An intelligent bias voltage regulation method and system for heteroepitaxial growth of single-crystal diamond

Abstract

The present invention relates to the technical field of diamond epitaxial growth, and discloses an intelligent bias voltage regulation method and system for diamond single crystal heteroepitaxial growth, including a data acquisition module, a prediction module, an intelligent optimization module, a gradient field compensation module, and a feedback and regulation module; by using a prediction model to predict future trends, potential growth anomalies such as a decrease in the nucleation probability in a certain area can be sensed in advance, making the control system forward-looking and avoiding the lag of passive compensation afterwards; the reinforcement learning controller comprehensively considers the cumulative effect and long-term goal during the strategy optimization process, making the bias voltage regulation more intelligent; the intelligent decision-making ability is far superior to the adjustment methods of manual experience or preset programs, making the growth process more robust to external disturbances such as air flow fluctuations and discharge state changes; maintaining the consistency of the crystal nuclei and growth rates in each area during the entire growth process is beneficial to suppressing the accumulation of thermal stress and defects in large-size thin films.

Description

Technical Field

[0001] The present invention relates to the technical field of diamond epitaxial growth, and more particularly to an intelligent bias voltage regulation method and system for heteroepitaxial growth of single crystal diamond. Background Art

[0002] Single crystal diamond has excellent properties in mechanics, acoustics, thermotics, optoelectronics, etc., and has broad application prospects in fields such as high-power electronic devices, quantum devices, and optical windows; growing large-size single crystal diamond films on substrates such as iridium (Ir) / yttrium-stabilized zirconia (YSZ) / silicon by heteroepitaxy is an important way to achieve large-area single crystal diamond at present; however, there are still many challenges in realizing large-size heteroepitaxial diamond in practice; for example, due to the large lattice mismatch and difference in thermal expansion coefficient between diamond and common substrate materials, high-density defects and internal stress up to several GPa are easily introduced in the diamond epitaxial layer with a thickness of hundreds of micrometers, which may cause cracking of the grown diamond film; in addition, when performing plasma chemical vapor deposition (CVD) growth on a large-size substrate, the plasma state and temperature field in the center and edge of the reaction zone are often uneven; this non-uniformity will result in a difference in nucleation density between the center region and the edge region of the substrate. If not controlled, it may lead to non-uniform thickness and quality of the grown diamond film, affecting the yield of device fabrication.

[0003] In order to improve the uniformity of heteroepitaxial single crystal diamond growth, researchers often use bias-enhanced nucleation (BEN) technology to apply a bias voltage to the substrate at the initial stage of growth to increase the nucleation density; however, traditional bias voltage regulation usually remains constant during the growth process or is adjusted according to a preset program, lacking the ability to feedback-regulate the real-time state of the growth process; when the substrate size expands to the inch level, the fixed bias voltage scheme is difficult to adapt to the plasma environment differences that change with time in the center and edge regions, resulting in a large deviation in nucleation density between different regions; according to existing experience estimates, without special control measures, the difference in nucleation density between the center and edge of a large-size substrate can reach about ±15%; this difference will further affect the consistency of grain growth and is not conducive to obtaining high-quality large-area single crystal diamond films.

[0004] In view of this, the present invention provides an intelligent bias voltage regulation method and system for heteroepitaxial growth of single crystal diamond, which can realize intelligent dynamic regulation of the bias voltage to significantly reduce the difference in nucleation density between the center and edge regions of the substrate and improve the uniformity and controllability of the epitaxial growth process. Summary of the Invention

[0005] To overcome the above-mentioned defects of the prior art, the present invention provides a method and system for intelligent bias voltage regulation in the heteroepitaxial growth of single-crystal diamond to solve the problems existing in the above-mentioned background technology.

[0006] The present invention provides the following technical solutions: A method for intelligent bias voltage regulation in the heteroepitaxial growth of single-crystal diamond, comprising the following steps:

[0007] Step S1: Collect characteristic spectral data of the plasma through an inductively coupled plasma optical emission spectrometer, and collect substrate surface thermal radiation data through a thermal radiation sensor to obtain characteristic information on the substrate state;

[0008] Step S2: Construct a bias voltage regulation model based on reinforcement learning, obtain a reward function using the characteristic information on the substrate state, and dynamically optimize the bias voltage output of each region of the substrate during the epitaxial growth process according to the reward function;

[0009] Step S3: Use a dual-channel LSTM network to fuse and process the characteristic spectral data and the substrate surface thermal radiation data to predict the nucleation probability of each local region of the substrate in real time;

[0010] Step S4: When it is detected that the electron density in the substrate edge region is lower than a preset threshold, automatically generate an asymmetric bias voltage waveform and apply a compensation adjustment to the edge region;

[0011] Step S5: Based on the real-time feedback information during the epitaxial growth process, perform closed-loop regulation on the bias voltage output.

[0012] Preferably, the reward function is defined as the total reward with a discount factor accumulated at each time step t, and the formula is expressed as:

[0013] ;

[0014] Where, represents the reward function, represents the immediate reward corresponding to the deviation between the nucleation density in the central region and the preset target value at time step t; represents the immediate reward corresponding to the deviation between the nucleation density in the edge region and the preset target value at time step t; and are the weight coefficients of the central region and the edge region respectively, represents the discount factor, and its value range satisfies ; is the total number of time steps.

[0015] Preferably, in the step S3, the dual-channel LSTM network prediction model has a dual-input channel structure. The first channel inputs the spectral intensity sequence of the plasma, and the second channel inputs the substrate temperature sequence. The hidden states of the two channels are fused in the network to output the nucleation probability estimation value at each spatial position corresponding to a certain time.

[0016] The prediction steps of the dual-channel LSTM network prediction model are as follows:

[0017] The first channel receives the characteristic spectral data of the plasma and extracts spectral features to predict the initial nucleation probability value of each local area.

[0018] The second channel receives the substrate surface thermal radiation data and extracts temperature features, combines them with the spectral features to optimize the prediction accuracy, and generates the nucleation probability prediction value for each local area.

[0019] Preferably, the expression of the dual-channel LSTM network prediction model is:

[0020] ;

[0021] Where, represents the predicted nucleation probability at the substrate position at the time step ; represents the LSTM network output function, which is used to map the input membership sequence to the current prediction output; represents the number of historical time steps; is the weight coefficient of the characteristic spectral data for the historical time step ; is the weight coefficient of the thermal radiation data for the historical time step ; represents the characteristic spectral data collected at the position at the time step ; represents the substrate surface thermal radiation data collected at the position at the time step ; represents the number of historical time steps considered for prediction.

[0022] Preferably, the asymmetric bias waveform in the step S4 can be expressed by the formula:

[0023] ;

[0024] Where, is the asymmetric bias waveform representation at the position at the time step ; represents the bias voltage value applied to the central region of the substrate at time step ; represents the compensation coefficient, and the value range is ; represents the change in the electron density of the edge region detected at time step ; is a position compensation function, which is used to dynamically adjust the compensation amplitude according to the specific position of the edge region.

[0025] An intelligent bias voltage regulation system for heteroepitaxial growth of single crystal diamond includes a data acquisition module, a prediction module, an intelligent optimization module, a gradient field compensation module, and a feedback and regulation module;

[0026] The data acquisition module is used to collect the characteristic spectral data of the plasma and the thermal radiation data of the substrate surface in real time through a collection device, and obtain the characteristic information of the substrate state;

[0027] The prediction module is used to receive the characteristic information output by the data acquisition module, and perform real-time processing on the characteristic information by using a dual-channel LSTM network to predict the nucleation probability of each local region of the substrate;

[0028] The intelligent optimization module is used to receive the characteristic information output by the data acquisition module, obtain the reward function of nucleation uniformity, and generate a dynamically optimized bias voltage regulation strategy;

[0029] The gradient field compensation module is used to automatically generate an asymmetric bias voltage waveform according to the change in the electron density of the edge region monitored in real time, and apply bias voltage compensation adjustment to the edge region of the substrate;

[0030] The feedback and regulation module is used to perform closed-loop regulation on the bias voltage output based on the real-time monitored data.

[0031] Preferably, the collection device includes a plasma emission spectrometer and a thermal radiation sensor; the plasma emission spectrometer is used to monitor the characteristic spectral intensity of the plasma in the growth reaction chamber in real time; the thermal radiation sensor is used to measure the temperature and radiation intensity of the substrate surface in real time.

[0032] Preferably, the gradient field compensation module is provided with an independent bias voltage application unit for applying an additional bias voltage signal to the edge region of the substrate; the gradient field compensation module outputs a bias voltage waveform to the edge bias voltage source according to a predetermined asymmetric waveform generation rule.

[0033] Preferably, the intelligent optimization module is pre-set with a reward function and a reinforcement learning algorithm strategy for measuring nucleation uniformity, and the control instructions output by the intelligent optimization module include adjustment suggestions for the bias voltage values of the central region and the edge region, and a decision signal on whether to trigger gradient field compensation.

[0034] Technical effects and advantages of the present invention:

[0035] The present invention is provided with step S2 and step S3, which is conducive to predicting future trends by adopting a prediction model, and can perceive potential growth anomalies in advance, such as a reduced nucleation probability in a certain area, so that the control system is forward-looking and avoids the lag of passive compensation afterwards; in addition, the reinforcement learning controller can comprehensively consider the cumulative effect and long-term goals in the process of strategy optimization, so that the bias regulation is more intelligent and adaptive; for example, when the edge area is in an unfavorable state for a long time, the system may increase the bias of the area in advance or even adjust the vapor deposition parameters in the entire substrate range to prevent the expansion of non-uniformity; this intelligent decision-making ability is far superior to the adjustment method of manual experience or preset programs, so that the growth process has stronger robustness to external disturbances such as airflow fluctuations, discharge state changes, etc.

[0036] The present invention is provided with step S4, which is beneficial to significantly reduce the difference in nucleation density between the center and edge areas of the substrate by real-time monitoring of the center and edge plasma states and dynamically adjusting the bias voltage. This improvement in uniformity ensures that the density of diamond nuclei on the entire substrate surface is more consistent, providing equal starting conditions for subsequent grain growth, thereby helping to obtain high-quality single crystal films with small thickness differences and smaller stress concentration areas; at the same time, stable single crystal diamond epitaxial growth can be achieved on a larger substrate; uniform nucleation density avoids the problem of local grain size being too large or too small, reduces the concentration of stress in the film from the source, and is expected to reduce the risk of cracks in the epitaxial layer; and maintains the consistency of nuclei and growth rates in each region throughout the growth process, which is beneficial to suppressing the accumulation of thermal stress and defects in large-size films. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 This is a flow chart of the intelligent bias control method for diamond single crystal heteroepitaxial growth of the present invention.

[0038] Figure 2 This is a structural diagram of the intelligent bias control system for diamond single crystal heteroepitaxial growth of the present invention.

[0039] Figure 3 This is a flow chart of the intelligent bias control system for diamond single crystal heteroepitaxial growth of the present invention.

[0040] Figure 4 This is the reward function and LSTM prediction graph of the present invention.

[0041] Figure 5 Schematic diagram of the gradient field compensation algorithm of the present invention.

[0042] Figure 6 It is an experimental comparison diagram of the present invention. DETAILED DESCRIPTION

[0043] The technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the present invention. In addition, the forms of the various structures described in the following embodiments are merely examples, and a method and system for intelligent bias regulation of diamond single crystal heteroepitaxial growth according to the present invention are not limited to the various structures described in the following embodiments. All other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

[0044] Example 1

[0045] As Figure 1 shown, the present invention provides a method for intelligent bias regulation of diamond single crystal heteroepitaxial growth, including the following steps:

[0046] Step S1: Collect characteristic spectral data of the plasma, and use a thermal radiation sensor to collect thermal radiation data on the substrate surface, so as to form characteristic information describing the substrate state;

[0047] The acquisition method of the characteristic spectral data of the plasma is as follows: Select characteristic emission peaks in the plasma. For example, the spectral line at a wavelength of about 405 nm is used as the monitoring object to obtain light intensity information reflecting plasma activity or electron density; the thermal radiation sensor is used to obtain thermal radiation data of the substrate to measure the temperature distribution or average temperature on the substrate surface;

[0048] Step S2: Build a bias regulation model based on reinforcement learning, calculate a reward function using the characteristic information describing the substrate state, and dynamically optimize the bias output of each region of the substrate during the epitaxial growth process according to the reward function; the purpose is to use the data collected in step S1, through the definition and solution of the reward function, to dynamically optimize the bias output of each region during the epitaxial growth process; in step S2, reinforcement learning can directly use the data collected in step S1, or can analyze the information output in step S3, so as to take into account both forward-looking and real-time when dynamically optimizing the bias output;

[0049] Input the real-time obtained characteristic information into the bias regulation model constructed by the reinforcement learning algorithm. The reinforcement learning controller uses the characteristic information as the environmental state, evaluates the quality of the growth uniformity under the current strategy according to the pre-set reward function, and adjusts the bias output strategy through continuous exploration and learning; among them, the reward function is designed to characterize the uniformity of the nucleation density on the substrate surface, and the reinforcement learning controller optimizes the bias control strategy with the goal of maximizing this reward function;

[0050] Step S3: Use a dual-channel LSTM network to fuse the characteristic spectral data and the substrate surface thermal radiation data, and predict the nucleation probability of each local area of the substrate in real time, so as to realize the anticipation and regulation of the epitaxial growth process; its purpose is that Step S3 focuses on using a dual-channel LSTM network to fuse the collected spectral data and thermal radiation data, and predict the nucleation probability of each local area of the substrate in real time, so as to anticipate the future growth trend in advance and provide a reference for subsequent regulation; Step S2 and Step S3 have independent functions. In the actual control process, Step S2 and Step S3 can be synchronously processed within the same control cycle, so as to realize the monitoring of the current growth state and the anticipation of the future trend;

[0051] Step S4: When it is detected that the electron density in the edge area of the substrate is lower than the preset threshold, an asymmetric bias waveform is automatically generated, and a compensation adjustment is applied to the edge area, so as to reduce the difference in nucleation density between the central area and the edge area;

[0052] When it is detected that the plasma state in the edge area of the substrate deteriorates significantly, for example, the electron density indicated by the spectral intensity is lower than the preset threshold, the gradient field compensation algorithm will be triggered to generate an asymmetric bias waveform to perform additional bias compensation adjustment on the edge area; the so-called "asymmetric bias waveform" refers to adding a bias component with different synchronization or different amplitude relative to the bias applied to the central area in the edge area, so as to differentially enhance the ion bombardment or electric field intensity in the edge area;

[0053] Step S5: According to the real-time feedback information in the epitaxial growth process, perform closed-loop adjustment on the bias output to ensure that the nucleation density on the substrate surface always remains uniform.

[0054] In this embodiment, it should be specifically noted that the reward function is defined as the total reward with a discount factor accumulated at each time step t, and the formula is expressed as:

[0055] ;

[0056] Among them, represents the reward function, represents the immediate reward corresponding to the deviation between the nucleation density in the central area and the preset target value at time step t; represents the immediate reward corresponding to the deviation between the nucleation density in the edge area and the preset target value at time step t; and are the weight coefficients of the central area and the edge area respectively, used to balance the influence of the two areas on the overall uniformity, satisfying ; represents the discount factor, and the value range satisfies , used to reflect the influence of future rewards on the current decision; is the total number of time steps, that is, the number of decision-making times considered in the entire reinforcement learning process;

[0057] When the nucleation density in the central region or edge region of the substrate deviates more from the preset target, the corresponding immediate reward function or is lower (the negative value of the deviation or the opposite of its absolute value can be used as the reward, so that the greater the deviation, the smaller the reward), thereby reducing the accumulated reward , and the reinforcement learning controller thus tends to take actions that can improve , that is, the bias control strategy that reduces the deviation of the nucleation density in each region.

[0058] In this embodiment, it should be specifically noted that the step S3 uses a dual-channel LSTM network prediction model to process the plasma spectral data and substrate surface thermal radiation data provided by the data acquisition module, so as to predict the nucleation probability at different positions of the substrate in real time; LSTM is a recurrent neural network that can capture long-term dependencies in time series. In this embodiment, it is designed as a dual-input channel structure, with the first channel inputting the spectral intensity sequence of the plasma and the second channel inputting the substrate temperature sequence; the hidden states of the two channels are fused in the network to output the nucleation probability estimation values at each spatial position corresponding to the time;

[0059] The prediction steps of the dual-channel LSTM network prediction model are as follows:

[0060] The first channel extracts the features of the 405nm characteristic spectral intensity changing with time to preliminarily predict the change trend of the nucleation probability;

[0061] The second channel extracts the features of the substrate temperature changing with time, and fuses it with the output features of the first channel to correct and optimize the prediction result of the nucleation probability;

[0062] Through this dual-source information fusion, the prediction accuracy of the complex growth process can be effectively improved.

[0063] In this embodiment, it should be specifically noted that the expression of the dual-channel LSTM network prediction model is:

[0064] ;

[0065] Among them, represents the predicted nucleation probability at the substrate position at time step ; represents the output function of the LSTM network, which is a non-linear output mapping function obtained by training the LSTM network and is used to map the input membership sequence to the current prediction output; represents the historical time steps; is the weight coefficient of the characteristic spectral data for the historical time step ; is the weight coefficient of the thermal radiation data for the historical time step ; The weight coefficient is comprehensively generated by the memory cells and gating mechanisms of the LSTM, reflecting the relative influence of each historical data on the current prediction; represents at the position at the time step the collected characteristic spectral data; represents at the position at the time step the collected substrate surface thermal radiation data; represents the number of historical time steps considered for prediction, that is, the number of past moment data included in the network input;

[0066] Adopting a dual-channel LSTM network prediction model, it can predict the probability of forming nuclei at different positions at the current moment based on the spectral and temperature information of the past N samplings, providing a forward-looking basis for the bias voltage regulation decision;

[0067] As Figure 4 shown, it shows the calculation process of the reward function and the prediction schematic diagram of the dual-channel LSTM network. The blue curve represents the change of the immediate reward corresponding to the deviation between the nucleation density in the central region and the preset target value over time, the green curve represents the change of the immediate reward corresponding to the deviation between the nucleation density in the edge region and the preset target value over time, and the red dashed line represents the total reward obtained after weighting. Figure 4 It shows how to evaluate the impact of the current bias voltage strategy on the nucleation uniformity through the reinforcement learning algorithm, and intuitively shows how the dual-channel LSTM uses historical data to predict the future nucleation probability;

[0068] In each control decision cycle, for example, a control update is performed every several seconds, preferably updating the plasma state vector and the calculation results of each module every 5 seconds. The reinforcement learning controller takes the latest sensing data and the LSTM prediction result as the current state input and outputs the corresponding bias voltage regulation action. This regulation action adjusts the bias voltage of different regions of the substrate to maximize the aforementioned reward function. Usually, the reinforcement learning controller will gradually approach the optimal strategy during the training or online learning process, so that the nucleation rates at the center and edge of the substrate tend to be consistent. The present invention introduces a gradient field compensation mechanism in the control process to further handle the situation of abnormal states in the edge region in extreme cases, such as significant attenuation of the plasma.

[0069] In this embodiment, it should be specifically noted that the asymmetric bias voltage waveform in step S4 can be expressed by the formula:

[0070] ;

[0071] wherein, is the representation of the asymmetric bias waveform at position at time step ; represents the bias value applied to the central region of the substrate at time step ; represents the compensation coefficient, with a value range of , which is used to adjust the amplitude of bias compensation; represents the change amount of the electron density in the edge region detected at time step , which can be defined as the deviation amount of the electron density in the edge region relative to the normal level or relative to the electron density in the central region; is the position compensation function, which is used to dynamically adjust the compensation amplitude according to the specific position of the edge region; the position compensation function can be a constant, a radial distance function, a sine function, etc., and the specific form of the position compensation function is determined by the substrate geometry;

[0072] When the plasma in the edge region becomes weak, relative to will be increased or adjusted to increase the actual effective bias at the edge; conversely, if the state of the edge region improves, the additional amount of the edge bias can also be reduced accordingly; the asymmetric bias waveform usually shows high-speed pulses or fluctuations superimposed on the reference bias, and its frequency and amplitude can be selected by those skilled in the art according to the actual situation, as long as it can improve the nucleation conditions in the edge region without significantly affecting the growth in the central region;

[0073] As Figure 5 shown, it shows the principle of the gradient field compensation algorithm. In the figure, the blue curve represents the reference bias maintained in the central region (e.g., -150V), and the orange curve represents the asymmetric bias waveform generated by superimposing high pulses through the compensation algorithm due to the detected decrease in electron density in the edge region. This figure intuitively shows how the gradient field compensation algorithm automatically generates an asymmetric bias waveform according to the detected change in electron density in the edge region, thereby performing additional compensation on the edge region and effectively improving the problem of insufficient edge nucleation.

[0074] In this embodiment, it should be specifically noted that the intelligent bias voltage regulation method for diamond single crystal heteroepitaxial growth provided in this embodiment can form a closed-loop control, continuously adjust the bias voltage distribution according to the feedback of the growth process, ensure that the nucleation density on the entire surface of the substrate is uniform and stable, and the method provided in this embodiment is applicable to the single crystal diamond growth on large-size hetero substrates such as 4 inches and above. It can still dynamically correct the center-edge deviation during a long growth time, thereby effectively improving the overall quality of the epitaxial film.

[0075] Embodiment 2

[0076] As Figure 2 shown, the present invention provides an intelligent bias voltage regulation system for diamond single crystal heteroepitaxial growth, including a data acquisition module, a prediction module, an intelligent optimization module, a gradient field compensation module, and a feedback and regulation module;

[0077] The data acquisition module is used to collect the characteristic spectral data of the plasma and the substrate surface thermal radiation data in real time through a collection device to form characteristic information describing the substrate state; the collection device includes an inductively coupled plasma optical emission spectrometer and a thermal radiation sensor, which work together to obtain the characteristic spectral data and the substrate surface thermal radiation data in real time; the inductively coupled plasma optical emission spectrometer is used to monitor the characteristic spectral intensity of the plasma in the growth reaction chamber in real time to reflect the change in the concentration of specific active substances or the electron density in the plasma; the thermal radiation sensor is used to measure the temperature or radiation intensity of the substrate surface in real time to reflect the heating situation of the substrate by the plasma. The data collected by the spectrometer and the temperature sensor are transmitted to the control unit through a high-speed data interface to form a plasma state vector for characterizing the current substrate state;

[0078] The prediction module is used to receive the characteristic information output by the data acquisition module, and use a dual-channel LSTM network to process the characteristic information in real time to predict the nucleation probability of each local area of the substrate, so as to realize the anticipation and regulation of the epitaxial growth process;

[0079] The intelligent optimization module is used to receive the characteristic information output by the data acquisition module, obtain a reward function for nucleation uniformity, and generate a dynamically optimized bias voltage regulation strategy; an enhanced learning controller is provided in the intelligent optimization module, and the enhanced learning algorithm is used to make decision control on the bias voltage output. Its optimization goal is to minimize the difference in nucleation density between the central area and the edge area of the substrate. The enhanced learning algorithm dynamically adjusts the bias voltage of each area by maximizing the reward function based on the nucleation density difference;

[0080] The gradient field compensation module is used to automatically generate an asymmetric bias voltage waveform according to the change in the electron density in the edge area monitored in real time, and apply bias voltage compensation adjustment to the edge area of the substrate;

[0081] The feedback and regulation module is used to perform closed-loop regulation on the bias voltage output based on the real-time monitored data to ensure the uniform nucleation density on the substrate surface.

[0082] In this embodiment, specifically, the intelligent optimization module is the decision-making center, which receives the feature information from the data acquisition module and the nucleation probability prediction result from the prediction module, and comprehensively uses them as the input of the current environmental state to the reinforcement learning controller; the intelligent optimization module is pre-set with a reward function and a reinforcement learning algorithm strategy for measuring nucleation uniformity, such as specific implementation methods like the Deep Q Network (DQN) or the policy gradient algorithm, which are not specifically limited in this embodiment; in each control cycle, the reinforcement learning controller quickly evaluates the immediate reward and long-term value of each optional control action according to the current state, such as different amplitude or different distribution schemes of adjusting the bias voltage, and selects the optimal action to execute the bias voltage adjustment strategy; the control instructions output by the intelligent optimization module include the adjustment suggestions for the bias voltage values in the central region and the edge region, as well as the decision signal on whether to trigger the gradient field compensation. When the reinforcement learning controller converges after sufficient training or online learning, its strategy will significantly reduce the difference in nucleation density in each region, thus maximizing the reward function. The introduction of the intelligent optimization module enables the intelligent bias voltage regulation system to autonomously adapt to complex and changeable growth conditions and avoid relying on human experience for adjustment.

[0083] In this embodiment, specifically, the prediction module is provided with a dual-channel LSTM prediction model for processing and analyzing the plasma state vector obtained by each update; the prediction module is used to output the nucleation probability prediction values of each monitoring region through three main steps: spectral channel feature extraction, temperature channel feature extraction, and feature fusion. For example, when the substrate is divided into two major regions: the center and the edge, the prediction module can output the "nucleation probability in the central region" and the "nucleation probability in the edge region". The design details of the prediction module are as described in part of Embodiment 1: preferably, an LSTM neural network is used to utilize the time series history to improve the prediction reliability; when it is found that the predicted nucleation probability in a certain region is continuously low, it indicates that there may be a problem of insufficient nucleation in this region in the future, which needs attention.

[0084] In this embodiment, specifically, the gradient field compensation module mainly compensates for the bias voltage for the special situation in the edge region; the gradient field compensation module is provided with an independent bias voltage application unit, such as a ring electrode or a controllable bias voltage source, which is specifically used to apply an additional bias voltage signal to the edge region of the substrate; once the intelligent optimization module issues a trigger compensation instruction (usually based on detecting that the edge plasma parameters drop below the threshold or the edge nucleation probability is too low), the gradient field compensation module outputs a bias voltage waveform to the edge bias voltage source according to the predetermined asymmetric waveform generation rule.

[0085] In this embodiment, it should be specifically noted that the feedback and regulation module includes the comparison and regulation units required for a closed-loop control circuit, which sends the bias voltage output strategy formulated by the intelligent optimization module to the actual bias voltage power supply / electrode to apply corresponding voltages to the central and edge regions in the growth reactor. At the same time, it continuously monitors the feedback from the data acquisition module and feeds back the new state to the intelligent optimization module when the next cycle arrives, thus completing the closed-loop. The feedback and regulation module is also responsible for some safety and stability strategies. For example, when abnormal peaks in the bias voltage or plasma parameters are detected, the output of the controller is temporarily locked or adjusted to avoid system instability.

[0086] As Figure 3 shown, a flowchart of the intelligent bias voltage regulation system for diamond single crystal heteroepitaxial growth is provided. The main data flow includes the characteristic spectral data and temperature data collected from the data acquisition module being simultaneously transmitted to the prediction module and the intelligent optimization module. After the prediction module processes the data using the dual-channel LSTM model, it transmits the nucleation probability prediction result to the intelligent optimization module. The intelligent optimization module calculates the reward function based on the prediction and real-time data, generates the bias voltage regulation strategy, and the strategy signal is respectively transmitted to the gradient field compensation module and the feedback and regulation module. Finally, the compensation module and the feedback and regulation module jointly adjust the actually applied bias voltage and feed back the new state data to the data acquisition module. Through the collaborative action of each module, the intelligent bias voltage regulation system provided by the present invention can achieve adaptive and high-precision control of the diamond epitaxial growth process.

[0087] Embodiment 3

[0088] The intelligent bias voltage regulation system for diamond single crystal heteroepitaxial growth provided by the present invention is applied to a set of 4-inch hetero-substrate single crystal diamond CVD growth equipment. The implementation methods of each module and their working conditions in the actual growth process are as follows:

[0089] A circular substrate with a size of 4 inches (with an iridium thin film deposited on a silicon substrate as an epitaxial growth substrate) is used. Inside the reaction chamber, the substrate is placed above the grounded electrode, and the reaction atmosphere is a typical diamond growth gas (such as a mixture of hydrogen and methane). Plasma is generated by microwave plasma CVD. An optical fiber probe is arranged directly above the growth surface of the substrate and connected to an inductively coupled plasma emission spectrometer to monitor the characteristic spectra in the plasma. In this embodiment, a spectral line closely related to hydrocarbon radicals or other relevant active particles in the plasma is selected as the monitoring index, and this spectral line is located near a wavelength of 405 nm. Through calibration, the change in the light intensity of this spectral line can reflect the change trend of the plasma electron density and the carbon material supply state near the substrate surface. At the same time, a high-temperature infrared sensor is installed on the side wall of the reactor, aimed at the substrate surface, to measure the thermal radiation intensity of the substrate. The signal detected by the infrared sensor can be converted by an algorithm to obtain an approximate temperature value of the substrate surface. The above two sensors collect data once every 5 seconds and transmit the data to the central control computer. The central control computer is pre-installed with the software of the diamond single crystal heteroepitaxial growth intelligent bias voltage regulation system provided by the present invention, which is used to execute the prediction algorithm and the reinforcement learning control algorithm.

[0090] Before the growth starts, the system is first initialized: the reaction chamber is evacuated and the reaction gas is introduced, the plasma is ignited and made to operate stably; then an initial bias voltage is applied to the substrate for plasma cleaning and pretreatment. In this embodiment, the initial bias voltage is set to a constant -150 V DC bias voltage in the central region of the substrate, and no additional signal is superimposed in the edge region. At this time, the data acquisition module starts to record the initial spectral intensity and temperature information, which is used as the reference starting point for the reinforcement learning algorithm.

[0091] After the formal growth stage starts, the intelligent bias voltage regulation system enters closed-loop control operation. In each control cycle, the data acquisition module sends the most recently monitored 405 nm spectral intensity value and substrate temperature value to the prediction module. The prediction module calls the dual-channel LSTM model to perform updated calculations on these two time series. In actual implementation, the LSTM model will maintain the data of the most recent several time points as the input state (in this embodiment, N = 5 is set, that is, the sensing data within the past 25 seconds is used to predict the current nucleation probability). Through the aforementioned dual-channel architecture, the LSTM model outputs the estimated nucleation probability values of the central region and the edge region of the substrate at the current moment. Suppose in a certain cycle, the prediction result shows that the nucleation probability in the edge region is significantly lower than that in the central region, which may be due to a decrease in the edge temperature or a decrease in the plasma intensity, then the intelligent optimization module will pay more attention to this.

[0092] The reinforcement learning controller within the intelligent optimization module then takes the state vector composed of the current sensing data and the prediction result as input, and evaluates the reward values of different bias voltage regulation actions. In this embodiment, the control actions are set to fine-tune the central bias voltage and the edge bias voltage, and their values can vary continuously within a certain range. To simplify the problem, the reinforcement learning controller can define the action as two real numbers and , representing the increments applied to the central and edge bias voltages respectively based on the current bias voltage. Since we hope to keep the nucleation densities at the center and the edge balanced, the controller will refer to the reward function to evaluate each alternative and combination. If an action is expected to increase the nucleation probability in the edge region and will not significantly reduce the nucleation probability in the central region, its reward value will be higher; conversely, if an action exacerbates the non-uniformity between the two regions, the reward value will be lower or even negative. The reinforcement learning controller selects the action with the highest expected cumulative reward from these candidate actions through an algorithm. For example, in the case where the nucleation probability in the edge is low as described above, the controller may choose to increase the edge bias voltage ( , that is, further increase the absolute value of the negative bias voltage) and slightly reduce the central bias voltage ( , reduce the absolute value of the central negative bias voltage) to tilt the plasma resources towards the edge. It should be noted that this decision is not based on a simple comparison of the current values, but comprehensively considers the impact on the next few steps - for example, excessive increase in the edge bias voltage may lead to too fast growth at the edge in the subsequent steps, so the controller will choose a compromise optimal strategy.

[0093] When the intelligent optimization module determines the optimized bias voltage strategy, the system will judge according to the actual situation whether it is necessary to trigger the gradient field compensation module. In most cases, if the state of the edge region is only slightly lower than that of the center, the decision of the reinforcement learning controller (the small-scale and as described above) can gradually correct the difference without additional compensation. However, when it is detected that the electron density in the edge region is significantly lower than the threshold (for example, the spectral intensity drops by more than a certain predetermined ratio, indicating a serious shortage of edge plasma), in addition to selecting and , the controller will issue a compensation trigger signal. After receiving this signal, the gradient field compensation module is activated to output an additional bias voltage waveform to the edge region. In this embodiment, the compensation module applies a high-frequency pulsed bias voltage with a frequency of 50 kHz to the annular electrode in the edge region, and its peak amplitude is calculated according to the previous formula. For example, when the 405 nm light intensity at the edge drops suddenly at a certain moment, corresponding to being -20 (normalized unit), the compensation coefficient per unit is selected, and assuming the position compensation function Taking a constant value of 1 for the entire edge, the amplitude of the applied pulsed bias voltage . This pulsed bias voltage is superimposed on the original -150V reference, so that the edge electrode reaches an instantaneous bias voltage of -250V at the negative peak, thereby greatly enhancing the electric field strength and plasma density at the edge. After several cycles of compensatory application, the spectral intensity in the edge region begins to recover. Once the monitored value returns above the threshold, the gradient field compensation module will gradually reduce the pulse amplitude until the output stops. The entire compensation process is supervised by the feedback and regulation module to avoid overshoot or oscillation.

[0094] The above decision-making and execution processes continue throughout the growth cycle to form a closed-loop control. Through continuous "sensing - prediction - decision-making - application - feedback", it can adapt to slow changes and even sudden fluctuations in growth conditions. For example, as the deposition progresses, the substrate temperature may gradually increase and the spectral intensity baseline may change. The reinforcement learning controller will dynamically adjust the ratio of the center and edge bias voltages to maintain balance. If there are external disturbances, such as a slight fluctuation in microwave power leading to a change in the overall plasma, since the controller focuses on relative uniformity, its output will correspondingly adjust the bias voltages in the two regions to maintain a stable difference level. In this embodiment, through the action of the intelligent bias voltage regulation system, the nucleation density on the entire 4-inch substrate surface remains highly consistent. After growing for a period of time and diagnosing the substrate, it is found that the number of nuclei per unit area in the center region and the edge region is almost the same, and the difference is less than 3% statistically. In contrast, if the intelligent regulation function of this system is turned off, for example, only using a constant bias voltage for a comparative experiment, the nucleation density in the edge region is about 15% lower than that in the center region on average, fully demonstrating the effectiveness of the present invention in improving the epitaxial growth uniformity of large substrates.

[0095] As Figure 6 shown, the experimental comparison data is presented through a histogram. The blue histogram represents the nucleation density distribution in the center region and the edge region under traditional bias voltage control, showing a difference of about ±15% between the two. The red histogram represents the nucleation density distribution after adopting the intelligent bias voltage regulation method of the present invention, showing that the difference is reduced to about ±3%. This figure provides a comparison of experimental data and intuitively demonstrates the remarkable effect of the present invention in improving the nucleation uniformity of the substrate.

[0096] The bias control model described in the embodiment of the present invention uses a reinforcement learning algorithm. The training process can be performed offline through a simulated environment before actual growth, or online learning can be performed with a safe small-amplitude trial in the early stage of growth to obtain a better initial strategy. In actual operation, considering safety and convergence speed, a supervised learning model can also be used to provide an initial strategy, and then handed over to reinforcement learning for fine-tuning. The dual-channel LSTM prediction model can be trained through historical experimental data, and its structure (including the number of hidden layer units, etc.) and hyperparameters can be adjusted by those skilled in the art according to actual needs. In addition, the present invention does not limit which specific reinforcement learning algorithm or neural network structure is used. Other types of neural networks, fuzzy control or expert systems can also be introduced as prediction or decision modules to improve the adaptive ability of bias control to complex environments. The hardware for bias application can be a single electrode combined with a waveform change to achieve equivalent area control, or multiple independent electrodes can be used to apply different bias voltages respectively; these are all equivalent variations of the present invention; the present invention realizes intelligent regulation of bias voltage during heteroepitaxial growth of single crystal diamond by innovatively combining advanced sensing technology, dual-channel LSTM prediction algorithm and reinforcement learning optimization control. Under the premise of ensuring growth stability, the uniformity of nucleation and growth of large-size substrates has been greatly improved, providing important technical support for the industrial preparation of diamond materials.

[0097] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

[0098] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. An intelligent bias voltage regulation method for diamond single crystal heteroepitaxial growth, characterized in that: It includes the following steps: Step S1: Collect the characteristic spectral data of the plasma by a plasma emission spectrometer, and collect the substrate surface thermal radiation data by a thermal radiation sensor to obtain the characteristic information of the substrate state; Step S2: Construct a bias voltage regulation model based on reinforcement learning, obtain a reward function using the characteristic information of the substrate state, and dynamically optimize the bias voltage output of each region of the substrate during epitaxial growth according to the reward function; Step S3: Use a dual-channel LSTM network to fuse and process the characteristic spectral data and the substrate surface thermal radiation data to predict the nucleation probability of each local region of the substrate in real time; Step S4: When it is detected that the electron density in the substrate edge region is lower than a preset threshold, automatically generate an asymmetric bias voltage waveform and apply a compensation adjustment to the edge region; Step S5: Based on the real-time feedback information during epitaxial growth, perform a closed-loop regulation on the bias voltage output; In the dual-channel LSTM network prediction model in step S3, it is a dual-input channel structure. The first channel inputs the spectral intensity sequence of the plasma, and the second channel inputs the substrate temperature sequence; the hidden states of the two channels are fused in the network to output the nucleation probability estimation value of each spatial position at the corresponding time; The prediction steps of the dual-channel LSTM network prediction model are: The first channel receives the characteristic spectral data of the plasma and extracts spectral features to predict the initial value of the nucleation probability of each local region; The second channel receives the substrate surface thermal radiation data and extracts temperature features, combines them with the spectral features to optimize the prediction accuracy, and generates the nucleation probability prediction value for each local region; The expression of the dual-channel LSTM network prediction model is: ； Among them, represents the predicted substrate position at time step with the nucleation probability; represents the output function of the LSTM network, which is used to map the input membership sequence to the current predicted output; represents the number of historical time steps; is the weight coefficient of the characteristic spectral data for the historical time step ; is the weight coefficient of the thermal radiation data for the historical time step ; represents the characteristic spectral data collected at position at time step ; represents the thermal radiation data of the substrate surface collected at position at time step ; represents the number of historical time steps considered for prediction. The asymmetric bias voltage waveform in step S4 can be expressed by the formula: ； Among them, is the representation of the asymmetric bias waveform at position at time step ; represents the bias value applied to the central region of the substrate at time step ; represents the compensation coefficient, and its value range is ; represents the change amount of the electron density in the edge region detected at time step ; is the position compensation function, which is used to dynamically adjust the compensation amplitude according to the specific position of the edge region.

2. The intelligent bias voltage regulation method for diamond single crystal heteroepitaxial growth according to claim 1, wherein: The reward function is defined as the total reward with a discount factor accumulated at each time step t, and is expressed by the formula: ； Among them, is expressed as the reward function, represents the immediate reward corresponding to the deviation between the nucleation density in the central region and the preset target value at time step t; represents the immediate reward corresponding to the deviation between the nucleation density in the edge region and the preset target value at time step t; and are the weight coefficients of the central region and the edge region respectively, represents the discount factor, and its value range satisfies ; is the total number of time steps.

3. A smart bias voltage regulation system for diamond single crystal heteroepitaxial growth, which is used to implement the method described in any one of the above claims 1-2, and is characterized in that: It includes a data acquisition module, a prediction module, an intelligent optimization module, a gradient field compensation module, and a feedback and regulation module; The data acquisition module is used to collect the characteristic spectral data of the plasma and the substrate surface thermal radiation data in real time through a collection device to obtain the characteristic information of the substrate state; The prediction module is used to receive the characteristic information output by the data acquisition module, and use a dual-channel LSTM network to process the characteristic information in real time to predict the nucleation probability of each local region of the substrate; The intelligent optimization module is used to receive the characteristic information output by the data acquisition module, obtain the reward function of nucleation uniformity, and generate a dynamically optimized bias voltage regulation strategy; The gradient field compensation module is used to automatically generate an asymmetric bias voltage waveform according to the change of the electron density in the edge region monitored in real time, and apply a bias voltage compensation adjustment to the substrate edge region; The feedback and regulation module is used to perform a closed-loop regulation on the bias voltage output based on the data monitored in real time.

4. The intelligent bias voltage regulation system for diamond single crystal heteroepitaxial growth according to claim 3, wherein: The collection device includes a plasma emission spectrometer and a thermal radiation sensor; the plasma emission spectrometer is used to monitor the characteristic spectral intensity of the plasma in the growth reaction chamber in real time; the thermal radiation sensor is used to measure the temperature and radiation intensity of the substrate surface in real time.

5. The intelligent bias voltage regulation system for diamond single crystal heteroepitaxial growth according to claim 4, characterized in that: The gradient field compensation module is provided with an independent bias voltage application unit for applying an additional bias voltage signal to the edge region of the substrate; the gradient field compensation module outputs a bias voltage waveform to the edge bias voltage source according to a predetermined asymmetric waveform generation rule.

6. The intelligent bias voltage regulation system for diamond single crystal heteroepitaxial growth according to claim 5, characterized in that: The intelligent optimization module is preset with a reward function and a reinforcement learning algorithm strategy for measuring nucleation uniformity. The control instructions output by the intelligent optimization module include adjustment suggestions for the bias voltage values in the central region and the edge region, as well as a decision signal on whether to trigger gradient field compensation.

Images