specifically, the parameter terms of the reward function include: a temperature gradient tracking term, which is used to penalize the measured gradient G and the deviation from the target G t , ensuring low-defect crystal growth; a power change penalty term, which is used to suppress drastic power fluctuations, reduce energy consumption, and extend the heater life; a safety constraint term, which is used to apply a penalty when the crucible bottom temperature T c exceeds the threshold T th , avoiding the melting damage of the quartz crucible.

[0102] In this embodiment, the agent continuously generates actions, and the reward function provides real-time feedback on the action quality, including:

[0103] Positive reward: When G → G t and ΔP is small, the R value increases, strengthening the current strategy;

[0104] Negative penalty: If T c > T th , the R value drops sharply, prompting the agent to avoid dangerous actions;

[0105] For example: in the scenario of diameter mutation, the agent may attempt to increase the power of the upper section and decrease the power of the lower section simultaneously. The reward function guides it to find a balance point through the comprehensive evaluation of α(G t - G)² and β(ΔP)².

[0106] Specifically, the weight coefficients α, β, γ essentially define the priorities of the control objectives. In different production stages, the control objectives are different, so different weight coefficient ratios need to be set. The agent learns the optimal strategies under different weights through offline training and dynamically switches strategies according to the growth stage during online deployment.

[0107] In actual production, crystal growth is divided into three stages, including: the seeding stage, the shoulder release stage, and the isodiameter stage. The weight coefficients in the reward function are dynamically adjusted according to the crystal growth stage as follows:

[0108] Seeding stage: The seed crystal makes initial contact with the melt, and the solid-liquid interface is formed. At this time, the diameter of the seed crystal is small (4.5 - 5.5 mm), the heat capacity is low, and slight fluctuations in the temperature gradient can cause melting or dislocation multiplication. Increase the value of α, allowing short-term power mutations to quickly respond to gradient deviations, sacrificing some energy consumption economy, and giving priority to ensuring the stability of growth initiation. At this time, the weight distribution is α = 0.6, β = 0.3, γ = 0.1. This stage requires gradient accuracy to be prioritized.

[0109] Shoulder release stage: The crystal diameter increases rapidly, resulting in a sharp reduction in heat demand. Drastic power adjustments are likely to cause mechanical vibrations and thermal stress cracks. Increase the value of β, restricting the power adjustment range of each section. At this time, the weight distribution is α = 0.2, β = 0.7, γ = 0.1, preventing power mutations;

[0110] Equal-diameter stage: The diameter remains constant. For long-term steady-state growth, both energy consumption (cost-sensitive) and equipment life (cumulative damage from power cycling) need to be considered. At this time, the weight distribution is α = 0.4, β = 0.4, γ = 0.2. The balance coefficient is used to synchronously optimize the temperature gradient and energy consumption. Moreover, the safety weight γ is increased to prevent the temperature drift at the bottom of the crucible during long-term operation (such as the slow increase caused by the aging of the quartz crucible).

[0111] S40. Based on the multi-segment temperature coupling model of the thermal field and the reward function, simulate the entire process of crystal growth and optimize the parameters of the agent's policy network.

[0112] Specifically, in the above step S10, the dynamic relationship between the power input of each heating segment and the temperature gradient is established through finite element analysis, providing a physical basis for the simulation. In the simulation, it is necessary to strictly align the input and output of the agent according to step S20 to ensure the compatibility of the policy network structure with the actual control. The reward function in step S30 is used as the optimization criterion to drive the agent to explore the optimal solution set.

[0113] Specifically, in order to cover as many situations and problems as possible that may occur in the actual process, when simulating the entire process of crystal growth, it should at least include:

[0114] The dynamic response process of the thermal field, including: the evolution of the multi-segment temperature field and the formation of the longitudinal temperature gradient. For example: the influence of the upper-segment annular heater on the heat flow at the solid-liquid interface, the cooperative temperature control of the middle-segment gradient heating ring, the regulation effect of the lower-segment auxiliary heater on the temperature at the bottom of the crucible, and the gradient tracking under normal conditions, as well as the gradient distortion caused by abnormal disturbances (such as sudden changes in melt convection).

[0115] The crystal morphology evolution process, including: the growth of the seed crystal in the seeding stage, the diameter expansion in the shoulder release stage, and the steady-state growth in the equal-diameter stage. Simulate the initial crystallization process after the seed crystal is immersed in the melt, the rapid growth process of the crystal diameter from millimeter level to the target size (such as 200mm), and simulate the crystal growth process with a constant diameter.

[0116] The thermal-mechanical-electrical coupling process, including: the interaction between the pulling speed v and the thermal field, the crucible lifting speed s and the melt level control, and the power regulation and energy consumption accumulation. Simulate the feedback effect of the change in the crystal pulling speed on the heat flow at the solid-liquid interface, and based on the dynamic synchronization relationship between the pulling speed v and the crucible lifting speed s, simulate the deviation of the thermal field boundary condition caused by the melt level deviation and the risk of melt depletion caused by abnormal crucible lifting speed (such as sticking, s = 0). Simulate the total energy consumption under different control strategies and optimize the β term in the reward function.

[0117] Abnormal conditions and fault tolerance processes, including: equipment failure simulation, process disturbance simulation, safety boundary testing. For example: thermal field compensation strategies when a heater failure causes ΔP = 0 in a certain section, changes in melt composition and fluctuations in cooling water, emergency responses when the bottom of the crucible overheats or the diameter gets out of control.

[0118] Long-term evolution and equipment life, including: heater aging simulation, quartz crucible corrosion. For example: simulating the power attenuation effect caused by the increase in the resistivity of the temperature-controlled heater over time, training the intelligent agent to adaptively adjust the initial power adjustment amount; simulating the impact of the reduction in the crucible wall thickness on the thermal field boundary to determine when to replace the crucible.

[0119] S50. Connect the trained intelligent agent to the thermal field, read sensor data in real time, and output power adjustment instructions for each section.

[0120] After ensuring that the policy network has learned the typical conditions of the thermal field, apply the intelligent agent to the actual process production, control the power adjustment instructions for each section to achieve the expected technical effects, realize the coordinated adjustment of the multi-section power of the thermal field, and improve the crystal quality and production efficiency through real-time closed-loop optimization.

[0121] S60. Dynamically correct the control strategy, set trigger conditions, automatically trigger strategy updates, and correct control parameters.

[0122] In this embodiment, the purpose of dynamically correcting the control strategy is to adaptively adjust control parameters and strategies by real-time sensing of system state changes to ensure the continuous optimal performance of the single-crystal thermal field control system in a complex dynamic environment. The trigger conditions can be set with different types and different thresholds according to actual production requirements, including: process parameter deviation, equipment performance degradation, energy consumption economy degradation, and safety boundary approximation, etc., and corresponding mechanisms at different levels can also be set according to actual trigger situations.

[0123] In this embodiment, two-level corresponding mechanisms are set, which are respectively used to produce different control strategies, where:

[0124] First-level response: The diameter deviation > 0.5mm and ≤ 1mm. Adjust the crystal pulling speed and preferentially adjust the power of the top heating section. The first-level response is an adjustment response mechanism, that is, the automatic adjustment of the strategy is realized according to the simulation training of the intelligent agent, belonging to the early warning of process deviation.

[0125] Second-level response: The diameter deviation > 1mm and the melt level deviation ≤ 1mm. Adjust the crystal pulling speed and preferentially adjust the power of the middle heating section. The second-level response is an adjustment response mechanism, that is, the automatic adjustment of the strategy is realized according to the simulation training of the intelligent agent, belonging to the process deviation warning during the production stage.

[0126] Level 3 response: When the deviation of the molten liquid level line > 1 mm and the temperature fluctuation at the bottom of the crucible > 15 °C, the lifting action, the crucible lifting action, and the crystal rotation and crucible rotation actions are suspended. The heating power of the top and middle parts is set to zero, and the constant temperature protection mode of the bottom heating section is started. The level 3 response is an emergency response mechanism that is independent of the agent, that is, for the emergency disposal of major process anomalies to avoid equipment damage and safety accidents.

[0127] Specifically, after each crystal growth is completed, the agent performs the following operations:

[0128] Drive the temperature control heaters of each heating section to increase the load step by step from 10% to 95% of the rated power, and record the temperature rise curve;

[0129] Compare with historical data. If the deviation of the temperature rise rate in a certain section > 15%, mark the temperature control heater in this section as the state to be repaired.

[0130] In this embodiment, through the step-by-step load increase test and temperature rise rate analysis, the online diagnosis and predictive maintenance of the health status of the temperature control heaters are realized. It can detect hidden faults such as the aging of the heater resistance wire and the deterioration of the heat insulation layer, quantify the degree of performance decay, and identify anomalies (such as local hot spots and poor contacts) before the heater completely fails, avoiding the scrapping of the entire furnace of crystals caused by sudden failures during the production process. Embodiment 2

[0131] Specifically, on the basis of the above Embodiment 1, in order to further illustrate how to implement the above heating control method, referring to Figure 2 as shown, the present invention also discloses a multi-section dynamic collaborative heating system for a single crystal thermal field to implement the method described in Embodiment 1, including:

[0132] Thermal field coupling modeling module: including: a storage unit for presetting a multi-section temperature coupling model of the thermal field and defining the dynamic relationship between the power input and the temperature gradient of n axial heating sections; a processing unit: calculating the power-temperature gradient mapping relationship based on the finite element analysis equation;

[0133] Agent control module: used to receive the temperatures T1~T n of each heating section, the crystal lifting speed v, the crucible lifting speed s, and the measured value G of the longitudinal temperature gradient of the thermal field, and output the power adjustment amounts ΔP1~ΔP n ;

[0134] Reward calculation module: preset the target temperature gradient G and the safety temperature threshold T th , and perform the reward function operation:

[0135] ;

[0136] Strategy optimization module: Simulate the whole process of crystal growth based on the thermal field coupling model and optimize the parameters of the agent policy network;

[0137] Real-time control module: Collect temperature, speed, and gradient data in real time, deploy the trained policy network, generate power adjustment instructions, and send adjustment signals of ΔP1 to ΔP to the temperature control heater; n

[0138] Strategy update module: Automatically trigger strategy updates according to preset trigger conditions to correct control parameters.

[0139] Obviously, the above embodiments are only examples given for clear illustration and are not limitations on the implementation manners. For those of ordinary skill in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to enumerate all the implementation manners here. And the obvious changes or modifications derived therefrom still fall within the protection scope of the present invention.​

Claims

1. A method for dynamically and cooperatively controlling the heating of multiple sections of a single-crystal silicon thermal field, characterized in that, Including the following steps: Construct a thermal field multi-segment temperature coupling model, divide the thermal field into n axial heating segments, and establish the dynamic relationship between the power input of each heating segment and the temperature gradient of crystal growth; The construction of the thermal field multi-segment temperature coupling model includes: based on the finite element analysis method, establish the dynamic relationship between the power input of each heating segment and the temperature gradient, satisfying: G(t)=f(P1(t - τ1), P2(t - τ2),..., P n (t - τ n )) + ε(t); Among them, τ1 to τ n are the heat conduction delay times of each heating section, and ε(t) is the environmental noise disturbance term; Design a deep reinforcement learning agent whose state space includes: real-time temperature values T1 to T of each heating section n , crystal pulling speed v, crucible lifting speed s, measured value G of the longitudinal temperature gradient of the thermal field; the action space is the power adjustment amount ΔP1 to ΔP of each heating section n ; Define the reward function: ； Among them, α, β, γ are weight coefficients related to G, ΔP, and T respectively, where G t : is the target temperature gradient, ΔP is the total power change, and T c : is the temperature at the bottom of the crucible, and T th : is the safety temperature threshold; Based on the thermal field multi-segment temperature coupling model and the reward function, simulate the whole process of crystal growth, and optimize the parameters of the intelligent agent policy network; Connect the trained intelligent agent to the thermal field, read the sensor data in real time, and output the power adjustment instructions for each segment; Dynamically correct the control strategy, set the trigger condition, automatically trigger the strategy update, and correct the control parameters.

2. The method for dynamically and cooperatively controlling multi-segment heating of a single-crystal silicon thermal field according to claim 1, wherein: Set multiple independent temperature-controlled heaters along the axial distribution of the thermal field, including: The upper-segment annular heater covers the area 50-100 mm above the solid-liquid interface and isolates the top water-cooled area; The middle-segment multi-stage gradient heater includes at least 3 concentric heating rings that can be independently temperature-controlled; The lower-segment bottom auxiliary heater isolates the bottom area.

3. The method for dynamically and cooperatively controlling the heating of multiple sections of a single-crystal silicon thermal field according to claim 1, wherein: Set up an infrared temperature measurement array at the crystal growth interface, the bottom of the crucible and multiple axial positions, including: Laser pyrometer: Detect the temperature of each heating segment; Contact temperature measurement module: Used to measure the temperature at the bottom of the crucible; Distributed optical fiber temperature sensor, used to collect axial temperature gradient data.

4. The method for multi-section dynamic collaborative heating control of a single-crystal silicon thermal field according to claim 1, wherein: The crystal pulling speed v and the crucible lifting speed s satisfy the dynamic synchronization relationship: s(t)=v(t)×(1 + k×dD(t) / dt) where k = 0.1-0.3 is the diameter compensation coefficient, D(t) is the real-time crystal diameter, and dD(t) / dt is the crystal diameter change rate.

5. The method for dynamically and cooperatively controlling the heating of multiple sections of a single-crystal silicon thermal field according to claim 1, wherein: The weight coefficient in the reward function is dynamically adjusted according to the crystal growth stage: Seeding stage: Increase the α value to give priority to stabilizing the temperature gradient; Shoulder release stage: Increase the β value to suppress power mutation; Constant diameter stage: Balance the coefficient and synchronously optimize the temperature gradient and energy consumption.

6. The method for controlling multi-segment dynamic collaborative heating of a single-crystal silicon thermal field according to claim 1, wherein: Based on the thermal field multi-segment temperature coupling model and the reward function, simulate the whole process of crystal growth, including at least: Thermal field dynamic response process, including: multi-segment temperature field evolution and longitudinal temperature gradient formation; Crystal morphology evolution process, including: seed crystal growth in the seeding stage, diameter expansion in the shoulder release stage, and steady growth in the constant diameter stage; Thermal-mechanical-electrical coupling process, including: interaction between the pulling speed v and the thermal field, crucible lifting speed s and melt level control, and power adjustment and energy consumption accumulation; Abnormal working conditions and fault tolerance process, including: equipment failure simulation, process disturbance simulation, safety boundary test; Long-term evolution and equipment life, including: heater aging simulation, quartz crucible corrosion.

7. The method for controlling multi-segment dynamic collaborative heating of a single-crystal silicon thermal field according to claim 1, characterized in that: The dynamic correction control strategy includes: First-level response: The diameter deviation > 0.5 mm and ≤ 1 mm, adjust the crystal pulling speed, and preferentially adjust the power of the top heating segment; Second-level response: The diameter deviation > 1 mm and the melt level deviation ≤ 1 mm, adjust the crystal pulling speed, and preferentially adjust the power of the middle heating segment; Third-level response: The melt level line deviation > 1 mm and the bottom temperature fluctuation of the crucible > 15 °C, pause the pulling action, the crucible lifting action, the crystal rotation and crucible rotation actions, set the top and middle heating powers to zero, and start the constant temperature protection mode of the bottom heating segment.

8. The method for multi-segment dynamic collaborative heating control of a single-crystal silicon thermal field according to claim 1, wherein: After each crystal growth is completed, the agent performs the following operations: Drive the temperature control heaters in each heating section to step up the load from 10% to 95% of the rated power, and record the temperature rise curve; Compare the historical data. If the temperature rise rate deviation in a certain section is > 15%, mark the temperature control heater in this section as the state to be repaired.

9. A multi-segment dynamic collaborative heating system for a single crystal silicon thermal field, characterized in that: Including: Thermal field coupling modeling module: including: a storage unit for presetting a multi-section temperature coupling model of the thermal field and defining the dynamic relationship between the power input and the temperature gradient in n axial heating sections; a processing unit: calculating the power-temperature gradient mapping relationship based on the finite element analysis equation; constructing a multi-section temperature coupling model of the thermal field includes: based on the finite element analysis method, establishing the dynamic relationship between the power input and the temperature gradient in each heating section, satisfying: G(t)=f(P1(t - τ1), P2(t - τ2),..., P n (t - τ n )) + ε(t); Among them, τ1~τ n are the heat conduction delay times of each heating section, and ε(t) is the environmental noise disturbance term; Agent control module: used to receive the temperatures T1 to T of each heating section n , the crystal pulling speed v, the crucible lifting speed s, and the measured value G of the longitudinal temperature gradient of the thermal field, and output the power adjustment amounts ΔP1 to ΔP n ; Reward calculation module: preset the target temperature gradient G and the safety temperature threshold T th , and perform the reward function operation: ； Strategy optimization module: Simulate the whole process of crystal growth based on the thermal field coupling model, and optimize the parameters of the agent's policy network; Real-time control module: Collect temperature, speed, and gradient data in real time, deploy the trained policy network, generate power adjustment instructions, and send adjustment signals of ΔP1 to ΔP to the temperature control heater; n ; Strategy update module: used to automatically trigger strategy update according to the preset trigger conditions to correct the control parameters.

Images