A method for coupling and controlling the production of p-type silicon carbide multiphase interfacial carrier precursors via liquefaction

Abstract

This invention belongs to the field of semiconductor material preparation technology and discloses a dynamic evolution coupling method for the composition of P-type silicon carbide molten liquid prepared by liquefaction. The invention employs partial differential equation discretization calculation, nonlinear constraint optimization iteration, nonequilibrium dynamic state space construction, multi-field coupling residual convergence solution, operator dimensionality reduction compression, and industrial incremental iterative learning. Based on high-temperature molten liquid multi-component global dynamic modeling, liquid-solid interface phase transition coupling mass transfer mechanism analysis, advanced extrapolation of doping dynamic imbalance trends, and adaptive dynamic matching of liquefaction process parameters, this invention suppresses local enrichment and component deficiency phenomena of P-type doping from the source of dynamic component transport, blocks the doping gradient heterogeneous expansion path caused by multi-dimensional coupling, stabilizes the silicon carbide crystal lattice growth quality and doping uniformity under continuous production conditions of liquefaction, and effectively improves the consistency of electrical parameters of P-type silicon carbide substrates.

Description

Technical Field

[0001] This invention belongs to the field of semiconductor materials technology, and in particular to a method for coupling and regulating the multiphase interface carrier precursor of P-type silicon carbide produced by liquefaction. Background Technology

[0002] Silicon carbide, as a wide-bandgap core semiconductor material, has become an irreplaceable substrate material for high-voltage, high-frequency power electronic devices due to its advantages of high breakdown field strength, high thermal conductivity, high temperature resistance, and radiation resistance. P-type silicon carbide, through doping with acceptor impurities such as boron atoms, forms a hole-conducting system, making it a core material for PN junction structures, bipolar power devices, and gate-isolated semiconductor devices. Compared to vapor transport and solid-state sintering methods, the liquefaction method relies on a high-temperature silicon-carbon eutectic liquid phase system to achieve liquid-to-solid directional crystallization. This method offers faster growth rates, lower preparation costs, better large-size crystal formability, and lower micro-defect density, making it the mainstream process for the commercial mass production of large-size P-type silicon carbide. The core principle of the liquefaction method for preparing P-type silicon carbide is as follows: a silicon-carbon composite molten system is constructed within a closed high-temperature furnace; a quantitative P-type doped precursor is introduced; and impurity particles migrate to the liquid-solid crystallization interface through molten convection mass transfer. Under the drive of a temperature gradient, impurity atoms achieve lattice solid solution, completing the P-type conductive modification of silicon carbide. However, the liquefaction growth system is a typical multi-component, multi-phase, and dynamically evolving nonlinear complex system. The continuous migration of the liquid-solid interface, the random fluctuation of the molten rheological state, and the continuous effect of the long-term high-temperature environment lead to strong coupling, time-varying, and non-stationary characteristics of impurity mass transfer and lattice solid solution process. It is very easy to generate hidden defects such as uneven doping distribution and local solid solution imbalance.

[0003] Currently, the industry's R&D and patent layout for liquefied silicon carbide focuses on physical process and hardware optimization, such as melt composition ratio, furnace thermal field structure optimization, crystal traction mechanical structure improvement, protective atmosphere purification, and macroscopic defect suppression. There is a lack of systematic mathematical modeling, multi-field coupling iterative solutions, and algorithmic control schemes for nonlinear mass transfer coupling at multiphase interfaces, dynamic lattice solid solution dissimilarity, and high-temperature lattice relaxation-induced doping imbalance. Conventional detection methods can only detect steady-state parameters such as average doping concentration and overall resistivity of the finished product, failing to capture trace solid solution shifts at micro-interfaces during crystallization. The multiphase melt is mixed with multiple particles, and the weak solid solution shift characteristics of lattice micro-regions cannot be nonlinearly decoupled and inverted. The deep coupling of four processes—multiphase rheological shear mass transfer, dynamic solid solution at the liquid-solid interface, micro-lattice curvature segregation, and high-temperature long-term lattice relaxation—lacks a unified numerical solution model. The fixed-value control of liquefaction growth parameters and the lack of lattice solid solution adaptation constraints induce continuous doping gradient dissimilarity. Summary of the Invention

[0004] This invention provides a method for coupling and controlling the multiphase interface carrier precursor in the liquefaction process of producing P-type silicon carbide. The method includes: simultaneously acquiring multi-source nonlinear signals from high-temperature multiphase molten particles, liquid-solid interface lattice solution, and high-temperature environmental disturbances; extracting microscopic weak solid solution migration features through a multiphase rheological manifold-constrained mixing regularized nonlinear purification model; constructing a lattice solution fit index by fusing multiple types of lattice solution heterogeneity parameters; solving the global doping degradation risk using a nonstationary time-domain iterative algorithm; building a rheological-solution-segregation-relaxation four-field coupled finite element model; using multi-scale nested discretization and step-by-step coupling iteration to achieve a unified solution of the global solute field; constructing a nonlinear lattice coupling evolution state-space model; using a nonlinear ensemble filtering algorithm to extrapolate the doping gradient heterogeneity trend over a long period; constructing a multi-objective constrained optimization model with lattice solution equilibrium as the goal; numerically solving and outputting adaptive control process parameters for the liquefaction method; deploying a lightweight four-field coupled model at the edge; dynamically updating the multiphase mass transfer correction coefficient through incremental adaptive iterative learning; and forming a closed-loop doping control throughout the entire P-type silicon carbide crystallization process.

[0005] Furthermore, the method includes: a multi-source signal synchronous acquisition module, a nonlinear feature decoupling operation module, a four-field coupling modeling and solving module, a nonlinear time series extrapolation and early warning module, a multi-objective parameter optimization and control module, and an incremental lightweight adaptive learning module.

[0006] Furthermore, the hybrid regularized nonlinear purification model is solved by a nonstationary alternating direction multiplier iterative algorithm, which relies on rheological manifold constraints and sparse constraints to adapt to the nonlinear mass transfer characteristics of multiphase systems, thereby achieving precise decoupling of microscopic and weak solid solution defects.

[0007] Furthermore, the four-field coupled finite element model adopts multi-scale nested adaptive non-uniform finite element discretization, and uses a multi-process step-by-step coupled nonlinear iterative algorithm to quantify and analyze the cross-scale coupling correlation of rheological mass transfer, interface solid solution, lattice segregation, and high-temperature relaxation.

[0008] Furthermore, the nonlinear ensemble filtering algorithm is based on an ensemble sampling iterative update mechanism, which is adapted to the non-Gaussian random fluctuations of multiphase mass transfer and the long-delay accumulation characteristics of lattice relaxation, avoids the simplification defects of linear time series model mechanisms, and improves the accuracy of long-period evolution inference.

[0009] Furthermore, the multi-objective constrained optimization model uses the doping precursor loading acceleration rate, crystal pulling rate, furnace temperature field gradient, and melt stirring intensity as decision variables, embeds the LSAI exponential threshold and high-temperature thermal stability hard constraints, and uses a sequential quadratic programming algorithm to solve the nonlinear optimal solution.

[0010] Furthermore, the lightweight processing is achieved through structured pruning of four-field coupling operators and nonlinear dimensionality reduction of high-dimensional solid solution features, which compresses the computational dimension of the model and adapts it to the low-computing-power real-time computing scenarios of semiconductor production line edge controllers.

[0011] Furthermore, the incremental adaptive iterative learning adopts an adaptive mini-batch gradient descent algorithm to dynamically update the interface mass transfer correction coefficient, lattice segregation compensation coefficient, and high-temperature relaxation fitting coefficient, adapting to raw material batch fluctuations and furnace equipment aging conditions.

[0012] Furthermore, the Lattice Solid Solution Adaptation Index (LSAI) is a continuous quantitative evaluation index for the entire crystal domain. It is used to hierarchically identify local impurity enrichment, solid solution vacancies, interfacial mass transfer hindrance, and latent defects of lattice relaxation differentiation, providing a quantitative basis for adaptive closed-loop control of liquefaction process parameters.

[0013] Beneficial effects:

[0014] This invention focuses on innovative technologies related to the dynamic non-equilibrium transport of multi-components in high-temperature molten liquids during liquefaction, selective segregation at the phase transition interface of P-type doped particles, and multi-dimensional dynamic coupling imbalances involving temperature, composition, flow rate, and phase transition. The core implementation path of this invention is based on discretization of partial differential equations, nonlinear constraint optimization iteration, construction of non-equilibrium dynamic state space, convergence solution of multi-field coupled residuals, operator dimensionality reduction compression, and industrial incremental iterative learning. This invention relies on global dynamic modeling of multi-components in high-temperature molten liquids, analysis of the mass transfer mechanism of phase transition coupling at the liquid-solid interface, advanced deduction of doping dynamic imbalance trends, and adaptive dynamic matching of liquefaction process parameters. It suppresses local enrichment and component loss of P-type doping at the source of dynamic component transport, blocks the doping gradient heterogeneous expansion path caused by multi-dimensional coupling, stabilizes the lattice growth quality and doping uniformity of silicon carbide crystals under continuous production conditions in liquefaction, effectively improves the consistency of electrical parameters of P-type silicon carbide substrates, and is widely applicable to the large-scale mass production of high-end devices such as high-voltage power electronics, new energy vehicle semiconductors, and high-frequency RF chips. Attached Figure Description

[0015] Figure 1 Flowchart. Detailed Implementation

[0016] Example 1

[0017] Implementation steps

[0018] 1. Deploy distributed multi-component sensing units, liquid-solid interface dynamic monitoring modules, and high-temperature temperature field dynamic disturbance acquisition modules in the high-temperature furnace of silicon carbide liquefaction process. Collect P-type doped component diffusion signals, silicon-carbon melt convection transport signals, interface phase change adsorption dynamic signals, and high-temperature environment dynamic disturbance noise signals in a layered and synchronous manner. Complete multi-source signal timing alignment, high-temperature baseline dynamic correction, and thermal radiation noise dynamic suppression preprocessing.

[0019] 2. Perform cross-scale nonstationary dynamic normalization, convection noise pre-filtering, and component baseline dynamic drift correction on coupled multi-component signals. Introduce multi-component transport manifold correlation constraint operator and micro-doping dynamic fluctuation sparse prior constraint to construct a hybrid regularized nonlinear robust optimization model. Clarify the multi-scale component separation constraint boundary, nonlinear dynamic iterative convergence condition, and computational error control range.

[0020] 3. A non-stationary dynamic alternating direction multiplier iterative algorithm is used to carry out multi-round constrained convergence calculations, and irrelevant components such as crucible high-temperature dissolution impurities, gas phase condensation dynamic interference, and melt convection shear noise are separated in layers. The steady-state dynamic transport components of the flux matrix, environmental dynamic interference noise components and weak doping dynamic fluctuation characteristic components of the interface micro-region are accurately decoupled and separated.

[0021] 4. Based on the dynamic evolution characteristic data of the purified high-fidelity components, four heterogeneous nonlinear dynamic parameters were calculated in sequence: micro-region doping dynamic fluctuation coefficient, interface component transport attenuation deviation, interface selective adsorption accumulation rate, and high-temperature component dynamic attenuation differentiation coefficient. This completed the systematic and quantitative characterization of the dynamic degradation characteristics of silicon carbide global doping.

[0022] 5. A spatiotemporal dual-dimensional adaptive variable weight fusion operator is introduced and substituted into the calculation model of the doping dynamic imbalance index. Combined with the non-stationary variable step size time-domain dynamic iterative recursion mechanism, the dynamic imbalance index of the crystal axial and radial regions is dynamically updated region by region, so as to realize the full-domain graded quantitative dynamic assessment of the component imbalance risk of the entire liquefaction growth process.

[0023] The core modeling innovation of this embodiment is a manifold-constrained hybrid regularized nonlinear dynamic decoupling architecture adapted to high-temperature multi-component nonequilibrium dynamic systems. This overcomes the inherent limitations of traditional linear static component decomposition models, which cannot simultaneously account for the continuity of molten convection dynamic transport, the abrupt changes in interface phase transition doping, and the sparse distribution of trace dynamic defects. By binding the dynamic transfer correlation logic of components in different phase states and spatial regions to multi-component transport manifold constraint operators, the convective transport dynamic mechanism is avoided from being fragmented after multi-component decomposition coupling. Relying on L1-L2 hybrid regularized layered constraints, it simultaneously achieves high-temperature strong noise suppression and complete preservation of low-amplitude trace doping dynamic fluctuation characteristics, perfectly matching the complex production conditions of multi-phase coexistence, interface dynamic evolution, and high-temperature strong dynamic disturbances in liquefaction processes. At the algorithm level, a non-stationary dynamic alternating direction multiplier iterative solution mechanism is adopted, possessing extremely strong high-temperature anti-interference and multi-component nonlinear dynamic decoupling capabilities. It can completely preserve the weak characteristics of early microscopic doping dynamic shifts under harsh industrial high-temperature dynamic environments, effectively solving the core technical defects of traditional linear algorithms such as over-smoothing, loss of microscopic doping dynamic details, and failure of defect dynamic prediction. The DDII index incorporates a spatiotemporally adaptive variable weight fusion mechanism, which can adapt to changes in the real-time crystal growth position, interface dynamics, and temporal high-temperature operating conditions. Combined with a variable-step-size time-domain dynamic iterative algorithm, it suppresses single-point instantaneous sampling fluctuations, ensuring the continuity and real-time stability of the global doping dynamic imbalance assessment. This embodiment completes the nonlinear dynamic decoupling of multi-component coupled signals and the global dynamic inversion of doping parameters from the algorithm's underlying layer, providing standardized, high-fidelity basic data support for subsequent four-dimensional coupling modeling, dynamic temporal deduction, and adaptive optimization of process parameters.

[0024] Example 2

[0025] Implementation steps

[0026] 1. Import the global component dynamic evolution characteristic data, furnace real-time temperature field dynamic parameters, crystal pulling motion parameters, and melt stirring operation dynamic parameters output from Example 1, and calibrate the multi-component melt convection transport physical property parameters, liquid-solid interface phase transition constraint conditions, doping dynamic gradient imbalance critical threshold, and high-temperature component decay and degradation constraint index.

[0027] 2. Four dynamic coupled evolution equations were established to construct the unsteady convective transport of molten liquid, phase change precipitation at the liquid-solid interface, selective adsorption at the interface, and dynamic decay of high-temperature components. The integrated construction of the four-dimensional dynamic coupled overall model and the precise calibration of key parameters of multi-component transport and interface evolution were completed.

[0028] 3. A multi-scale nested adaptive non-uniform mesh generation strategy is adopted to refine the micro-mesh in the active region of component exchange at the liquid-solid crystallization interface, the dynamic sensitive region of crystal edge doping, and the region of severe molten convection disturbance. The stable transport region of the molten matrix adopts a conventional mesh layout to balance the computational load of the overall coupled model with the micro-dynamic solution accuracy. A variable time step is set to adapt to the instantaneous interface phase transition and the slow evolution characteristics of long-term component dynamic decay.

[0029] 4. A multi-process step-by-step coupled dynamic iterative solution mechanism is adopted to sequentially complete the dynamic update of convective transport parameters, interface phase change doping correction, selective adsorption iterative calculation, and high-temperature component decay differentiation quantitative analysis. The dynamic iterative error range is constrained by multi-level residual convergence criteria to ensure the global stability and evolution mechanism integrity of the strongly nonlinear multi-dimensional coupled solution.

[0030] 5. Full-domain output of multi-component dynamic distribution law, interface doping dynamic gradient distribution, interface adsorption enrichment region, and long-term high-temperature component decay and differentiation range, accurately marking key crystallization sections that are prone to P-type doping dynamic imbalance, lattice mismatch and dynamic degradation of electrical properties under continuous high-temperature growth conditions.

[0031] This embodiment innovatively constructs a four-dimensional integrated multi-field dynamic coupling numerical solution framework encompassing convection, phase change, adsorption, and attenuation. It abandons the fragmented and singular static simulation analysis mode of existing technologies, and fully establishes a nonlinear closed-loop quantitative mapping system between dynamic control parameters of the liquefaction process, dynamic liquid-solid interface state, multi-component dynamic transport response, and dynamic degradation evolution of doping. The core innovation of the modeling lies in introducing a multi-scale convection dynamic correction operator and a high-temperature component attenuation linkage evolution operator. This breaks down the scale barrier between macroscopic growth dynamic control and microscopic interface doping dynamic evolution, accurately characterizing the driving effect of unsteady-state convection on doped component migration, the determining role of dynamic interfaces on phase change doping distribution, the inducing effect of microscopic interface characteristics on selective adsorption, and the weakening effect of long-term high temperatures on melt component stability. This overcomes the significant shortcomings of traditional macroscopic steady-state simulations, which neglect the dynamic coupling mechanism of microscopic multi-components and suffer from one-sided modeling. At the algorithmic level, relying on adaptive finite element discretization and step-by-step coupled nonlinear dynamic iterative algorithms, it effectively adapts to the complex characteristics of multi-dimensional coupled systems, such as strong nonlinearity, multi-component cross-dynamic interference, spatial heterogeneous distribution, and long-term slow variable dynamic accumulation. It avoids problems such as dynamic mechanism distortion and doping parameter calculation deviation caused by linear approximate steady-state solutions. Based on pure algorithmic dynamic iterative solutions, it quantitatively analyzes the intrinsic mechanism of doping imbalance driven by multi-dimensional coupling. With nonlinear numerical calculation as the core, it deeply reveals the underlying dynamic mechanism of the discretization of electrical parameters of P-type silicon carbide induced by multi-component dynamic coupling, providing rigorous and complete multi-scale dynamic numerical model support for intelligent control of liquefaction process under component dynamic equilibrium constraints.

[0032] Example 3

[0033] Implementation steps

[0034] 1. Continuously collect multi-dimensional time-series data on DDII doping dynamic imbalance index, micro-area doping dynamic fluctuation amplitude, global component transport cumulative gradient, and high-temperature component decay differentiation coefficient under the combined conditions of long-term high-temperature steady-state growth, furnace temperature field dynamic fluctuation, and molten convection random disturbance. Construct a multi-scenario component evolution sample set covering stable mass production, parameter dynamic fluctuation, and raw material batch differences.

[0035] 2. Based on the multi-dimensional coupling correlation law determined by the four-dimensional dynamic coupling model, a non-equilibrium nonlinear component evolution state space equation is constructed to clarify the nonlinear mapping relationship between the external observable doping dynamic parameters, the internal interface component transport hidden state variables, and the multi-field coupled dynamic perturbation variables.

[0036] 3. A nonlinear ensemble sampling dynamic iterative filtering algorithm is adopted to dynamically update the ensemble sampling weights and nonlinear component state correction gains, and adapt in real time to the inherent evolution characteristics of non-Gaussian random fluctuations in multi-component convective transport, long-delay accumulation of high-temperature component attenuation, and superposition of dynamic disturbances from multiple fields coupled together.

[0037] 4. By combining real-time online multi-component dynamic monitoring data of the liquefaction production line, real-time iterative correction of multi-dimensional coupled state variables is completed. The real-time evolution trend of dynamic alienation and spread of global doping gradient and dynamic expansion of local component imbalance is deduced in long-term recursive deduction under the coupled conditions of continuous high temperature growth, dynamic interface migration, and molten convection fluctuation.

[0038] 5. Based on the DDII dynamic imbalance index grading threshold, a multi-level dynamic early warning mechanism is set up. When the dynamic degradation of doping is identified in the time series analysis, the dynamic adjustment command of the process parameters is issued in advance. The liquefaction growth control system is linked to adaptively optimize the operating parameters, thereby blocking the dynamic expansion path of doping defects from the source of dynamic evolution of composition.

[0039] This embodiment addresses the complex characteristics of multi-component non-equilibrium dynamic evolution in liquefaction processes, including strong nonlinearity, multi-dimensional coupled disturbance superposition, long-period slow accumulation, and delayed dynamic differentiation. It innovatively employs a nonlinear ensemble time-series dynamic estimation algorithm, overcoming the inherent defects of traditional linear time-series models that forcibly simplify coupled correlations and fail to fit the historical dependence of molten convection random fluctuations and high-temperature component decay. The modeling innovation lies in comprehensively incorporating multi-dimensional external dynamic disturbance variables, such as furnace temperature field dynamic fluctuations, real-time disturbances in stirring intensity, batch-to-batch material property differences, and furnace equipment aging and drift, into a nonlinear dynamic state-space modeling system. This ensures that the long-period dynamic extrapolation model fully conforms to the complex dynamic conditions of industrial mass production, completely eliminating problems such as prediction distortion, simplification of mechanisms, and insufficient industrial adaptability caused by idealized single-variable steady-state modeling. The algorithm's efficiency enhancement principle focuses on the core characteristics of doping dynamic defects: concealment, gradual dynamic accumulation, and irreversible expansion. It achieves high-precision advanced prediction of microscopic doping dynamic shifts, interfacial component transport blockages, and nascent gradient dynamic imbalances, distinguishing it from the lagging control mode of post-product inspection and defect removal and rectification in existing technologies. By relying on nonlinear ensemble statistical dynamic iteration to accurately characterize the random uncertainty of multi-component transport and the hysteresis characteristics of high-temperature component decay, the response time window of dynamic control of liquefaction process is effectively extended. Active dynamic intervention and control are completed before component imbalance defects are solidified on a large scale, the lattice doping structure undergoes permanent changes, and the electrical parameters show obvious dynamic differentiation. Relying on algorithm dynamic modeling and time series extrapolation, the risk of dynamic imbalance of P-type silicon carbide doping in liquefaction process is prevented in advance, accurately and in real time.

[0040] Example 4

[0041] Implementation steps

[0042] 1. Real-time reading of the dynamic imbalance index distribution of doping across the entire crystal, the spatial distribution of local doping dynamic enrichment and component deficiency, and the dynamic monitoring data of component transport gradient at the liquid-solid interface. Combined with the dynamic fluctuation of silicon carbide raw material batch properties and the dynamic operating conditions of furnace high-temperature aging, the safe dynamic adjustment boundaries of furnace temperature dynamic gradient, real-time crystal pulling rate, dynamic intensity of melt stirring, and dynamic acceleration rate of doping source are defined.

[0043] 2. Taking the three core optimization objectives of improving the dynamic homogenization of global P-type doping composition, accurately suppressing the dynamic imbalance of local doping, and the dynamic steady-state constraint of the interfacial composition transport gradient as the three core optimization objectives, we embed the dynamic imbalance index safety threshold, the thermal stability dynamic range of high-temperature melt, and the lower limit of crystal directional crystallization morphology hard constraint conditions to construct a multi-field collaborative nonlinear optimization model for composition dynamic equilibrium.

[0044] 3. A sequential quadratic programming nonlinear constraint optimization algorithm is adopted to carry out multi-round iterative convergence numerical solution to accurately calculate the optimal combination of process parameters that adapts to the current dynamic state of the interface, multi-component convection conditions, and high-temperature dynamic conditions;

[0045] 4. The dynamic parameters for component balance control obtained from the optimization solution are sent in real time to the dynamic doping module, the dynamic crystal traction mechanism, the dynamic furnace temperature control unit, and the melt stirring drive module to achieve adaptive dynamic adjustment of differentiated parameters in different growth stages and different interface regions of the crystal, and balance the dynamic transport efficiency of components across the entire domain.

[0046] 5. Deploy a lightweight four-dimensional dynamic coupling model at the edge to collect dynamic data of equipment operation and online multi-component monitoring data during continuous mass production of P-type silicon carbide using the liquefaction method. Through an adaptive small-batch gradient descent incremental iterative learning algorithm, dynamically update the interface component transport correction coefficient, interface selective adsorption compensation coefficient, and high-temperature component attenuation fitting correction coefficient in real time. Adapt to the dynamic changes in mass production conditions to achieve closed-loop intelligent control of the entire process of dynamic balance of global components and long-term suppression of doping degradation in the preparation of P-type silicon carbide by the liquefaction method.

[0047] This embodiment completely breaks through the traditional static fixed-value open-loop control mode of liquefaction method. For the first time, it takes the dynamic balance of melt components and the dynamic steady-state constraint of doping gradient as the core optimization guide, and constructs a multi-field synergistic nonlinear dynamic optimization modeling system of convection-thermal field-interface-attenuation. It realizes the deep coupling of macroscopic process dynamic control logic and microscopic multi-component doping nonlinear dynamic evolution mechanism, and comprehensively fills the modeling gap of dynamic controllable doping and adaptive algorithm control of P-type impurities in liquefaction method. The modeling innovation lies in establishing a full-chain nonlinear optimization mapping relationship of dynamic control parameters of liquefaction method, multi-component convection transport state, global doping dynamic distribution, and electrical dynamic stability of P-type silicon carbide semiconductor, breaking through the limitation of traditional process optimization that only focuses on static quality indicators. The algorithm efficiency enhancement principle is to adjust the differentiated dynamic parameters in real time, accurately match the dynamic consumption of components in different crystallization regions, the real-time update of melt and the doping requirements of interface phase change, balance the impurity distribution from the source of multi-component dynamic transport, suppress the problem of local dynamic enrichment of doping and component deficiency, stabilize the dynamic gradient of interface component transport, and block the core evolution path of doping gradient alienation induced by multi-dimensional coupled dynamic effects. By combining multi-field coupling operator pruning and compression with high-dimensional dynamic feature nonlinear dimensionality reduction technology, this method effectively compresses the computational dimension of complex coupled models, reduces the computing power consumption of semiconductor production line edge terminals, and meets the real-time dynamic computational needs of continuous industrial mass production. Furthermore, it incorporates an incremental adaptive iterative learning mechanism to continuously adapt to complex variables such as raw material batch dynamic fluctuations, furnace equipment high-temperature aging parameter drift, and dynamic disturbances in the mass production environment. This ensures long-term stable maintenance of algorithm computational accuracy and control stability for multi-component dynamic feature inversion, long-cycle doping evolution dynamic prediction, and multi-objective optimization solutions. From the perspective of intelligent algorithm dynamic optimization closed-loop control, it comprehensively improves the global doping dynamic uniformity, resistance to high-temperature component decay, and dynamic consistency of electrical parameters in batch products produced by the liquefaction method for P-type silicon carbide crystals. This provides long-term assurance of parameter stability and long-term service reliability of P-type silicon carbide substrates in high-end application scenarios such as high-voltage power devices and high-frequency semiconductor chips.

[0048] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for coupling the dynamic evolution of components in a liquefaction-based p-type silicon carbide melt, characterized in that, include: Simultaneously, multi-source nonlinear signals from high-temperature multi-component melts, dynamic phase transitions at the liquid-solid interface, and dynamic disturbances in the high-temperature environment are acquired. Weak characteristics of microscopic doping dynamic fluctuations are extracted through a multi-component transport manifold-constrained, regularized, nonlinear purification model. A doping dynamic imbalance index is constructed by fusing dynamic evolution heterogeneous parameters of multiple components, and a non-stationary time-domain dynamic iterative algorithm is used to solve the global doping dynamic degradation risk. A four-dimensional dynamically coupled finite element model of convection-phase transition-adsorption-attenuation is built, and a unified solution for the global component field is achieved using multi-scale nested discretization and step-by-step coupled dynamic iteration. A nonlinear dynamic state-space model of nonequilibrium component evolution is constructed, and the dynamic alienation evolution trend of doping gradient is extrapolated over a long period using a nonlinear ensemble filtering algorithm. A multi-objective constrained optimization model is constructed with the dynamic equilibrium of melt components as the core, and the numerical solution outputs the dynamic adaptive control process parameters of the liquefaction method. Deploy a lightweight four-dimensional coupling model at the edge and dynamically update the multi-component transport correction coefficients through incremental adaptive iterative learning to form a closed-loop dynamic control of the entire doping process in the P-type silicon carbide crystallization process.

2. A coupled modeling and intelligent solution control system for the dynamic evolution of components in p-type silicon carbide melt prepared by liquefaction, characterized in that, The method for performing the method of claim 1 includes: a multi-source signal synchronous acquisition module, a nonlinear dynamic feature decoupling operation module, a four-dimensional dynamic coupling modeling and solving module, a nonlinear component time series extrapolation and early warning module, a multi-objective dynamic parameter optimization and control module, and an incremental lightweight adaptive learning module.

3. The method according to claim 1, characterized in that: The hybrid regularized nonlinear purification model is solved by a nonstationary dynamic alternating direction multiplier iterative algorithm. It relies on the multi-component transport manifold constraint and sparse constraint to adapt to the non-equilibrium nonlinear dynamic transport characteristics of high-temperature multiphase systems, and achieves precise decoupling of microscopic weak doping dynamic defects.

4. The method according to claim 1, characterized in that: The four-dimensional dynamic coupled finite element model adopts a multi-scale nested adaptive non-uniform finite element discretization method. Through a multi-process step-by-step coupled dynamic iterative algorithm, it quantifies and analyzes the cross-scale dynamic coupling correlation of convection transport, interface phase change, selective adsorption, and high-temperature component decay.

5. The method according to claim 1, characterized in that: The nonlinear ensemble filtering algorithm is based on a dynamic iterative update mechanism of ensemble sampling, which is adapted to the non-Gaussian random fluctuations of multi-component transport and the long-delay accumulation characteristics of high-temperature component decay. It avoids the simplification defects of traditional linear time series models and improves the accuracy of long-period dynamic inference.

6. The method according to claim 1, characterized in that: The multi-objective constrained optimization model uses the dynamic gradient of furnace temperature, real-time crystal pulling rate, dynamic intensity of melt stirring, and dynamic acceleration rate of dopant source as decision variables, embedding dynamic imbalance index threshold and high-temperature thermal stability hard constraints, and employs a sequential quadratic programming algorithm to solve for the nonlinear optimal solution.

7. The method according to claim 1, characterized in that: The lightweight processing is achieved through four-dimensional coupling operator structured pruning and nonlinear dimensionality reduction of high-dimensional component dynamic evolution characteristics, compressing the model's computational dimension and adapting it to low-computing-power real-time dynamic computation scenarios of semiconductor production line edge controllers.

8. The method according to claim 1, characterized in that: The incremental adaptive iterative learning adopts an adaptive mini-batch gradient descent algorithm to dynamically update the interface component transport correction coefficient, interface adsorption compensation coefficient, and high-temperature component attenuation fitting coefficient, adapting to the dynamic fluctuations of raw material batches and the drift of high-temperature aging conditions in the furnace equipment.

9. The method according to claim 3, characterized in that: The doping dynamic imbalance index is a continuous dynamic quantitative evaluation index for the entire crystal domain. It is used to identify local doping dynamic enrichment, component transport hindrance, interfacial selective adsorption anomalies, and latent defects such as dynamic decay of high-temperature components in a hierarchical manner, providing a quantitative basis for adaptive closed-loop dynamic control of liquefaction process parameters.

Images