Accelerator design optimization method based on asynchronous deep reservoir and digital twinning

Abstract

This invention proposes an accelerator design optimization method based on asynchronous deep reservoir and digital twin. The steps are as follows: obtaining a training sample set; constructing and training an asynchronous deep reservoir model; and obtaining the optimization results of the accelerator. This invention optimizes the accelerator using an asynchronous deep reservoir model. The operating state parameters of the digital twin accelerator are input into the asynchronous deep reservoir model for forward propagation, outputting the optimized accelerator design parameters. This efficiently and dynamically maps the accelerator design parameters to performance, improving reliability. By using a training sample set containing various operating conditions to train the asynchronous deep reservoir model, optimization of the accelerator under multiple different operating conditions can be completed, avoiding manual intervention during model adjustment and improving efficiency.

Description

Technical Field

[0001] This invention belongs to the field of accelerator design technology and relates to an accelerator design optimization method. Specifically, it relates to an accelerator design optimization method based on asynchronous deep reservoir and digital twin, which can be used in the fields of accelerator design and simulation. Background Technology

[0002] A particle accelerator is an electromagnetic device that uses various forms of electric fields to accelerate different types of charged particles to higher energies. Particle accelerators play an important role in basic scientific research, medical treatment, and materials processing.

[0003] Particle accelerators, depending on their energy, range in length from tens of meters to tens of kilometers and contain a huge number of components. Building an accelerator is a complex and time-consuming project. Verification of its design is usually based on various software simulations, but the verification results are often difficult to directly feed back into the accelerator design for correction.

[0004] Digital twin technology is a technique that uses computer simulation to accurately replicate a physical system in digital space. It enables online, 1:1 3D modeling of all equipment on an accelerator facility. This allows for simulating the actual installation process before construction, identifying potential installation and integration problems in advance, and iteratively optimizing the design to ensure compatibility between the various components of the accelerator.

[0005] For example, the Institute of Modern Physics, Chinese Academy of Sciences, disclosed an optimization method for particle accelerators in its patent application, "A Method and System for Optimizing Particle Accelerator Engineering" (Patent Application No.: CN202010645144.8, Publication No.: CN111651903A). The method includes the following steps: S1. In a collaborative modeling platform, a particle accelerator engineering product structure tree is built, which includes several layers of skeleton models; S2. Elements of each layer of skeleton models in the product structure tree are created; S3. Geometric elements of each layer of skeleton models are defined; S4. Sub-models of each layer are designed based on the elements and geometric elements of the skeleton models; S5. The collaborative modeling platform saves the designed sub-models of each layer and establishes a general three-dimensional model of the particle accelerator engineering based on them; S6. The particle accelerator engineering is optimized based on the three-dimensional model. This method establishes strict positional and logical relationships between models, and the modeling and assembly of components at each level are only related to skeleton elements, facilitating model updates and adjustments and improving modeling efficiency.

[0006] However, this method optimizes the particle accelerator project based on a three-dimensional model of the particle accelerator project. It mainly focuses on the position and logical relationship of components at each level in the three-dimensional model, and lacks the ability to directly model complex nonlinear dynamic characteristics. It does not consider the relationship between accelerator design parameters and performance during optimization, resulting in low reliability of accelerator optimization. Furthermore, this method adopts a top-down modeling and structure tree update approach, which requires a lot of manual intervention when adjusting the model. The model update is still cumbersome and the optimization efficiency is low. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and propose an accelerator design optimization method based on asynchronous deep reservoir and digital twin, which is used to solve the technical problems of low reliability and efficiency in the prior art.

[0008] To achieve the above objectives, the technical solution adopted by the present invention includes the following steps:

[0009] (1) Obtain the training sample set:

[0010] Obtain the digital twin accelerator operating state parameters U and accelerator design parameters Y at K time steps under multiple different operating conditions, and combine the operating state vector and design parameter vector u at the k-th time step. k and y k The training samples are composed to obtain a training sample set containing K training samples, where K≥30;

[0011] (2) Constructing an asynchronous deep reserve pool model:

[0012] Construct an asynchronous deep reservoir model O consisting of an input layer, an output layer, and L stacked reservoirs, with delay links loaded between adjacent reservoirs, where L ≥ 2, and the number of neurons in the l-th reservoir is N. l ;

[0013] (3) Training the asynchronous deep reserve pool model:

[0014] The asynchronous deep reservoir model is trained using a training sample set to obtain the trained asynchronous deep reservoir model O. * ;

[0015] (4) Obtain the optimization results of the accelerator:

[0016] Through the trained asynchronous deep reserve pool model O * The accelerator is optimized to obtain a better accelerator design.

[0017] Compared with the prior art, the present invention has the following advantages:

[0018] (1) This invention optimizes the accelerator by using an asynchronous deep reservoir model. The operating status parameters of the digital twin accelerator are input into the asynchronous deep reservoir model for forward propagation and the optimized accelerator design parameters are output. This efficiently maps the accelerator design parameters and performance, thereby improving reliability.

[0019] (2) This invention uses a training sample set containing various working conditions to train the asynchronous deep reservoir model, which can optimize the accelerator under various working conditions, avoids manual intervention during model adjustment and improves efficiency. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the implementation of the present invention. Detailed Implementation

[0021] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] Reference Figure 1 The present invention includes the following steps:

[0023] Step 1) Obtain the training sample set:

[0024] Obtain the digital twin accelerator operating state parameters U and accelerator design parameters Y at K time steps under multiple different operating conditions, and combine the operating state vector and design parameter vector u at the k-th time step. k and y k The training samples are composed to obtain a training sample set containing K training samples, where K≥30;

[0025] Running state parameter vector u k This includes beam intensity, beam energy, beam pulse width, beam mass, beam focusing degree, electric field strength, magnetic field strength, distribution of electric and magnetic fields, total input power required by the accelerator, power of the accelerator output beam, temperature of equipment such as the accelerating cavity, beam tube, and magnets, coolant temperature, beam stability, vacuum pressure, and ambient temperature and humidity.

[0026] Design parameter vector y k These correspond to the physical and geometric parameters in the accelerator design scheme, including the size of the accelerating cavity, electric field strength, magnetic field strength, electromagnetic wave frequency, magnet type, magnetic field gradient, magnet size, total length of the accelerator, cavity spacing, beam pipe diameter, type and number of vacuum pumps, coolant flow rate, power supply capacity, and control system accuracy.

[0027] Step 2) Construct an asynchronous deep reserve pool model:

[0028] Construct an asynchronous deep reservoir model O consisting of an input layer, L stacked reservoirs, with delay links between adjacent reservoirs, and an output layer, where L ≥ 2, and the number of neurons in the l-th reservoir is N. l In this implementation example, L is set to 4, the number of neurons in the first reserve pool is 1024, the number of neurons in the second reserve pool is 512, the number of neurons in the third reserve pool is 256, and the number of neurons in the fourth reserve pool is 128.

[0029] Asynchronous deep reservoir computing is a computational model that combines reservoir computing and deep neural networks. It focuses on the processing of time series data and dynamic system modeling. By introducing an asynchronous update mechanism, it optimizes computational efficiency and enhances the ability to capture complex time dependencies. It can efficiently model the dynamic mapping mechanism between accelerator design parameters and performance, and improve the reliability of the optimization process.

[0030] Step 3) Train the asynchronous deep reservoir model:

[0031] The asynchronous deep reservoir model is trained using a training sample set to obtain the trained asynchronous deep reservoir model O. * ;

[0032] The asynchronous deep reserve pool model is trained by a training sample set containing K time steps under multiple different operating conditions. The trained model can optimize the accelerator under various different operating conditions, avoiding manual intervention during model adjustment and improving efficiency. At the same time, the asynchronous deep reserve pool model establishes a high-dimensional mapping between the accelerator time series operating parameters and the design parameters, which improves reliability.

[0033] The steps to train the asynchronous deep reservoir model are as follows:

[0034] (3a) The initial weight matrix of the asynchronous deep reserve pool model is the weight matrix W. O ;

[0035] (3b) Each storage pool is connected to each operating state parameter vector u k Calculate your own state And through the weight matrix W O State vector of asynchronous deep reservoir model By performing a linear combination, we obtain the optimized accelerator design parameter vector prediction value y. k =W O T s k W O T The output weight matrix W represents... O The transpose operation;

[0036] Each layer of the reserve pool is accessed through each operational state parameter vector u. k Calculate your own state The calculation formula is:

[0037] When l = 1:

[0038]

[0039] When l > 1:

[0040]

[0041] Where f is a nonlinear activation function. and Let be the input weight matrix and the internal connection matrix of the l-th layer reservoir, respectively. The size is A l ×N l , The size is N l ×N l A l This indicates the input to the l-th layer of the reserve pool. Dimensions It is the class concept matrix C and the internal connection matrix of the first level. product The internal connectivity matrix after feature recombination, Λ C It is by The size of the singular values ​​is N 1 ×N 1 diagonal matrix, U C It is of size N 1 ×N 1 of Eigenvector matrix, M is the number of singular values, U C,M×M and Λ C,M×M They are U C and Λ C The matrix consisting of M×M elements in the upper left corner. It is randomly generated with a size of M×N 1 The input weight matrix;

[0042] Introducing a class concept matrix C to reduce redundant information in the reserve pool using an elastic weight consolidation method, where C is a matrix of size N. 1 ×N 1 A matrix, whose function is to manage long-term neural memory, can permanently store different time patterns in a storage pool by learning C, establishing a high-dimensional mapping between accelerator time-series operating parameters and design parameters, thereby improving reliability. The concept of a matrix C is achieved by... The minimization index E where the singular values ​​largely approach zero CWhat was obtained:

[0043]

[0044] C = (R) X +β 2 I C ) -1 R X (6)

[0045] in, This is the input weight matrix before feature recombination in the first-layer reserve pool, ||·|| F It is the Frobenius norm, R X yes The covariance matrix, β is the regularization coefficient, f -1 It is the inverse function of f, I C It is dimension and R X The same identity matrix;

[0046] Using singular value decomposition pairs By performing feature recombination and selecting a small number of singular values ​​and eigenvectors, the most important features can be extracted. In accelerator design optimization, this helps to reduce the amount of computation while efficiently understanding the relationship between design parameters and accelerator performance, thereby improving the efficiency and reliability of optimization.

[0047] (3c) Predict the set of values ​​Y using the optimized digital twin accelerator design parameters Y at K time steps and the optimized accelerator design parameter vector. K Calculate the loss value E = ||Y|| for the asynchronous deep pool model. K -Y||+α 2 ||W O ||, and use the ridge back algorithm to apply E to W O Update the model to obtain the trained asynchronous deep reservoir model;

[0048] The loss value E of the asynchronous deep pool model is calculated using the following formula:

[0049] E = ||Y K -Y||+α 2 ||W O || (7)

[0050] Through E to W O The update is performed using the following formula:

[0051] W O * =(S K S K T +α 2 I) -1 SK (Y) T (8)

[0052] Where ||·|| represents the modulo operation, α represents the regularization coefficient, and S K =[s k ] 1≤k≤K This represents a matrix consisting of state vectors of an asynchronous deep reserve pool model under K time steps of different operating conditions;

[0053] Step 4) Obtain the accelerator optimization results:

[0054] Through the trained asynchronous deep reserve pool model O * The optimized accelerator design is obtained by optimizing the accelerator. The steps are as follows:

[0055] (4a) Initialize the number of iterations to t, the maximum number of iterations to T, T≥200, and let t=1;

[0056] (4b) Obtain the operating state parameters U of the digital twin model of the accelerator to be optimized, which includes multiple different operating conditions and K time steps. * and take it as O * The input is propagated forward to obtain the optimized design parameter set Y. t ;

[0057] (4c) Based on the optimized accelerator design parameter set Y t Create a design parameter value of y in the collaborative design verification platform. K ∈Y t Digital twin model;

[0058] (4d) Determine whether t = T holds true. If so, obtain the optimized design parameter value y. K ∈Y T If the accelerator design result is not obtained, otherwise, the digital twin model constructed in step (4c) is used as the digital twin model of the accelerator to be optimized, and t = t + 1 is set, and step (4b) is executed.

[0059] The design parameter value is y. K ∈Y t The digital twin model, containing design parameter value y K ∈Y t The online 1:1 3D modeling of key hardware devices on the accelerator device can accurately simulate the operating status of the accelerator under different operating conditions. Key hardware devices include, but are not limited to, magnets, vacuum pipes, acceleration chambers, radio frequency sources, and power supplies.

Claims

1. An accelerator design optimization method based on asynchronous deep reservoir and digital twin, characterized in that, Includes the following steps: (1) Obtain the training sample set: Obtain the digital twin accelerator operating state parameters U and accelerator design parameters Y at K time steps under multiple different operating conditions, and combine the operating state vector and design parameter vector u at the k-th time step. k and y k The training samples are composed to obtain a training sample set containing K training samples, where K≥30; (2) Constructing an asynchronous deep reserve pool model: Construct an asynchronous deep reservoir model O consisting of an input layer, L stacked reservoirs, with delay links between adjacent reservoirs, and an output layer, where L ≥ 2, and the number of neurons in the l-th reservoir is N. l ; (3) Training the asynchronous deep reserve pool model: (3a) Initialize the output weight matrix of the asynchronous deep reserve pool model as W O ; (3b) Each storage pool is connected to each operating state parameter vector u k Calculate your own state The output layer outputs the weight matrix W. O State vector of asynchronous deep reservoir model By performing a linear combination, the optimized accelerator design parameter vector prediction values ​​are obtained. Among them W O T The output weight matrix W represents O The transpose operation; (3c) The set of predicted values ​​of the accelerator design parameters Y and the optimized accelerator design parameter vector over K time steps. Calculate the loss value of the asynchronous deep reservoir model. And the ridge back algorithm is used to analyze W through E. O The trained asynchronous deep reservoir model O is then updated. * ; (4) Obtain the design optimization results of the accelerator: (4a) Initialize the number of iterations to t, the maximum number of iterations to T, T≥200, and let t=1; (4b) Obtain the operating state parameters U of the digital twin model of the accelerator to be optimized, which includes multiple different operating conditions and K time steps. * and take it as O * The input is propagated forward to obtain the optimized design parameter set. (4c) Based on the optimized accelerator design parameter set Create design parameter values ​​in the collaborative design verification platform Digital twin model; (4d) Determine whether t = T holds true. If so, obtain the optimized design parameter values. If the accelerator design result is not obtained, otherwise, the digital twin model constructed in step (4c) is used as the digital twin model of the accelerator to be optimized, and t = t + 1 is set, and step (4b) is executed.

2. The method according to claim 1, characterized in that, The running state vector and design parameter vector u at the k-th time step mentioned in step (1) k and y k ,in: Running state parameter vector u k This includes beam intensity, beam energy, beam pulse width, beam mass, beam focusing degree, electric field strength, magnetic field strength, distribution of electric and magnetic fields, total input power required by the accelerator, power of the accelerator output beam, temperature of equipment such as the accelerating cavity, beam tube, and magnet, coolant temperature, beam stability, vacuum pressure, and ambient temperature and humidity. Design parameter vector y k These correspond to the physical and geometric parameters in the accelerator design scheme, including the accelerator cavity size, electric field strength, magnetic field strength, electromagnetic wave frequency, magnet type, magnetic field gradient, magnet size, total accelerator length, accelerator cavity spacing, beam pipe diameter, vacuum pump type and quantity, coolant flow rate, power supply capacity, and control system accuracy.

3. The method according to claim 2, characterized in that, Each layer of the storage pool described in step (3b) is accessed through each operating state parameter vector u. k Calculate your own state The calculation formula is: When l = 1: When l>1: Where f is a nonlinear activation function. and Let be the input weight matrix and the internal connection matrix of the l-th layer reservoir, respectively. The size is A l ×N l , The size is N l ×N l A l This indicates the input to the l-th layer of the reserve pool. Dimensions It is the class concept matrix C and the internal connection matrix of the first level. product The internal connectivity matrix after feature recombination, Λ C It is by The size of the singular values ​​is N 1 ×N 1 diagonal matrix, U C It is of size N 1 ×N 1 of Eigenvector matrix, M is the number of singular values, U C,M×M and Λ C,M×M They are U C and Λ C The matrix consisting of M×M elements in the upper left corner. It is randomly generated with a size of M×N 1 The input weight matrix.

4. The method according to claim 3, characterized in that, The aforementioned class concept matrix C is achieved by making The minimization index E where the singular values ​​largely approach zero C What was obtained: C=(R X +β 2 I C ) -1 R X in, This is the input weight matrix before feature recombination in the first-layer reserve pool, ||·|| F It is the Frobenius norm, R X yes The covariance matrix, β is the regularization coefficient, f -1 It is the inverse function of f, I C It is dimension and R X The same identity matrix.

5. The method according to claim 3, characterized in that, The loss value E of the asynchronous deep reservoir model described in step (3c) is calculated using the following formula: Where ||·|| represents the modulo operation, and α represents the regularization coefficient.

6. The method according to claim 3, characterized in that, The step (3c) described above involves using E to process W. O The update is performed using the following formula: Among them, S K =[s k ] 1≤k≤K I represents a matrix consisting of state vectors of an asynchronous deep reservoir model under K time steps of different operating conditions. R ×N R The identity matrix, N R It represents the total number of neurons in the asynchronous deep reservoir model.

Images