Multi-chamber vertical furnace for heat treatment in semiconductor industry

Abstract

The invention relates to the field of semiconductor manufacturing equipment, in particular to a multi-chamber vertical furnace for heat treatment in the semiconductor industry. The conveying mechanism comprises a bracket capable of rotating around a central shaft, a movement mechanism for driving the bracket to lift and a plurality of furnace doors corresponding to the process cavities in number, and each furnace door is used for bearing one wafer boat; the conveying mechanism rotates the furnace door bearing the wafer boat to the position below the target process cavity, the furnace door is driven to ascend to be in butt joint with the target process cavity, a closed processing cavity is formed, after the process is completed, the furnace door descends and rotates to the next cavity or the unloading position, and seamless transfer of the wafer boat among different process environments is achieved; the process intelligent scheduling system dynamically allocates wafers to a target process cavity based on a reinforcement learning algorithm; the gas pollution prevention and control system adopts a neural network model to predict the residual gas concentration, and automatically triggers a cleaning process when the residual gas concentration exceeds a preset threshold value. The production efficiency and the product quality are improved.

Description

Technical Field

[0001] This invention relates to the field of semiconductor manufacturing equipment, and more particularly to a multi-chamber vertical furnace for heat treatment in the semiconductor industry. Background Technology

[0002] In semiconductor manufacturing processes, vertical furnaces are widely used as key heat treatment equipment in wafer oxidation, diffusion, annealing, and other processes. As semiconductor device feature sizes continue to shrink and process complexity continues to increase, higher demands are being placed on the precision control, production efficiency, and pollution prevention of heat treatment equipment.

[0003] Existing vertical furnace technology mainly adopts a single-chamber structure, using a lifting mechanism to send the wafer-carrying boat into the reaction chamber for heat treatment. Chinese patent CN103673582B discloses a method for controlling the temperature of the loading area during the boat lowering process in a vertical furnace, reducing particle adhesion and wafer contamination through segmented boat lowering control. Chinese patent CN118571775A discloses a furnace tube process control method, improving the problem of uneven temperature distribution within the furnace tube through temperature pre-adjustment. Chinese patent CN117198931A proposes a vertical reaction furnace device that uses a partition wall to allow two furnace bodies to share a single wafer carrier processing unit. Chinese patent CN114360997B discloses a multi-chamber cleaning method, using a gas detection device to cyclically monitor the cleaning status of each chamber. Chinese patent CN220169915U proposes a vertical furnace with an auxiliary furnace door structure, achieving temperature control by sealing the process tubes through the auxiliary furnace door.

[0004] However, existing technologies still have the following technical shortcomings: First, multi-device cluster solutions pose a risk of interface contamination during wafer transfer. Perfect vacuum locking cannot be achieved when wafers are transferred between different process equipment. Chamber disruptions and robotic arm movement during transfer can cause instantaneous contact between the wafer surface and oxygen- and water-containing molecular environments, forming natural oxide layers or introducing impurities at critical interfaces, severely degrading the electrical performance of the device. Simultaneously, each independent process equipment requires independent preheating, processing, cleaning, and cooling cycles, accumulating significant non-value-adding time and resulting in low production efficiency. Second, single-device multi-step process solutions suffer from process temperature conflicts and gas cross-contamination. Traditional vertical furnaces have slow heating and cooling processes, making it impossible to complete contradictory temperature requirements in the same process, leading to uncontrolled thermal budgets. When switching between different chemical reaction gases, precursor and byproduct residues are difficult to completely remove, causing cross-contamination and affecting film purity and electrical performance. Furthermore, existing technologies lack intelligent process scheduling and contamination prediction mechanisms, failing to dynamically optimize the production process based on real-time process status and unable to predict and control gas contamination risks in advance. Summary of the Invention

[0005] To address the issues of interface contamination risk and low production efficiency in existing semiconductor vertical furnace technologies, the process temperature conflict and gas cross-contamination problems in single-equipment multi-step process schemes, and the irreconcilable contradiction between efficiency, compatibility, and control precision caused by the inherent "one chamber, one state" model, this invention provides a multi-chamber vertical furnace for heat treatment in the semiconductor industry. This aims to improve interface quality and device performance, enhance process compatibility and integration capabilities, increase production efficiency and flexibility, and achieve precise thermal budgeting and doping control.

[0006] This invention provides a multi-chamber vertical furnace for heat treatment in the semiconductor industry, comprising: multiple independent process chambers, each equipped with an independent temperature control system and atmosphere control system; a transfer mechanism, including a bracket rotatable around a central axis, a motion mechanism for driving the bracket to rise and fall, and multiple furnace doors disposed on the bracket corresponding to the number of process chambers, each furnace door for carrying a wafer boat; wherein, the transfer mechanism is configured to: rotate the furnace door carrying the wafer boat to below the target process chamber, and drive the furnace door to rise and dock with the corresponding target process chamber, locking the furnace door to form a sealed processing chamber; after completing the target process, drive the furnace door to fall and rotate to the next process chamber or unloading position, realizing seamless transfer of the wafer boat between different process environments, while the wafer remains in a controlled environment and does not need to be exposed to the atmosphere; a process intelligent scheduling system for dynamically allocating wafers to target process chambers based on reinforcement learning algorithms; and a gas pollution control system for predicting the residual gas concentration of the next process cycle using an LSTM neural network model, and automatically triggering a cleaning process when the predicted value exceeds a preset threshold.

[0007] Optionally, the state space of the intelligent process scheduling system consists of the following parameters: real-time temperature of each process chamber, process gas concentration in each process chamber, three-dimensional position coordinates of the crystal boat, priority of the batch of wafers to be processed, remaining process time of each process chamber, historical process quality indicators, and equipment health score.

[0008] Optionally, the action space of the intelligent process scheduling system includes selecting a target process cavity for the target crystal boat, planning a transmission path, and loading a preset set of process parameters.

[0009] Optionally, the intelligent process scheduling system adopts the dual-delay deep deterministic strategy gradient TD3 algorithm and performs multi-objective optimization of throughput, process quality and energy consumption through a reward function.

[0010] Optionally, the input vector of the LSTM neural network model includes: current residual gas concentration, cavity temperature, process gas partial pressure, gas flow rate, process duration, cavity material coefficient, and historical contamination accumulation.

[0011] Optionally, the historical pollution accumulation is obtained by weighted summation of residual gas concentrations over the past N process cycles, with the weights decreasing with each process cycle during the weighted summation process.

[0012] Optionally, when the predicted value exceeds a preset threshold, the gas pollution control system triggers a composite cleaning process that includes the following steps executed in sequence: inert gas purging, argon plasma cleaning, and oxygen plasma cleaning combined with vacuum extraction.

[0013] Optionally, the bracket consists of a cantilever fixed to the motion mechanism and a turntable disposed at the end of the cantilever. The turntable has multiple branches evenly distributed around its circumference, and each branch can support and release or clamp the furnace door. The turntable is driven to rotate by a motor or a rotary cylinder.

[0014] Optionally, the crystal boat is a vertically placed frame structure made of quartz or silicon carbide, used to safely support the wafer at high temperatures and keep the wafer in a precise and undeformed position, as well as to prevent the wafer from sticking or slipping.

[0015] Optionally, a main frame is also included to support the process chamber and facilitate the docking and unloading of the crystal boat.

[0016] The advantages and beneficial effects of this invention are as follows: Through the seamless transfer function of multiple independent process cavities combined with the transfer mechanism, all interconnected process steps are completed within a single vacuum cycle, completely eliminating the natural oxide layer and interface contamination caused by atmospheric exposure between different processes on the wafer, achieving an atomically clean and perfect interface, significantly improving device carrier mobility, reliability, and yield; successfully resolving the inherent contradiction of the inability to integrate high-temperature processes and low-temperature sensitive processes in traditional equipment, enabling seamless integration of previously incompatible process combinations such as "atomic layer deposition + oxidation annealing" and "high-temperature thermal oxidation and low-temperature dielectric deposition" in the same equipment; and integrating multiple process steps. By integrating wafers into a single device, the system eliminates the time spent transferring, queuing, and preparing wafers between different devices. The dynamic shuttle mechanism enables rapid switching of thermal processes, significantly reducing non-process waiting time, shortening product manufacturing cycles, and increasing equipment utilization by over 40%. Independent temperature zones and rapid shuttle capabilities enable localized rapid heating and cooling of wafers, avoiding unnecessary heat accumulation caused by the slow heating and cooling of traditional vertical furnaces, and precisely controlling dopant activation and diffusion. The intelligent process scheduling system achieves a 25% efficiency improvement and a 40% reduction in conflict rate. The gas contamination prevention and control system achieves a 95% contamination prediction accuracy rate and a cleaning time of less than 20 seconds, effectively avoiding cross-contamination caused by precursor residues. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the internal structure of the multi-chamber vertical furnace of the present invention.

[0018] Figure 2 This is a schematic diagram of the furnace door lock and the process cavity forming a sealed processing cavity according to the present invention.

[0019] Figure 3 This is a schematic diagram of the process cavity of the present invention.

[0020] Figure 4 This is a partially enlarged view of the process cavity of the present invention.

[0021] Figure 5 This is a schematic diagram of the conveying mechanism of the present invention.

[0022] Figure 6 This is a schematic diagram of the bracket structure of the present invention.

[0023] Figure 7 This is a schematic diagram of the structure of the crystal boat of the present invention.

[0024] The components include: process chamber 1, heating assembly 11, furnace shell 12, insulation layer 13, quartz tube 14, conveying mechanism 2, bracket 21, motion mechanism 22, furnace door 23, cantilever 211, turntable 212, crystal boat 3, and main frame 4. Detailed Implementation

[0025] The embodiments of this application will now be described in detail. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application. Furthermore, the following embodiments and features can be combined with each other unless otherwise specified. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0026] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection between two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0027] Example 1

[0028] This embodiment provides a multi-chamber vertical furnace for heat treatment in the semiconductor industry. The multi-chamber vertical furnace mainly consists of multiple independent process chambers 1 and a conveying mechanism 2.

[0029] Multiple process chambers 1 are arranged in a ring, each equipped with an independent temperature control system and atmosphere control system, enabling simultaneous execution of different heat treatment processes, such as oxidation, annealing, and diffusion. The chambers do not interfere with each other, ensuring precise control of process parameters. It should be noted that the temperature control system and atmosphere control system involved in this invention are existing technologies, as long as they can achieve temperature control (e.g., heating element 11) and atmosphere control (e.g., NH3 gas source) within the process chamber 1.

[0030] The process chamber 1 involved in this invention has a similar structure to existing products on the market, mainly composed of a heating component 11, a furnace shell 12, an insulation layer 13, a quartz tube 14, etc. The difference lies in that after the furnace door 23 is connected, it is locked to prevent it from detaching. The process chamber 1 is fixed to the main frame 4 in the vertical furnace equipment, serving to heat and insulate the wafer, and to introduce process gases, thereby enabling processes such as oxidation, diffusion, annealing, and chemical vapor deposition. (See attached diagram.) Figure 3 and attached Figure 4 As shown.

[0031] Reference Appendix Figure 5 The conveying mechanism 2 is the core component of the entire system, including a bracket 21 that can rotate around a central axis, a motion mechanism 22 that drives the bracket 21 to rise and fall, and multiple furnace doors 23 mounted on the bracket 21 and corresponding to the number of process chambers 1. The number of furnace doors 23 is typically 2 to 5. The bracket 21 consists of a cantilever 211 fixed to the motion mechanism 22 and a turntable 212 mounted on the end of the cantilever 211. The turntable 212 has multiple branches evenly distributed circumferentially, each branch supporting and releasing or clamping the furnace door 23. The turntable 212 is driven by a motor or rotary cylinder and can rotate around the central axis to any angle. (Refer to the attached diagram.) Figure 6 The motion mechanism 22 includes a transmission device consisting of a lead screw, guide rod, motor, etc., which can drive the crystal boat 3 to move stably and accurately in the vertical direction. Each furnace door 23 is used to carry one crystal boat 3, and after docking, it forms a complete sealed processing chamber with the target process cavity 1.

[0032] The wafer carrier 3, carried by the transfer mechanism 2, is loaded from the wafer storage unit Stocker into one or more wafer carriers 3 at different locations via one or more docking wafer transfer ROBOTs during its travel. After loading, the following docking process is performed: the drive carriage 21 rotates around its central axis to the process cavity 1 corresponding to the desired process position. This process can simultaneously load multiple wafer carriers 3 and rotate them synchronously into the corresponding process cavity 1. Then, the motion mechanism 22 of the transfer mechanism 2 drives the carriage 21 and the wafer carriers 3 on it to move upwards and dock with the target process cavity 1. After docking, the furnace door 23 is locked (it should be noted that the mechanical / pneumatic locking structure of the furnace door after docking is existing technology, including but not limited to the braking mechanism mentioned in CN116403945A), forming a complete sealed processing cavity with the target process cavity 1. (See attached diagram.) Figure 2 The furnace door 23 separates from the carrier 21, and the motion mechanism 22 drives the carrier 21 downwards, then enters a standby state. The process chamber 1 executes the process flow according to the process design. The above docking process is repeated as needed, or the unloading process is executed after docking with a process chamber 1 that has completed the process. After unlocking the furnace door 23, the furnace door 23 and the wafer boat 3 are brought to the lower position of the stroke. Cooling is performed as needed, or the carrier 21 rotates to the next process chamber 1 corresponding to the next process before executing the docking process. The docking, process, and unloading processes are repeated until all processes are completed. The wafer is then unloaded by the wafer transfer ROBOT. While there is a work cycle for the docking, process, and unloading processes for a single wafer boat 3, different work cycles can be performed simultaneously during the overall equipment operation. The transfer mechanism 2 flows between multiple process chambers 1. Compared to the original one-to-one configuration of chambers and transfer mechanisms 2, although docking and unloading processes are executed more frequently, the number of transfer mechanisms 2 and equipment idle time are reduced, and equipment utilization is improved. At the same time, this design avoids the risk of wafer contamination during transport, improving product yield.

[0033] Reference Appendix Figure 7 The crystal boat 3 is a vertically placed frame structure made of quartz or silicon carbide. It is used to safely support the wafer at high temperatures and keep the wafer in a precise and undeformed position, as well as to prevent the wafer from sticking or slipping.

[0034] Reference Appendix Figure 1 The entire multi-chamber vertical furnace also includes a main frame 4, which supports the process chamber 1 and facilitates the docking and unloading of the crystal boat 3. The main frame 4 is constructed from welded stainless steel square tubes and plates to form a support skeleton, providing excellent rigidity and stability to ensure the equipment maintains its accuracy during long-term operation.

[0035] In summary, this multi-chamber vertical furnace, through parallel processing in multiple chambers (i.e., multiple process chambers 1 set up in the same equipment) and a transfer mechanism 2 that can transport the wafer boat 3 between different process chambers 1 as a whole, enables the wafer to be quickly transferred to another environment with a different temperature and process atmosphere after the process is completed in one temperature zone, without the need for loading and unloading the wafer between the wafer boat 3 and the transfer box, effectively reducing the time spent on transfer and waiting between equipment. This further reduces non-value-added time, improves product output efficiency, and at the same time reduces the number of transfer devices, lowering the factory's equipment ownership costs.

[0036] Compared to existing equipment that executes multiple processes sequentially within a single vertical furnace chamber by changing the type and order of introduced gases and adjusting the chamber temperature, the multi-chamber vertical furnace provided by this invention avoids contamination caused by residual precursors when switching between different process gases. It also enables the completion of doping and undoping processes within a single unit, facilitating the production of various process products and improving equipment flexibility. Furthermore, it provides the possibility of achieving process combinations that are impossible with existing equipment architectures, and avoids discrepancies between different units during the execution of experimental processes.

[0037] Example 2

[0038] This embodiment, based on Embodiment 1, further provides a process intelligent scheduling system to achieve dynamic load balancing among multiple process cavities 1, further optimizing the overall process cycle. This process intelligent scheduling system dynamically allocates wafers to target process cavities 1 based on a reinforcement learning algorithm, maximizing production efficiency. The system's state space consists of the following parameters: real-time temperature of each process cavity 1, process gas concentration within each process cavity 1, three-dimensional position coordinates of the wafer boat 3, priority of the wafer batch to be processed, remaining process time for each process cavity 1, historical process quality indicators, and equipment health score. The action space includes selecting the target process cavity 1 for the target wafer boat 3, planning the transmission path, and loading a preset set of process parameters. The system employs a dual-delay deep deterministic strategy gradient TD3 algorithm, using a reward function to perform multi-objective optimization of throughput, process quality, and energy consumption, maximizing equipment utilization and reducing energy consumption while ensuring process quality.

[0039] The workflow of this intelligent scheduling system includes the following steps.

[0040] 1. State-space modeling: Define a 7-dimensional state vector. ,in: It is a 7-dimensional state vector. The real-time temperature (0-1 normalized value) of each process chamber 1. The process gas concentration (ppm, normalized value) within each process chamber 1. These are the three-dimensional position coordinates (in a three-dimensional Cartesian coordinate system) of Crystal Boat 3. Priority of the wafer batch to be processed (discrete value from level 1 to 5). The remaining process time (time-standardized value) for each process cavity 1. Historical process quality indicators (σ value normalized). Rate the health of the equipment (0-100% continuous value).

[0041] 2. Motion space design: Discrete motion sets are used. Where: A is a discrete action set, Let n be the nth action, and each action corresponds to an action space, defined as follows: ,in: For the i-th action, its action space is selected by the target crystal boat 3 and the target process cavity 1 is numbered. Planning transmission paths and load the preset process parameter group constitute.

[0042] 3. Reward Function Construction: The multi-objective weighted reward function is defined as follows: ,in: For multi-objective weighted total reward, The throughput reward is calculated based on the number of tasks or data processed per unit of time. This is a process quality reward, which is obtained by evaluating product quality indicators after task completion. Energy consumption rewards are calculated based on power consumption data during operation, with a weighting coefficient. , , The initial calibration values ​​were determined to be 0.6, 0.3, and 0.1 through orthogonal experiments.

[0043] 4. Learning Algorithm Design: The double-delay deep deterministic strategy gradient TD3 algorithm is adopted. The core formulas of this algorithm include: (1) The formula for calculating the target Q value is: Where: y is the target Q value, and r is the immediate reward. This is a discount factor used to calculate future returns. For the next state, The action output by the target policy network, after adding noise, becomes... , For the target policy network, The noise term follows a truncated normal distribution and is used to introduce smoothness in the target policy. For the target value network, the parameters are as follows: and , used to estimate state-action value. (2) The policy network update formula is: The network parameters are updated using the value network gradient update strategy, where: For the updated policy network parameters, For the current policy network parameters, The learning rate is updated for the policy network. For a policy network, s represents the current state, and the parameters are... , For a value network, 'a' represents the output action in the current state 's'. The policy network outputs actions relative to the policy network parameters. gradient, Let be the gradient of the Q-function of the value network with respect to action a. (3) The formula for soft network update is: The target network parameters are slowly synchronized with the main network parameters through soft updates, wherein: Let θ be the parameters of the target network, and θ be the parameters of the policy network. The soft update coefficient is 0.001 to 0.001. (4) The formula for the Critic loss function is: ,in: This represents the Critic loss value. Let y be the output of the i-th value network, and y be the target Q value, which is calculated using the formula (1) above. (5) The formula for policy gradient update is: ,in: These are the parameters after the policy network is updated. For the first value network The output gradient is relative to action a. The action output by the policy network. For the parameters of the policy network, For the output of the first value network, The policy network outputs actions relative to the policy network parameters. The gradient.

[0044] In one embodiment of the present invention, the workflow of the intelligent process scheduling system is further illustrated. The specific steps include: 1. Initialization stage: (1) Deploy multiple temperature sensors, multiple gas concentration detectors and 6-axis position encoders in each process chamber 1; (2) Establish a process knowledge base containing multiple standard process recipes; (3) Train the initial strategy network. 2. Online scheduling stage: Real-time acquisition of the temperature in process chamber 1A and the NH3 concentration in process chamber 1B, for example: the temperature in process chamber 1A is 852℃ (normalized value 0.71), and the NH3 concentration in process chamber 1B is 1200ppm; the current position of crystal boat 3 is detected as (2.1m, 0.3m, 1.8m), the current batch priority is 3, and the remaining process time is 12min for process chamber 1A and 25min for process chamber 1B; the strategy network outputs the action as follows. The target strategy actions for the next state include: selecting target process cavity 1 (number B) from target crystal boat 3, and planning the transmission path. Load the preset process parameter set 3. Continuous optimization: (1) Update the strategy network parameters every 24 hours; (2) Adapt the new process formula through transfer learning; Historical data shows that the average equipment utilization rate can be increased from 68% to 89%, which can maximize equipment utilization and reduce energy consumption while ensuring process quality.

[0045] Example 3

[0046] This embodiment, based on Embodiment 1 above, further provides a gas pollution control system to address the problem of cross-contamination of gases from multiple processes. This gas pollution control system uses an LSTM neural network model to predict the residual gas concentration for the next process cycle and automatically triggers a composite cleaning process when the predicted value exceeds a preset threshold. The input vector of the LSTM neural network model includes: current residual gas concentration, chamber temperature, process gas partial pressure, gas flow rate, process duration, chamber material coefficient, and historical accumulated pollution. The historical accumulated pollution is obtained by weighted summation of residual gas concentrations from the past N process cycles. The weights in the weighted summation process decrease with each process cycle, making recent pollution data have a greater impact on the prediction results. When the predicted value exceeds the preset threshold, the composite cleaning process triggered by the system includes the following steps executed sequentially: inert gas purging, argon plasma cleaning, and oxygen plasma cleaning combined with vacuum extraction, ensuring that the chamber cleanliness meets the requirements of the next process.

[0047] The workflow of this gas pollution control system includes the following steps.

[0048] 1. Data Acquisition and Feature Engineering: Establishing a 7-dimensional input vector ,in: Given a 7-dimensional input vector, This is the standardized value of the current residual gas concentration (in ppm); This is a standardized value for the cavity temperature (in °C). Standardized values ​​for process gas partial pressure (in Pa); This is a standardized value for gas flow rate (in sccm). This is a standardized value for the process duration (in minutes). This is a standardized value for the cavity material coefficient (e.g., quartz = 0.8, silicon carbide = 0.5). This is a standardized value for the historical cumulative pollution amount, and The calculation formula is: Weight .

[0049] 2. LSTM Network Architecture: A 3-layer stacked LSTM structure is adopted, with 128 units in each layer; the time step is set to 5, and the input window covers 5 consecutive process cycles; the output layer uses the Sigmoid activation function to predict the residual concentration in the next cycle. The loss function used is Huber loss, and the formula is: ,in: This represents Huber's loss value. The true value represents the actual value of the contaminated data. These are predicted values, representing the predicted results of pollution data. This is a threshold parameter with a value of 1.0, used to control the smoothness and robustness of the loss function.

[0050] 3. Composite cleaning triggering mechanism: When ≥10 The three-stage cleaning process is initiated as follows: (1) Primary cleaning: nitrogen purging, with parameters set to a purging duration of 10 seconds and a nitrogen flow rate of 20 standard liters per minute; (2) Secondary cleaning: argon plasma, with parameters set to a duration of argon plasma acting on the material surface of 5 seconds and a plasma generator output power of 300 watts; (3) Deep cleaning: oxygen plasma cleaning combined with vacuum extraction, with parameters set to a duration of oxygen plasma acting on the material surface of 3 seconds, a plasma generator output power of 500 watts, and a vacuum extraction duration of 2 seconds. The cleaning efficiency model is defined as follows: Where η is the cleaning efficiency, ranging from 0 to 1, and k is the cleaning coefficient, with a value of 2.3. This refers to the actual cleaning time. The standard time is 15 seconds. 4. Online learning optimization: The weights of each layer of the LSTM neural network are updated every 50 process cycles; the adaptive moment estimation (Adam) optimizer is used, and the learning rate can be set to 0.001; an early stopping mechanism is set: when the validation set loss does not decrease for 3 consecutive times, the training of the LSTM neural network is terminated.

[0051] In one embodiment of the present invention, the workflow of the intelligent process scheduling system is further illustrated. Specific steps include: after continuously executing the SiH4-based CVD process, firstly, seven relevant data points are collected, including: current residual gas concentration (e.g., 8.2 ppm), chamber temperature (e.g., 650°C), process gas partial pressure (e.g., 200 Pa), gas flow rate (e.g., 120 sccm), process duration (e.g., 45 min), chamber material coefficient (e.g., 0.8), and historical contamination accumulation (e.g., 6.7 ppm); subsequently, the standardized 7-dimensional input vector is input into an LSTM neural network model to predict the residual gas concentration for the next process cycle, and the model outputs... The value is 11.3 ppm. If the preset residual gas concentration threshold for the next process cycle is 10 ppm, then the model output will be 11.3 ppm. If the residual gas concentration exceeds the threshold of the next process cycle, the cleaning process is automatically triggered. The specific cleaning process includes: (1) Nitrogen purging for 10 seconds and the detected value drops to 5.8ppm; (2) Argon plasma treatment for 5 seconds and the value drops to 2.1ppm; (3) Oxygen plasma treatment for 3 seconds and the residual value is 0.3ppm. The total time of the entire cleaning process is 18 seconds, and the cleaning efficiency η is calculated to be 97.6%. Finally, the LSTM neural network model is updated. The specific model update process includes: (1) Recording the actual residual value of 0.3ppm after cleaning as a new sample; (2) Calculating the Huber loss value to be 0.42; (3) Adjusting the weight of the last layer of the LSTM neural network model and the gradient descent step size to 0.0007.

[0052] In a preferred embodiment, the LSTM neural network model, trained through deep learning, can accurately predict the trend of residual gas concentration changes with a prediction accuracy of over 95%, effectively avoiding cross-contamination problems.

[0053] In summary, this multi-chamber vertical furnace, through multi-chamber parallel processing, intelligent scheduling, and predictive maintenance, can achieve high efficiency, high quality, and low pollution in semiconductor heat treatment, thereby significantly improving production efficiency and product quality.

[0054] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A multi-chamber vertical furnace for heat treatment in the semiconductor industry, characterized in that, include: Multiple independent process chambers, each equipped with an independent temperature control system and atmosphere control system; The transfer mechanism includes a bracket rotatable about a central axis, a motion mechanism for driving the bracket to rise and fall, and multiple furnace doors mounted on the bracket corresponding to the number of process cavities. Each furnace door is used to carry one wafer boat. The transfer mechanism is configured to: rotate the furnace door carrying the wafer boat to below the target process cavity, drive the furnace door to rise and dock with the corresponding target process cavity, and lock the furnace door to form a sealed processing cavity; after the target process is completed, drive the furnace door to fall and rotate to the next process cavity or unloading position, realizing seamless transfer of the wafer boat between different process environments, while the wafer is always in a controlled environment and does not need to be exposed to the atmosphere. A process intelligent scheduling system is used to dynamically allocate wafers to target process cavities based on reinforcement learning algorithms; The gas pollution control system uses an LSTM neural network model to predict the residual gas concentration in the next process cycle and automatically triggers the cleaning process when the predicted value exceeds a preset threshold.

2. The multi-chamber vertical furnace for heat treatment in the semiconductor industry as described in claim 1, characterized in that, The state space of the intelligent process scheduling system consists of the following parameters: real-time temperature of each process chamber, process gas concentration in each process chamber, three-dimensional position coordinates of the wafer boat, priority of the wafer batch to be processed, remaining process time of each process chamber, historical process quality indicators, and equipment health score.

3. A multi-chamber vertical furnace for heat treatment in the semiconductor industry as described in claim 1, characterized in that, The action space of the intelligent process scheduling system includes selecting the target process cavity for the target crystal boat, planning the transmission path, and loading the preset process parameter set.

4. A multi-chamber vertical furnace for heat treatment in the semiconductor industry as described in claim 1, characterized in that, The intelligent process scheduling system adopts the dual-delay deep deterministic strategy gradient TD3 algorithm and performs multi-objective optimization of throughput, process quality and energy consumption through a reward function.

5. A multi-chamber vertical furnace for heat treatment in the semiconductor industry as described in claim 1, characterized in that, The input vector of the LSTM neural network model includes: current residual gas concentration, cavity temperature, process gas partial pressure, gas flow rate, process duration, cavity material coefficient, and historical accumulated contamination.

6. A multi-chamber vertical furnace for heat treatment in the semiconductor industry as described in claim 5, characterized in that, The historical pollution accumulation is obtained by weighted summation of residual gas concentrations over the past N process cycles, with the weights decreasing with each process cycle.

7. A multi-chamber vertical furnace for heat treatment in the semiconductor industry as described in claim 1, characterized in that, When the predicted value exceeds a preset threshold, the gas pollution control system triggers a composite cleaning process that includes the following steps executed in sequence: inert gas purging, argon plasma cleaning, and oxygen plasma cleaning combined with vacuum extraction.

8. A multi-chamber vertical furnace for heat treatment in the semiconductor industry as described in claim 1, characterized in that, The bracket consists of a cantilever fixed to the motion mechanism and a turntable set at the end of the cantilever. The turntable has multiple branches evenly distributed around its circumference. Each branch can support and release or clamp the furnace door. The turntable is driven to rotate by a motor or a rotary cylinder.

9. A multi-chamber vertical furnace for heat treatment in the semiconductor industry as described in claim 1, characterized in that, The crystal boat is a vertically placed frame structure made of quartz or silicon carbide. It is used to safely support the wafer at high temperatures, keep the wafer in a precise position without deformation, and prevent the wafer from sticking or slipping.

10. A multi-chamber vertical furnace for heat treatment in the semiconductor industry as described in claim 1, characterized in that, It also includes a main frame, which supports the process chamber and facilitates the docking and unloading of the crystal boat.

Images