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

Abstract

The invention provides an accelerator design optimization method based on an asynchronous deep reservoir and digital twinning. The method comprises the implementation steps of obtaining a training sample set; constructing an asynchronous deep reservoir model and training the asynchronous deep reservoir model; and obtaining an optimization result of the accelerator. According to the method, the accelerator is optimized through the asynchronous deep reservoir model, the operation state parameters of the digital twin accelerator are input into the asynchronous deep reservoir model for forward propagation, the design parameters of the optimized accelerator are output, efficient dynamic mapping is carried out on the design scheme parameters and performance of the accelerator, the reliability is improved, and the method is suitable for industrial production. The asynchronous deep reservoir model is trained by using the training sample set comprising a plurality of different working conditions, so that the accelerator under the plurality of different working conditions can be optimized, manual intervention during model adjustment is avoided, and the efficiency is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of accelerator design, and relates to an accelerator design scheme optimization method, and specifically to an accelerator design optimization method based on an asynchronous deep reserve pool and digital twins, which can be used in technical fields such as accelerator design and simulation. Background Art

[0002] Particle accelerators are electromagnetic devices that use artificial methods to accelerate various types of charged particles to higher energies with the help of various forms of electric fields. Particle accelerators play an important role in basic scientific research, medical treatment and material processing.

[0003] Depending on the energy, the length of a particle accelerator is about tens of meters to tens of kilometers, and the number of parts is huge. The construction of an accelerator is a time-consuming and labor-intensive complex project. The verification of its design is usually based on multiple software simulations, but the verification results are usually difficult to directly feed back into the feedback correction of the accelerator design.

[0004] Digital twin technology is a technology that uses computer simulation technology to accurately replicate physical systems in digital space. Digital twin technology can achieve online one-to-one 3D modeling of all equipment on the accelerator device, simulate the actual installation process before the accelerator device is built, perform online installation simulation, and pre-identify potential installation and integration problems, so as to iteratively optimize the design plan and ensure the compatibility between various devices in the accelerator device.

[0005] For example, the Institute of Modern Physics of the Chinese Academy of Sciences disclosed a particle accelerator optimization method in its patent application "A particle accelerator engineering optimization method and system" (patent application number: CN202010645144.8, application publication number: CN111651903A), which includes the following steps: S1 builds a particle accelerator engineering product structure tree in the collaborative modeling platform, and the product structure tree includes several layers of skeleton models; S2 creates elements of skeleton models at each level in the product structure tree; S3 limits the geometric elements of skeleton models at each level; S4 designs sub-models at each level according to the elements and geometric elements of the skeleton model; S5 The collaborative modeling platform saves the designed sub-models at each level and establishes the overall particle accelerator engineering three-dimensional model based on it, and S6 optimizes the particle accelerator engineering according to the particle accelerator engineering three-dimensional model. This method establishes a strict position and logical relationship between models, and the modeling and assembly of parts at each level are only related to the skeleton elements, which facilitates the update and adjustment of the model and improves the modeling efficiency.

[0006] However, this method optimizes the particle accelerator project based on the three-dimensional model of the particle accelerator project, and mainly focuses on the position and logical relationship of the components at each level in the three-dimensional model. It lacks the ability to directly model complex nonlinear dynamic characteristics. The relationship between accelerator design parameters and performance is not considered during optimization. The accelerator optimization reliability is low. This method adopts a top-down modeling and structure tree update method, which requires more 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 the present invention is to overcome the defects of the prior art and propose an accelerator design optimization method based on an asynchronous deep reserve pool and digital twins to solve the technical problems of low reliability and efficiency existing in the prior art.

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

[0009] (1) Obtain training sample set:

[0010] Obtain the digital twin accelerator operating state parameters U and accelerator design parameters Y including multiple K time steps of different working conditions, and convert the kth time step operating state vector and design parameter vector u k and k Composing training samples, obtaining a training sample set including K training samples, where K ≥ 30;

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

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

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

[0014] The asynchronous deep reservoir model is trained by the 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 reservoir model O * The accelerator is optimized to obtain an optimized accelerator design result.

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

[0018] (1) The present invention optimizes the accelerator through an asynchronous deep reservoir model, inputs the operating status parameters of the digital twin accelerator into the asynchronous deep reservoir model for forward propagation, and outputs the design parameters of the optimized accelerator, thereby performing efficient dynamic mapping between the accelerator design parameters and performance, thereby improving reliability.

[0019] (2) The present invention utilizes a training sample set containing a variety of different working conditions to train the asynchronous deep reserve pool model, which can optimize the accelerator under a variety of different working conditions, avoids manual intervention during model adjustment, and improves efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a flow chart for implementing the present invention. DETAILED DESCRIPTION

[0021] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0022] Reference Figure 1 , the present invention comprises 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 including multiple K time steps of different working conditions, and convert the kth time step operating state vector and design parameter vector u k and k Composing training samples, obtaining a training sample set including K training samples, where K ≥ 30;

[0025] Running state parameter vector u k Including beam intensity, beam energy, beam pulse width, beam quality, beam focusing, electric field strength, magnetic field strength, distribution of electric and magnetic fields, total input power required by the accelerator, output beam power of the accelerator, temperature of the accelerating cavity, beam tube, magnet and other equipment, coolant temperature, beam stability, vacuum pressure, ambient temperature and humidity;

[0026] Design parameter vector y k , corresponding to the physical and geometric parameters in the accelerator design, including acceleration cavity size, electric field strength, magnetic field strength, electromagnetic wave frequency, magnet type, magnetic field gradient, magnet size, total accelerator length, acceleration cavity spacing, beam pipe diameter, vacuum pump type and quantity, coolant flow, power supply capacity, and control system accuracy;

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

[0028] Construct an asynchronous deep pool model O including an input layer, L stacked pools with delay links between adjacent pools, and an output layer, where L ≥ 2, and the number of neurons in the lth pool is N l ; In this embodiment, L is taken as 4, the number of neurons in the first layer reserve pool is 1024, the number of neurons in the second layer reserve pool is 512, the number of neurons in the third layer reserve pool is 256, and the number of neurons in the fourth layer reserve pool is 128;

[0029] Asynchronous deep reservoir computing is a computing model that combines reservoir computing and deep neural networks, focusing on time series data processing and dynamic system modeling. It optimizes computing efficiency by introducing an asynchronous update mechanism, while enhancing 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 by the training sample set to obtain the trained asynchronous deep reservoir model O * ;

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

[0033] To train the asynchronous deep reservoir model, the steps are as follows:

[0034] (3a) Initialize the weight matrix of the asynchronous deep reservoir model as the weight matrix W O ;

[0035] (3b) Each layer of the reserve pool is connected through each operating state parameter vector u k Calculate your own status And through the weight matrix W O State vector for the asynchronous deep reservoir model Perform linear combination to obtain the optimized accelerator design parameter vector prediction value y k =W O T s k , where W O T Represents the output weight matrix W O The transpose operation of

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

[0037] When l = 1:

[0038]

[0039] When l>1:

[0040]

[0041] Where f is a nonlinear activation function, and are the input weight matrix and internal connection matrix of the l-th layer reserve pool, respectively. The size is A l ×N l , The size is N l ×N l , A l Represents the input to the l-th layer storage pool The dimension of is the class concept matrix C and the internal connection matrix of the first layer The product of The internal connection matrix after feature reorganization, Λ C Is The size of the singular values ​​of 1 ×N 1 Diagonal matrix, U C is of size N 1 ×N 1 of The eigenvector matrix, M is the number of singular values, U C,M×M and Λ C,M×M U C and Λ C The matrix consisting of M×M elements in the upper left corner, is randomly generated with size M×N 1 The input weight matrix of

[0042] Introduce the class concept matrix C to reduce the redundant information in the reserve pool by elastic weight consolidation method. C is a matrix of size N. 1 ×N 1 The function of the matrix is ​​to manage neural long-term memory. By learning C, some different time patterns can be permanently stored in the reserve pool, and a high-dimensional mapping between the accelerator time series operating parameters and the design scheme parameters can be established to improve reliability. The class concept matrix C is obtained by using The singular values ​​of the minimization index E tend to zero in large numbers CObtained:

[0043]

[0044] C=(R X +β 2 I C ) -1 R X (6)

[0045] in, is the input weight matrix before the first layer of reservoir features are reorganized, ||·|| F is the Frobenius norm, R X yes The covariance matrix of , β is the regularization coefficient, f -1 is the inverse function of f, I C is the dimension and R X Same identity matrix;

[0046] Using singular value decomposition 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 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) The digital twin accelerator design parameters Y under the optimized K time steps and the optimized accelerator design parameter vector prediction value set Y K , calculate the loss value E = || Y of the asynchronous deep reservoir model K -Y||+α 2 ||W O ||, and use the ridge back algorithm to calculate W through E O Update and obtain the trained asynchronous deep reservoir model;

[0048] The loss value E of the asynchronous deep reserve pool model is calculated as:

[0049] E=||Y K -Y||+α 2 ||W O || (7)

[0050] By E to W O Update, the update formula is:

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

[0052] Among them, ||·|| represents the modulo operation, α represents the regularization coefficient, S K =[s k ] 1≤k≤K Represents the matrix composed of the state vectors of the asynchronous deep reservoir model under different working conditions K time steps;

[0053] Step 4) Get the optimization results of the accelerator:

[0054] Through the trained asynchronous deep reservoir model O * The accelerator is optimized to obtain an optimized accelerator design scheme, and the implementation steps are as follows:

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

[0056] (4b) Obtain the operating state parameters U of the digital twin model of the accelerator to be optimized under multiple different working conditions K time steps * and use it as O * The input is forward propagated 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 y in the collaborative design verification platform K ∈Y t Digital twin models of

[0058] (4d) Determine whether t=T is true. If so, the optimized design parameter value is y K ∈Y T accelerator design result, 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 to execute step (4b);

[0059] The design parameter value is y K ∈Y t The digital twin model of K ∈Y t The online one-to-one 3D modeling of key hardware equipment on the accelerator device can accurately simulate the operating status of the accelerator under different working conditions. The key hardware equipment includes but is not limited to magnets, vacuum pipes, acceleration chambers, RF sources and power supplies.

Claims

1. An accelerator design optimization method based on asynchronous deep reservoir and digital twin, characterized in that: The following steps are involved: (1) Obtain training sample set: Obtain the digital twin accelerator operating state parameters U and accelerator design parameters Y including multiple K time steps of different working conditions, and convert the kth time step operating state vector and design parameter vector u k and k Composing training samples, obtaining a training sample set including K training samples, where K ≥ 30; (2) Constructing an asynchronous deep reserve pool model: Construct an asynchronous deep pool model O including an input layer, L stacked pools with delay links between adjacent pools, and an output layer, where L ≥ 2, and the number of neurons in the lth pool is N l ; (3) Training the asynchronous deep reservoir model: The asynchronous deep reservoir model is trained by the training sample set to obtain the trained asynchronous deep reservoir model O * ; (4) Obtain the design optimization results of the accelerator: Through the trained asynchronous deep reservoir model O * The accelerator is optimized to obtain an optimized accelerator design result.

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

3. The method according to claim 1, characterized in that The training of the asynchronous deep reservoir model described in step (3) is implemented as follows: (3a) Initialize the output weight matrix of the asynchronous deep reservoir model to W O ; (3b) Each layer of the reserve pool is connected through each operating state parameter vector u k Calculate your own status The output layer outputs the weight matrix W O State vector for the asynchronous deep reservoir model Perform linear combination to obtain the optimized accelerator design parameter vector prediction value y k =W O T s k , where W O T Represents the output weight matrix W O The transpose operation of (3c) The accelerator design parameter Y at K time steps and the optimized accelerator design parameter vector prediction value set Y K , calculate the loss value E = || Y of the asynchronous deep reservoir model K -Y||+α 2 ||W O ||, and use the ridge back algorithm to calculate W by E O Update and obtain the trained asynchronous deep reservoir model.

4. The method according to claim 3, characterized in that Each layer of the reserve pool described in step (3b) is represented by each operating state parameter vector u k Calculate your own status The calculation formula is: When l = 1: When l>1: Where f is a nonlinear activation function, and are the input weight matrix and internal connection matrix of the l-th layer reserve pool, respectively. The size is A l ×N l , The size is N l ×N l , A l Represents the input to the l-th layer storage pool The dimension of is the class concept matrix C and the internal connection matrix of the first layer The product of The internal connection matrix after feature reorganization, Λ C Is The size of the singular values ​​of 1 ×N 1 Diagonal matrix, U C is of size N 1 ×N 1 of The eigenvector matrix, M is the number of singular values, U C,M×M and Λ C,M×M U C and Λ C The matrix consisting of M×M elements in the upper left corner, is randomly generated with size M×N 1 The input weight matrix of .

5. The method according to claim 4, characterized in that The class concept matrix C is obtained by using The singular values ​​of the minimization index E tend to zero in large numbers C Obtained: in, is the input weight matrix before the first layer of reservoir features are reorganized, ||·|| F is the Frobenius norm, R X yes The covariance matrix of , β is the regularization coefficient, f -1 is the inverse function of f, I C is the dimension and R X Same identity matrix.

6. The method according to claim 4, characterized in that The loss value E of the asynchronous deep reservoir model described in step (3c) is calculated as: E=||Y K -Y||+α 2 ||W O || Here, ||·|| represents a modulo operation, and α represents a regularization coefficient.

7. The method according to claim 4, characterized in that Step (3c) described in step (3c) O Update, the update formula is: Among them, S K =[s k ] 1≤k≤K Represents the matrix composed of the state vectors of the asynchronous deep reservoir model under different working conditions K time steps, I is the size of N R ×N R The identity matrix, N R is the total number of neurons in the asynchronous deep reservoir model.

8. The method according to claim 3, characterized in that The asynchronous deep reservoir model O trained in step (4) * To optimize the accelerator, the implementation steps are as follows: (4a) Initialize the number of iterations to t, the maximum number of iterations to T, T ≥ 200, and set t = 1; (4b) Obtain the operating state parameters U of the digital twin model of the accelerator to be optimized under multiple different working conditions K time steps * and use it as O * The input is forward propagated to obtain the optimized design parameter set Y t ; (4c) Based on the optimized accelerator design parameter set Y t , create a design parameter value y in the collaborative design verification platform K ∈Y t Digital twin models of (4d) Determine whether t=T is true. If so, the optimized design parameter value is y K ∈Y T 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 to execute step (4b).

Images