AI-based thermal field control methods and systems

Abstract

This application provides an AI-based thermal field control method and system, relating to the field of Czochralski crystal growth. The method includes: collecting operational data of the preparation system; determining the target power of multiple heaters based on the operational data and a pre-trained thermal field control model; wherein the thermal field control model is pre-optimized to achieve the goals of oxygen reduction, crystal optimization, and energy saving through a reward function; and controlling multiple heaters based on the target power of the multiple heaters. By introducing a pre-optimized thermal field control model obtained through a multi-objective reward function, this application unifies the mutually restrictive process objectives of oxygen reduction, crystal optimization, and energy saving under a single optimization framework, achieving multi-objective dynamic coordination and adaptive control. This significantly reduces the oxygen content of the crystal, improves the crystallization rate and crystal quality, and reduces system energy consumption.

Description

Technical Field

[0001] This application relates to the field of Czochralski crystal growth, and more specifically, to an AI-based thermal field control method and system. Background Technology

[0002] The Czochralski method, a core process in the semiconductor industry for preparing single-crystal silicon rods, has developed a relatively mature system for thermal field control technology since the 1950s. Modern single-crystal furnaces generally use resistance heating to construct the thermal field, maintaining the temperature gradient of molten silicon by precisely controlling the power distribution of the heaters, thus ensuring the stability of the crystal growth interface. With the increasing demand for large-size silicon wafers (300mm and above) and low-defect-density crystals, the thermal field structure has gradually evolved from the early single-heater configuration to dual-heater and even triple-heater configurations to achieve zoned temperature control of different areas of the quartz crucible.

[0003] In existing technologies, dual-heater systems typically employ a combination of a sidewall main heater and a bottom auxiliary heater, using a PID control algorithm to adjust their power ratio and maintain a relatively stable melt temperature field. To address the processing requirements of larger crucible sizes (e.g., 32 inches and above), some advanced equipment introduces a three-heater vertical layered structure, adding an auxiliary side heater below the main heater to form a "side-main-side-auxiliary-bottom" three-section heating layout. This design can alter the melt convection pattern by adjusting the power of the auxiliary side heater, thereby affecting the distribution of oxygen impurities in the crystal. Furthermore, multi-zone independent temperature control systems based on resistance differences are also used for thermal field optimization, suppressing melt turbulence and improving solid-liquid interface smoothness through differentiated power distribution.

[0004] In terms of control strategies, existing technologies mainly employ PID control with preset parameters, lookup table control based on process curves, or composite control methods combining fuzzy logic. Some systems attempt to introduce expert rule bases to address abnormal operating conditions during the growth process, such as controlling oxygen content through forced cooling of auxiliary heaters. However, these control methods rely on fixed parameter sets or empirical rules, making it difficult to adapt to dynamic changes in heat demand at different crucible loading rates and crystal growth stages (crystal introduction, shoulder formation, constant diameter, and tailing). Especially when dealing with the coupling effect of melt convection and thermal radiation, traditional control methods cannot adjust power distribution in real time to compensate for thermal field disturbances, leading to increased fluctuations in liquid surface temperature and failing to effectively reduce the oxygen content in the crystal. Summary of the Invention

[0005] The purpose of this application is to provide an AI-based thermal field control method and system to achieve active temperature control, significantly reduce the oxygen content in crystals, improve crystallization rate and crystal quality, and reduce system energy consumption.

[0006] Firstly, an AI-based thermal field control method is provided. This method is applied to a single-crystal silicon fabrication system, which includes multiple heaters and a quartz crucible. The multiple heaters are distributed in different regions of the quartz crucible. The method includes: Collect operational data from the preparation system; The target power of multiple heaters is determined based on operational data and a pre-trained thermal field control model. The thermal field control model is pre-optimized to achieve the goals of oxygen reduction, crystal optimization, and energy saving through a reward function. Multiple heaters are controlled based on their target power.

[0007] The above scheme introduces a thermal field control model that is pre-guided by a reward function to optimize multiple objectives such as oxygen reduction, crystal optimization, and energy saving. This unifies the originally mutually restrictive process objectives into the same optimization framework, so that the target power output by the model is no longer the result of a single temperature tracking, but a coupled power allocation that integrates the global optimality of oxygen reduction, crystal optimization, and energy saving, thus realizing multi-objective dynamic coordination and adaptive control.

[0008] Optionally, the operating data includes thermal field parameters and process parameters; the thermal field parameters include the surface temperature of the melt in the quartz crucible; the process parameters include the current time, process stage, crucible rise, crucible rotation, argon flow rate, and actual furnace pressure; the target power of multiple heaters is determined based on the operating data and the pre-trained thermal field control model, including: inputting the current thermal field parameters and the current process parameters into the optimized thermal field control model of the previous time, and outputting the target power of multiple heaters at the current time.

[0009] The above scheme uses gas phase process parameters such as argon flow rate and furnace pressure as model inputs, enabling the control system to have the basis for coordinated regulation of gas-thermal-electric multi-physics fields. Argon flow rate can accelerate the carry-out of oxygen from the melt surface, and furnace pressure affects the partial pressure and diffusion rate of oxygen. Introducing these two into the AI ​​model can actively intervene in the volatilization and diffusion behavior of oxygen, rather than relying solely on passive cooling, thereby blocking the oxygen dissolution path at the source.

[0010] Optionally, the thermal field control model can also be integrated with bidirectional LSTM and graph neural network to not only extract the temporal features of process parameters evolving over time to capture dynamic thermal demand, but also to model the spatial thermal coupling relationship between multiple heaters through graph neural network, thereby accurately predicting oxygen evolution hot spot areas and convection structures, so that the output power distribution can accurately match the thermal field decoupling requirements of different areas.

[0011] Optionally, the operating data also includes: heater power data and crystal quality parameters; the crystal quality parameters include historical batch crystal oxygen content detection values ​​and the current solid-liquid interface shape; the reward function guides the thermal field control model to optimize the θ direction towards the goals of oxygen reduction, crystal optimization, and energy saving through the heater power data, crystal quality parameters, and thermal field parameters.

[0012] The above scheme breaks through the closed-loop blind spot of temperature and mass decoupling in traditional control by incorporating quality parameters such as historical batch oxygen content and real-time solid-liquid interface shape into the reward function's guidance system. This allows the model parameter updates to be directly driven by the real quality feedback of the crystal, enabling proactive prediction and intervention based on quality results.

[0013] Alternatively, by constructing a multi-dimensional reward function that covers temperature deviation, oxygen content, energy consumption, and defect penalty, the three mutually constraining physical objectives of oxygen reduction, crystal optimization, and energy saving can be quantified into calculable reward values. The policy gradient mechanism can then be used to guide the model to evolve toward maximizing the comprehensive reward, effectively solving the dynamic trade-off problem among multiple objectives.

[0014] Optionally, the weighting coefficients can be dynamically adjusted at different growth stages to match the core physical requirements of each stage: maximizing the oxygen reduction weight in the melting stage to suppress oxygen evolution in the quartz crucible, maximizing the temperature stabilization weight in the crystal pulling and shoulder forming stage to maintain the solid-liquid interface, balancing the temperature and oxygen reduction weights in the constant diameter stage to ensure stable oxygen content, and maximizing the anti-breakage weight in the finishing stage to avoid breakage during lifting, thereby achieving refined multi-objective decoupled control of the entire process.

[0015] Optionally, the target power of multiple heaters is determined based on the current operating data and the pre-trained thermal field control model, including: determining the initial target power of multiple heaters based on the current operating data and the pre-trained thermal field control model; determining whether the initial target power of multiple heaters meets the predetermined constraints; if it meets the constraints, the initial target power of multiple heaters is determined as the final target power of multiple heaters; if it does not meet the constraints, the initial target power of multiple heaters is corrected to obtain the final target power of multiple heaters.

[0016] The above solution constructs a safety protection layer by introducing a constraint verification and correction mechanism after the AI ​​model outputs, avoiding the risk of the pure data-driven model outputting out-of-bounds or drastic power changes under extreme operating conditions or data anomalies, and ensuring the physical safety and process continuity of thermal field control.

[0017] Optionally, the constraints include single heater power boundary constraints, power mutation constraints, and stage ratio constraints; single heater power boundary constraints are used to ensure that the target power of a single heater does not exceed the boundary; power mutation constraints are used to ensure that the change in target power does not exceed the maximum allowable step size; stage ratio constraints are used to ensure that each heater adjusts its power according to a preset ratio.

[0018] The above solution limits and smooths the AI ​​output from three dimensions: equipment safety, thermal stability, and process specifications by setting three-dimensional constraints of boundaries, mutations, and proportions. This not only prevents local overheating and hardware damage, but also avoids violent fluctuations in thermal convection caused by power steps, ensuring a smooth transition of the crystal growth interface.

[0019] Optionally, boundary constraints and mutation constraints can be transformed into precisely calculable engineering algorithms through specific truncation functions and signed step size correction formulas, so that any AI output that exceeds the physical limit or step size threshold can be truncated or smoothly transitioned in an instant, completely eliminating process accidents caused by power runaway.

[0020] Secondly, an AI-based thermal field control system is provided. The thermal field control system includes a monocrystalline silicon preparation system, an edge computing device deployed locally on the monocrystalline silicon preparation system, a data acquisition module, and a controller. The monocrystalline silicon preparation system includes multiple heaters and a quartz crucible, with the multiple heaters distributed in different areas of the quartz crucible. The data acquisition module is used to collect operational data from the preparation system. Edge computing devices are used to determine the target power of multiple heaters based on operational data and a pre-trained thermal field control model; the thermal field control model is pre-optimized to achieve the goals of oxygen reduction, crystal optimization, and energy saving through a reward function. A controller used to control multiple heaters based on the target power of multiple heaters.

[0021] The above solution, by deploying edge computing devices locally, brings the AI ​​inference and control closed loop down to the production line side, eliminating the constraints of cloud communication latency on real-time thermal field control, ensuring millisecond-level power response speed, while retaining the model's self-learning and self-evolution capabilities, and achieving a deep integration of intelligent control and real-time execution. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a schematic diagram of the AI-based thermal field control method provided in an embodiment of this application; Figure 2 This is a schematic diagram of the AI-based thermal field control system structure provided in an embodiment of this application; Figure 3 A schematic diagram illustrating the operating principle of the thermal field control model provided in this application embodiment; Figure 4 A comparison diagram of the power timing control curves of the three heaters provided in the embodiments of this application; Figure 5 This is a schematic diagram of the structure of an edge computing device provided in an embodiment of this application. Detailed Implementation

[0024] The technical solutions in the embodiments of this application will now be described with reference to the accompanying drawings.

[0025] Figure 1 This is a schematic diagram of an AI-based thermal field control method provided in an embodiment of this application. The method is applied to a single-crystal silicon preparation system, which includes multiple heaters and a quartz crucible, with the multiple heaters distributed in different areas of the quartz crucible.

[0026] like Figure 1 As shown, the method may include the following steps: Step S110: Collect the operating data of the preparation system.

[0027] Operational data is a digital representation of the multi-physics state during the growth of single-crystal silicon. It encompasses major categories of thermal field parameters reflecting the thermal field distribution and major categories of process parameters reflecting process intervention methods. For example, thermal field parameters may include the surface temperature of the melt in the quartz crucible; process parameters may include the current time, the current process stage, crucible rise, crucible rotation, argon flow rate, and actual furnace pressure. It should be understood that the above examples are illustrative and not restrictive. In actual production environments, operational data may also include heater power data and crystal quality parameters; these crystal quality parameters include historical batch oxygen content measurements and the current solid-liquid interface shape.

[0028] Step S120: Determine the target power of multiple heaters based on the operating data and the pre-trained thermal field control model; wherein, the thermal field control model is pre-optimized to achieve the goals of oxygen reduction, crystal optimization, and energy saving through a reward function.

[0029] Specifically, the core of this embodiment lies in establishing a top-level logical closed loop from passive temperature control to active optimization control. Traditional single-objective PID control or empirical lookup table methods have a single optimization objective, usually only tracking the stability of the liquid surface temperature. This local optimum often leads to the loss of control over other process indicators. For example, running the main heater at high load for a long time to maintain the temperature not only causes local overheating and energy waste, but also exacerbates the oxygen evolution reaction in the quartz crucible, causing the oxygen content to soar. The thermal field control model introduced in this embodiment unifies the three physically interdependent process objectives of oxygen reduction, crystal optimization, and energy saving into the same optimization framework through a reward function. This makes the target power output of the model no longer a derivative of single temperature tracking, but an optimal power allocation strategy that integrates the global multi-physics coupling effect. This enables the control system to perform dynamic trade-offs and adaptive optimization in complex thermal convection and oxygen evolution dynamics environments.

[0030] Step S130: Control the multiple heaters based on the target power of the multiple heaters.

[0031] Specifically, the controller converts the target power into actual physical control signals, such as adjusting the output current or voltage of the power supply to each heater, thereby accurately executing the power coupling allocation scheme given by the thermal field control model.

[0032] The phrase "multiple heaters distributed in different areas of the quartz crucible" in this embodiment can refer to a three-heater structure or a dual-heater system. For example, the preparation system can be a traditional layout consisting only of a side main heater and a bottom heater; or it can be a multi-zone independent temperature control system, for example, dividing the side wall heaters into upper, middle, and lower zones, or even more micro-zones, each with an independent power adjustment module. These heaters are spatially distributed in different heights or radial regions of the quartz crucible, enabling stratified intervention of the melt temperature gradient and convection structure. This distributed spatial layout is the physical basis for achieving power coupling decoupling control and precise suppression of local oxygen evolution hotspots.

[0033] Through the above scheme, this embodiment utilizes a thermal field control model pre-optimized by a multi-objective reward function to transform the originally mutually restrictive demands for oxygen reduction, crystal optimization, and energy saving into calculable optimization directions, achieving a leap from passively tracking a single temperature to actively coordinating multiple physical objectives.

[0034] In some embodiments, traditional control methods often treat the liquid surface temperature T as a single input variable, attempting to passively track the target temperature by adjusting the heating power. However, the oxygen content fluctuation during the preparation of single-crystal silicon is not solely determined by the thermal field temperature. The argon flow rate F directly affects the oxygen carrying efficiency from the melt surface. High argon flow rate can accelerate the removal of volatilized silica gas from the melt surface from the furnace, reducing the residence time and re-dissolution probability of oxygen in the melt. The actual furnace pressure P, according to Henry's Law, directly affects the solubility and partial pressure of oxygen in the melt. Reducing the furnace pressure can actively reduce the oxygen partial pressure, promoting oxygen diffusion from the melt to the gas phase, thereby inhibiting oxygen dissolution at the source. If only temperature control is relied upon, even if the liquid surface temperature is stable, under high furnace pressure or low argon flow rate, a large amount of silica thermal decomposition products on the quartz crucible wall will still remain in the melt, leading to uncontrolled oxygen content. Therefore, this embodiment uses the argon flow rate F, actual furnace pressure P, and liquid surface temperature T as synchronized model inputs, enabling the control system to have the foundation for coordinated regulation of multiple physical fields including gas, heat, and electricity. This allows for proactive prediction and intervention of the entire oxygen evolution, convection, and volatilization process, rather than relying solely on passive cooling. It should be understood that although this embodiment lists the specific parameters mentioned above, in other embodiments, the process parameters can be further extended to any parameters that can characterize the gas phase or mechanical intervention methods, such as crystal pulling speed and cooling water flow rate, as long as these parameters can provide physical state inputs for multi-objective optimization.

[0035] Furthermore, based on the above operating data, the step of determining the target power of multiple heaters based on the operating data and the pre-trained thermal field control model may specifically include: inputting the current thermal field parameters and the current process parameters into the optimized thermal field control model of the previous moment, and outputting the target power of multiple heaters at the current moment.

[0036] Regarding the internal architecture of the thermal field control model, combined with Figure 3 As shown, the thermal field control model is determined according to the following formula: ; in, , The feature vector input at time t; For the current moment, For the process stage, This is the actual furnace pressure. The liquid surface temperature, For the pot rise, For the pot to turn, Argon flow rate; For neural network parameters The defined deep learning model is a nonlinear function, including bidirectional LSTM networks and graph neural networks, and its internal dynamics can be approximated as follows: ;in, Let t be the target power of the multiple heaters output at time t; Here is the weight matrix of the graph neural network; For bias terms; Used to normalize the output, ensuring a reasonable distribution of target power among multiple heaters; It is a bidirectional LSTM network used for extraction. The temporal characteristics; include and .

[0037] Specifically, the growth process of single-crystal silicon is a physical process with strong time dependence and delay effects. Bidirectional LSTM, by simultaneously extracting historical and future trend time-series features, can anticipate, for example, the temperature decrease trend during the transition from the melting stage to the crystal-leading stage, thus achieving smooth power prediction and transition. Simultaneously, this embodiment introduces a graph neural network weight matrix W to model the spatial thermal coupling relationship between heaters. Fully connected layers typically treat all heaters as independent nodes, neglecting the local heat conduction and radiation relationships of the heaters in the quartz crucible space. However, the graph neural network constructs topological connections between heaters through the weight matrix W, allowing the heat conduction and convection coupling relationships between the main heater, auxiliary heater, and bottom heater to be explicitly expressed at the algorithm level, thereby accurately predicting local oxygen evolution hotspot regions and convection structures.

[0038] Furthermore, the Softmax normalization operation ensures a reasonable allocation of target power among multiple heaters, rather than independent output. Since each heater physically shares the same thermal field space, the total heat input must match the heat demand of the current process stage. Softmax transforms the original score output by the model into a probabilistically distributed power allocation, allowing the power increase or decrease of each heater to be coupled and adjusted under the global total power constraint, thus avoiding local overheating and energy waste caused by blindly increasing the power of a single heater.

[0039] Finally, to link the optimization direction of the above model architecture with the actual crystal quality, the reward function guides the θ-axis optimization of the thermal field control model towards the goals of oxygen reduction, crystal optimization, and energy saving through heater power data, crystal quality parameters, and thermal field parameters.

[0040] Specifically, traditional control loops often form local loops between temperature tracking and power output, resulting in a disconnect between temperature and final crystal quality. This leads to seemingly stable temperatures but substandard oxygen content or interface shape. This embodiment breaks this loop blind spot by incorporating historical batch oxygen content detection values ​​and current solid-liquid interface shape, among other real quality parameters, into the reward function's guiding source. Although historical oxygen content detection values ​​are not real-time back-end quality inspection data, they reflect the long-term oxygen reduction effect under specific power ratios and gas phase parameter combinations, providing the model with a global optimization direction across batches. The solid-liquid interface shape reflects the degree of disturbance of the crystal growth interface by current thermal convection and temperature gradient. By guiding the optimization of θ together with these quality parameters, power data, and thermal field parameters, the model parameter updates are no longer driven solely by temperature deviations but directly by real crystal quality feedback. This enables proactive prediction and intervention based on quality results, ensuring the true implementation of multiple objectives such as oxygen reduction, crystal optimization, and energy saving. It should be understood that crystal quality parameters can also include any parameters that can characterize the intrinsic quality of the crystal, such as dislocation density and resistivity. This embodiment is for illustrative purposes only and not restrictive.

[0041] As an example, the reward function is defined as: ; in, Let be the value of the reward function at time t; : Let t be the current liquid surface temperature. The target liquid surface temperature; The maximum permissible temperature deviation threshold is preset. The predicted oxygen content at time t is determined based on the historical batches of crystal oxygen content detection values; This represents the upper limit of oxygen content. The energy consumption per unit time at time t is determined based on the power data of the heater at time t. The preset energy consumption benchmark value per unit time; The penalty for a broken wire or broken section event is determined based on the shape of the solid-liquid interface. , where is the weight coefficient of each term in the reward function; through For the thermal field control model To reward or punish, the thermal field control model should be adjusted in the next time step. Prefers output that can produce high Target power .

[0042] Specifically, this embodiment quantifies the three interdependent physical objectives of oxygen reduction, crystal optimization, and energy conservation into calculable reward values ​​by constructing a multi-dimensional reward function encompassing temperature deviation, oxygen content, energy consumption, and defect penalties. To make this quantification system operable, specific physical boundary examples must be provided for each parameter. For example, The maximum permissible temperature deviation threshold represents the deviation of the liquid surface temperature from the target value during the production process. The acceptable upper limit can be set according to the sensitivity of different process stages. When the deviation exceeds this threshold, the bonus value of the temperature term will approach zero or even be negative, thus severely punishing the power output strategy that causes temperature runaway. This serves as the upper limit for oxygen content; it is a red line that must never be exceeded in production. Once the oxygen content is predicted... As the model approaches this value, the reward incentive for the oxygen reduction term will drop sharply, forcing it to seek a more aggressive oxygen reduction power ratio. As a benchmark for energy consumption per unit time, it serves as a target energy consumption benchmark for process optimization, encouraging models to explore energy-saving potential for low-power operation while meeting temperature and oxygen reduction requirements. As a penalty term for line breakage or fracture events, its value is determined by the degree to which the solid-liquid interface shape deviates from flatness. When the interface shape is severely distorted, resulting in an extremely high risk of fracture, Dt is set to 1; when the interface is flat and safe, it is set to 0. This 0-1 penalty mechanism directly links catastrophic process accidents with negative rewards, causing the model to have an absolute tendency to avoid the risk of line breakage when optimizing parameter θ. It should be understood that the specific values ​​mentioned above are only interpretive and not restrictive, as long as these thresholds can reflect the process constraint boundaries of the current furnace.

[0043] Furthermore, the weighting coefficients of each term in the reward function are subject to differentiated control strategies at different growth stages: the weighting strategy for the melting stage is as follows: The weighting strategy for the crystal introduction and shoulder formation stages is as follows: The weighting strategy for the equal diameter growth stage is as follows: The weighting strategy for the closing phase is as follows: .

[0044] Specifically, during the melting stage, the core challenge is the severe oxygen evolution in the quartz crucible. At this point, the melt is in extensive contact with the crucible wall, and if an incorrect weighting distribution is used, such as... To achieve absolute temperature stability, the model tends to maintain high power output of the main heater to prevent the liquid surface temperature from dropping. However, this results in the crucible wall temperature remaining in a high-temperature range for an extended period, leading to uncontrolled thermal decomposition of silica and a surge in oxygen content. Even if the temperature curve appears stable, a large amount of oxygen impurities have dissolved into the melt, creating a potential problem for subsequent crystal growth. Therefore, it is necessary to maximize the oxygen reduction weight. This allows the model to proactively increase the auxiliary heater power and reduce the main heater load, even with slight deviations from the target temperature, thus keeping the crucible wall temperature within a safe range and suppressing oxygen evolution at its source. During the crystal pulling and shoulder formation stages, the core challenge shifts to maintaining extreme stability at the solid-liquid interface to prevent wire breakage. At this stage, the crystal is newly formed and extremely sensitive to thermal shock; if high temperatures are incorrectly maintained... The model may frequently fine-tune the power ratio to further reduce oxygen, leading to disturbances in the thermal convection structure and causing interface jitter, which can easily trigger dislocations and wire breakage accidents; therefore, temperature stability is a crucial factor at this stage. Defect prevention weight With this becoming the dominant factor, the model's primary task becomes smoothing power output to lock the interface shape, while oxygen reduction and energy saving become secondary. During the constant-diameter growth stage, the crystal constitutes the majority of the length, requiring a balance between temperature stability and oxygen content control. Much larger This can lead to a situation where, although the oxygen content in the head is low, the oxygen content in the tail rebounds due to insufficient compensation. Much larger Then the diameter fluctuations become uncontrollable, therefore ≈ The balancing strategy is the only solution for maintaining stable long-term growth. In the final stage, preventing crystal tail pull-out and breakage is the absolute priority; if at this point... or The model, in its pursuit of maintaining a final degree of temperature stability or reducing oxygen, might overlook the thermal stress concentration caused by the rapid shrinkage of the interface, leading to a fracture accident that renders the entire crystal rod unusable. Therefore... The weight must be much greater than other weights, and the model will sacrifice temperature accuracy and energy consumption indicators to fully implement the power contraction strategy to prevent lift-off.

[0045] Combination Figure 4The comparison chart of the power timing control curves of the three heaters clearly shows that, guided by the aforementioned differentiated weighting strategy, the power curve under AI control exhibits a smooth transition during stage switching. When the model senses an impending stage switch, it predicts the direction of the weighting strategy shift using a bidirectional LSTM, allowing for advance power fine-tuning. In contrast, traditional PID control, lacking a multi-objective weighting mechanism, relies solely on manual intervention or rigid table lookups at stage switching points, resulting in significant jumps and violent fluctuations in the power curve. This abrupt change is the physical root cause of severe thermal convection disturbances and solid-liquid interface collapse. This embodiment demonstrates the physical inevitability of weight changes with stage and contrasts the process accidents caused by incorrect weight allocation, establishing a robust defense and explaining why this differentiated strategy cannot be easily replaced by other alternatives. The above weighting strategies and parameter values ​​are interpretive rather than restrictive. Those skilled in the art can make adaptive adjustments based on specific furnace types and product specifications, as long as the principle of prioritizing the core physical pain points of each stage is followed.

[0046] After determining the target power step, this embodiment further introduces a constraint verification and correction sub-process for the target power to establish a security protection layer for the data-driven model.

[0047] Specifically, although the thermal field control model possesses multi-objective optimization capabilities guided by the reward function, under extreme operating conditions or abnormal input data, the purely data-driven AI model still carries the risk of outputting out-of-bounds or drastically changing power. For example, when sensor malfunctions lead to missing or distorted liquid surface temperature data, the model may output power commands exceeding the physical limits of the heater in an attempt to forcibly compensate for the temperature, resulting in the burnout of the graphite heater or overload damage to the power module. Similarly, during process stage switching, the model may output a step-like drastic power change, causing violent fluctuations in the thermal convection structure within the thermal field, potentially leading to solid-liquid interface collapse and wire breakage accidents. Therefore, it is necessary to introduce a constraint verification mechanism based on physical mechanisms and engineering safety specifications to provide a secondary safety net for the AI ​​model's output.

[0048] Based on this, the target power of multiple heaters is determined based on the current operating data and the pre-trained thermal field control model, including: determining the initial target power of multiple heaters based on the current operating data and the pre-trained thermal field control model; determining whether the initial target power of multiple heaters meets the predetermined constraints; if it meets the constraints, the initial target power of multiple heaters is determined as the final target power of multiple heaters; if it does not meet the constraints, the initial target power of multiple heaters is corrected to obtain the final target power of multiple heaters.

[0049] Regarding the specific composition of the above constraints, the constraints include single heater power boundary constraints, power mutation constraints, and stage ratio constraints; single heater power boundary constraints are used to ensure that the target power of a single heater does not exceed the boundary; power mutation constraints are used to ensure that the change in target power does not exceed the maximum allowable step size; stage ratio constraints are used to ensure that each heater adjusts its power according to a preset ratio.

[0050] These three-dimensional constraints constitute a comprehensive defense system covering equipment safety, thermal field stability, and process specifications. The specific engineering algorithms for each constraint will be detailed below.

[0051] First, the power boundary constraints for a single heater are determined based on the following formula: ; in, For the first One heater at time The initial target power, , For the number of heaters, To constrain the minimum value of a single heater, To constrain the maximum value of a single heater.

[0052] For example, in a three-heater system of a 12-inch single crystal furnace, 20kW can be selected, which is the safety limit for maintaining the minimum heating life of the heater graphite body and preventing cold shock. A power rating of 80kW is acceptable; this is the physical upper limit of the power module's rated current and the graphite's oxidation-resistant fuse. This is based on the initial target power output of the model. =85kW, which is excessive. If the boundary is not satisfied, direct execution will directly cause hardware damage. Therefore, if the boundary constraint is not met, Pt,i will be truncated to the boundary value: ;in, For the first One heater at time The final target power; This is a truncation operation.

[0053] By using the clipping operation, the aforementioned 85kW exceeding power limits will be forcibly reduced to 80kW, ensuring that the safe operating boundary of the equipment is not breached. It should be understood that 20kW and 80kW are merely illustrative examples; other values ​​can be used for furnace types of different power ratings, as long as these thresholds reflect the physical limits of the current hardware system.

[0054] Secondly, the power mutation constraint is determined based on the following formula: If the threshold is exceeded, the adjustment will be made according to the maximum allowable step size. ; in, For the first One heater at time The final target power; The threshold for power variation; To determine the direction of power adjustment, i.e., to decide the first... A heater in Whether the temperature is rising or falling.

[0055] Specifically, a thermal field system is a physical object with large inertia and thermal delay. A step-like change in power will induce violent transient convection vortices within the melt. These vortices can directly distort the solid-liquid interface shape and may break the newly formed crystal rod. For example, A step size of 2kW can be chosen, representing the maximum power change that the thermal field can absorb within a single control cycle. If the final power at the previous moment... The initial target power Pt,i is 55kW, while the current model outputs 50kW. The change of 5kW exceeds the 2kW threshold, so a step size correction is necessary. In this case, sign(55-50) is positive, indicating that the adjustment direction is towards increasing temperature, and the final target power after correction is... =50 + 1·2 = 52kW; conversely, if The power is 45kW, and sign(45-50) is negative, indicating that the adjustment direction is cooling. (The corrected value is...) =50-1·2=48kW. The sign of the sign function directly determines whether the heater physically increases current to raise the temperature or decreases current to lower the temperature. Through this signed maximum allowable step size correction, the originally steep power jump is smoothed into a gradual, ramp-like transition, completely eliminating the violent fluctuations in heat convection caused by sudden power changes. It should be understood that... =2kW / step is just an example; in systems with higher sampling frequencies or greater thermal inertia, Alternatively, you can choose 1kW / step or 5kW / step, as long as the threshold can match the transient absorption capacity of the current thermal field.

[0056] Finally, the stage proportioning constraint is used to ensure that each heater adjusts its power according to a pre-set ratio. For example, in the constant diameter growth stage, the process specification may require that the power ratio of the main heater, auxiliary heater, and bottom heater be maintained within a preset range of 6:3:1 to maintain a specific axial temperature gradient and convection structure. If the AI ​​model outputs an illegal ratio of 7:2:1 due to local data disturbances, although the individual power values ​​may not exceed the boundary and mutation threshold, the proportioning relationship disrupts the thermal field symmetry defined by the process specification. In this case, the stage proportioning constraint will redistribute the total power according to the preset target ratio of 6:3:1, forcibly correcting the illegal initial target power to the final target power that conforms to the process specification, thereby ensuring the consistency of crystal diameter during the long-cycle growth process.

[0057] Through the aforementioned three-dimensional correction mechanism of boundary truncation, step size smoothing, and ratio redistribution, this embodiment superimposes a deterministic safety protection layer on top of the AI ​​model's proactive control capabilities. This ensures that any AI output exceeding physical limits, thermal field inertia, or process specifications can be intercepted and corrected in real time, guaranteeing a smooth transition of the crystal growth interface and absolute safety of the process loop. The above constraint parameters and correction algorithms are interpretive rather than restrictive. Those skilled in the art can make adaptive adjustments based on the hardware parameters and process specifications of specific furnace types, as long as the three-dimensional defense principles of equipment safety, thermal field stability, and process specifications are followed.

[0058] Figure 2 As shown, this embodiment provides an AI-based thermal field control system, a monocrystalline silicon preparation system, an edge computing device 202 deployed locally on the monocrystalline silicon preparation system, a data acquisition module 201, and a controller 203. The monocrystalline silicon preparation system includes multiple heaters and a quartz crucible, with the multiple heaters distributed in different areas of the quartz crucible. Acquisition module 201 is used to acquire the operating data of the preparation system; Edge computing device 202 is used to determine the target power of multiple heaters based on operating data and a pre-trained thermal field control model; wherein the thermal field control model is pre-optimized to the goals of oxygen reduction, crystal optimization, and energy saving by a reward function. Controller 203 is used to control multiple heaters based on the target power of multiple heaters.

[0059] Specifically, this embodiment provides a physical foundation for the timing logic of the aforementioned method embodiments from the perspectives of hardware architecture and spatial connectivity. Combined with... Figure 2As shown, the core thermal field hardware structure of the single-crystal silicon preparation system includes a quartz crucible for holding silicon material and multiple heaters distributed in different areas of the quartz crucible. In a typical three-heater layout, the main heater is located on the upper side wall of the quartz crucible and is the main heat source input, responsible for maintaining the basic temperature field of the melt surface; the auxiliary heater is located below the main heater, on the middle side of the quartz crucible, and is used to adjust the temperature gradient of the melt in the middle of the quartz crucible. This is a key physical intervention point for suppressing local oxygen evolution hotspots on the crucible wall and interfering with the melt convection structure; the bottom heater is located at the bottom of the quartz crucible and is used to adjust the axial temperature gradient, accelerate the bottom melting efficiency, and compensate for bottom heat loss. It should be understood that although... Figure 2 The illustration shows a three-section vertical layered layout of main side, auxiliary side, and bottom. However, in other embodiments, the multiple heaters can also be multi-zone independently temperature-controlled heating bands distributed around the quartz crucible. As long as these heaters can spatially form a layered intervention capability for the temperature gradient of different regions of the melt, they fall within the protection scope of this application. This distributed spatial layout is not only a prerequisite for achieving power decoupling control, but also the physical root cause of the strong local heat conduction and radiation relationship between the heaters. This spatial thermal coupling relationship is precisely the physical object modeled by the graph neural network weight matrix W in the aforementioned embodiments.

[0060] The data acquisition module 201 can consist of a sensor group composed of various sensors, including but not limited to a non-contact visual brightness and temperature sensor for acquiring the surface temperature of the molten metal in the quartz crucible, a pressure sensor for acquiring the actual furnace pressure, a flow meter for acquiring the argon flow rate, and an encoder for acquiring mechanical parameters such as crucible lifting and rotation. These sensors maintain real-time data connectivity with edge computing devices via industrial fieldbus or high-speed Ethernet, ensuring that multi-source heterogeneous operating data can be synchronously input to the control core with millisecond-level latency.

[0061] Specifically, in this embodiment, the edge computing device 202 is physically deployed locally on the monocrystalline silicon preparation system, or a cloud server architecture can be used. For example, an industrial edge server installed next to the monocrystalline furnace runs the thermal field control model detailed in the previous embodiment. The end-to-end latency from data acquisition module to model inference to command output is compressed to milliseconds or even microseconds, completely eliminating the constraints of cloud communication latency and ensuring the real-time response of the thermal field control model to the multi-objective reward function and millisecond-level power closed-loop execution. After receiving the current operating data, the edge computing device immediately performs feature extraction and coupled inference using its internal bidirectional LSTM and graph neural network, and outputs the target power ratio of multiple heaters at the current moment through Softmax normalization. Subsequently, this target power data is directly transmitted to the controller through the local high-speed bus.

[0062] The controller 203 is the core hardware module that performs the physical power conversion, typically a high-precision thyristor power regulator or pulse-width modulation power module. After receiving the target power command from the edge computing device, the controller converts it into actual physical current or voltage output signals, which are then applied to the power supply circuits of the main heater, auxiliary heater, and bottom heater, respectively. For example, when the edge computing device determines that the current strategy of prioritizing oxygen reduction is to increase the power of the auxiliary heater and reduce the load on the main heater, the controller will correspondingly increase the output current of the auxiliary heater circuit while decreasing the output current of the main heater circuit. Through this differentiated allocation of physical current, the temperature gradient and convection structure within the quartz crucible space are reconstructed, thereby suppressing the crucible wall temperature within a safe range that inhibits the thermal decomposition of silica. It should be understood that the controller can employ not only current regulation mode, but also voltage regulation mode or phase regulation mode in other embodiments, as long as it can accurately execute the physical output of the target power.

[0063] Through the above system architecture, this embodiment brings the AI ​​inference and control closed loop down to the production line side, constructing a millisecond-level local closed loop from sensor data acquisition and real-time inference of the edge model to physical execution by the controller. This not only preserves the model's self-learning and self-evolution capabilities based on a multi-objective reward function, but also achieves deep integration of intelligent control and real-time execution, completely eliminating the risk of thermal runaway caused by cloud latency. This provides solid hardware support for multi-objective collaborative optimization of oxygen reduction, crystal quality, and energy saving in monocrystalline silicon. The above hardware layout and connection relationships are illustrative and not restrictive. Those skilled in the art can make adaptive adjustments according to the specific furnace type and site conditions, as long as the principles of local deployment to eliminate latency and multi-zone, layered intervention in the thermal field are followed.

[0064] Figure 5 This is a schematic diagram of the structure of an edge computing device provided in an embodiment of this application. The edge computing device includes: a processor 50, a memory 51, a bus 52 and a communication interface 53. The processor 50, the communication interface 53 and the memory 51 are connected through the bus 52. The processor 50 is used to execute executable modules, such as computer programs, stored in the memory 51.

[0065] The memory 51 may include high-speed random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 33 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc.

[0066] The bus 52 can be an ISA bus, PCI bus, or EISA bus, etc. Buses can be divided into address buses, data buses, control buses, etc. For ease of representation, Figure 3 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0067] The memory 51 is used to store the program. After receiving the execution instruction, the processor 50 executes the program. The method executed by the device for the flow process definition disclosed in any of the foregoing embodiments of this application can be applied to the processor 50 or implemented by the processor 50.

[0068] The processor 50 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of the processor 50 or by instructions in software form. The processor 50 may be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it may also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly manifested as execution by a hardware decoding processor, or execution by a combination of hardware and software modules in the decoding processor. The software module can reside in a mature storage medium in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory 51, and the processor 50 reads the information from memory 51 and, in conjunction with its hardware, completes the steps of the above method.

[0069] The computer program product of the readable storage medium provided in the embodiments of this application includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods in the foregoing method embodiments. For specific implementation, please refer to the foregoing method embodiments, which will not be repeated here.

[0070] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0071] Finally, it should be noted that the above embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The protection scope of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the technical scope disclosed in this application. Such modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be covered within the protection scope of this application. Therefore, the protection scope of this application should be determined by the protection scope of the claims.

Claims

1. An AI-based thermal field control method, characterized by, The method is applied to a single-crystal silicon fabrication system, the fabrication system including multiple heaters and a quartz crucible, the multiple heaters being distributed in different regions of the quartz crucible, the method comprising: Collect the operational data of the preparation system; The target power of multiple heaters is determined based on the operational data and the pre-trained thermal field control model; wherein, the thermal field control model is pre-optimized to achieve the goals of oxygen reduction, crystal optimization, and energy saving through a reward function. The multiple heaters are controlled based on the target power of the multiple heaters.

2. The method of claim 1, wherein, The operating data includes thermal field parameters and process parameters; the thermal field parameters include the surface temperature of the melt in the quartz crucible; the process parameters include the current time, process stage, crucible rise, crucible rotation, argon flow rate, and actual furnace pressure. The determination of the target power of multiple heaters based on the operational data and a pre-trained thermal field control model includes: The current thermal field parameters and the current process parameters are input into the optimized thermal field control model from the previous time step, and the target power of the multiple heaters at the current time step is output.

3. The method of claim 2, wherein, The thermal field control model is determined according to the following formula: ； wherein, , is the input feature vector at time t; is the current time, is the process phase, is the actual furnace pressure, is the liquid surface temperature, is the ladle lift, is the ladle rotation, is the argon flow rate; For neural network parameters The defined deep learning model is a nonlinear function, including bidirectional LSTM networks and graph neural networks, and its internal dynamics can be approximated as follows: ； in, The target power of the multiple heaters output at time t; Here is the weight matrix of the graph neural network; For bias terms; Used to normalize the output, ensuring a reasonable distribution of target power among the multiple heaters; It is a bidirectional LSTM network used for extraction. The temporal characteristics; include and .

4. The method according to claim 3, characterized in that, The operational data also includes: Heater power data and crystal quality parameters; the crystal quality parameters include historical batch crystal oxygen content detection values ​​and current solid-liquid interface shape; The reward function guides the thermal field control model using the heater's power data, the crystal quality parameters, and the thermal field parameters. Optimize towards the goals of reducing oxygen, improving crystal quality, and saving energy.

5. The method according to claim 4, characterized in that, The reward function is defined as follows: ； in, Let be the value of the reward function at time t; : Let t be the current liquid surface temperature. The target liquid surface temperature; The maximum permissible temperature deviation threshold is preset. The predicted oxygen content at time t is determined based on the historical batches of crystal oxygen content detection values; This represents the upper limit of oxygen content. The energy consumption per unit time at time t is determined based on the power data of the heater at time t. The preset energy consumption benchmark value per unit time; The penalty for a broken wire or broken section event is determined based on the shape of the solid-liquid interface. , where is the weight coefficient of each term in the reward function; Through the above For the thermal field control model Rewards or penalties are applied to make the thermal field control model function in the next time step. Prefers output that can produce high Target power .

6. The method according to claim 5, characterized in that, in, The weight coefficients of each term in the reward function are subject to differentiated control strategies at different growth stages: The weighting strategy for the melting phase is as follows: ; The weighting strategy for the crystal pulling and shoulder formation stages is as follows: ; The weighting strategy for the equal diameter growth stage is as follows: ; The weighting strategy for the closing phase is as follows: .

7. The method according to claim 1, characterized in that, Based on the current operating data and a pre-trained thermal field control model, the target power of multiple heaters is determined, including: The initial target power of the multiple heaters is determined based on the current operating data and the pre-trained thermal field control model. Determine whether the initial target power of the plurality of heaters meets the predetermined constraints; If the conditions are met, the initial target power of the plurality of heaters is determined as the final target power of the plurality of heaters; If the target power is not met, the initial target power of the multiple heaters is corrected to obtain the final target power of the multiple heaters.

8. The method according to claim 7, characterized in that, The constraints include single heater power boundary constraints, power mutation constraints, and stage ratio constraints; the single heater power boundary constraints are used to ensure that the target power of a single heater does not exceed the boundary; the power mutation constraints are used to ensure that the change in target power does not exceed the maximum allowable step size; the stage ratio constraints are used to ensure that each heater adjusts its power according to a preset ratio.

9. The method according to claim 8, characterized in that, The power boundary constraints for a single heater are determined based on the following formula: ； in, For the first One heater at time The initial target power, , For the number of heaters, To constrain the minimum value of a single heater, To constrain the maximum value of a single heater; If not satisfied, then Truncate to boundary value: ； in, For the first One heater at time The final target power; For truncation operation; Power mutation constraints are determined based on the following formula: ； If the threshold is exceeded, the adjustment will be made according to the maximum allowed step size: ； in, For the first One heater at time The final target power; The threshold for power variation; To determine the direction of power adjustment, i.e., to decide the first... A heater in Whether the temperature is rising or falling.

10. An AI-based thermal field control system, characterized in that, The thermal field control system includes a monocrystalline silicon preparation system, an edge computing device deployed locally on the monocrystalline silicon preparation system, a data acquisition module, and a controller. The monocrystalline silicon preparation system includes multiple heaters and a quartz crucible, with the multiple heaters distributed in different areas of the quartz crucible. The acquisition module is used to acquire the operating data of the preparation system; The edge computing device is used to determine the target power of multiple heaters based on the operating data and a pre-trained thermal field control model; wherein the thermal field control model is pre-optimized to achieve the goals of oxygen reduction, crystal optimization, and energy saving through a reward function. The controller is used to control the plurality of heaters based on the target power of the plurality of heaters.

Images