CN-122021345-A - Superconducting cavity cooling structure optimization method based on multitask learning agent model

CN122021345ACN 122021345 ACN122021345 ACN 122021345ACN-122021345-A

Abstract

The invention discloses a superconducting cavity cooling structure optimization method based on a multi-task learning agent model, which comprises the steps of adjusting initial structural parameters of a superconducting cavity cooling structure, constructing a superconducting cavity cooling structure model according to each group of structural parameters, performing simulation calculation to obtain transient isothermal indexes STI and steady state margin indexes HMI corresponding to the structural parameters, screening a group of key parameters from the initial structural parameters according to the influence degree of each parameter on the transient isothermal indexes STI and the steady state margin indexes HMI, generating a training sample set training optimization multi-task learning agent model according to the group of key parameters, wherein each training sample comprises a group of key parameter values, the corresponding transient isothermal indexes and steady state margin indexes, taking the group of key parameters as optimization variables, and determining a group of optimal parameter values as the structural parameters of the superconducting cavity cooling structure by utilizing the optimized multi-task learning agent model. The invention can obtain the cooling structure with optimal cooling effect.

Inventors

Chang Zhengze
GE RUI
LIU XIAO
ZHOU JIANRONG
XU MIAOFU
SHA PENG
MA CHANGCHENG

Assignees

中国科学院高能物理研究所

Dates

Publication Date: 20260512
Application Date: 20260407

Claims (10)

1. A superconducting cavity cooling structure optimization method based on a multitask learning agent model comprises the following steps: The method comprises the steps of adjusting initial structural parameters of a superconducting cavity cooling structure, constructing a superconducting cavity cooling structure model according to each group of structural parameters, and obtaining transient isothermal indexes STI and steady state margin indexes HMI corresponding to the group of structural parameters through simulation calculation; Screening a group of key parameters from the initial structure parameters according to the influence degree of each parameter on the transient isothermal index STI and the steady state margin index HMI; Generating a training sample set training optimization multi-task learning agent model according to the group of key parameters, wherein each training sample comprises a group of key parameter values and a transient isothermal index STI and a steady state margin index HMI corresponding to the key parameter values; And determining a group of optimal parameter values as structural parameters of the superconducting cavity cooling structure by using the group of key parameters as optimization variables and utilizing the optimized multi-task learning agent model.
2. The method of claim 1, wherein the transient isothermal indicator STI is used to measure control energy of a temperature difference across the superconducting cavity when crossing the superconducting transition temperature region, and the steady state margin indicator HMI is used to measure a magnitude of a maximum temperature of the superconducting cavity in steady state and a margin of a relative quench upper limit.
3. The method of claim 1, wherein the multi-task learning agent model comprises an input module, a shared trunk, and a dual task output head, and wherein the method for generating a training sample set training optimized multi-task learning agent model based on the set of key parameters comprises: Generating N candidate sample points in a d-dimensional space by adopting a Sobol low-difference sequence, performing feasibility discrimination on each candidate sample point, and taking the candidate sample points meeting feasibility as effective sample points, wherein each candidate sample point corresponds to a group of normalized key parameter values, and d is the number of parameters in the group of key parameters; simulating and calculating a corresponding transient isothermal index STI and a steady state margin index HMI as labels of the effective sample points under the same thermal boundary condition to form a training sample set; The input module receives a group of key parameters in a training sample and sends the key parameters to the shared trunk, the shared trunk extracts intermediate characteristics h (x) commonly related to transient isothermal indexes STI and steady-state margin indexes HMI according to the group of key parameters, a first task head outputs predicted values of the transient isothermal indexes STI according to the intermediate characteristics h (x), a second task head outputs predicted values of the steady-state margin indexes HMI according to the intermediate characteristics h (x), then transient isothermal index loss and steady-state margin index loss are calculated according to the predicted values and labels of the training sample, and the multi-task learning agent model is optimized according to the loss values.
4. The method of claim 3, wherein the shared trunk uses two fully connected networks each having 64 hidden units and using a ReLU activation function, and the task output head uses a fully connected layer having 32 hidden units and a linear output layer.
5. A method according to claim 3, wherein the transient isothermal index STI and the steady state margin index HMI corresponding to each candidate sample point are calculated and substituted into the objective function Y based on the average value of the integrated objective function Y And standard deviation of Constructing confidence lower bound criteria And selecting the point with the smallest LCB (x) as a newly added simulation point from the candidate sample point set.
6. The method of claim 1, wherein a Sobol sampling method is adopted to generate a plurality of sample points in k-dimensional space, k is the number of key parameters, then a transient isothermal index STI and a steady state margin index HMI corresponding to each sample point are calculated according to an objective function Y by utilizing a multi-task learning agent model and substituted into the objective function Y to obtain an objective function value, a plurality of sample points with the minimum objective function value and central points and 1-2 angular points in the sample points are selected to serve as local optimization starting points, and a Nelder-Mead method is adopted to conduct local optimization to determine a group of optimal parameter values to serve as structural parameters of the superconducting cavity cooling structure.
7. The method of claim 1, wherein the initial structural parameters of the superconducting cavity cooling structure are adjusted using a Morris sensitivity analysis method to screen a set of key parameters.
8. The superconducting cavity cooling structure optimizing system based on the multi-task learning agent model is characterized by comprising an index calculating module, a parameter screening module, a model training module and an optimizing module; The index calculation module is used for constructing a superconducting cavity cooling structure model according to each group of structural parameters of the superconducting cavity cooling structure and obtaining a transient isothermal index STI and a steady state margin index HMI corresponding to the group of structural parameters through simulation calculation; The parameter screening module is used for screening a group of key parameters from the structural parameters according to the influence degree of each parameter on the transient isothermal index STI and the steady state margin index HMI; the model training module is used for generating a training sample set training optimization multi-task learning agent model according to the group of key parameters, wherein each training sample comprises a group of key parameter values and a transient isothermal index STI and a steady state margin index HMI corresponding to the key parameter values; the optimizing module is used for taking the group of key parameters as optimization variables, and determining a group of optimal parameter values as structural parameters of the superconducting cavity cooling structure by utilizing the optimized multi-task learning agent model.
9. A computing device comprising a processor, a memory storing a computer program which, when executed by the processor, performs the method of any one of claims 1 to 7.
10. A computer readable storage medium storing instructions which, when run on a computer, cause the computer to perform the method of any one of claims 1 to 7.

Description

Superconducting cavity cooling structure optimization method based on multitask learning agent model Technical Field The invention belongs to the field of heat design and structure optimization of a conduction cooling structure of a superconducting cavity, and relates to a superconducting cavity cooling structure optimization method based on a multi-task learning agent model. The method is suitable for optimizing the cooling structure parameters of the superconducting cavity under the conditions of vertical test, horizontal modules and different nonuniform heat leakage. Background The radio frequency superconducting module taking the conduction cooling superconducting cavity as a core is regarded as a key technology for pushing the radio frequency superconducting technology to be applied to industrial accelerators (such as electron beam irradiation, medical sterilization, environmental management and the like). The method has the core advantages of avoiding the complicated construction and operation cost of a large liquid helium low-temperature system and further improving the miniaturization and modularization of the superconducting particle accelerator. However, due to factors such as limited cooling capacity of a small refrigerator and uneven distribution of module leakage heat, the thermal stability and the thermal uniformity of a superconducting cavity often become key bottlenecks of a conduction cooling scheme, and the problems are mainly that, on one hand, when the temperature difference of the surface of the cavity is too large in a transient cooling process crossing a superconducting transition temperature zone, thermal current can be induced and an additional magnetic field can be formed, the risk of magnetic flux capture is increased, and therefore residual resistance is increased and the quality factor is reduced, and on the other hand, when the superconducting cavity runs stably, the thermal stability margin is compressed due to the fact that the local hot spot temperature is too high, and the risk of quench is increased. Therefore, how to optimize the geometric parameters of the cooling structure under given assembly and manufacturing constraints to achieve both thermal uniformity and thermal stability is a key issue in the engineering design of conduction-cooled superconducting cavities. The main scheme of the current conduction cooling structure comprises (1) copper hoop type, namely, a high heat conduction copper clamp is adopted on the surface of a superconducting cavity to take away heat, (2) surface spraying type, namely, a cold ring is formed on the surface of the superconducting cavity in a way of electroplating or cold spraying copper, and (3) niobium ring type is welded, namely, niobium material is welded at the positions of the equator of the superconducting cavity and a beam tube to serve as the cold ring. The welded niobium ring type cooling structure with the equatorial cooling ring and the beam tube cooling ring as cores has been widely used in vertical test and horizontal modules because of the advantages of simple structure, convenient processing, controllable thermal resistance and the like. For a vertical test environment with uniform static heat load, a relatively simplified and symmetrically arranged cold ring structure can be adopted, and in a horizontal module, the non-uniformity of heat leakage of the parts such as an input coupler, a support piece, a single side of a beam tube and the like is obviously enhanced, so that the arrangement and the geometric parameters of the cold ring are required to be re-optimized for an application scene. If the experimental iteration or full-parameter violent scanning is directly relied on, the calculation cost is high, and multi-objective thermal performance optimization is difficult to achieve. In this context, there is a great need for a method of optimizing a cooling structure that is guided by quantifiable criteria and is streamlined and efficient, so as to obtain a geometric solution with better thermal stability and thermal uniformity given mechanical and thermal boundary constraints. Disclosure of Invention Aiming at the problems existing in the prior art, the invention aims to provide a superconducting cavity cooling structure optimization method based on a multi-task learning agent model, the method fully considers the measurement index of the cooling effect, and combines Morris sensitivity screening, a multitask learning agent model and a preferred method, thereby obtaining the cooling structure with the optimal cooling effect. The invention provides an optimization method aiming at an index definition, parameter screening, proxy modeling and model optimizing of a conduction cooling structure, which comprises the following steps of (1) constructing transient isothermal indexes STI and steady-state margin indexes HMI, wherein STI is used for measuring control energy of temperature difference on a superconducting cavity when t