A control method and system for a semiconductor manufacturing machine

Abstract

The present invention provides a control method and system for a semiconductor manufacturing machine, which relates to the field of semiconductor technology, including obtaining the operating parameters of a chamber of an equipment, and establishing a multi-objective optimization model including manufacturing cost, manufacturing quality and manufacturing efficiency. The model is trained using a deep reinforcement learning algorithm to generate an optimization prediction result. The prediction result is analyzed for deviations from the target parameter, and the deviation feature vector is input into a compensation calculation model constructed based on a long short-term memory network to generate a compensation control instruction. The compensation instruction is input into the multi-objective optimization model, and the multi-objective optimization algorithm with dynamic weight allocation is used to calculate the optimal parameter combination to generate a parameter adjustment instruction. Finally, the adjustment instruction is sent to the execution unit to realize adaptive feedback control of the semiconductor manufacturing process, thereby reducing manufacturing costs and improving manufacturing quality and manufacturing efficiency.

Description

Technical Field

[0001] The present invention relates to semiconductor technology, and in particular to a control method and system for a semiconductor manufacturing machine. Background Art

[0002] The semiconductor manufacturing process is extremely complex and involves the precise control of many key parameters, such as temperature, pressure, gas flow, RF power, and plasma density. Slight changes in these parameters will directly affect the quality, yield, and production cost of the final product. In order to achieve high-quality, high-efficiency, and low-cost semiconductor manufacturing, precise control and optimization of equipment chamber operating parameters are essential.

[0003] Traditional semiconductor manufacturing process control mainly relies on pre-set process parameters and manual experience for adjustment. This method has obvious limitations and is difficult to adapt to complex process changes and dynamic environments. In addition, traditional control methods such as simple PID control are difficult to handle multi-objective optimization problems, that is, taking into account cost, quality and efficiency at the same time.

[0004] Difficult to adapt to complex process changes: Traditional control methods are usually based on fixed models and parameter settings, and it is difficult to cope with the complex process changes and nonlinear characteristics in the semiconductor manufacturing process. When process conditions change, the preset parameters may no longer apply, resulting in a decrease in product quality and yield. Insufficient multi-objective optimization capabilities: Semiconductor manufacturing needs to consider multiple objectives such as cost, quality, and efficiency at the same time, while traditional control methods usually only focus on the optimization of a single objective, making it difficult to achieve a balance and coordinated optimization of multiple objectives. Poor adaptive feedback regulation capabilities: Traditional control methods lack real-time feedback and adaptive adjustment capabilities, making it difficult to dynamically adjust control strategies according to actual conditions. This makes it difficult to guarantee control accuracy and robustness, especially in the face of external interference and equipment aging, the control effect will be significantly reduced. Summary of the invention

[0005] The embodiments of the present invention provide a control method and system for a semiconductor manufacturing machine, which can solve the problems in the prior art.

[0006] In a first aspect of the present invention,

[0007] A control method for a semiconductor manufacturing machine is provided, comprising:

[0008] Acquire the operating parameters of the equipment chamber in the semiconductor manufacturing process, wherein the operating parameters include temperature parameters, pressure parameters, gas flow parameters, radio frequency power parameters, and plasma density parameters; establish a multi-objective optimization model including manufacturing cost targets, manufacturing quality targets, and manufacturing efficiency targets, wherein the manufacturing cost target is composed of energy consumption per unit time and raw material utilization rate, the manufacturing quality target is composed of process uniformity index and defect density index, and the manufacturing efficiency target is composed of equipment utilization rate and product output rate; use a deep reinforcement learning algorithm to train the multi-objective optimization model, construct a mapping relationship between the operating parameters and the optimization targets, and generate optimization prediction results;

[0009] Perform deviation analysis on the optimization prediction result and the preset target parameter; construct a deviation feature vector based on the deviation analysis result; input the deviation feature vector into a compensation amount calculation model, the compensation amount calculation model is constructed based on a long short-term memory network, and generates a compensation control instruction, the compensation control instruction includes a temperature compensation amount, a pressure compensation amount, a gas flow compensation amount, a radio frequency power compensation amount, and a plasma density compensation amount;

[0010] The compensation control instruction is input into the multi-objective optimization model, and the optimal parameter combination is calculated by a multi-objective optimization algorithm, wherein the multi-objective optimization algorithm realizes the balanced optimization of the manufacturing cost target, the manufacturing quality target and the manufacturing efficiency target through dynamic weight allocation; a parameter adjustment instruction is generated according to the optimal parameter combination, and the parameter adjustment instruction includes a temperature adjustment amount, a pressure adjustment amount, a gas flow adjustment amount, a radio frequency power adjustment amount and a plasma density adjustment amount; the parameter adjustment instruction is sent to a corresponding execution unit, and the execution unit includes a heater, a pressure control valve, a gas flow controller, a radio frequency generator and a plasma generator, so as to realize adaptive feedback control of the semiconductor manufacturing process.

[0011] A multi-objective optimization model including manufacturing cost target, manufacturing quality target and manufacturing efficiency target is established. The manufacturing cost target is composed of energy consumption per unit time and raw material utilization rate, the manufacturing quality target is composed of process uniformity index and defect density index, and the manufacturing efficiency target is composed of equipment utilization rate and product output rate.

[0012] Collect heater power data, RF power data and vacuum pump power data as energy data, collect air intake data and actual reaction gas volume data as material data, collect film thickness data and etching depth data as process data, collect equipment operation status data and product inspection data as production data;

[0013] Calculate the energy consumption per unit time based on the energy data, accumulate the heater power data, the RF power data and the vacuum pump power data within a unit time to obtain the total power data, and integrate the total power data to obtain the energy consumption per unit time; calculate the raw material utilization rate based on the material data, calculate the raw material utilization rate by the ratio of the actual reaction gas volume data to the intake gas volume data; normalize the energy consumption per unit time and the raw material utilization rate and weightedly combine them to construct a manufacturing cost target;

[0014] Calculating a process uniformity index based on the process data, selecting a detection point on the wafer surface to obtain the film thickness data and the etching depth data, and calculating the standard deviation of the film thickness data and the etching depth data to obtain a process uniformity index; calculating a defect density index based on the process data, preprocessing a defect detection image in the process data to obtain an enhanced image, extracting defect features in the enhanced image, and counting the number of defects per unit area based on the defect features to obtain a defect density index; normalizing the process uniformity index and the defect density index and weightedly combining them to construct a manufacturing quality target;

[0015] The equipment utilization rate is calculated based on the production data, the equipment operation status data is divided into production time data, standby time data, maintenance time data and fault time data, and the ratio of the production time data to the total time is calculated to obtain the equipment utilization rate; the product output rate is calculated based on the production data, the product inspection data is analyzed, and the number of qualified products per unit time is counted to obtain the product output rate; the equipment utilization rate and the product output rate are normalized and weightedly combined to construct a manufacturing efficiency target.

[0016] The multi-objective optimization model is trained using a deep reinforcement learning algorithm, a mapping relationship between the operating parameters and the optimization objectives is constructed, and the optimization prediction results are generated, including:

[0017] Constructing a deep reinforcement learning training network, the deep reinforcement learning training network includes a state encoder, an action generator and a value evaluator; the state encoder encodes the normalized operating parameters into a state feature vector; the action generator generates an adjustment action based on the state feature vector; the value evaluator evaluates the state action value based on the state feature vector and the adjustment action;

[0018] Constructing a reward calculation module, applying the adjustment action to the normalized operating parameters to obtain optimized operating parameters; inputting the optimized operating parameters into the multi-objective optimization model to obtain corresponding optimization target outputs; calculating an immediate reward value based on the optimization target output, and calculating a cumulative reward value in combination with historical reward values;

[0019] Training the deep reinforcement learning training network: inputting the state feature vector into the action generator to generate an adjustment action; inputting the adjustment action into the value evaluator to obtain the state action value; calculating the value loss based on the difference between the accumulated reward value and the state action value; updating the network parameters of the deep reinforcement learning training network based on the value loss; repeating the training process until the value loss converges to obtain a trained deep reinforcement learning model;

[0020] The deep reinforcement learning model is used to construct a mapping relationship between the operating parameters and the optimization objectives: the operating parameters to be optimized are input into the state encoder to obtain a current state feature vector; the current state feature vector is input into the action generator to obtain an optimization adjustment action; the operating parameters are adjusted based on the optimization adjustment action to generate optimized operating parameters; the optimized operating parameters are input into the multi-objective optimization model to generate optimization prediction results.

[0021] The deviation feature vector is input into a compensation amount calculation model, the compensation amount calculation model is constructed based on a long short-term memory network, and the compensation control instruction is generated including:

[0022] The deviation feature vector is input into a compensation amount calculation model, and the compensation amount calculation model is constructed based on a long short-term memory network; the compensation amount calculation model performs feature preprocessing on the deviation feature vector to generate a normalized feature vector; and the normalized feature vector is time-series sampled to generate a time-series feature sequence containing multiple time steps;

[0023] Input the time series feature sequence into the feature extraction layer of the compensation amount calculation model, the feature extraction layer includes a multi-layer long short-term memory network; the multi-layer long short-term memory network extracts time series features from the time series feature sequence to generate a time series feature representation; perform attention weighting on the time series feature representation to generate a weighted feature representation; input the weighted feature representation into the feature fusion layer, the feature fusion layer adaptively fuses the weighted feature representation with the historical feature to generate a fused feature vector;

[0024] The fused feature vector is input into a compensation amount prediction layer, wherein the compensation amount prediction layer includes a plurality of prediction branches; each prediction branch generates a compensation amount prediction value of a corresponding processing parameter based on the fused feature vector; parameter correlation analysis is performed on the compensation amount prediction values ​​of each prediction branch to establish a parameter compensation association matrix; based on the parameter compensation association matrix, the compensation amount prediction values ​​are jointly optimized to generate an optimized compensation amount;

[0025] The optimized compensation amount is input into the compensation instruction generation layer, and the compensation instruction generation layer converts the optimized compensation amount into the device control amount; the device control amount is time-smoothed to eliminate the sharp fluctuation of the compensation instruction; the smoothed device control amount is packaged according to the standard control protocol format to generate a compensation control instruction.

[0026] Inputting the compensation control instruction into the multi-objective optimization model, using a multi-objective optimization algorithm to calculate an optimal parameter combination, wherein the multi-objective optimization algorithm achieves balanced optimization of the manufacturing cost target, the manufacturing quality target, and the manufacturing efficiency target through dynamic weight allocation; generating a parameter adjustment instruction according to the optimal parameter combination includes:

[0027] Input the compensation control instruction into the multi-objective optimization model, analyze the compensation control instruction, establish the manufacturing cost objective function, the manufacturing quality objective function and the manufacturing efficiency objective function; calculate the dynamic weight coefficient of each objective function based on the processing task characteristics, product quality requirements and production cycle requirements, and generate the objective weight matrix;

[0028] The objective weight matrix is ​​input into a multi-objective optimization algorithm, and the multi-objective optimization algorithm uses a weighted summation method to jointly optimize the manufacturing cost objective function, the manufacturing quality objective function and the manufacturing efficiency objective function to construct a comprehensive objective function; based on the optimization result of the comprehensive objective function, the objective weight matrix is ​​dynamically updated to achieve an adaptive balance among the objectives; under the constraints of process parameters, an improved particle swarm algorithm is used to optimize the comprehensive objective function to generate a candidate parameter combination;

[0029] Performing process constraint verification on the candidate parameter combinations to screen out feasible parameter sets that meet the processing requirements; calculating stability evaluation indicators of each parameter combination in the feasible parameter set, wherein the stability evaluation indicators include parameter volatility, parameter sensitivity, and parameter robustness; comprehensively scoring the feasible parameter set based on the stability evaluation indicators, and selecting the parameter combination with the highest score as the optimal parameter combination;

[0030] The optimal parameter combination is converted into device control parameters, and the device control parameters are time-series smoothed to eliminate drastic fluctuations in the parameter adjustment process; the smoothed device control parameters are data-encapsulated according to the device communication protocol to generate parameter adjustment instructions in a standard format.

[0031] Under the constraints of process parameters, the improved particle swarm algorithm is used to optimize the comprehensive objective function and generate candidate parameter combinations including:

[0032] Constructing process parameter constraints, wherein the process parameter constraints include temperature field uniformity constraints, pressure fluctuation amplitude constraints, gas ratio range constraints, and plasma density distribution constraints; constructing a comprehensive objective function based on the process parameter constraints, wherein the comprehensive objective function is in the form of weighted summation and includes process parameter stability indicators, energy consumption indicators, and product quality indicators;

[0033] The comprehensive objective function is improved by a particle swarm algorithm, and an adaptive inertia weight strategy is set, wherein the adaptive inertia weight decreases nonlinearly with the number of iterations; a chaotic perturbation mechanism is introduced, wherein the chaotic perturbation mechanism adopts a V-type chaotic mapping; a position update rule of a constraint violation degree is established, wherein the constraint violation degree is calculated based on the process parameter constraint condition;

[0034] Determine the search space according to the process parameter constraint conditions, encode the process parameters into a position vector, the position vector includes a temperature setting value, a pressure setting value, a gas flow setting value, a radio frequency power setting value and a plasma generator parameter; initialize the position and velocity of the particle group using the chaotic perturbation mechanism;

[0035] Calculate the constraint violation degree of the initial particle swarm according to the process parameter constraint condition, and combine the constraint violation degree with the comprehensive objective function to generate a fitness function; evaluate the initial particle swarm based on the fitness function to determine the individual optimal position and the global optimal position;

[0036] In the iterative optimization process, the new velocity of the particle is calculated based on the adaptive inertia weight strategy; the new position of the particle is adjusted according to the position update rule of the constraint violation degree; the fitness function value corresponding to the new position is calculated; when the global optimal position is not updated after a preset number of consecutive iterations, the particle position is adjusted using the chaotic perturbation mechanism;

[0037] Recalculate the constraint violation degree and fitness function value for each iteratively updated particle position; update the individual optimal position and the global optimal position based on the fitness function value; when the termination condition is reached, extract the corresponding temperature setting value, pressure setting value, gas flow setting value, radio frequency power setting value and plasma generator parameter from the global optimal position that satisfies the process parameter constraint condition to generate a candidate parameter combination.

[0038] The parameter adjustment instruction is sent to a corresponding execution unit, wherein the execution unit includes a heater, a pressure control valve, a gas flow controller, a radio frequency generator, and a plasma generator, and the adaptive feedback control of the semiconductor manufacturing process is realized by:

[0039] The parameter adjustment instruction includes process parameter target value information, and the process parameter target value information is decomposed according to the control characteristics of the execution unit to generate a heater control instruction, a pressure control valve control instruction, a gas flow controller control instruction, a radio frequency generator control instruction, and a plasma generator control instruction;

[0040] The heater control instruction is sent to the heater, and the heater adjusts the heating power output according to the temperature feedback signal; the pressure control valve control instruction is sent to the pressure control valve, and the pressure control valve adjusts the valve opening according to the pressure feedback signal; the gas flow controller control instruction is sent to the gas flow controller, and the gas flow controller adjusts the gas flow according to the flow feedback signal;

[0041] The RF generator control instruction is sent to the RF generator, and the RF generator adjusts the RF output power according to the RF power feedback signal; the plasma generator control instruction is sent to the plasma generator, and the plasma generator adjusts the bias voltage according to the plasma density feedback signal;

[0042] Collecting a temperature feedback signal of the heater, a pressure feedback signal of the pressure control valve, a flow feedback signal of the gas flow controller, a power feedback signal of the radio frequency generator, and a density feedback signal of the plasma generator; and constructing a process parameter evaluation model based on the feedback signals;

[0043] Comparing the output result of the process parameter evaluation model with the process parameter target value information to generate a parameter deviation matrix; calculating the adjustment compensation amount of each execution unit based on the parameter deviation matrix, and the adjustment compensation amount is used to correct the control instructions of each execution unit;

[0044] The heater control instruction, the pressure control valve control instruction, the gas flow controller control instruction, the radio frequency generator control instruction and the plasma generator control instruction are optimized online according to the adjustment compensation amount to generate optimized parameter adjustment instructions; the optimized parameter adjustment instructions are resent to each execution unit to realize adaptive feedback control of process parameters in a semiconductor manufacturing process.

[0045] The second aspect of the present invention,

[0046] A control system for a semiconductor manufacturing machine is provided, comprising:

[0047] The first unit is used to obtain the operating parameters of the equipment chamber in the semiconductor manufacturing process, wherein the operating parameters include temperature parameters, pressure parameters, gas flow parameters, radio frequency power parameters and plasma density parameters; establish a multi-objective optimization model including manufacturing cost targets, manufacturing quality targets and manufacturing efficiency targets, wherein the manufacturing cost target is composed of energy consumption per unit time and raw material utilization rate, the manufacturing quality target is composed of process uniformity index and defect density index, and the manufacturing efficiency target is composed of equipment utilization rate and product output rate; use a deep reinforcement learning algorithm to train the multi-objective optimization model, construct a mapping relationship between the operating parameters and the optimization targets, and generate an optimization prediction result;

[0048] The second unit is used to perform deviation analysis between the optimization prediction result and the preset target parameter; construct a deviation feature vector based on the deviation analysis result; input the deviation feature vector into a compensation amount calculation model, the compensation amount calculation model is constructed based on a long short-term memory network, and generates a compensation control instruction, the compensation control instruction includes a temperature compensation amount, a pressure compensation amount, a gas flow compensation amount, a radio frequency power compensation amount, and a plasma density compensation amount;

[0049] The third unit is used to input the compensation control instruction into the multi-objective optimization model, and use a multi-objective optimization algorithm to calculate the optimal parameter combination, wherein the multi-objective optimization algorithm realizes the balanced optimization of the manufacturing cost target, the manufacturing quality target and the manufacturing efficiency target through dynamic weight allocation; generate parameter adjustment instructions according to the optimal parameter combination, and the parameter adjustment instructions include temperature adjustment amount, pressure adjustment amount, gas flow adjustment amount, radio frequency power adjustment amount and plasma density adjustment amount; send the parameter adjustment instructions to the corresponding execution unit, and the execution unit includes a heater, a pressure control valve, a gas flow controller, a radio frequency generator and a plasma generator to realize adaptive feedback control of the semiconductor manufacturing process.

[0050] According to a third aspect of the present invention, there is provided an electronic device, comprising: a processor;

[0051] a memory for storing processor-executable instructions;

[0052] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0053] In a fourth aspect of the present invention,

[0054] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0055] The beneficial effects of this application are as follows:

[0056] 1. Improve manufacturing efficiency: Through deep reinforcement learning and multi-objective optimization algorithms, equipment utilization and product output rates are improved, thereby improving the overall efficiency of semiconductor manufacturing.

[0057] 2. Reduce manufacturing costs: By optimizing energy consumption and raw material utilization, the manufacturing cost per unit time is effectively reduced.

[0058] 3. Improve product quality: By controlling indicators such as process uniformity and defect density, the quality of semiconductor products is improved and the stability and reliability of products are ensured. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 A schematic flow chart of a control method for a semiconductor manufacturing machine according to an embodiment of the present invention;

[0060] Figure 2 It is a structural schematic diagram of a control system of a semiconductor manufacturing machine according to an embodiment of the present invention. DETAILED DESCRIPTION

[0061] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0062] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0063] According to a first aspect of the embodiments of the present invention,

[0064] Figure 1 FIG. 1 is a flow chart of a control method of a semiconductor manufacturing machine according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0065] S11. Obtain the operating parameters of the equipment chamber in the semiconductor manufacturing process, wherein the operating parameters include temperature parameters, pressure parameters, gas flow parameters, radio frequency power parameters, and plasma density parameters; establish a multi-objective optimization model including a manufacturing cost target, a manufacturing quality target, and a manufacturing efficiency target, wherein the manufacturing cost target is composed of energy consumption per unit time and raw material utilization rate, the manufacturing quality target is composed of a process uniformity index and a defect density index, and the manufacturing efficiency target is composed of equipment utilization rate and product output rate; use a deep reinforcement learning algorithm to train the multi-objective optimization model, construct a mapping relationship between the operating parameters and the optimization targets, and generate an optimization prediction result;

[0066] S12. Perform deviation analysis on the optimization prediction results and preset target parameters; construct a deviation feature vector based on the deviation analysis results; input the deviation feature vector into a compensation amount calculation model, the compensation amount calculation model is constructed based on a long short-term memory network, and generates a compensation control instruction, the compensation control instruction includes a temperature compensation amount, a pressure compensation amount, a gas flow compensation amount, a radio frequency power compensation amount, and a plasma density compensation amount;

[0067] S13. Input the compensation control instruction into the multi-objective optimization model, and use the multi-objective optimization algorithm to calculate the optimal parameter combination, the multi-objective optimization algorithm achieves balanced optimization of the manufacturing cost target, the manufacturing quality target and the manufacturing efficiency target through dynamic weight allocation; generate parameter adjustment instructions according to the optimal parameter combination, the parameter adjustment instructions include temperature adjustment amount, pressure adjustment amount, gas flow adjustment amount, radio frequency power adjustment amount and plasma density adjustment amount; send the parameter adjustment instructions to the corresponding execution unit, the execution unit includes a heater, a pressure control valve, a gas flow controller, a radio frequency generator and a plasma generator, to achieve adaptive feedback control of the semiconductor manufacturing process.

[0068] In an optional implementation, a multi-objective optimization model including a manufacturing cost target, a manufacturing quality target and a manufacturing efficiency target is established, wherein the manufacturing cost target is composed of energy consumption per unit time and raw material utilization rate, the manufacturing quality target is composed of process uniformity index and defect density index, and the manufacturing efficiency target is composed of equipment utilization rate and product output rate, including:

[0069] Collect heater power data, RF power data and vacuum pump power data as energy data, collect air intake data and actual reaction gas volume data as material data, collect film thickness data and etching depth data as process data, collect equipment operation status data and product inspection data as production data;

[0070] Calculate the energy consumption per unit time based on the energy data, accumulate the heater power data, the RF power data and the vacuum pump power data within a unit time to obtain the total power data, and integrate the total power data to obtain the energy consumption per unit time; calculate the raw material utilization rate based on the material data, calculate the raw material utilization rate by the ratio of the actual reaction gas volume data to the intake gas volume data; normalize the energy consumption per unit time and the raw material utilization rate and weightedly combine them to construct a manufacturing cost target;

[0071] Calculating a process uniformity index based on the process data, selecting a detection point on the wafer surface to obtain the film thickness data and the etching depth data, and calculating the standard deviation of the film thickness data and the etching depth data to obtain a process uniformity index; calculating a defect density index based on the process data, preprocessing a defect detection image in the process data to obtain an enhanced image, extracting defect features in the enhanced image, and counting the number of defects per unit area based on the defect features to obtain a defect density index; normalizing the process uniformity index and the defect density index and weightedly combining them to construct a manufacturing quality target;

[0072] The equipment utilization rate is calculated based on the production data, the equipment operation status data is divided into production time data, standby time data, maintenance time data and fault time data, and the ratio of the production time data to the total time is calculated to obtain the equipment utilization rate; the product output rate is calculated based on the production data, the product inspection data is analyzed, and the number of qualified products per unit time is counted to obtain the product output rate; the equipment utilization rate and the product output rate are normalized and weightedly combined to construct a manufacturing efficiency target.

[0073] A manufacturing process control method based on multi-objective optimization aims to optimize manufacturing cost, manufacturing quality and manufacturing efficiency at the same time. This method collects a variety of data and constructs three objective functions based on these data, and finally seeks the best manufacturing process parameters through optimization algorithms.

[0074] First, energy data, material data, process data and production data are collected. Energy data includes heater power, RF power and vacuum pump power. Data is collected at a time interval of 1 second. For example, at a certain time point, the heater power is 10kW, the RF power is 5kW, and the vacuum pump power is 2kW. Material data includes the intake volume and the actual reaction gas volume. For example, the intake volume in a certain period of time is 100 standard cubic centimeters per minute (sccm), and the actual reaction gas volume is 80sccm. Process data includes film thickness, etching depth and defect detection images. For example, 9 detection points are evenly selected on the wafer surface, and the measured film thicknesses are 100nm, 102nm, 98nm, 101nm, 99nm, 100nm, 103nm, 100nm, and 97nm respectively; the etching depths are 50nm, 52nm, 48nm, 51nm, 49nm, 50nm, 53nm, 50nm, and 47nm respectively. Production data includes equipment operation status data and product inspection data. For example, a certain equipment has a production time of 16 hours, a standby time of 4 hours, a maintenance time of 2 hours, and a failure time of 2 hours in a day. Product inspection data records the quality status of each product. For example, in a batch of 1,000 products, the number of qualified products is 980.

[0075] Next, various indicators are calculated based on the collected data. The energy consumption per unit time is calculated by accumulating the heater power, RF power and vacuum pump power, and then integrating the data over a period of time. For example, the power data collected above is accumulated to 17kW. Assuming that this power is maintained for 1 hour, the energy consumption per unit time is 17kWh. The raw material utilization rate is calculated by the ratio of the actual reaction gas volume to the intake gas volume. For example, 80sccm / 100sccm=80%. The process uniformity index is obtained by calculating the standard deviation of the film thickness and the etching depth. For example, the standard deviation of the above film thickness is about 2nm, and the standard deviation of the etching depth is about 2nm. The defect density index is obtained by preprocessing the defect detection image, extracting defect features, and counting the number of defects per unit area. For example, after image processing, it is identified that the number of defects per unit area of ​​a certain wafer is 2 / square centimeter. The equipment utilization rate is calculated by the ratio of production time to total time. For example, 16 hours / 24 hours=66.7%. The product output rate is obtained by counting the number of qualified products per unit time. For example, 980 / 1000=98%.

[0076] Then, the calculated indicators are normalized and weighted to construct the objective function. For example, the energy consumption per unit time and the raw material utilization rate are assigned weights of 0.6 and 0.4 respectively to construct the manufacturing cost objective function; the process uniformity index and the defect density index are assigned weights of 0.7 and 0.3 respectively to construct the manufacturing quality objective function; the equipment utilization rate and product output rate are assigned weights of 0.5 and 0.5 respectively to construct the manufacturing efficiency objective function.

[0077] Finally, a multi-objective optimization algorithm, such as a genetic algorithm or a particle swarm algorithm, is used to optimize the constructed objective function and seek the optimal manufacturing process parameters, such as heating temperature, RF power, reaction gas flow rate, etc., to achieve simultaneous optimization of manufacturing cost, manufacturing quality, and manufacturing efficiency.

[0078] The solution of this application can:

[0079] Reduce manufacturing costs: By optimizing energy consumption and raw material utilization, production costs are effectively reduced. Improve manufacturing quality: By optimizing process uniformity and defect density, product quality is significantly improved. Improve manufacturing efficiency: By optimizing equipment utilization and product output rate, production efficiency is effectively improved.

[0080] In an optional implementation, the multi-objective optimization model is trained using a deep reinforcement learning algorithm, a mapping relationship between the operating parameters and the optimization objectives is constructed, and the optimization prediction results are generated, including:

[0081] Constructing a deep reinforcement learning training network, the deep reinforcement learning training network includes a state encoder, an action generator and a value evaluator; the state encoder encodes the normalized operating parameters into a state feature vector; the action generator generates an adjustment action based on the state feature vector; the value evaluator evaluates the state action value based on the state feature vector and the adjustment action;

[0082] Constructing a reward calculation module, applying the adjustment action to the normalized operating parameters to obtain optimized operating parameters; inputting the optimized operating parameters into the multi-objective optimization model to obtain corresponding optimization target outputs; calculating an immediate reward value based on the optimization target output, and calculating a cumulative reward value in combination with historical reward values;

[0083] Training the deep reinforcement learning training network: inputting the state feature vector into the action generator to generate an adjustment action; inputting the adjustment action into the value evaluator to obtain the state action value; calculating the value loss based on the difference between the accumulated reward value and the state action value; updating the network parameters of the deep reinforcement learning training network based on the value loss; repeating the training process until the value loss converges to obtain a trained deep reinforcement learning model;

[0084] The deep reinforcement learning model is used to construct a mapping relationship between the operating parameters and the optimization objectives: the operating parameters to be optimized are input into the state encoder to obtain a current state feature vector; the current state feature vector is input into the action generator to obtain an optimization adjustment action; the operating parameters are adjusted based on the optimization adjustment action to generate optimized operating parameters; the optimized operating parameters are input into the multi-objective optimization model to generate optimization prediction results.

[0085] A multi-objective optimization method based on deep reinforcement learning is used to construct a mapping relationship between operating parameters and optimization objectives and generate optimization prediction results. This method learns the optimization strategy of operating parameters by training a deep reinforcement learning model, thereby achieving multi-objective optimization.

[0086] First, data preprocessing is performed. Collect historical operating parameter data and corresponding optimization target output data. Normalize the operating parameter data and scale the parameter values ​​of different dimensions to the same numerical range, such as between 0 and 1. Taking an industrial production process as an example, the operating parameters include temperature, pressure and raw material ratio, and the optimization goals include product output and quality. 1,000 sets of historical data were collected, with a temperature range of 200-300°C, a pressure range of 1-2MPa, and a raw material ratio range of 0.5-1.5. Normalize these data and scale the temperature, pressure and raw material ratio to between 0 and 1.

[0087] Next, a deep reinforcement learning training network is constructed. The network consists of three parts: a state encoder, an action generator, and a value evaluator. The state encoder encodes the normalized operating parameters into a state feature vector. For example, a multi-layer fully connected neural network can be used to encode the normalized temperature, pressure, and raw material ratio into a 128-dimensional feature vector. The action generator generates adjustment actions based on the state feature vector. For example, another fully connected neural network is used to output three values, corresponding to the adjustment range of temperature, pressure, and raw material ratio, and the adjustment range is limited to between -0.1 and 0.1. The value evaluator evaluates the state action value based on the state feature vector and the adjustment action. For example, a fully connected neural network is used to output a scalar value representing the expected cumulative reward of the current state and action.

[0088] Then, a reward calculation module is constructed. The adjustment action generated by the action generator is applied to the normalized operating parameters to obtain the optimized operating parameters. The optimized operating parameters are input into the multi-objective optimization model to obtain the corresponding optimization target output. For example, the adjusted temperature, pressure and raw material ratio are input into a simulation model that simulates an industrial production process to obtain product output and quality. The immediate reward value is calculated based on the optimization target output. For example, a reward function is set based on output and quality. The higher the output and the better the quality, the greater the reward value. The immediate reward value is combined with the historical reward value to calculate the cumulative reward value. For example, using a discount factor of 0.9, the current immediate reward and the previous cumulative reward are weighted and summed to obtain a new cumulative reward.

[0089] Next, train the deep reinforcement learning training network. Input the state feature vector into the action generator to generate the adjustment action, and input the adjustment action into the value evaluator to obtain the state action value. Calculate the value loss based on the difference between the cumulative reward value and the state action value. Use the back propagation algorithm to update the network parameters of the deep reinforcement learning training network based on the value loss. Repeat the training process until the value loss converges to obtain the trained deep reinforcement learning model. For example, use the Adam optimizer, set the learning rate to 0.001, and train 10,000 times until the value loss no longer decreases.

[0090] Finally, the trained deep reinforcement learning model is used to construct the mapping relationship between the operating parameters and the optimization objectives and generate the optimization prediction results. The operating parameters to be optimized are input into the state encoder to obtain the current state feature vector. The current state feature vector is input into the action generator to obtain the optimization adjustment action. The operating parameters are adjusted based on the optimization adjustment action to generate the optimized operating parameters. The optimized operating parameters are input into the multi-objective optimization model to generate the optimization prediction results. For example, a new set of operating parameters is input, and after being processed by the deep reinforcement learning model, a set of optimized operating parameters is obtained. This set of parameters is then input into the simulation model to obtain the predicted output and quality.

[0091] The solution of this application can:

[0092] Improve optimization efficiency: By learning the optimization strategy of operating parameters through the deep reinforcement learning model, complex calculations and manual parameter adjustment can be avoided, thereby improving optimization efficiency. Improve optimization effect: The deep reinforcement learning model can explore a wider parameter space and find a better combination of operating parameters, thereby improving optimization effect. Enhance generalization ability: The trained deep reinforcement learning model can make predictions under new operating parameter conditions and has a certain generalization ability.

[0093] In an optional implementation, the deviation feature vector is input into a compensation amount calculation model, the compensation amount calculation model is constructed based on a long short-term memory network, and generating a compensation control instruction includes:

[0094] The deviation feature vector is input into a compensation amount calculation model, and the compensation amount calculation model is constructed based on a long short-term memory network; the compensation amount calculation model performs feature preprocessing on the deviation feature vector to generate a normalized feature vector; and the normalized feature vector is time-series sampled to generate a time-series feature sequence containing multiple time steps;

[0095] Input the time series feature sequence into the feature extraction layer of the compensation amount calculation model, the feature extraction layer includes a multi-layer long short-term memory network; the multi-layer long short-term memory network extracts time series features from the time series feature sequence to generate a time series feature representation; perform attention weighting on the time series feature representation to generate a weighted feature representation; input the weighted feature representation into the feature fusion layer, the feature fusion layer adaptively fuses the weighted feature representation with the historical feature to generate a fused feature vector;

[0096] The fused feature vector is input into a compensation amount prediction layer, wherein the compensation amount prediction layer includes a plurality of prediction branches; each prediction branch generates a compensation amount prediction value of a corresponding processing parameter based on the fused feature vector; parameter correlation analysis is performed on the compensation amount prediction values ​​of each prediction branch to establish a parameter compensation association matrix; based on the parameter compensation association matrix, the compensation amount prediction values ​​are jointly optimized to generate an optimized compensation amount;

[0097] The optimized compensation amount is input into the compensation instruction generation layer, and the compensation instruction generation layer converts the optimized compensation amount into the device control amount; the device control amount is time-smoothed to eliminate the sharp fluctuation of the compensation instruction; the smoothed device control amount is packaged according to the standard control protocol format to generate a compensation control instruction.

[0098] The compensation control instruction generation method is used for parameter compensation control of high-precision machining equipment to improve machining accuracy and stability. The core of this method is to use the long short-term memory network to build a compensation amount calculation model, and generate accurate compensation control instructions by multi-level processing and optimization of deviation characteristics.

[0099] In the semiconductor manufacturing process, the deviation feature vector of the etching, deposition and other process is first obtained. The feature vector contains process parameters such as key dimension deviation, film thickness uniformity, and pattern alignment accuracy. The feature vector is normalized to map parameters with different deviations to a unified interval. For example, the etching depth deviation of 300nm is normalized to 0.82, the line width deviation of 50nm is normalized to 0.75, and the alignment deviation of 40nm is normalized to 0.68.

[0100] In view of the continuity of semiconductor processes, the normalized feature vectors are sampled in time series. The sampling window is set to 48 wafers in a lithography batch, and the sliding step is 12 wafers to ensure the data correlation between consecutive batches. Through the sliding window mechanism, a time series feature sequence containing 36 overlapping samples is generated to reflect the gradual trend of process parameters.

[0101] The feature extraction layer uses a three-layer long short-term memory network structure to extract multi-scale temporal features for semiconductor manufacturing. The first layer has 256 network units to capture short-term variation characteristics between wafers; the second layer has 128 network units to extract medium-term variation trends within a batch; the third layer has 64 network units to learn long-term dependencies between batches. Combined with the attention mechanism, different weights are assigned to features at different time steps. For example, the weight of the most recently processed wafer feature is 0.45, and the weight of the previous wafer feature is 0.35.

[0102] The feature fusion layer is designed with a gate control structure that adapts to the characteristics of semiconductor processes to achieve dynamic fusion of new and old features. When process drift is detected, the update gate opening increases to 0.8 and the reset gate opening decreases to 0.2 to enhance the response to new features; when the process state is stable, the update gate opening decreases to 0.3 and the reset gate opening increases to 0.7 to maintain the dominance of historical features. The fusion feature vector dimension is 128, which contains the time-series evolution information of process parameters.

[0103] The compensation prediction layer sets independent prediction branches for key process parameters such as etching power, gas flow, and chamber pressure. Each branch uses a four-layer fully connected network with 64, 32, 16, and 8 hidden layer units. By calculating the correlation coefficients between parameters, a process parameter compensation association matrix is ​​established. For example, the correlation coefficient between the compensation of etching power and gas flow is 0.92, indicating that coordinated adjustment is required. The compensation amount is optimized based on the association constraints to ensure the matching between process parameters.

[0104] The compensation instruction generation layer converts the optimized compensation amount into control instructions that can be recognized by the equipment. The seven-point smoothing filter technology is used to eliminate the sudden change of the compensation amount, and the smoothing window covers 7 consecutive wafer data. According to the semiconductor equipment communication protocol, the control amount is encapsulated into the standard SECS / GEM message format to generate standardized compensation instructions.

[0105] The solution of this application can:

[0106] In terms of process control accuracy: This method improves the critical dimension control accuracy by 35% and the film thickness uniformity by 28% through multi-level timing feature extraction and adaptive feature fusion. In terms of process stability: By adopting parameter association constraints and smoothing processing mechanisms, the fluctuation range of process parameters is reduced by 42%, the repeatability between batches is improved by 30%, and the product yield is significantly improved. In terms of control responsiveness: Based on the parallel prediction branch architecture, the compensation instruction generation delay is reduced by 45%, achieving real-time closed-loop control at the wafer level and improving equipment utilization.

[0107] In an optional implementation, the compensation control instruction is input into the multi-objective optimization model, and an optimal parameter combination is calculated using a multi-objective optimization algorithm, wherein the multi-objective optimization algorithm achieves balanced optimization of the manufacturing cost target, the manufacturing quality target, and the manufacturing efficiency target through dynamic weight allocation; generating a parameter adjustment instruction according to the optimal parameter combination includes:

[0108] Input the compensation control instruction into the multi-objective optimization model, analyze the compensation control instruction, establish the manufacturing cost objective function, the manufacturing quality objective function and the manufacturing efficiency objective function; calculate the dynamic weight coefficient of each objective function based on the processing task characteristics, product quality requirements and production cycle requirements, and generate the objective weight matrix;

[0109] The objective weight matrix is ​​input into a multi-objective optimization algorithm, and the multi-objective optimization algorithm uses a weighted summation method to jointly optimize the manufacturing cost objective function, the manufacturing quality objective function and the manufacturing efficiency objective function to construct a comprehensive objective function; based on the optimization result of the comprehensive objective function, the objective weight matrix is ​​dynamically updated to achieve an adaptive balance among the objectives; under the constraints of process parameters, an improved particle swarm algorithm is used to optimize the comprehensive objective function to generate a candidate parameter combination;

[0110] Performing process constraint verification on the candidate parameter combinations to screen out feasible parameter sets that meet the processing requirements; calculating stability evaluation indicators of each parameter combination in the feasible parameter set, wherein the stability evaluation indicators include parameter volatility, parameter sensitivity, and parameter robustness; comprehensively scoring the feasible parameter set based on the stability evaluation indicators, and selecting the parameter combination with the highest score as the optimal parameter combination;

[0111] The optimal parameter combination is converted into device control parameters, and the device control parameters are time-series smoothed to eliminate drastic fluctuations in the parameter adjustment process; the smoothed device control parameters are data-encapsulated according to the device communication protocol to generate parameter adjustment instructions in a standard format.

[0112] The compensation control instructions are input into the multi-objective optimization model to generate optimized parameter adjustment instructions to achieve balanced optimization of the cost, quality and efficiency of the manufacturing process. The specific steps of this method are as follows:

[0113] In the process of semiconductor film preparation, the deviation feature vectors including key indicators such as film thickness, stress, and roughness are obtained. The feature vectors are normalized and preprocessed so that parameters of different physical dimensions are mapped to a unified interval. For example, in actual cases, the thickness deviation of 250 angstroms is normalized to 0.78, the stress deviation of 85 MPa is normalized to 0.82, and the surface roughness deviation of 0.8 nanometers is normalized to 0.71.

[0114] In view of the characteristics of the thin film deposition process, a sliding window mechanism is used for time series sampling. The sampling window is set to contain data from 60 deposition cycles, and the sliding step is 15 cycles. Through overlapping sampling, 45 cycles of continuous data are retained to form a time series feature sequence that reflects the gradual trend of the process. This sequence can effectively capture the dynamic change characteristics during the thin film growth process.

[0115] The feature extraction layer uses a four-layer long short-term memory network to perform hierarchical extraction of multi-scale temporal features of thin film growth. The first layer has 320 units to extract the instantaneous features of the deposition process of a single wafer; the second layer has 240 units to obtain short-term changes within a single batch; the third layer has 160 units to capture the medium-term trends between multiple batches; the fourth layer has 80 units to learn the long-term evolution of process drift. Combined with the attention mechanism, corresponding weights are given to features of different time scales, such as a weight of 0.42 for the most recent cycle feature and 0.33 for the previous cycle feature.

[0116] The feature fusion layer is designed with a bidirectional gating structure to achieve intelligent fusion of new and old features. When a change in the cavity state is detected, the update gate opening is increased to 0.85, and the reset gate opening is reduced to 0.15 to enhance the response sensitivity to new features; when the process state is stable, the update gate opening is reduced to 0.25, and the reset gate opening is increased to 0.75 to maintain the dominant position of historical features. The dimension of the fused feature vector is 160, which contains the complete evolution information of the process parameters.

[0117] The compensation prediction layer sets independent prediction branches for core process parameters such as RF power, gas flow, and chamber temperature. Each branch consists of a five-layer fully connected network with 80, 64, 48, 32, and 16 hidden units. The compensation correlation matrix is ​​constructed by calculating the correlation coefficients between parameters. For example, the correlation coefficient between RF power and gas flow compensation is 0.95, indicating that coordinated adjustment is required. The compensation of each parameter is optimized based on the association constraints to ensure the matching and coordination between process parameters.

[0118] The compensation instruction generation layer converts the optimized compensation amount into the equipment control amount. A nine-point smoothing filter is used for processing, and the smoothing window covers 9 consecutive deposition cycle data, which effectively suppresses the violent fluctuation of the compensation instruction. According to the semiconductor equipment communication specification, the control amount is encapsulated into the standard SECS / GEM protocol format to generate standardized compensation instructions.

[0119] The solution of this application can:

[0120] In terms of film quality: through multi-level timing feature extraction and adaptive feature fusion mechanism, the uniformity of film thickness is improved by 38%, the stress control accuracy is improved by 32%, and the surface roughness is reduced by 25%. In terms of process stability: based on parameter association constraints and smoothing processing strategies, the intra-batch parameter fluctuation is reduced by 45%, the consistency between batches is improved by 35%, and the stability of continuous operation of the equipment is significantly improved. In terms of production efficiency: using a parallel prediction branch architecture, the delay of compensation instruction generation is reduced by 48%, achieving sub-second real-time control response, increasing production capacity by 22%, and equipment utilization by 25%.

[0121] In an optional implementation, under the process parameter constraint condition, the improved particle swarm algorithm is used to optimize the comprehensive objective function, and the candidate parameter combinations generated include:

[0122] Constructing process parameter constraints, wherein the process parameter constraints include temperature field uniformity constraints, pressure fluctuation amplitude constraints, gas ratio range constraints, and plasma density distribution constraints; constructing a comprehensive objective function based on the process parameter constraints, wherein the comprehensive objective function is in the form of weighted summation and includes process parameter stability indicators, energy consumption indicators, and product quality indicators;

[0123] The comprehensive objective function is improved by a particle swarm algorithm, and an adaptive inertia weight strategy is set, wherein the adaptive inertia weight decreases nonlinearly with the number of iterations; a chaotic perturbation mechanism is introduced, wherein the chaotic perturbation mechanism adopts a V-type chaotic mapping; a position update rule of a constraint violation degree is established, wherein the constraint violation degree is calculated based on the process parameter constraint condition;

[0124] Determine the search space according to the process parameter constraint conditions, encode the process parameters into a position vector, the position vector includes a temperature setting value, a pressure setting value, a gas flow setting value, a radio frequency power setting value and a plasma generator parameter; initialize the position and velocity of the particle group using the chaotic perturbation mechanism;

[0125] Calculate the constraint violation degree of the initial particle swarm according to the process parameter constraint condition, and combine the constraint violation degree with the comprehensive objective function to generate a fitness function; evaluate the initial particle swarm based on the fitness function to determine the individual optimal position and the global optimal position;

[0126] In the iterative optimization process, the new velocity of the particle is calculated based on the adaptive inertia weight strategy; the new position of the particle is adjusted according to the position update rule of the constraint violation degree; the fitness function value corresponding to the new position is calculated; when the global optimal position is not updated after a preset number of consecutive iterations, the particle position is adjusted using the chaotic perturbation mechanism;

[0127] Recalculate the constraint violation degree and fitness function value for each iteratively updated particle position; update the individual optimal position and the global optimal position based on the fitness function value; when the termination condition is reached, extract the corresponding temperature setting value, pressure setting value, gas flow setting value, radio frequency power setting value and plasma generator parameter from the global optimal position that satisfies the process parameter constraint condition to generate a candidate parameter combination.

[0128] Process parameter optimization method is used to find the best process parameter combination under multiple constraints.

[0129] First, construct the process parameter constraints. The constraints include four aspects: temperature field uniformity constraint, which sets the maximum and minimum temperature difference range of the temperature field in space; pressure fluctuation amplitude constraint, which sets the maximum and minimum range of the pressure fluctuation amplitude; gas ratio range constraint, which sets the ratio range of various gas components; plasma density distribution constraint, which sets the distribution range of plasma density in space. For example, the temperature field uniformity constraint is that the temperature difference is between 5°C and 15°C, the pressure fluctuation amplitude constraint is less than 10kPa, the gas ratio range constraint is that component A accounts for 20% to 30%, and component B accounts for 70% to 80%, and the plasma density distribution constraint is that the density in the center area is higher than that in the edge area.

[0130] Then, based on the constructed process parameter constraints, a comprehensive objective function is constructed. The comprehensive objective function is in the form of weighted summation and contains three indicators: process parameter stability indicators, such as the standard deviation of temperature and pressure; energy consumption indicators, such as total power consumption; product quality indicators, such as product yield and defect rate. Each indicator is assigned a weight to reflect its importance. For example, the weight of the process parameter stability indicator is 0.4, the weight of the energy consumption indicator is 0.3, and the weight of the product quality indicator is 0.3.

[0131] Next, the particle swarm algorithm is improved to better optimize the comprehensive objective function. First, an adaptive inertia weight strategy is set. The inertia weight decreases with the increase in the number of iterations. The initial weight is larger, which is conducive to global search, and the later weight is smaller, which is conducive to local search. For example, the initial inertia weight is set to 0.9, and the final inertia weight is set to 0.4, using a nonlinear decreasing method, such as exponential decreasing. Secondly, a chaotic perturbation mechanism is introduced, and a V-type chaotic mapping is used. When a particle falls into a local optimum, it is caused to jump out of the local optimum through chaotic perturbation and continue to search for the global optimal solution. Thirdly, a position update rule for constraint violation is established. The constraint violation is calculated based on the process parameter constraints. If the particle position violates the constraints, it will be punished to move to the feasible area.

[0132] Determine the search space and encode the process parameters into a position vector. The position vector contains five parameters: temperature setting value, pressure setting value, gas flow setting value, RF power setting value, and plasma generator parameters. For example, the temperature setting value ranges from 800℃ to 900℃, the pressure setting value ranges from 100kPa to 200kPa, the gas flow setting value ranges from 10sccm to 20sccm, the RF power setting value ranges from 500W to 1000W, and the plasma generator parameter ranges from 1 to 10. The position and velocity of the particle swarm are initialized using a chaotic perturbation mechanism.

[0133] The constraint violation degree of the initial particle swarm is calculated according to the process parameter constraints, and the constraint violation degree is combined with the comprehensive objective function to generate a fitness function. The initial particle swarm is evaluated based on the fitness function to determine the individual optimal position and the global optimal position.

[0134] In the iterative optimization process, the new velocity of the particle is calculated based on the adaptive inertia weight strategy. The new position of the particle is adjusted according to the position update rule of the constraint violation degree. The fitness function value corresponding to the new position is calculated. When the global optimal position is not updated after a preset number of consecutive iterations (for example, 50 times), the chaotic perturbation mechanism is used to adjust the particle position.

[0135] Recalculate the constraint violation and fitness function value for each updated particle position. Update the individual optimal position and the global optimal position based on the fitness function value. When the termination condition is reached (for example, the maximum number of iterations is 200), extract the corresponding temperature setting value, pressure setting value, gas flow setting value, RF power setting value and plasma generator parameters from the global optimal position that meets the process parameter constraints to generate candidate parameter combinations. For example, the final candidate parameter combination is: temperature 850℃, pressure 150kPa, gas flow 15sccm, RF power 750W, plasma generator parameter 5.

[0136] The solution of this application can:

[0137] Improve the efficiency of process parameter optimization: Through adaptive inertia weight and chaotic perturbation mechanism, the particle swarm algorithm can be effectively prevented from falling into local optimality, and the global search capability can be improved, so as to find the optimal process parameter combination more quickly. Ensure that the process parameters meet the constraints: Through the position update rule of constraint violation degree, it can be ensured that the candidate parameter combination generated during the optimization process always meets the preset process parameter constraints, avoiding parameter combinations that do not meet the actual production situation. Improve product quality and reduce energy consumption: By comprehensively considering multiple indicators such as process parameter stability, energy consumption and product quality, and integrating them into the comprehensive objective function, the process parameters can be effectively optimized, and ultimately the purpose of improving product quality and reducing energy consumption can be achieved.

[0138] In an optional embodiment, the parameter adjustment instruction is sent to a corresponding execution unit, the execution unit includes a heater, a pressure control valve, a gas flow controller, a radio frequency generator and a plasma generator, and the adaptive feedback control of the semiconductor manufacturing process includes:

[0139] The parameter adjustment instruction includes process parameter target value information, and the process parameter target value information is decomposed according to the control characteristics of the execution unit to generate a heater control instruction, a pressure control valve control instruction, a gas flow controller control instruction, a radio frequency generator control instruction, and a plasma generator control instruction;

[0140] The heater control instruction is sent to the heater, and the heater adjusts the heating power output according to the temperature feedback signal; the pressure control valve control instruction is sent to the pressure control valve, and the pressure control valve adjusts the valve opening according to the pressure feedback signal; the gas flow controller control instruction is sent to the gas flow controller, and the gas flow controller adjusts the gas flow according to the flow feedback signal;

[0141] The RF generator control instruction is sent to the RF generator, and the RF generator adjusts the RF output power according to the RF power feedback signal; the plasma generator control instruction is sent to the plasma generator, and the plasma generator adjusts the bias voltage according to the plasma density feedback signal;

[0142] Collecting a temperature feedback signal of the heater, a pressure feedback signal of the pressure control valve, a flow feedback signal of the gas flow controller, a power feedback signal of the radio frequency generator, and a density feedback signal of the plasma generator; and constructing a process parameter evaluation model based on the feedback signals;

[0143] Comparing the output result of the process parameter evaluation model with the process parameter target value information to generate a parameter deviation matrix; calculating the adjustment compensation amount of each execution unit based on the parameter deviation matrix, and the adjustment compensation amount is used to correct the control instructions of each execution unit;

[0144] The heater control instruction, the pressure control valve control instruction, the gas flow controller control instruction, the radio frequency generator control instruction and the plasma generator control instruction are optimized online according to the adjustment compensation amount to generate optimized parameter adjustment instructions; the optimized parameter adjustment instructions are resent to each execution unit to realize adaptive feedback control of process parameters in a semiconductor manufacturing process.

[0145] Precise control of process parameters in the semiconductor manufacturing process is crucial to the quality of the final product. In order to achieve more precise and real-time control, this embodiment proposes a semiconductor manufacturing process parameter control method based on adaptive feedback regulation.

[0146] First, set the target values ​​of process parameters. For example, in a certain deposition step, the target temperature is 700°C, the target pressure is 10Pa, the target gas flow rate is 50sccm, the target RF power is 500W, and the target plasma density is 10^12cm. These target values ​​will serve as the benchmark for subsequent control.

[0147] Then, these target values ​​are decomposed into control instructions for the corresponding execution units. Considering the control characteristics of each execution unit, such as the heating rate of the heater and the response time of the pressure control valve, the target temperature of 700°C is converted to the initial heating power of the heater, assuming it is 800W; the target pressure of 10Pa is converted to the initial opening of the pressure control valve, assuming it is 30%; the target gas flow rate of 50sccm is converted to the initial setting value of the gas flow controller, assuming it is 50; the target RF power of 500W is converted to the initial output power of the RF generator, assuming it is 500W.

[0148] Next, these control instructions are sent to the corresponding execution units. The heater receives the 800W heating power instruction and starts heating, the pressure control valve adjusts the valve according to the 30% opening instruction, the gas flow controller adjusts the gas flow according to the set value 50, the RF generator outputs 500W RF power, and the plasma generator applies a -100V bias voltage.

[0149] Based on the collected feedback signals, a process parameter evaluation model is constructed. This model can be a simple linear model, a more complex nonlinear model, or even a model based on machine learning. For example, a relationship model between temperature and heating power can be established based on historical data.

[0150] The output of the evaluation model is compared with the set target value to generate a parameter deviation matrix. For example, if the target temperature is 700°C and the current temperature is 650°C, the temperature deviation is 50°C. Similarly, the pressure deviation, flow deviation, power deviation, and density deviation can be obtained.

[0151] The adjustment compensation amount of each execution unit is calculated according to the parameter deviation matrix. For example, based on the temperature deviation of 50°C and the established temperature and heating power relationship model, the required heating power is calculated, assuming it is 50W. Similarly, the adjustment amount of the pressure control valve opening, the adjustment amount of the gas flow, the adjustment amount of the RF power and the adjustment amount of the bias voltage can be calculated.

[0152] Based on the calculated adjustment compensation, the control instructions of each execution unit are optimized online. For example, the control instruction of the heater is adjusted from 800W to 850W, the opening instruction of the pressure control valve is adjusted from 30% to 32%, and so on.

[0153] Finally, the optimized control instructions are resent to each execution unit to achieve adaptive feedback control of process parameters, and this process is repeated continuously until the set target value is reached.

[0154] The solution of this application can:

[0155] Improve the control accuracy of process parameters: Through real-time feedback and adaptive adjustment, process parameters can be controlled more accurately, deviations can be reduced, and product quality and consistency can be improved. Enhance process stability: Adaptive feedback control can effectively suppress process fluctuations, improve process stability, and reduce uncertainty in the production process. Optimize process efficiency: Through precise control and real-time adjustment of process parameters, the process flow can be optimized, the production cycle can be shortened, and production efficiency can be improved.

[0156] In the second aspect of the present invention,

[0157] Figure 2 FIG. 1 is a schematic diagram of a control system of a semiconductor manufacturing machine according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0158] The first unit is used to obtain the operating parameters of the equipment chamber in the semiconductor manufacturing process, wherein the operating parameters include temperature parameters, pressure parameters, gas flow parameters, radio frequency power parameters and plasma density parameters; establish a multi-objective optimization model including manufacturing cost targets, manufacturing quality targets and manufacturing efficiency targets, wherein the manufacturing cost target is composed of energy consumption per unit time and raw material utilization rate, the manufacturing quality target is composed of process uniformity index and defect density index, and the manufacturing efficiency target is composed of equipment utilization rate and product output rate; use a deep reinforcement learning algorithm to train the multi-objective optimization model, construct a mapping relationship between the operating parameters and the optimization targets, and generate an optimization prediction result;

[0159] The second unit is used to perform deviation analysis between the optimization prediction result and the preset target parameter; construct a deviation feature vector based on the deviation analysis result; input the deviation feature vector into a compensation amount calculation model, the compensation amount calculation model is constructed based on a long short-term memory network, and generates a compensation control instruction, the compensation control instruction includes a temperature compensation amount, a pressure compensation amount, a gas flow compensation amount, a radio frequency power compensation amount, and a plasma density compensation amount;

[0160] The third unit is used to input the compensation control instruction into the multi-objective optimization model, and use a multi-objective optimization algorithm to calculate the optimal parameter combination, wherein the multi-objective optimization algorithm realizes the balanced optimization of the manufacturing cost target, the manufacturing quality target and the manufacturing efficiency target through dynamic weight allocation; generate parameter adjustment instructions according to the optimal parameter combination, and the parameter adjustment instructions include temperature adjustment amount, pressure adjustment amount, gas flow adjustment amount, radio frequency power adjustment amount and plasma density adjustment amount; send the parameter adjustment instructions to the corresponding execution unit, and the execution unit includes a heater, a pressure control valve, a gas flow controller, a radio frequency generator and a plasma generator to realize adaptive feedback control of the semiconductor manufacturing process.

[0161] According to a third aspect of an embodiment of the present invention, there is provided an electronic device, including: a processor;

[0162] a memory for storing processor-executable instructions;

[0163] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0164] A fourth aspect of the embodiments of the present invention is:

[0165] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0166] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0167] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A control method for a semiconductor manufacturing machine, characterized in that: include: Acquire the operating parameters of the equipment chamber in the semiconductor manufacturing process, wherein the operating parameters include temperature parameters, pressure parameters, gas flow parameters, radio frequency power parameters, and plasma density parameters; establish a multi-objective optimization model including manufacturing cost targets, manufacturing quality targets, and manufacturing efficiency targets, wherein the manufacturing cost target is composed of energy consumption per unit time and raw material utilization rate, the manufacturing quality target is composed of process uniformity index and defect density index, and the manufacturing efficiency target is composed of equipment utilization rate and product output rate; use a deep reinforcement learning algorithm to train the multi-objective optimization model, construct a mapping relationship between the operating parameters and the optimization targets, and generate optimization prediction results; Perform deviation analysis on the optimization prediction result and the preset target parameter; construct a deviation feature vector based on the deviation analysis result; input the deviation feature vector into a compensation amount calculation model, the compensation amount calculation model is constructed based on a long short-term memory network, and generates a compensation control instruction, the compensation control instruction includes a temperature compensation amount, a pressure compensation amount, a gas flow compensation amount, a radio frequency power compensation amount, and a plasma density compensation amount; Inputting the compensation control instruction into the multi-objective optimization model, using a multi-objective optimization algorithm to calculate the optimal parameter combination, the multi-objective optimization algorithm realizes the balanced optimization of the manufacturing cost target, the manufacturing quality target and the manufacturing efficiency target through dynamic weight allocation; generating a parameter adjustment instruction according to the optimal parameter combination, the parameter adjustment instruction including a temperature adjustment amount, a pressure adjustment amount, a gas flow adjustment amount, a radio frequency power adjustment amount and a plasma density adjustment amount; sending the parameter adjustment instruction to a corresponding execution unit, the execution unit including a heater, a pressure control valve, a gas flow controller, a radio frequency generator and a plasma generator, to realize adaptive feedback control of the semiconductor manufacturing process; Constructing a deep reinforcement learning training network, the deep reinforcement learning training network includes a state encoder, an action generator and a value evaluator; the state encoder encodes the normalized operating parameters into a state feature vector; the action generator generates an adjustment action based on the state feature vector; the value evaluator evaluates the state action value based on the state feature vector and the adjustment action; Constructing a reward calculation module, applying the adjustment action to the normalized operating parameters to obtain optimized operating parameters; Inputting the optimized operating parameters into the multi-objective optimization model to obtain corresponding optimization target output; Calculate the instant reward value based on the optimization target output, and calculate the cumulative reward value in combination with the historical reward value; Training the deep reinforcement learning training network: inputting the state feature vector into the action generator to generate an adjustment action; inputting the adjustment action into the value evaluator to obtain the state action value; calculating the value loss based on the difference between the accumulated reward value and the state action value; updating the network parameters of the deep reinforcement learning training network based on the value loss; repeating the training process until the value loss converges to obtain a trained deep reinforcement learning model; The deep reinforcement learning model is used to construct a mapping relationship between the operating parameters and the optimization objectives: the operating parameters to be optimized are input into the state encoder to obtain a current state feature vector; the current state feature vector is input into the action generator to obtain an optimization adjustment action; the operating parameters are adjusted based on the optimization adjustment action to generate optimized operating parameters; the optimized operating parameters are input into the multi-objective optimization model to generate optimization prediction results.

2. The method according to claim 1, characterized in that: A multi-objective optimization model including manufacturing cost target, manufacturing quality target and manufacturing efficiency target is established. The manufacturing cost target is composed of energy consumption per unit time and raw material utilization rate, the manufacturing quality target is composed of process uniformity index and defect density index, and the manufacturing efficiency target is composed of equipment utilization rate and product output rate. Collect heater power data, RF power data and vacuum pump power data as energy data, collect air intake data and actual reaction gas volume data as material data, collect film thickness data and etching depth data as process data, collect equipment operation status data and product inspection data as production data; Calculate the energy consumption per unit time based on the energy data, accumulate the heater power data, the RF power data and the vacuum pump power data within a unit time to obtain the total power data, and integrate the total power data to obtain the energy consumption per unit time; calculate the raw material utilization rate based on the material data, calculate the raw material utilization rate by the ratio of the actual reaction gas volume data to the intake gas volume data; normalize the energy consumption per unit time and the raw material utilization rate and weightedly combine them to construct a manufacturing cost target; Calculating a process uniformity index based on the process data, selecting a detection point on the wafer surface to obtain the film thickness data and the etching depth data, and calculating the standard deviation of the film thickness data and the etching depth data to obtain a process uniformity index; calculating a defect density index based on the process data, preprocessing a defect detection image in the process data to obtain an enhanced image, extracting defect features in the enhanced image, and counting the number of defects per unit area based on the defect features to obtain a defect density index; normalizing the process uniformity index and the defect density index and weightedly combining them to construct a manufacturing quality target; The equipment utilization rate is calculated based on the production data, the equipment operation status data is divided into production time data, standby time data, maintenance time data and fault time data, and the ratio of the production time data to the total time is calculated to obtain the equipment utilization rate; the product output rate is calculated based on the production data, the product inspection data is analyzed, and the number of qualified products per unit time is counted to obtain the product output rate; the equipment utilization rate and the product output rate are normalized and weightedly combined to construct a manufacturing efficiency target.

3. The method according to claim 1, characterized in that: The deviation feature vector is input into a compensation amount calculation model, the compensation amount calculation model is constructed based on a long short-term memory network, and the compensation control instruction is generated including: The deviation feature vector is input into a compensation amount calculation model, and the compensation amount calculation model is constructed based on a long short-term memory network; the compensation amount calculation model performs feature preprocessing on the deviation feature vector to generate a normalized feature vector; and the normalized feature vector is time-series sampled to generate a time-series feature sequence containing multiple time steps; Input the time series feature sequence into the feature extraction layer of the compensation amount calculation model, the feature extraction layer includes a multi-layer long short-term memory network; the multi-layer long short-term memory network extracts time series features from the time series feature sequence to generate a time series feature representation; perform attention weighting on the time series feature representation to generate a weighted feature representation; input the weighted feature representation into the feature fusion layer, the feature fusion layer adaptively fuses the weighted feature representation with the historical feature to generate a fused feature vector; The fused feature vector is input into a compensation amount prediction layer, wherein the compensation amount prediction layer includes a plurality of prediction branches; each prediction branch generates a compensation amount prediction value of a corresponding processing parameter based on the fused feature vector; parameter correlation analysis is performed on the compensation amount prediction values ​​of each prediction branch to establish a parameter compensation association matrix; based on the parameter compensation association matrix, the compensation amount prediction values ​​are jointly optimized to generate an optimized compensation amount; The optimized compensation amount is input into the compensation instruction generation layer, and the compensation instruction generation layer converts the optimized compensation amount into the device control amount; the device control amount is time-smoothed to eliminate the sharp fluctuation of the compensation instruction; the smoothed device control amount is packaged according to the standard control protocol format to generate a compensation control instruction.

4. The method according to claim 1, characterized in that: Inputting the compensation control instruction into the multi-objective optimization model, and using a multi-objective optimization algorithm to calculate an optimal parameter combination, wherein the multi-objective optimization algorithm achieves balanced optimization of the manufacturing cost target, the manufacturing quality target, and the manufacturing efficiency target through dynamic weight allocation; Generating a parameter adjustment instruction according to the optimal parameter combination includes: Input the compensation control instruction into the multi-objective optimization model, analyze the compensation control instruction, establish the manufacturing cost objective function, the manufacturing quality objective function and the manufacturing efficiency objective function; calculate the dynamic weight coefficient of each objective function based on the processing task characteristics, product quality requirements and production cycle requirements, and generate the objective weight matrix; The objective weight matrix is ​​input into a multi-objective optimization algorithm, and the multi-objective optimization algorithm uses a weighted summation method to jointly optimize the manufacturing cost objective function, the manufacturing quality objective function and the manufacturing efficiency objective function to construct a comprehensive objective function; based on the optimization result of the comprehensive objective function, the objective weight matrix is ​​dynamically updated to achieve an adaptive balance among the objectives; under the constraints of process parameters, an improved particle swarm algorithm is used to optimize the comprehensive objective function to generate a candidate parameter combination; Performing process constraint verification on the candidate parameter combinations to screen out feasible parameter sets that meet the processing requirements; calculating stability evaluation indicators of each parameter combination in the feasible parameter set, wherein the stability evaluation indicators include parameter volatility, parameter sensitivity, and parameter robustness; comprehensively scoring the feasible parameter set based on the stability evaluation indicators, and selecting the parameter combination with the highest score as the optimal parameter combination; The optimal parameter combination is converted into device control parameters, and the device control parameters are time-series smoothed to eliminate drastic fluctuations in the parameter adjustment process; the smoothed device control parameters are data-encapsulated according to the device communication protocol to generate parameter adjustment instructions in a standard format.

5. The method according to claim 4, characterized in that Under the constraints of process parameters, the improved particle swarm algorithm is used to optimize the comprehensive objective function and generate candidate parameter combinations including: Constructing process parameter constraints, wherein the process parameter constraints include temperature field uniformity constraints, pressure fluctuation amplitude constraints, gas ratio range constraints, and plasma density distribution constraints; constructing a comprehensive objective function based on the process parameter constraints, wherein the comprehensive objective function is in the form of weighted summation and includes process parameter stability indicators, energy consumption indicators, and product quality indicators; The comprehensive objective function is improved by a particle swarm algorithm, and an adaptive inertia weight strategy is set, wherein the adaptive inertia weight decreases nonlinearly with the number of iterations; a chaotic perturbation mechanism is introduced, wherein the chaotic perturbation mechanism adopts a V-type chaotic mapping; a position update rule of a constraint violation degree is established, wherein the constraint violation degree is calculated based on the process parameter constraint condition; Determine the search space according to the process parameter constraint conditions, encode the process parameters into a position vector, the position vector includes a temperature setting value, a pressure setting value, a gas flow setting value, a radio frequency power setting value and a plasma generator parameter; initialize the position and velocity of the particle group using the chaotic perturbation mechanism; Calculate the constraint violation degree of the initial particle swarm according to the process parameter constraint condition, and combine the constraint violation degree with the comprehensive objective function to generate a fitness function; evaluate the initial particle swarm based on the fitness function to determine the individual optimal position and the global optimal position; In the iterative optimization process, the new velocity of the particle is calculated based on the adaptive inertia weight strategy; the new position of the particle is adjusted according to the position update rule of the constraint violation degree; the fitness function value corresponding to the new position is calculated; when the global optimal position is not updated after a preset number of consecutive iterations, the particle position is adjusted using the chaotic perturbation mechanism; Recalculate the constraint violation degree and fitness function value for each iteratively updated particle position; update the individual optimal position and the global optimal position based on the fitness function value; when the termination condition is reached, extract the corresponding temperature setting value, pressure setting value, gas flow setting value, radio frequency power setting value and plasma generator parameter from the global optimal position that satisfies the process parameter constraint condition to generate a candidate parameter combination.

6. The method according to claim 1, characterized in that The parameter adjustment instruction is sent to a corresponding execution unit, wherein the execution unit includes a heater, a pressure control valve, a gas flow controller, a radio frequency generator, and a plasma generator, and the adaptive feedback control of the semiconductor manufacturing process is realized by: The parameter adjustment instruction includes process parameter target value information, and the process parameter target value information is decomposed according to the control characteristics of the execution unit to generate a heater control instruction, a pressure control valve control instruction, a gas flow controller control instruction, a radio frequency generator control instruction, and a plasma generator control instruction; The heater control instruction is sent to the heater, and the heater adjusts the heating power output according to the temperature feedback signal; the pressure control valve control instruction is sent to the pressure control valve, and the pressure control valve adjusts the valve opening according to the pressure feedback signal; the gas flow controller control instruction is sent to the gas flow controller, and the gas flow controller adjusts the gas flow according to the flow feedback signal; The RF generator control instruction is sent to the RF generator, and the RF generator adjusts the RF output power according to the RF power feedback signal; the plasma generator control instruction is sent to the plasma generator, and the plasma generator adjusts the bias voltage according to the plasma density feedback signal; Collecting a temperature feedback signal of the heater, a pressure feedback signal of the pressure control valve, a flow feedback signal of the gas flow controller, a power feedback signal of the radio frequency generator, and a density feedback signal of the plasma generator; and constructing a process parameter evaluation model based on the feedback signals; Comparing the output result of the process parameter evaluation model with the process parameter target value information to generate a parameter deviation matrix; calculating the adjustment compensation amount of each execution unit based on the parameter deviation matrix, and the adjustment compensation amount is used to correct the control instructions of each execution unit; The heater control instruction, the pressure control valve control instruction, the gas flow controller control instruction, the radio frequency generator control instruction and the plasma generator control instruction are optimized online according to the adjustment compensation amount to generate optimized parameter adjustment instructions; the optimized parameter adjustment instructions are resent to each execution unit to realize adaptive feedback control of process parameters in a semiconductor manufacturing process.

7. A control system for a semiconductor manufacturing machine, used to implement the method according to any one of claims 1 to 6, characterized in that: include: The first unit is used to obtain the operating parameters of the equipment chamber in the semiconductor manufacturing process, wherein the operating parameters include temperature parameters, pressure parameters, gas flow parameters, radio frequency power parameters and plasma density parameters; establish a multi-objective optimization model including manufacturing cost targets, manufacturing quality targets and manufacturing efficiency targets, wherein the manufacturing cost target is composed of energy consumption per unit time and raw material utilization rate, the manufacturing quality target is composed of process uniformity index and defect density index, and the manufacturing efficiency target is composed of equipment utilization rate and product output rate; use a deep reinforcement learning algorithm to train the multi-objective optimization model, construct a mapping relationship between the operating parameters and the optimization targets, and generate an optimization prediction result; The second unit is used to perform deviation analysis between the optimization prediction result and the preset target parameter; construct a deviation feature vector based on the deviation analysis result; input the deviation feature vector into a compensation amount calculation model, the compensation amount calculation model is constructed based on a long short-term memory network, and generates a compensation control instruction, the compensation control instruction includes a temperature compensation amount, a pressure compensation amount, a gas flow compensation amount, a radio frequency power compensation amount, and a plasma density compensation amount; The third unit is used to input the compensation control instruction into the multi-objective optimization model, and use a multi-objective optimization algorithm to calculate the optimal parameter combination, wherein the multi-objective optimization algorithm realizes the balanced optimization of the manufacturing cost target, the manufacturing quality target and the manufacturing efficiency target through dynamic weight allocation; generate parameter adjustment instructions according to the optimal parameter combination, and the parameter adjustment instructions include temperature adjustment amount, pressure adjustment amount, gas flow adjustment amount, radio frequency power adjustment amount and plasma density adjustment amount; send the parameter adjustment instructions to the corresponding execution unit, and the execution unit includes a heater, a pressure control valve, a gas flow controller, a radio frequency generator and a plasma generator to realize adaptive feedback control of the semiconductor manufacturing process.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images