A periodic structure for accelerating the simulation of electromagnetic field scattering.

By constructing the transverse wave vector matrix and the Fourier convolution matrix, and utilizing the even function properties of the matrix cosine function and the matrix sinc function, combined with the scaling-doubling-angle recursion mechanism, the problem of low parallel efficiency of eigenvalue decomposition in RCWA calculation is solved, and efficient and stable electromagnetic field simulation calculation is achieved.

CN122864182BActive Publication Date: 2026-11-10SIXING SEMICON
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611282792.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-08-24
Publication Date
2026-11-10
Estimated Expiration
2046-08-24

AI Technical Summary

Technical Problem

Existing RCWA computation suffers from problems such as low parallel efficiency of eigenvalue decomposition, heavy computational paths for the full matrix exponent, high storage requirements for temporary matrices, numerical instability in thick or strongly fleeting modes, and excessive GPU memory usage for matrix multiplication.

Method used

The system employs a master control node and multiple computing nodes interconnected via a network interface. By constructing a transverse wave vector matrix and a Fourier convolution matrix, and utilizing the even function properties of the matrix cosine function and the matrix sinc function, combined with a scaling-doubling-angle recursion mechanism, parallel computation is performed. This avoids eigenvalue decomposition and matrix square root calculation, optimizes the computation path, and reduces memory usage.

Benefits of technology

It improves the efficiency and throughput of computing devices, reduces computational load and memory usage, ensures stable convergence over a wide parameter range, supports higher simulation scale and computational throughput, and enhances overall simulation efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122864182B_ABST
    Figure CN122864182B_ABST
Patent Text Reader

Abstract

This invention discloses an accelerated simulation computing device for scattering electromagnetic fields with a periodic structure, relating to the field of computer digital data processing technology. The device includes a master control node and multiple computing nodes. The master control node distributes simulation parameters and tasks to each computing node and cascades the input interface scattering matrix, the scattering matrices of each layer, and the output interface scattering matrix to obtain the global scattering matrix. The computing nodes construct Fourier convolution matrices and RCWA coupling matrix pairs, obtain the matrix cosine function and matrix sinc function through scaling, joint polynomial evaluation, and double-angle recursion, then construct the layer transfer matrices in blocks and convert them into scattering matrices, which are then returned to the master control node. The scattering matrices of each layer are constructed in parallel by multiple GPUs. This invention significantly reduces the amount of GPU matrix multiplication operations, lowers the peak memory usage, and enables the device to support higher truncation orders or more layers of parallel computing with the same memory capacity.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of computer digital data processing technology, and in particular to a periodic structure-based scattering electromagnetic field acceleration simulation computing device. Background Technology

[0002] Rigorous Coupled-Wave Analysis (RCWA) is the mainstream method for numerical simulation of light scattering in periodic media, widely used in semiconductor manufacturing for optical defect detection, critical dimension measurement, and lithography process simulation optimization. The basic principle of RCWA is to expand the periodic dielectric constant distribution and electromagnetic field into spatial Fourier series, establish the coupled-wave equations using Maxwell's equations, calculate the characteristic states of the electric and magnetic fields, then calculate the single-layer scattering matrix, and finally connect the single-layer scattering matrices to obtain the global scattering matrix.

[0003] The most time-consuming part of traditional RCWA methods is solving large-scale eigenvalue problems. This step accounts for more than half of the computation time of the entire RCWA simulation, becoming a bottleneck restricting simulation efficiency. Although the parallel computing architecture of modern GPUs has provided significant acceleration for regular operations such as matrix multiplication and inversion, the eigenvalue solving algorithm has a strict data dependency chain. Subsequent calculation steps must wait for the completion of the preceding steps, which severely limits the parallel potential. At the same time, conditional branches such as convergence judgment can cause GPU threads to diverge, significantly reducing the utilization of the execution units of the GPU architecture.

[0004] To overcome the aforementioned bottlenecks, existing technologies attempt to replace eigenvalue decomposition with matrix exponent approximation. For example, Chinese invention patent publication CN121328142A discloses an improved RCWA scheme using a matrix exponent scaling-flattening method combined with a GPU computing cluster. However, this scheme expands the complete transfer matrix exponent as a whole, failing to fully utilize the inherent structure of matrix functions to simplify computation, resulting in a large number of dense matrix multiplications and high GPU memory consumption. Another example is the use of polynomial approximation of matrix exponents to replace the most time-consuming eigenvalue decomposition step in traditional RCWA, thereby accelerating the simulation of arbitrary two-dimensional periodic structures. However, this scheme directly cascades the transfer matrix in thick layers or strongly evanescent modes, easily introducing numerical overflow or ill-conditioned amplification. Furthermore, existing technologies have proposed a differentiable RCWA implementation based on the square root of a matrix. While this bypasses explicit eigenvector solving, the square root of the matrix itself still needs to handle issues such as branch selection, iterative convergence, and spectral distribution, resulting in a long computational path and high engineering implementation complexity. Summary of the Invention

[0005] In view of the above-mentioned defects or deficiencies in the prior art, the present invention provides a periodic structure scattering electromagnetic field accelerated simulation computing device to solve the technical problems existing in the current RCWA calculation, such as low parallel efficiency of eigenvalue decomposition, heavy full matrix exponent calculation path, high temporary matrix storage, numerical instability in thick layer or strong evanescent mode, and high memory resource consumption of matrix multiplication.

[0006] One aspect of the present invention provides an accelerated simulation computing device for scattered electromagnetic fields with a periodic structure, comprising: The master control node and multiple computing nodes are interconnected via a network interface; The master control node is used to acquire the electromagnetic parameters of the input medium of the incident light, the electromagnetic parameters of the output medium of the transmitted light, the electromagnetic parameters of the gap medium, and the parameters of each layer of the periodic structure. Based on the structural period and Fourier harmonics in the parameters of each layer, it constructs the transverse wave vector matrix of each layer. The task of constructing the electromagnetic parameters of the input medium of the incident light, the electromagnetic parameters of the output medium of the transmitted light, the electromagnetic parameters of the gap medium, the parameters of each layer of the periodic structure, the transverse wave vector matrix of each layer, and the scattering matrix of each layer is distributed to each computing node for parallel execution. The node then summarizes the scattering matrices of each layer, the input interface scattering matrix, and the output interface scattering matrix returned by each computing node. Finally, the node concatenates the input interface scattering matrix, the scattering matrices of each layer, and the output interface scattering matrix sequentially to obtain the global scattering matrix. Each computing node, in response to the scattering matrix construction task sent by the master node, constructs the Fourier convolution matrix of each layer based on the dielectric constant and permeability distributions in the allocated layer parameters. It defines the normalized thickness parameter of each layer based on the layer thickness and incident light wavelength in the allocated layer parameters. Based on the Fourier convolution matrix and the transverse wave vector matrix of each layer, it constructs RCWA coupling matrix pairs for each layer and calculates the product of the RCWA coupling matrix pairs. It determines the scaling parameter based on the norm of the matrix to be scaled, determined by the product of the RCWA coupling matrix pairs and the normalized thickness parameter. It then scales the matrix to be scaled and obtains the matrix cosine through joint polynomial evaluation at the scaling scale. The function and matrix sinc function are then recursively restored to the original scale using the same scaling parameter through a doubling angle, yielding the matrix cosine function and matrix sinc function at the original scale. Each assigned layer is considered to be surrounded by a gap medium on both sides. Based on the matrix cosine function and matrix sinc function, the transmission matrix blocks of each assigned layer are constructed, and these blocks are converted into scattering matrices for each layer. An input interface scattering matrix is ​​constructed based on the electromagnetic parameters of the input medium and the gap medium of the incident light. An output interface scattering matrix is ​​constructed based on the electromagnetic parameters of the output medium and the gap medium of the transmitted light. Finally, the scattering matrices of each layer, the input interface scattering matrix, and the output interface scattering matrix are fed back to the master control node.

[0007] Furthermore, each computing node includes a CPU and multiple GPUs connected via a bus. The GPUs collaborate on computation and exchange data via a high-speed interconnect protocol. The CPU of the computing node, in response to the layer scattering matrix construction task sent by the master node, constructs the Fourier convolution matrix of each layer based on the dielectric constant and permeability of the allocated layer parameters. It defines the normalized thickness parameter of each layer based on the layer thickness and incident light wavelength of the allocated layer parameters. Based on the Fourier convolution matrix and transverse wave vector matrix of each layer, it constructs RCWA coupling matrix pairs for each layer and calculates the product of these RCWA coupling matrix pairs. Then, it sends the product of the RCWA coupling matrix pairs, the normalized thickness parameter of each layer, the electromagnetic parameters of the input medium of the incident light, the electromagnetic parameters of the output medium of the transmitted light, and the electromagnetic parameters of the gap medium to the multiple GPUs of the node. Each GPU of the computing node is used to construct the RCWA coupling matrix pairs and the normalized thickness parameter of the allocated layer based on the product of the allocated layer RCWA coupling matrix pairs and the normalized thickness parameter. The scaling parameters are determined by the norm of the matrix to be scaled. The matrix is ​​scaled and the matrix cosine function and matrix sinc function are obtained by joint polynomial evaluation at the scaling scale. Then, the original scale is restored by double-angle recursion with the same scaling parameters to obtain the matrix cosine function and matrix sinc function at the original scale. The two sides of the assigned layer are regarded as being surrounded by the gap medium. The transmission matrix blocks of the assigned layer are constructed according to the matrix cosine function and matrix sinc function. The transmission matrix blocks are converted into the scattering matrix of the layer. The input interface scattering matrix is ​​constructed according to the electromagnetic parameters of the input medium of the incident light and the electromagnetic parameters of the gap medium. The output interface scattering matrix is ​​constructed according to the electromagnetic parameters of the output medium of the transmitted light and the electromagnetic parameters of the gap medium. The scattering matrix of each layer, the input interface scattering matrix, and the output interface scattering matrix are transmitted back to the master node via the CPU of the local node. The multiple GPUs in the computing node compute the scattering matrix, input interface scattering matrix, and output interface scattering matrix of each layer in parallel.

[0008] Furthermore, the CPU of the computing node is also used to determine the lateral spatial resolution of each layer based on the structural period and Fourier harmonics number in the allocated layer parameters, and to determine the thickness threshold of each layer based on the lateral spatial resolution. If the thickness of a certain layer exceeds the thickness threshold, the layer is uniformly divided into multiple sub-layers, and the thickness of each sub-layer is used to replace the original layer thickness, and the normalized thickness parameter of the layer is redefined. The GPU of the computing node is also used to merge the scattering matrices of all sub-layers into the scattering matrix of the layer after obtaining the scattering matrix of each sub-layer of the layer.

[0009] Furthermore, the evaluation of the joint polynomial uses 12th or 18th order Taylor polynomials as an approximation of the matrix cosine function and the matrix sinc function; wherein, the 12th or 18th order Taylor polynomial evaluation process pre-calculates and reuses low-order matrix powers, rearranging the high-order matrix polynomials corresponding to the Taylor polynomials into linear combinations, products, and nested combinations of multiple low-order matrix polynomials, so as to reduce the number of matrix multiplications.

[0010] Furthermore, the GPU of the computing node is also used to construct the magnetic field feature matrix of the gap medium when constructing the transfer matrix block of each layer, define an auxiliary matrix by using the relationship that the inverse matrix of the magnetic field feature matrix of the gap medium is equal to its own negative, and construct the transfer matrix block by linear combination of the auxiliary matrices.

[0011] Furthermore, the GPU of the computing node is also used to calculate only the reflection block and the transmission block when converting the transmission matrix into scattering matrices for each layer, so as to obtain the complete layer scattering matrix; wherein, the reflection block is the lower left block of the layer scattering matrix, representing the amplitude and phase of the electromagnetic wave reflected on the same side when it is incident from above the layer; the transmission block is the upper left block of the layer scattering matrix, representing the amplitude and phase of the electromagnetic wave transmitted on the other side when it is incident from above the layer.

[0012] Furthermore, the GPU of the computing node is also used to: perform matrix multiplication in any of the following computation processes using a batch general matrix multiplication function: scaling the matrix to be scaled, calculating the matrix cosine function and matrix sinc function, restoring to the original scale by double-angle recursion with the same scaling parameters, constructing the scattering matrix of each layer, constructing the input interface scattering matrix, and constructing the output interface scattering matrix; and to solve the linear equation system in the process of converting the transmission matrix into the scattering matrix of the layer by a batch LU decomposition algorithm, as well as the linear equation system in the process of cascading to obtain the global scattering matrix.

[0013] Furthermore, the simulation computing device employs parallel computing strategies, including data parallel mode, model parallel mode, or hybrid parallel mode. When using data parallel mode for parallel computing, RCWA simulation computing tasks with different incident wavelengths or different incident angles are assigned to different GPUs, and each GPU independently executes the complete simulation computing. When using model parallel mode for parallel computing, it includes vertical layering mode and horizontal partitioning mode. In vertical layering mode, the scattering matrix construction tasks corresponding to different physical layers of the periodic structure are assigned to different GPUs for parallel execution. In horizontal partitioning mode, the computing domain of the same physical layer is divided into multiple sub-regions, and each sub-region is assigned to a different GPU to independently execute the scattering matrix calculation task corresponding to that sub-region. When using hybrid parallel mode for parallel computing, the computing architecture is constructed by combining the aforementioned data parallel mode and model parallel mode.

[0014] Furthermore, the periodic structure is a periodic structure on a semiconductor wafer, wafer mask, diffractive optical element, metasurface, photonic crystal, or grating waveguide.

[0015] Furthermore, the master control node is also used to calculate and output any one of the following information based on the global scattering matrix: the reflectivity, transmittance, or diffraction efficiency of the diffracted waves of different orders diffracted in different directions by the periodic structure under incident light illumination; and the spatial distribution of the electromagnetic field inside and near the periodic structure.

[0016] The present invention provides an accelerated simulation calculation device for scattered electromagnetic fields of a periodic structure, which has the following beneficial effects: (1) This invention utilizes the even function properties of the matrix cosine function and the matrix sinc function to directly represent the layer response as a matrix polynomial of the product of coupled matrices without solving for the square root of the matrix. This transforms the calculation into an operation form dominated by batch matrix multiplication, replacing the eigenvalue decomposition step in the traditional scheme. This avoids the GPU execution unit from being idle due to serial dependence, reduces the computational load and memory usage, and improves the device's computational efficiency.

[0017] (2) This invention divides the ultra-thick layer into multiple sub-layers and merges the scattering matrices of the sub-layers into the scattering matrix of the original layer, thus avoiding exponential overflow. At the same time, it combines a scaling-doubling-angle recursion mechanism to control the convergence of the matrix cosine function and the matrix sinc function. The two together constitute a dual stability guarantee, enabling the simulation to converge stably over a wide parameter range, thereby avoiding the recalculation overhead and waste of computing resources caused by numerical anomalies.

[0018] (3) The present invention combines the calculation of matrix cosine function and matrix sinc function into a single joint evaluation. The two functions can be obtained simultaneously with one approximation without the need to perform Taylor expansion on the two functions separately. This eliminates repeated matrix multiplication, reduces the time occupied by GPU computing units, shortens the single-layer computing time, and improves the device throughput.

[0019] (4) This invention scales the matrix using scaling parameters to control its norm within the region of convergence, then performs low-order polynomial evaluation on the scaled scale, and finally restores it to the original scale through double-angle recursion. The above mechanism requires only a small number of matrix multiplications to meet the conventional accuracy requirements under the optimized 12th-order Taylor polynomial expansion, and can meet the high-precision requirements under the 18th-order Taylor polynomial expansion.

[0020] (5) This invention only requires maintaining two functions, matrix cosine and matrix sinc, which significantly reduces the peak temporary storage amount, thereby alleviating the peak memory occupation pressure of the GPU and enabling the device to support higher truncation order or execute more layers in parallel under the same memory, thus improving the simulation scale and computing throughput.

[0021] (6) Conventional cascading of scattering matrices requires calling the linear solver separately for two different coefficient matrices, and the next call can only begin after the previous one is completed. This invention uses matrix identities to merge the two sets of solutions into one LU decomposition and batch solution, so that each cascading step only needs to call the solver function once, reducing the calling overhead and improving the computational efficiency of the cascading process.

[0022] (7) The core calculations of this invention are mainly based on batch dense matrix multiplication and batch linear equation solving, which can be directly implemented by calling the GPU's underlying high-performance computing library. Since the tasks of each layer, wavelength, and incident angle are independent of each other, the device can distribute the tasks to multiple GPUs in parallel for execution, so that the computing resources can be linearly expanded with the number of GPUs, significantly improving the overall simulation efficiency.

[0023] (8) The present invention divides the computing task into data preparation and scheduling on the CPU side and intensive computing on the GPU side. The task allocation, data distribution and result aggregation mechanism between the master control node and the computing node makes full use of the high-speed network interface and bus bandwidth of the device, so that multiple GPU computing resources are fully utilized and the overall simulation computing efficiency is significantly improved. Attached Figure Description

[0024] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a schematic diagram of an electromagnetic wave scattering detection scenario for a periodic structure provided in one embodiment of this application; Figure 2This is a schematic diagram of the system structure of a periodic structure electromagnetic field scattering acceleration simulation computing device provided in one embodiment of this application; Figure 3 This is a flowchart illustrating an embodiment of the RCWA simulation calculation method based on trigonometric polynomial evaluation provided in this application. Figure 4 This is a schematic diagram of multi-GPU parallel task allocation provided in one embodiment of this application. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] To facilitate understanding of this invention, the following defines and explains important technical terms appearing in the specification and claims.

[0027] Interstitial medium: In RCWA simulation, a zero-thickness vacuum medium layer is embedded on both sides of each layer of a periodic structure to serve as a reference for constructing the scattering matrix of each layer.

[0028] Magnetic field characteristic matrix: refers to the transformation matrix between the transverse component of the magnetic field and the amplitude of the intrinsic wave in a uniform medium (including but not limited to gap medium, input medium or output medium).

[0029] Matrix cosine function: refers to the cosine function with a matrix as its independent variable, defined as follows: ,in It is an arbitrary square matrix.

[0030] Matrix sinc function: refers to the sinc function that takes a matrix as its independent variable, and is defined as follows: ,in It is an arbitrary square matrix.

[0031] Scaling parameter: refers to the number of scaling rounds used to control the matrix norm, so that the scaled matrix norm is within the region of convergence.

[0032] Double-angle recursion: refers to the recursive operation of doubling the arguments of the matrix cosine function and the matrix sinc function, used to restore the calculation results under the scale to the original scale.

[0033] Matrix to be scaled: refers to the product of the coupling matrices. With normalized thickness parameters The matrix formed by multiplying the squares of , i.e. The norm of this matrix is ​​used to determine the scaling parameters, and it is itself the object being scaled.

[0034] Transmission matrix partitioning: refers to single-layer transmission matrix The bottom two of the four sections and It is used to fully restore the symmetric layer scattering matrix.

[0035] Fourier convolution matrix: refers to the Toeplitz matrix obtained by performing a Fourier series expansion on the dielectric constant or permeability distribution of a periodic structure. It is used to convert the dielectric constant or permeability distribution in the spatial domain into a diagonal matrix operation in the frequency domain.

[0036] Transverse wave vector matrix: refers to the diagonal matrix formed by combining the transverse components of the incident wave vector and the lattice reciprocal. and Its diagonal elements are structurally periodic. , The Fourier harmonic order is determined.

[0037] RCWA coupling matrix pair: refers to two matrices in a first-order coupled system of rigorous coupled-wave analysis (RCWA). and It is constructed by combining the Fourier convolution matrix and the transverse wave vector matrix, which together describe the coupling relationship of the transverse electromagnetic field in the multilayer structure.

[0038] Normalized thickness parameter: refers to the physical layer thickness and incident light wavelength Combined into a single dimensionless parameter , which represents the ratio of layer thickness to wavelength.

[0039] Scattering matrix: A matrix that describes the relationship between the input and output fields when an electromagnetic wave passes through an optical structure, used to analyze reflection and transmission characteristics.

[0040] Input interface scattering matrix: refers to the scattering matrix that describes the reflection and transmission characteristics of incident light when it enters the gap medium (i.e., the interface before entering the periodic structure) from the input medium.

[0041] Output interface scattering matrix: refers to the scattering matrix that describes the reflection and transmission characteristics of transmitted light when it enters the output medium (i.e., the interface after passing through the periodic structure) from the gap medium.

[0042] Global scattering matrix: refers to the scattering matrix obtained by cascading the scattering matrix of the input interface, the scattering matrices of all layers of the periodic structure, and the scattering matrix of the output interface in sequence. It fully describes the overall scattering characteristics of incident light from the input medium to the output medium.

[0043] Structural periodicity: refers to the physical spacing of a periodic structure that repeats in a two-dimensional transverse direction, also known as lattice periodicity.

[0044] Fourier harmonics: refers to the number of harmonic terms retained in both directions when performing a Fourier series expansion on a periodic structure.

[0045] Example 1 This embodiment uses optical detection as an example to illustrate the electromagnetic wave scattering detection scenario of a periodic structure, so as to clarify the physical model targeted by the accelerated simulation computing device for scattered electromagnetic fields.

[0046] Figure 1 A two-dimensional periodic structure is shown. for The structural period in the axial direction, for A schematic diagram of optical scattering under plane wave incidence (with a periodic structure along the axial direction). In the diagram, the input layer refers to the layer containing the input medium of the incident light, and the output layer refers to the layer containing the output medium of the transmitted light. Between the input and output layers is a multi-layered structure with a periodic structure (layer 1 to layer 2). Layers, with thicknesses of respectively to Incident light strikes the surface of a periodic structure from the input layer at a certain angle. Multiple reflections and transmissions occur within the periodic structure. Part of the light is reflected back to the input layer, forming reflected light; the rest is transmitted through the periodic structure into the output layer, forming transmitted light. This detection scenario is widely used in engineering applications such as semiconductor wafer inspection, optical critical dimension measurement, photolithography simulation optimization, diffractive optical element design, and metasurface analysis.

[0047] For the electromagnetic scattering problem of the aforementioned periodic structure, this invention employs Rigorous Coupled-Wave Analysis (RCWA) as the numerical solution method. RCWA is a mainstream method for numerical simulation of light scattering in periodic media. Its core principle is to expand the dielectric constant and permeability distributions of the periodic structure into Fourier series in the transverse direction, and also expand the electromagnetic field into corresponding spatial harmonics. This transforms Maxwell's equations from a continuous problem in the spatial domain into a discrete algebraic eigenvalue problem in the frequency domain. Furthermore, by solving for the eigenmodes of each layer and matching the interlayer boundary conditions, the reflection and transmission characteristics of the entire structure can be obtained.

[0048] Specifically, in Figure 1 In the periodic multilayer structure shown, each layer is arranged in a repeating pattern on the transverse plane (the plane containing the surface of the periodic structure), and the dielectric constant distribution of each layer is as follows. and permeability distribution Both are periodic functions of the transverse plane coordinates, and therefore can be expanded into a Fourier series. When the incident light is incident at a certain angle, its transverse wave vector component combines with the lattice inverse, along with the retained Fourier harmonics. and The transverse wave vector matrix of each layer can be constructed. , and the coupling matrix pair in the first-order coupled system of RCWA and Therefore, the electromagnetic field propagation behavior of each layer is compressed into a layer scattering matrix, which describes the reflection and transmission characteristics of incident light on both sides (above and below) of the layer. The global scattering matrix of the entire periodic structure is obtained by cascading the input interface scattering matrix, the scattering matrices of each layer, and the output interface scattering matrix. Finally, based on the global scattering matrix and the incident light amplitude, the reflectivity, transmittance, and diffraction efficiency of each diffraction order after the incident light passes through the periodic structure are calculated, and the diffraction efficiency spectrum and the electromagnetic field distribution characteristics of the space near the periodic structure are output.

[0049] The most time-consuming step in traditional RCWA methods is eigenvalue decomposition of the coupling matrix at each layer. This step is difficult to fully utilize parallel computing resources on GPU architectures due to strong data dependencies and conditional branches. This invention addresses this bottleneck by providing a method for constructing RCWA layer scattering matrices based on trigonometric polynomial evaluation and an accelerated computing device. By compressing the layer propagation response into a joint polynomial evaluation of matrix cosine and matrix sinc functions, and combining mechanisms such as scaling-doubling-angle recursion, thick layer partitioning, and interstitial substrate transformation, eigenvalue decomposition and matrix square root calculation are completely avoided. This also reduces peak memory usage, transforming the core computation into a computational form primarily based on batch dense matrix multiplication and batch linear equation solving, making it suitable for efficient parallel execution on GPU clusters.

[0050] Figure 1 The physical model shown is the application object of the simulation computing device described in this invention. Figure 1 The incident light wavelength, angle, polarization state, and geometric and material parameters of the periodic structure obtained by measurement are the simulation input parameters acquired by the information acquisition module in the simulation computing device described in this invention. The system architecture and specific calculation methods of the device of this invention will be further explained below with reference to subsequent embodiments.

[0051] Example 2 See Figure 2 This embodiment provides a periodic structure scattering electromagnetic field acceleration simulation computing device 100, including a master control node 101 and multiple computing nodes 102 interconnected via a network interface.

[0052] The simulation computing device 100 receives simulation tasks transmitted from an external input device through the master control node 101. Optionally, the external input device is a user computer 103, which includes a monitor, host, keyboard, and other input devices. The simulation task includes simulation pattern information and simulation optical state information. The simulation pattern information includes one or more of the following: pattern structure, material parameters of each layer, thickness of each layer, and structural period. The simulation optical state information includes incident wavelength, incident angle, and incident amplitude.

[0053] The master control node 101 is equipped with a CPU and a high-speed network interface 104. The CPU communicates with multiple computing nodes 102 through the high-speed network interface 104. The high-speed network interface 104 can be an InfiniBand network card, an Ethernet network card, or other network interfaces that support high-speed data transmission. After receiving simulation tasks transmitted from external input devices, the master control node 101 divides and schedules the simulation tasks according to the number of parameters to be calculated, the load of the computing nodes, and the GPU resource status of the computing nodes.

[0054] Specifically, the CPU of the master control node 101 is used to acquire the electromagnetic parameters of the input medium of the incident light, the electromagnetic parameters of the output medium of the transmitted light, the electromagnetic parameters of the gap medium, and the parameters of each layer of the periodic structure. The parameters of each layer of the periodic structure include the dielectric constant distribution, magnetic permeability distribution, layer thickness, and the structural periodicity in two two-dimensional directions. and Fourier harmonics in two two-dimensional directions and and the wavelength of the incident light Among them, structural periodicity and The physical spacing of a periodic structure arranged in a repeating pattern in a two-dimensional transverse direction. Fourier harmonics. and The number of harmonic terms retained in two directions when performing a Fourier series expansion on a periodic structure, satisfying the following conditions: ,in This represents the total number of Fourier harmonics.

[0055] The CPU of master node 101 is also used to construct a transverse wave vector matrix based on the structure period and the Fourier harmonic number. Its diagonal element is obtained by normalizing the incident wave vector, lattice inverse vector, and vacuum wavenumber. Subsequently, the master node 101 distributes the construction tasks of the electromagnetic parameters of the input medium of the incident light, the electromagnetic parameters of the output medium of the transmitted light, the electromagnetic parameters of the interstitial medium, the parameters of each layer of the periodic structure, the transverse wave vector matrix of each layer, and the scattering matrix of each layer to each computing node 102 for parallel execution.

[0056] Each computing node 102 includes a central processing unit (CPU), memory, a PCIe high-speed bus 105, and a GPU array 106. The CPU performs data preprocessing, local task management, GPU task scheduling, and result processing. The GPU array 106 includes multiple GPUs, each connected to the CPU via the PCIe high-speed bus 105, and performs parallel computing tasks such as matrix multiplication, matrix polynomial evaluation, solving linear equations, and constructing layer scattering matrices. Multiple GPUs within the same computing node 102 communicate with each other via a high-speed GPU interconnect protocol. This high-speed GPU interconnect protocol can employ NVLink, xGMI, UALink, XeLink, etc., to enable matrix data exchange, multi-GPU collaborative computing, and intermediate result transmission between GPUs.

[0057] The master node 101 distributes the scattering matrix construction tasks for each layer to the GPUs of multiple computing nodes 102 for parallel execution. After receiving the tasks distributed by the master node 101, the CPU of each computing node 102, in response to the scattering matrix construction tasks for each layer sent by the master node, performs the following operations according to the allocated layer parameters: The CPU of compute node 102 calculates the dielectric constant distribution in the parameters allocated to each layer. and permeability distribution Fourier expansions are performed on each layer to construct the Fourier convolution matrix of the dielectric constant for each layer. (The double brackets indicate the Toeplitz matrix consisting of Fourier coefficients, not just the Fourier coefficients.) A certain scalar value, The Fourier convolution matrix of dielectric constant and permeability (The double brackets indicate the Toeplitz matrix consisting of Fourier coefficients, not just the Fourier coefficients.) A certain scalar value, (where is the permeability). In the implementation using the inverse factorization rule, it is also possible to construct... and Inverse material convolution matrix.

[0058] The CPU of compute node 102 determines the layer thickness based on the layer parameters allocated to it. and incident light wavelength Define the normalized thickness parameters for each layer. .

[0059] The CPU of computing node 102 is based on the Fourier convolution matrix and the transverse wave vector matrix of each layer. Construct the RCWA coupling matrix pairs for each layer and , And calculate the product of the RCWA coupling matrix pairs of each layer. .

[0060] When the layer thickness is large, the CPU of computing node 102 is also used to determine the structural period in the parameters allocated to each layer. and Fourier harmonics Determine the lateral spatial resolution of each layer and The thickness threshold of each layer is determined based on the lateral spatial resolution. If the thickness of a certain layer Thickness greater than the threshold Then the CPU of computing node 102 will uniformly divide this layer into multiple sub-layers, with the number of sub-layers being: ,in , This indicates the thickness of each sublayer. Not exceeding the stability threshold The upper limit of the required number of partitioning rounds. Thickness of each sublayer. The original layer thickness is replaced by the thickness of each sublayer, and the normalized thickness parameter of the layer is redefined.

[0061] The CPU of node 102 computes the product of the RCWA coupling matrix pairs of each layer. Normalized thickness parameters of each layer The electromagnetic parameters of the input medium of the incident light, the output medium of the transmitted light, and the gap medium are sent to multiple GPUs on this node, and the construction and calculation process of the scattering matrix of each layer is executed in parallel by the multiple GPUs.

[0062] Each GPU is based on the product of the RCWA coupling matrix pairs assigned to its layer. and normalized thickness parameters The determined matrix to be scaled norm Determine scaling parameters : in, Indicates to Round up, i.e., take the value greater than or equal to the nearest integer. The smallest integer.

[0063] GPU on the matrix to be scaled Scale the matrix and calculate the product of the scaled matrices. The matrix cosine function is obtained by joint polynomial evaluation under scaling conditions. Sinc function of the sum matrix Then through The wheel is recursively restored to its original scale using a multiple-angle approach.

[0064] Specifically, the process of double-angle recursion is as follows: Let the initial values ​​of the matrix cosine function and the matrix sinc function be... and : right Perform the following update steps in sequence: go through After round recursion, we get: Due to scaling parameters satisfy Finally, the matrix cosine function and matrix sinc function at the original scale are obtained: in, The sinc function represents the matrix at the current scale; Represents the matrix cosine function at the current scale; Represents the identity matrix; For scaling parameters, This is the normalized thickness parameter.

[0065] The above-described doubling-angle recursion can be described as follows: Starting from the joint evaluation result under the scaling scale, using the matrix cosine function and matrix sinc function under the scaling scale as initial values, each round of recursion sequentially updates the matrix sinc function and the matrix cosine function, approximately doubling the argument; this process is repeated. After the round, the number of arguments changes from the scaling factor. Restored to the original scale This yields the matrix cosine and sinc functions at the original scale. Scaling parameters This simultaneously determines the compression factor during scaling and the number of recursive iterations during restoration; the two are mathematically perfectly symmetrical. Through the above... The round-double-angle recursion achieves accurate recovery from the scaled scale to the original scale, and only requires two matrix multiplications per round, resulting in high computational efficiency.

[0066] The GPU treats both sides of the assigned layer as being surrounded by a gap medium, and constructs the magnetic field characteristic matrix of the gap medium. : in, This represents the transverse wave vector matrix in the x-direction; This represents the transverse wave vector matrix in the y-direction; This represents the wave vector matrix in the z-direction of the gap medium; for The inverse matrix; Represents the identity matrix.

[0067] GPUs utilize relational expressions Define auxiliary matrix as follows: in, It is the matrix cosine function; The imaginary unit; The thickness parameter is normalized. This represents the characteristic matrix of the magnetic field of the interstitial medium. The sinc function is used for matrices. This is the first coupling matrix in the RCWA coupling matrix pair; This is the second coupling matrix in the RCWA coupling matrix pair; for The inverse matrix.

[0068] Through the auxiliary matrix and Linear combination of transmission matrix blocks and : The GPU constructs the input interface scattering matrix based on the electromagnetic parameters of the input medium and the interstitial medium of the incident light. The output interface scattering matrix is ​​constructed based on the electromagnetic parameters of the output medium and the interstitial medium. For the construction of the interface scattering matrix, and The explicit expression is: in, and These are the magnetic field characteristic matrices of the input and output media, respectively; This represents the characteristic matrix of the magnetic field of the interstitial medium. It is the relative permittivity of the input medium; It is the relative permeability of the input medium; It is the relative permittivity of the output medium. It is the relative permeability of the output medium; For input medium Directional wave vector matrix; For output medium Directional wave vector matrix; for The inverse matrix; for The inverse matrix; It is an identity matrix.

[0069] The GPU divides the transfer matrix into blocks and converts them into the scattering matrix of that layer. In the interstitial medium embedding form, the layer scattering matrix has a symmetric structure. and Therefore, the GPU only needs to compute two blocks: in, The transmission block of the layer scattering matrix represents the amplitude and phase of an electromagnetic wave transmitted from above the layer to the other side (i.e., below), and is located in the upper left block of the layer scattering matrix. The reflection block of the layer scattering matrix represents the amplitude and phase of an electromagnetic wave reflected from the bottom of the layer on the same side (i.e., below). It is located in the upper right block of the layer scattering matrix. The lower left block of the transmission matrix; This is the inverse matrix of the lower right block of the transfer matrix. The inverse matrix operation is achieved by solving a system of linear equations through LU decomposition, ultimately yielding the complete layer scattering matrix. : If the layer is split into multiple sub-layers, the GPU obtains the scattering matrix of each sub-layer and then merges the scattering matrices of all sub-layers into the scattering matrix of the layer through Redheffer auto-multiplication.

[0070] The GPU transmits the scattering matrices of each layer, the input interface scattering matrix, and the output interface scattering matrix back to the master node 101 via the CPU and high-speed network interface of this node. The master node 101 then aggregates the scattering matrices of each layer, the input interface scattering matrix, and the output interface scattering matrix transmitted back from each compute node.

[0071] The master node 101 will scatter the input interface matrix Scattering matrices of each layer and output interface scattering matrix The global scattering matrix is ​​obtained by cascading the scattering sequentially and performing optimized Redheffer star multiplication. : In the Redheffer cascade process, master node 101 uses matrix identities to combine the solutions of two sets of linear equations into a single LU decomposition and a batch right-hand side solution. Specifically, for the two scattering matrices... and The cascaded structure, with master node 101 pair of coefficient matrices Perform a single LU decomposition and solve for the two right-hand sides in batch: Wherein, the coefficient matrix The coefficient matrix that needs to be inverted in Redheffer star multiplication is composed of a block combination of two scattering matrices; It is the identity matrix; Scattering matrix The reflective block; Scattering matrix The reflective block; Coefficient matrix The inverse matrix; Scattering matrix Transmission block; Scattering matrix Transmission block; For the first right-hand term, This is the solution for the second right-hand term.

[0072] Master node 101 uses the above-described optimized cascading method to sequentially cascade the input interface scattering matrix, the scattering matrices of all layers, and the output interface scattering matrix into a global scattering matrix. And based on the global scattering matrix The parameters such as incident light amplitude are used to calculate one or more of the following parameters for each diffraction order: reflectivity, transmittance, diffraction efficiency, and near-field electromagnetic field distribution (this calculation process is a conventional technique in the field and will not be described in detail in this embodiment).

[0073] The master control node 101 is also connected to the simulation result output unit 107, which receives one or more of the following calculated by the master control node 101: the global scattering matrix, reflectivity, transmittance, diffraction efficiency of each diffraction order, and near-field electromagnetic field distribution. The simulation result output unit 107 can be a display, a data storage device, or a data interface connected to an external simulation application device.

[0074] It should be noted that, Figure 2The number of computing nodes 102 and GPUs shown are for illustrative purposes only and do not constitute a limitation of the present invention. In practical applications, the number of computing nodes, the number of GPUs within each computing node, and the interconnection method between GPUs can be expanded or adjusted according to the computational scale of the problem to be simulated.

[0075] This invention only requires maintaining the matrix cosine function during the construction of the single-layer scattering matrix. Sinc function of the sum matrix Two intermediate matrices are required, whereas existing full matrix exponentiation schemes require maintaining multiple intermediate matrices corresponding to each level of the matrix exponentiation expansion. Due to the reduction in the number of intermediate matrices, the peak temporary storage requirement of this invention during single-layer construction is approximately half that of the full matrix exponentiation scheme. The practical effect of this storage advantage is that, with the same GPU memory capacity, the device can use the saved memory resources to increase the Fourier harmonic truncation order. (i.e., using computation with higher spatial resolution), or loading computational tasks with more layers and more parameter combinations simultaneously into GPU memory for parallel processing. Therefore, this invention is particularly suitable for large-scale 3D RCWA simulations and batch simulation scenarios that require scanning a large number of wavelengths, incident angles, or structural parameters, while maintaining high computational throughput on GPU devices with limited memory resources.

[0076] Example 3 See Figure 3 This embodiment provides an RCWA simulation calculation method based on trigonometric function polynomial evaluation, which is applied to the simulation calculation device in Embodiment 1.

[0077] The following explanation uses a multi-layered periodic structure of a single target to be simulated as an example. For batch simulations of multiple wavelengths, multiple incident angles, or multiple targets to be tested, the same procedure can be followed for each simulation.

[0078] Step S101: Obtain the parameters of each layer and the thick layer partition, construct the RCWA coupling matrix pair and calculate the product.

[0079] For each layer of the target periodic structure, obtain the dielectric constant distribution of that layer. Permeability distribution Layer thickness and incident light wavelength .right and Perform Fourier expansions on each side to construct the Fourier convolution matrix of the dielectric constant. (The double brackets indicate the Toeplitz matrix consisting of Fourier coefficients, not just the Fourier coefficients.) A certain scalar value, The Fourier convolution matrix of dielectric constant and permeability (The double brackets indicate the Toeplitz matrix consisting of Fourier coefficients, not just the Fourier coefficients.) A certain scalar value, (where is the permeability); in the implementation using the inverse factorization rule, it is also possible to construct... and The inverse material convolution matrix. A normalized thickness parameter is defined for each layer. .

[0080] Let the total number of two-dimensional Fourier harmonics retained in this layer be . The diagonal matrix of the transverse wave vector is Its diagonal elements are obtained by normalizing the incident wave vector, lattice reciprocal, and vacuum wavenumber. The coupling matrix pair in the first-order coupled RCWA system can be constructed from the material convolution matrix and the transverse wave vector matrix. and A commonly used method for constructing isotropic materials is as follows: in and This represents the equivalent effect achieved by inverting the corresponding convolution matrix or solving a system of linear equations, without requiring an explicit inverse matrix. For non-magnetic optical materials, we can take... For one-dimensional gratings or implementations using different Fourier factorization rules, and The block representation can be simplified or replaced accordingly, but their common feature is that they are all determined by the material convolution matrix and the transverse wave vector matrix. Since the transverse electromagnetic field contains two components, therefore... and The matrix has both rows and columns. ,in This represents the total number of Fourier harmonics. The meanings of the other parameters are the same as in Example 1, and will not be repeated in this example.

[0081] When calculating a single-layer transmission matrix, this invention uses the exponential form of the transmission matrix. Let's begin. After expanding the exponent of this matrix into a power series, a pattern emerges: even powers appear only in the diagonal partition of the matrix, while odd powers appear only in the anti-diagonal partition. Based on this structural characteristic, this invention decomposes the exponent operation into diagonal and anti-diagonal partition processing, respectively. Specifically, the diagonal partition corresponds to the matrix cosine function, and the anti-diagonal partition corresponds to the matrix sinc function. Both functions contain only... The even power of . Therefore, there is no need to calculate the square root of the matrix. This bypasses the computationally intensive and numerically sensitive step, simplifying the calculation process.

[0082] Due to the matrix cosine function Sinc function of the sum matrix It can be directly expressed as The matrix function will convert the transmission matrix into its exponential form. Summing by even powers and odd powers separately yields the following: , , and The compact block representation: Therefore, the present invention only requires and Two matrix functions, without having to evaluate the entire transfer matrix as a whole.

[0083] Furthermore, when the layer thickness is large, the matrix to be scaled... norm It may far exceed the order of magnitude of a single unit, resulting in a scaling wheel count. The resolution is too large. Therefore, a thick-layer partitioning mechanism is introduced. The lateral spatial resolution is defined. and Take the thickness threshold. The selection criterion for the thickness threshold stems from the relationship between lateral spatial resolution and the propagation operator norm. Specifically, the higher the Fourier harmonic truncation order and the smaller the structural period, the higher the lateral spatial resolution, and the thinner the sublayer should be to control the scaling rounds of the single-block matrix. If the layer thickness... Then the original layer will be evenly divided into Each sub-layer, among which Thickness of each sublayer First, regarding the thickness... Construct a single-block scattering matrix Then through Redheffer's self-star ride Restore the original layer response. Each round of auto-multiplication doubles the equivalent layer number. The original thickness was obtained exactly after the wheel. The corresponding scattering matrix.

[0084] Step S102: Scaling-double angle recursion and joint polynomial evaluation to obtain matrix cosine and matrix sinc functions.

[0085] This step constitutes the core calculation process of the RCWA simulation calculation method based on trigonometric function polynomial evaluation.

[0086] First, based on the coupling matrix, the product... The 1-norm automatically determines the scaling parameters. This controls the truncation error of subsequent low-order polynomial expansions.

[0087] Scaling parameters It can be represented as: in, Indicates to Round up, i.e., take the value greater than or equal to the nearest integer. The smallest integer.

[0088] Next, the scaled matrix parameters are calculated. Then calculate the matrix cosine function under the scaling factor. Sinc function of the sum matrix ; Then, by recursively restoring the original scale using double-angle recursion, the matrix cosine function at the original scale is obtained. Sinc function of the sum matrix For details, please refer to the calculation process in Example 1, which will not be repeated here.

[0089] The following describes the implementation process of the optimized 12th-order Taylor polynomial expansion. This process starts with the optimized 12th-order Taylor polynomial expansion and is used to approximate the matrix exponential function under scaling conditions. .

[0090] According to Euler's formula, the matrix exponential function It can be decomposed into a combination of the matrix cosine function and the matrix sinc function. Specifically, in the case of... Quadratic algebraic extension over the generated matrix polynomial algebra In this context, the decomposition is represented as: in, The imaginary unit, for The square root of the matrix. Using this relationship, the approximation of the matrix exponent is equivalent to the joint approximation of the two matrix functions, providing a theoretical basis for the subsequent separation of even and odd powers. Furthermore, the multiplication and addition operations of all quadratic algebraic extensions can be directly represented by their basis expansions, without requiring explicit calculation of the matrix square root.

[0091] Specifically, the first step is to index the four intermediate variables. Construct even function blocks and normalized odd function blocks : in , , , The optimized 12th-order Taylor expansion coefficients are shown in Table 1. It is the identity matrix. Even function block. Only contains Zero and first power terms, normalized odd function block They have the same structural form. Because... and The expressions do not contain The construction process here does not require calculating the square root of the matrix. .

[0092] Based on this, the aforementioned basic blocks are combined into intermediate combinatorial quantities to aggregate higher-order terms in the Taylor expansion: in, Aggregated even function blocks , and normalized odd function blocks Contribution Aggregated normalized odd function blocks , and The purpose of this aggregation operation is to pre-combine the powers that would otherwise need to be calculated one by one into a few intermediate quantities, thereby reducing the number of subsequent matrix multiplication calls.

[0093] Finally, based on the aforementioned basic blocks and intermediate combinations, the matrix cosine function under the scaling factor is jointly calculated. Sinc function of the sum matrix : The two formulas above integrate the basic blocks and intermediate combinations into a unified evaluation expression, allowing the matrix cosine function under scaling to be obtained in a single calculation. Sinc function of the sum matrix Since the entire evaluation process is performed under scaling, the matrix... The norm of is contained within the region of convergence, so a low-order Taylor expansion can guarantee sufficient approximation accuracy.

[0094] After performing the above joint evaluation at the scaling scale, with and As initial value, execute The recursive calculation of the rounds restores the original scale: each round proceeds sequentially according to... and Update, among which This is the identity matrix; for the meaning of other parameters, please refer to Implementation 1. After... After round-recursion, the number of cases is derived from... Recovering by doubling in each round Finally, the matrix cosine function at the original scale is obtained. Sinc function of the sum matrix ,in This represents the normalized thickness parameter; for the meanings of other parameters, please refer to Example 1.

[0095] Table 1: Optimized 12th-order Taylor expansion coefficients For applications requiring higher approximation accuracy, an optimized 18th-order Taylor expansion can be used. The 18th-order scheme employs the same idea as the 12th-order scheme, introducing... The coefficients cover higher-order Taylor terms. The coefficients of the 18th-order Taylor expansion are shown in Table 2. The key steps of the 18th-order scheme are as follows: (1) Determine the number of scaling rounds based on the product norm of the coupling matrix. ,structure ; (2) To structure and ; (3) Calculation: (4) Calculate the matrix cosine function under the scaling scale. Sinc function of the sum matrix : (5) Set initial values and (6) Execution Rotation update and .

[0096] The 18th-order scheme does not change the overall process of the method in this embodiment, only replacing the polynomial evaluation order and coefficient set in step S102; subsequent steps S103 to S106 are consistent with the 12th-order scheme.

[0097] Table 2: Optimized 18th-order Taylor expansion coefficients Step S103: Transform the interstitial medium substrate to construct the block structure of the layer transmission matrix.

[0098] When using the interstitial medium as the reference medium for constructing the matrix, that is, when constructing the transmission matrix, the two sides of the layer are first regarded as being surrounded by a vacuum interstitial medium with zero thickness, so that the scattering matrix of the layer can be constructed independently under a unified reference and is not affected by the material parameters of adjacent layers.

[0099] The magnetic field characteristic matrix of the interstitial medium is: Using Relations Define auxiliary matrix and : Subsequently, the transmission matrix is ​​divided into blocks after being embedded in the gap medium. and From auxiliary matrix and A linear combination yields: The meaning of each parameter in the above matrix is ​​explained in Example 1.

[0100] Step S104: Layer scattering matrix transformation, utilizing symmetry to calculate only the reflective and transmissive blocks.

[0101] The transmission matrix and the scattering matrix are two equivalent descriptive forms. This embodiment uses the scattering matrix for interlayer cascading to maintain numerical stability. In the interstitial medium embedding configuration, the layer scattering matrix has a symmetric structure. and Therefore, only two blocks need to be calculated: The inverse matrix operation in the formula is implemented in the GPU by solving the system of linear equations through LU decomposition, without explicitly forming the inverse matrix, to improve numerical stability. The final output of this layer... Scattering matrix , This represents the total number of Fourier harmonics. The meaning of each block in the above scattering matrix is ​​explained in Example 1, and will not be repeated here.

[0102] Step S105: Traverse all layers, construct the scattering matrix of each layer in parallel, and cascade them by accumulating them level by level.

[0103] Before starting the iteration, first increment the variable. Initialized to the input interface scattering matrix And set the layer number. Input interface scattering matrix It describes the reflection and transmission characteristics of incident light as it enters the gap medium from the input medium, serving as the starting point for cascading.

[0104] Then the main loop is entered, and the current layer number is checked. Is it less than or equal to the total number of floors? If so, then execute the scattering matrix construction process for that layer (i.e., the aforementioned steps S102 to S104) to obtain the scattering matrix of that layer. After obtaining the result, the current accumulated result is calculated using Redheffer star multiplication. With the scattering matrix of this layer Cascade, to obtain This cascading operation is physically equivalent to merging the propagation behavior of incident light in previously traversed layers with that in the current layer into an equivalent scattering process. After cascading, the layer numbers are... Increasing Return to the loop and check again until all layers have been processed. Then the loop exits. It includes scattering information for the input interface and all layers.

[0105] Since the construction of the single-layer scattering matrix of different layers depends only on the parameters of that layer itself (coupling matrix pair) , and layer thickness Since there are no data dependencies between layers, the scattering matrices of each layer can be constructed in parallel on different GPUs in the GPU computing cluster. This parallel characteristic does not conflict with the sequential requirement of cascading layer by layer in the above loop, because the scattering matrices of each layer can be constructed independently in parallel first, and then the master node can cascade them sequentially. The cascaded operations are executed sequentially.

[0106] Step S106: Optimize the Redheffer star product cascade to obtain the global scattering matrix and output it.

[0107] After the layer traversal is complete ( ), the last accumulated Scattering matrix at the output interface Cascaded, the global scattering matrix is ​​obtained. Input interface scattering matrix and output interface scattering matrix The scattering characteristics of the input medium-gap medium interface and the gap medium-output medium interface are described separately, and their explicit expressions are as follows: ; in, and These are the magnetic field characteristic matrices of the input and output media, respectively; This represents the characteristic matrix of the magnetic field of the interstitial medium. It is the relative permittivity of the input medium; It is the relative permeability of the input medium; It is the relative permittivity of the output medium. It is the relative permeability of the output medium; For input medium Directional wave vector matrix; For output medium Directional wave vector matrix; for The inverse matrix; for The inverse matrix; Let be the identity matrix. The standard definition of the Redheffer star product of two scattering matrices is: in, The transmission block represents the cascaded scattering matrix and describes the transmission characteristics when incident from above the cascaded structure. Represents the reflective block of the cascaded scattering matrix, describing the reflection characteristics when incident from below the cascaded structure; Represents the reflective block of the cascaded scattering matrix, describing the reflection characteristics when incident from above the cascaded structure; The transmission block represents the cascaded scattering matrix and describes the transmission characteristics when incident from below the cascaded structure. Represents the scattering matrix The transmission block, described from the matrix Transmission characteristics when incident from below the corresponding layer; Represents the scattering matrix The reflection block, describing the matrix The reflection characteristics when incident from below the corresponding layer; Represents the scattering matrix The reflection block, describing the matrix The reflection characteristics when incident from above the corresponding layer; Represents the scattering matrix The transmission block, described from the matrix Transmission characteristics when incident from below the corresponding layer; Represents the scattering matrix The transmission block, described from the matrix Transmission characteristics when incident from above the corresponding layer; Represents the scattering matrix The reflection block, describing the matrix The reflection characteristics when incident from below the corresponding layer; Represents the scattering matrix The reflection block, describing the matrix The reflection characteristics when incident from above the corresponding layer; Represents the scattering matrix The transmission block, described from the matrix Transmission characteristics when incident from below the corresponding layer.

[0108] In the above situation, it is necessary to solve two sets of linear equations with different coefficients. This invention utilizes matrix identities. Each step only requires adjusting the coefficient matrix. Perform a single LU decomposition and solve for the two right-hand sides in batch. and Thus we get: The above optimization reduces the number of linear solver calls compared to inverting the standard formula block by block, and the two right-hand terms can be combined into a single batch solution, improving GPU execution efficiency.

[0109] The final global scattering matrix is ​​obtained by Redheffer multiplication of the input interface scattering matrix, the scattering matrices of each layer, and the output interface scattering matrix. The reflectivity, transmittance, and diffraction efficiency of each diffraction order are then calculated from the global scattering matrix, outputting the diffraction efficiency spectrum and the near-field electromagnetic field distribution.

[0110] The RCWA simulation calculation method based on trigonometric polynomial evaluation of the present invention, although its computational complexity is on the same order of magnitude as the traditional RCWA method, retains a higher number of Fourier harmonics. When the number increases, the calculation amount is still based on the total number of Fourier harmonics. The cubic growth, but significantly reduced the proportionality constant in the growth curve, i.e., for the same total number of Fourier harmonics. The required number of matrix multiplications is less, and the peak memory usage is also lower. This means that for simulation problems of the same scale, this invention can complete the solution with fewer computational steps and lower storage requirements; in complex scenarios such as multi-layer structures or higher-order Fourier truncation, the computational efficiency advantage of this invention will be even more significant.

[0111] Table 3 provides a comparison of the main operands for constructing a single-layer scattering matrix.

[0112] Table 3: Comparison of major operands ( (To scale the number of wheels) As can be seen from Table 3, if the present invention and the prior art use the same number of scaling wheels... Therefore, the number of matrix multiplications is reduced in both the 12th-order and 18th-order schemes of the present invention compared to the complete block schemes of the prior art. Next. In typical cases For example, in the 12th-order scheme, the number of matrix multiplications is reduced from 28 to 13, a reduction of over 50%; in the 18th-order scheme, it is reduced from 30 to 15, a reduction of 50%. The linear solution of this invention requires only 2 steps, which is slightly more than the 1 step required by the existing complete block-based scheme. However, the computational cost of matrix multiplication on a GPU is far greater than that of linear solution, and the additional linear solution added in this invention can be parallelized on a GPU using batch processing techniques. The extra overhead is far less than the benefit of significantly reducing the number of matrix multiplications. Therefore, as the layer thickness or the norm of the propagation operator increases, the computational advantage of this invention will gradually expand.

[0113] Furthermore, this invention only requires maintaining the matrix cosine function during the construction of the single-layer scattering matrix. Sinc function of the sum matrix The present invention uses two intermediate matrices, whereas existing complete block-based schemes require maintaining multiple intermediate matrices corresponding to each level of the matrix exponent expansion. Therefore, the peak active matrix number of the simulation calculation method of this invention is approximately half that of the complete block-based scheme. Fewer intermediate matrices not only reduce memory usage but also decrease memory read / write operations, memory allocation pressure, and kernel function activation frequency, which is of great significance in large-scale GPU computing.

[0114] Based on the above explanation using a multi-layered periodic structure of a single target to be simulated as an example, this invention also supports a variety of parallel computing strategies to fully utilize the computing resources of GPU clusters.

[0115] Data Parallel Mode: When batch simulations of the scattering characteristics of the same periodic structure under different incident wavelengths or angles are required, the master node distributes the RCWA simulation calculation tasks for different wavelengths or angles to different GPUs, and each GPU independently executes the complete simulation calculation. This mode is suitable for parameter scanning scenarios and can achieve linear acceleration.

[0116] Model Parallel Mode: This includes a vertically layered mode and a horizontally partitioned mode. In the vertically layered mode, since the construction of the scattering matrices for each layer is independent, the master node allocates the scattering matrix construction tasks corresponding to different physical layers of the periodic structure to different GPUs for parallel execution. After each GPU completes the scattering matrix construction for its assigned layer, it returns the results to the master node for cascading. This mode is the default parallel method used in this embodiment. In the horizontally partitioned mode, when the single-layer computational domain is large, the master node divides the computational domain of the same physical layer into multiple sub-regions horizontally. Each sub-region is allocated to a different GPU to independently execute the scattering matrix calculation task corresponding to that sub-region. After the GPUs execute in parallel, the calculation results of each sub-region are merged.

[0117] Hybrid Parallelism: This invention integrates data parallelism and model parallelism to construct a multi-layered computing architecture. When batch simulation tasks and multi-layered structures coexist, the master node first allocates tasks of different wavelengths or incident angles to different GPU groups according to data parallelism. Within each GPU group, the multi-layered structure is then allocated to different GPUs within the group for execution according to model parallelism, forming a two-level parallel structure of inter-group data parallelism and intra-group model parallelism. This mode fully leverages the inter-layer independence of periodic structures and the parallelism of parameter scanning tasks, maximizing the utilization of the device's computing resources.

[0118] The two embodiments of the present invention provide an accelerated simulation calculation device for the scattering electromagnetic field of periodic structures and an RCWA simulation calculation method based on trigonometric function polynomial evaluation, which are applicable to electromagnetic scattering simulation scenarios of semiconductor wafer or mask optical inspection, optical critical dimension measurement, photolithography simulation optimization, diffractive optical element design, metasurface analysis, and other layered or periodic structures.

[0119] The above description is merely a preferred embodiment of the present invention. Those skilled in the art should understand that the scope of disclosure in this invention is not limited to the specific combination of the above-described technical features, but should also cover other technical solutions formed by any combination of the above-described technical features or their equivalents without departing from the above-described concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this invention.

Claims

1. A periodic structure-based accelerated simulation computing device for scattering electromagnetic fields, characterized in that, include: The master control node and multiple computing nodes are interconnected via a network interface; The master control node is used to acquire the electromagnetic parameters of the input medium of the incident light, the electromagnetic parameters of the output medium of the transmitted light, the electromagnetic parameters of the gap medium, and the parameters of each layer of the periodic structure. Based on the structural period and Fourier harmonics in the parameters of each layer, it constructs the transverse wave vector matrix of each layer. The task of constructing the electromagnetic parameters of the input medium of the incident light, the electromagnetic parameters of the output medium of the transmitted light, the electromagnetic parameters of the gap medium, the parameters of each layer of the periodic structure, the transverse wave vector matrix of each layer, and the scattering matrix of each layer is distributed to each computing node for parallel execution. Summarize the scattering matrices of each layer, the input interface scattering matrix, and the output interface scattering matrix returned by each computing node; Furthermore, the input interface scattering matrix, the scattering matrices of each layer, and the output interface scattering matrix are concatenated sequentially to obtain the global scattering matrix; Each computing node, in response to the layer scattering matrix construction task sent by the master node, constructs the Fourier convolution matrix of each layer based on the dielectric constant and permeability distributions in the allocated layer parameters. It defines the normalized thickness parameter of each layer based on the layer thickness and incident light wavelength in the allocated layer parameters. Based on the Fourier convolution matrix and the transverse wave vector matrix of each layer, it constructs RCWA coupling matrix pairs for each layer and calculates the product of the RCWA coupling matrix pairs. The scaling parameter is determined by the norm of the matrix to be scaled, which is determined by the scaling parameter. The matrix to be scaled is scaled and the matrix cosine function and matrix sinc function are obtained by joint polynomial evaluation at the scaling scale. Then, the original scale is restored by double-angle recursion with the same scaling parameter to obtain the matrix cosine function and matrix sinc function at the original scale. The two sides of each assigned layer are regarded as being surrounded by the gap medium. The transmission matrix blocks of each assigned layer are constructed according to the matrix cosine function and matrix sinc function. The transmission matrix blocks of each layer are converted into the scattering matrix of each layer. The input interface scattering matrix is ​​constructed based on the electromagnetic parameters of the input medium and the gap medium of the incident light. The output interface scattering matrix is ​​constructed based on the electromagnetic parameters of the output medium and the gap medium of the transmitted light. In addition, the scattering matrices of each layer, the input interface scattering matrix, and the output interface scattering matrix are transmitted back to the master control node.

2. The periodic structure electromagnetic field scattering acceleration simulation calculation device according to claim 1, characterized in that: Each computing node includes a CPU and multiple GPUs connected via a bus. The multiple GPUs achieve collaborative computing and data exchange through a high-speed interconnect protocol. The CPU of the computing node is used to respond to the scattering matrix construction task of each layer sent by the master node. Based on the dielectric constant and permeability of each layer's parameters, it constructs the Fourier convolution matrix of each layer. Based on the layer thickness and incident light wavelength of each layer's parameters, it defines the normalized thickness parameter of each layer. Based on the Fourier convolution matrix and transverse wave vector matrix of each layer, it constructs the RCWA coupling matrix pair of each layer and calculates the product of the RCWA coupling matrix pairs of each layer. Then, it sends the product of the RCWA coupling matrix pairs of each layer, the normalized thickness parameter of each layer, the electromagnetic parameters of the input medium of the incident light, the electromagnetic parameters of the output medium of the transmitted light, and the electromagnetic parameters of the gap medium to multiple GPUs of this node. Each GPU of the computing node is used to determine the scaling parameters based on the norm of the matrix to be scaled, determined by the product of the RCWA coupling matrix pairs of the assigned layer and the normalized thickness parameter. The matrix to be scaled is then scaled, and the matrix cosine function and matrix sinc function are obtained by joint polynomial evaluation at the scaling scale. The original scale is then restored using the same scaling parameters through a doubling-angle recursion to obtain the matrix cosine function and matrix sinc function at the original scale. The assigned layer is considered to be surrounded by a gap medium on both sides. A transmission matrix block of the assigned layer is constructed based on the matrix cosine function and matrix sinc function, and the transmission matrix block is converted into the scattering matrix of that layer. An input interface scattering matrix is ​​constructed based on the electromagnetic parameters of the input medium of the incident light and the electromagnetic parameters of the gap medium. An output interface scattering matrix is ​​constructed based on the electromagnetic parameters of the output medium of the transmitted light and the electromagnetic parameters of the gap medium. Finally, the scattering matrices of each layer, the input interface scattering matrix, and the output interface scattering matrix are transmitted back to the master node via the CPU of this node. In this process, at least one of the multiple GPUs within the computing node constructs the input interface scattering matrix and the output interface scattering matrix; the scattering matrix of each layer is constructed in parallel by multiple GPUs layer by layer.

3. The periodic structure scattering electromagnetic field acceleration simulation calculation device according to claim 2, characterized in that: The CPU of the computing node is also used to determine the lateral spatial resolution of each layer based on the structural period and Fourier harmonic number in the parameters of each layer allocated to it, and to determine the thickness threshold of each layer based on the lateral spatial resolution. If the thickness of a certain layer exceeds the thickness threshold, the layer is uniformly divided into multiple sub-layers, and the thickness of each sub-layer is used to replace the original layer thickness, and the normalized thickness parameter of the layer is redefined. The GPU of the computing node is also used to merge the scattering matrices of all sub-layers into the scattering matrix of the layer after obtaining the scattering matrices of each sub-layer.

4. The periodic structure electromagnetic field scattering acceleration simulation calculation device according to claim 2, characterized in that: The joint polynomial evaluation uses 12th or 18th order Taylor polynomial evaluation as an approximation of the matrix cosine function and the matrix sinc function; The evaluation process of the 12th or 18th order Taylor polynomial pre-calculates and reuses the lower-order matrix powers, rearranging the higher-order matrix polynomials corresponding to the Taylor polynomials into linear combinations, products, and nested combinations of multiple lower-order matrix polynomials to reduce the number of matrix multiplications.

5. The periodic structure electromagnetic field scattering acceleration simulation calculation device according to claim 2, characterized in that: The GPU of the computing node is also used to construct the magnetic field feature matrix of the gap medium when constructing the transfer matrix block of each layer, define an auxiliary matrix by using the relationship that the inverse matrix of the magnetic field feature matrix of the gap medium is equal to its own negative, and construct the transfer matrix block by linear combination of the auxiliary matrices.

6. The periodic structure electromagnetic field scattering acceleration simulation calculation device according to claim 2, characterized in that: The GPU of the computing node is also used to calculate only one of the two reflection blocks and one of the two transmission blocks when converting the transmission matrix into scattering matrices for each layer, so as to obtain the complete layer scattering matrix; Wherein, according to the symmetry relationship under the interstitial medium, the reflective block is the lower left or upper right block of the layer scattering matrix, representing the amplitude and phase of the electromagnetic wave being reflected on the same side when it is incident from above or below the layer; the transmission block is the upper left or lower right block of the layer scattering matrix, representing the amplitude and phase of the electromagnetic wave being transmitted on the other side when it is incident from above or below the layer.

7. The periodic structure electromagnetic field scattering acceleration simulation computing device according to claim 2, characterized in that, The GPU of the computing node is also used for: The following matrix multiplication processes are performed using the batch general matrix multiplication function: calculating the matrix cosine function and the matrix sinc function, restoring to the original scale by double-angle recursion with the same scaling parameter, constructing the scattering matrix of each layer, constructing the input interface scattering matrix, and constructing the output interface scattering matrix. The linear equations in the process of dividing the transmission matrix into blocks and converting them into the scattering matrix of that layer are solved by the batch LU decomposition algorithm, as well as the linear equations in the process of cascading to obtain the global scattering matrix are solved.

8. The periodic structure electromagnetic field scattering acceleration simulation calculation device according to claim 2, characterized in that: The simulation computing equipment adopts a parallel computing strategy, including data parallel mode, model parallel mode, or hybrid parallel mode; When using the data parallel mode for parallel computing, RCWA simulation computing tasks with different incident wavelengths or different incident angles are assigned to different GPUs, and each GPU independently executes the complete simulation computing. When using model-parallel mode for parallel computing, there are two modes: vertical layering mode and horizontal partitioning mode. In vertical layering mode, the task of constructing scattering matrices corresponding to different physical layers of a periodic structure is assigned to different GPUs for parallel execution. In horizontal partitioning mode, the computational domain of the same physical layer is divided into multiple sub-regions, and each sub-region is assigned to a different GPU to independently execute the scattering matrix calculation task corresponding to that sub-region. When using a hybrid parallel mode for parallel computing, the computing architecture is constructed by combining the data parallel mode and the model parallel mode.

9. A periodic structure-based electromagnetic field scattering acceleration simulation computing device according to claim 1 or 2, characterized in that, The periodic structure is a periodic structure on a semiconductor wafer, wafer mask, diffractive optical element, metasurface, photonic crystal, or grating waveguide.

10. A periodic structure-based electromagnetic field scattering acceleration simulation computing device according to claim 1 or 2, characterized in that, The master node is also used to calculate and output at least one of the following information based on the global scattering matrix: The reflectivity, transmittance, or diffraction efficiency of diffracted waves of various orders in different directions under incident light illumination from a periodic structure. And the spatial distribution of electromagnetic fields inside and near the periodic structure.

Citation Information

Patent Citations

  • RCWA simulation calculation method of target structure and scattering electromagnetic field acceleration calculation equipment

    CN121328142A

  • Electromagnetic field characteristic simulation method and electromagnetic field characteristic simulation system

    US20260202755A1