Internal target design method based on physical constraint neural network

Abstract

The invention discloses an inner target design method based on a physical constraint neural network, and relates to the technical field of treatment instrument improvement, and the method mainly comprises the steps: firstly, constructing and training a physical constraint neural network model, and in the training process, taking one or more physical field control equations controlling the work of a target system as constraint embedding models, specifically, physical equation residual errors corresponding to geometric data output by a model are calculated, and the residual errors are incorporated into a loss function. Then, in actual design application, a design instruction of a user is coded into a standard input vector, and the standard input vector is input into the trained model; the model can perform real-time reasoning and output corresponding optimized target surface geometric data through single forward propagation operation, and finally the data is converted and packaged into a three-dimensional model file which can be directly used. According to the method, the traditional inner target design mode depending on iterative simulation is broken through, and the second-level real-time generation of the physically reliable optimized target surface is realized.

Description

Technical Field

[0001] This invention relates to the field of therapeutic mechanical improvement technology, specifically to an internal target design method based on a physically constrained neural network. Background Technology

[0002] In cyclotrons, especially in internal target systems used for the production of medical radioisotopes, the geometry of the target surface directly determines the beam utilization efficiency, system power handling capacity, and the purity of the final nuclear products, making it a core physical component affecting overall performance. Traditional internal target design methods essentially follow a "geometric matching-numerical optimization" paradigm, which involves locally matching the target surface geometry with the beam trajectory and relying on computational fluid dynamics and particle transport programs for iterative iterations and simulation verification. While this approach has achieved some success in early stages and specific scenarios, it has gradually revealed systemic technical bottlenecks when facing higher power, stricter purity requirements, and more complex operating modes. On the one hand, its design process heavily relies on high-fidelity multiphysics numerical simulations. Each parameter adjustment or design iteration requires re-execution of time-consuming simulation calculations, resulting in a cycle of weeks or even months from conceptual design to final scheme confirmation, which cannot meet the development needs of rapid iteration. On the other hand, traditional methods are mostly based on pre-set geometric parameterized models (such as fixed circular arcs or planar configurations). Their search range and optimization results are highly dependent on the initial experience settings of engineers, which not only easily leads to local optima but also makes it difficult to explore high-performance geometric configurations with counterintuitive characteristics. In addition, existing solutions are mostly isolated processes built independently for specific scenarios such as single-loop or multi-loop designs, lacking a unified intelligent framework that can adapt to various design goals. Moreover, at the level of physical consistency, purely data-driven optimization models often lack explicit adherence to underlying physical laws, which may output physically infeasible or unstable design schemes.

[0003] Although artificial intelligence technologies such as physical information neural networks have made it possible to integrate prior physical knowledge in recent years, how to build an intelligent design system that can strictly follow physical laws and achieve end-to-end generation from high-level targets to geometric shapes remains a technical problem that urgently needs to be solved in the field of particle accelerator target design, which combines high-precision engineering with complex multi-physics coupling. Summary of the Invention

[0004] To construct an internal target design system that strictly adheres to physical laws and achieves end-to-end generation from high-level targets to geometric shapes, this invention proposes an internal target design method based on a physically constrained neural network, comprising the following steps: S1: Define a standardized input vector for receiving programmable design instructions and construct a neural network model that outputs target surface geometric parameterized data, thereby constructing and training a physically constrained neural network model; S2: During the training process of the physical constraint neural network model, the control equations of one or more physical fields that govern the operation of the target system are used as constraints. The physical equation residuals corresponding to the geometric data output by the physical constraint neural network model are calculated and incorporated into the loss function for embedding. S3: Encode the user-configured design instructions into standardized input vectors and input them into the trained physical constraint neural network model; S4: Through a single forward propagation operation of the physical constraint neural network model, the corresponding optimized target surface geometry data is inferred in real time based on the standardized input vector; S5: Convert and package the output optimized target surface geometry data into a 3D model file.

[0005] This invention achieves real-time generation of physically reliable optimized target surfaces in seconds by embedding physical control equations as constraints into a neural network and establishing an end-to-end mapping from programmable design instructions to target surface geometry, fundamentally breaking through the traditional design method that relies on iterative simulation.

[0006] Furthermore, in step S1, the standardized input vector includes: a pattern identifier for identifying the type of design task, beam parameters, material and cooling parameters, and preference weights for weighing conflicting physical objectives.

[0007] Furthermore, the pattern identifier is used to refer to at least one design task type among single-cycle optimization design, multi-cycle collaborative design, and robust design.

[0008] Furthermore, when the pattern identifier indicates a multi-loop collaborative design, the standardized input vector also includes beam parameters of multiple loops and loop-specific preference weights; the physical constraint neural network model adopts an encoder-decoder structure with an attention mechanism to generate a target geometry that uniformly coordinates the physical targets of multiple loops.

[0009] Furthermore, when the mode identifier indicates a robust design, the beam parameters include random variables characterizing uncertainty; in step S2, a target geometry insensitive to parameter fluctuations is generated by injecting noise into the network and using Monte Carlo sampling to estimate random biases in the physical equations.

[0010] Furthermore, in step S1, the physical constraint neural network model adopts a multi-task learning architecture. The output layer of this multi-task learning architecture includes a main output head for generating target surface geometric parameterization data, and an auxiliary output head for predicting one or more performance indicators.

[0011] Furthermore, in step S2, the physical field governing the operation of the target system includes at least one of the following: a beam dynamic field and a thermal conduction field.

[0012] Furthermore, step S2 also includes: designing activation functions, output layer transformations, or projection operations to make the geometric data output by the network automatically meet preset geometric boundary conditions.

[0013] Furthermore, after step S4, the following steps are also included: using a lightweight proxy model to evaluate the performance of the output optimized target surface geometry data; if the evaluation result does not meet the preset performance index, the design instruction adjustment process is triggered and the process returns to step S3; if the evaluation result meets the preset performance index, the process proceeds to step S5.

[0014] Furthermore, the preset performance indicators include at least one of the following: target output, maximum allowable temperature, peak-to-average power ratio, and beam interception efficiency.

[0015] Compared with the prior art, the present invention has at least the following beneficial effects: (1) The present invention proposes an internal target design method based on a physical constraint neural network. By embedding the control equations of multiple physical fields such as beam dynamics and heat conduction into the neural network training process in the form of residual constraints, the generated target geometry strictly follows the physical laws, thus ensuring the rationality and reliability of the design from the source. (2) With the help of programmable standardized input vectors, users can express diverse design intentions through design instructions. The system can then efficiently and in real time reason and output optimized target surface geometric parameter data, which can be automatically converted into a three-dimensional model file that can be directly used for manufacturing, thus shortening the design cycle and improving design efficiency. (3) In the multi-ring collaborative scenario, by introducing an encoder-decoder structure with attention mechanism and ring-specific preference weights, the model can coordinate conflicting physical targets between different rings and generate a system-optimal unified target geometry. In the robust design scenario, by injecting noise into the network and combining it with Monte Carlo sampling, a robust geometric configuration that is insensitive to beam parameter fluctuations and has good fault tolerance is trained and generated. Attached Figure Description

[0016] Figure 1 This is a step diagram of an internal target design method based on a physically constrained neural network. Detailed Implementation

[0017] Traditional internal target design methods heavily rely on high-fidelity multiphysics numerical simulations. Each design iteration requires re-executing time-consuming simulation calculations, resulting in design cycles lasting weeks or even months. Furthermore, traditional methods are typically based on fixed geometric parameterized models, leading to narrow search spaces, susceptibility to local optima during optimization, and a high dependence on the designer's experience. More importantly, while purely data-driven neural network models possess strong fitting capabilities, their outputs lack physical consistency, potentially resulting in physically unreasonable or unrealizable design flaws. To address these issues, this invention proposes a novel design method that explicitly embeds physical principles into a neural network, achieving a direct, fast, and physically consistent mapping from high-level design objectives to target geometry. This improves design efficiency while ensuring the reliability and innovation of the proposed solution. Figure 1 As shown, this design method mainly includes the following steps: S1: Define a standardized input vector for receiving programmable design instructions and construct a neural network model that outputs target surface geometric parameterized data, thereby constructing and training a physically constrained neural network model; S2: During the training process of the physical constraint neural network model, the control equations of one or more physical fields that govern the operation of the target system are used as constraints. The physical equation residuals corresponding to the geometric data output by the physical constraint neural network model are calculated and incorporated into the loss function for embedding. S3: Encode the user-configured design instructions into standardized input vectors and input them into the trained physical constraint neural network model; S4: Through a single forward propagation operation of the physical constraint neural network model, the corresponding optimized target surface geometry data is inferred in real time based on the standardized input vector; S5: Convert and package the output optimized target surface geometry data into a 3D model file.

[0018] The technical solution of the present invention will be described in detail below through several specific embodiments. First, a basic and complete implementation process will be introduced to demonstrate the core steps and general framework of the method of the present invention. In this basic embodiment, the single-loop target surface depth optimization commonly used in medical cyclotrons is taken as a scenario. The design goal is to minimize the peak-to-average power ratio on the target surface while ensuring thermal safety, thereby achieving more uniform energy deposition and temperature distribution.

[0019] The implementation process begins with the construction of the hardware and software platforms. On the hardware side, a workstation or server cluster with at least one high-performance graphics processing unit is required to provide the necessary computing power for neural network training and real-time inference. The software framework integrates mainstream deep learning libraries, scientific computing tools, multiphysics simulation software, and a computer-aided design (CAD) kernel. Deep learning libraries, such as PyTorch or TensorFlow, are used to build and train neural network models; scientific computing libraries like NumPy and SciPy are used for data processing and mathematical operations; the multiphysics simulation toolchain is used to generate the labeled data required for training and to perform high-fidelity verification of the final design; and the CAD kernel is responsible for automatically converting the parametric geometric data output by the neural network into 3D model files suitable for manufacturing.

[0020] After the platform is built, data preparation is the first step. The design task is defined as "single-cycle depth optimization," with specific physical objectives set, such as "minimizing peak-to-average power ratio and ensuring the highest temperature is below the material's melting point." A series of geometrically diverse and physically feasible initial target surface samples are generated based on existing successful design cases in the field or through traditional parametric methods. For each sample, high-fidelity multiphysics simulations are performed to obtain its true performance indicators. This typically involves using particle tracking programs to calculate the energy deposition distribution of the beam on the target surface, then using this energy deposition distribution as a heat source, importing it into computational fluid dynamics software for conjugate heat transfer simulations to obtain key data such as the steady-state temperature field and thermal stress distribution. After the simulations are completed, high-performing samples are selected, and their geometric parameters and corresponding simulation performance indicators are recorded. Finally, the design instructions, corresponding optimized geometric data, and their performance indicators are encapsulated into structured training sample pairs.

[0021] The design instructions need to be encoded into a standardized input vector. This input vector is a multi-dimensional data structure that includes at least a mode identifier, beam parameters, material and cooling parameters, and physical objectives and preference weights. The mode identifier indicates the current design task type, such as "single-loop optimization." Beam parameters include particle energy and beam intensity. Material and cooling parameters cover target properties, cooling channel geometry, and flow parameters. Physical objectives and preference weights define specific optimization objectives and trade-offs between conflicting objectives; for example, the weight of "peak-to-average power ratio" can be set to 1.0, while the weight of "maximum temperature" can be set to 0.8.

[0022] Next, we construct the neural network model itself. The number of neurons in the input layer matches the dimension of the normalized input vector. The output layer needs to be designed to directly generate parametric data representing the target geometry. A preferred implementation is to output a fixed-dimensional vector representing the coordinates of a series of control points defining the major axis profile of the target. For example, outputting a 16-dimensional vector can represent 8 control points. Coordinates, connected by spline curves, generate a smooth target surface contour curve. The main body of the model employs a deep fully connected network, for example, containing 5 hidden layers with 128 neurons per layer, and using non-linear activation functions such as Swish to enhance the model's expressive power. A key design feature of this invention is the use of a multi-task learning architecture. That is, in addition to the main output head used to generate geometry, the network also sets up several auxiliary output heads in parallel. These auxiliary output heads are used to directly predict key performance indicators, such as peak-to-average power ratio, maximum temperature, and beam interception efficiency. This design not only provides users with real-time performance predictions during inference, but more importantly, during training, gradient feedback from the auxiliary tasks helps the main task (geometry generation) better learn the complex relationship between geometry and performance, thereby improving the overall optimization capability of the model.

[0023] Furthermore, the embedding mechanism of physical constraints is the core innovation of this invention, distinguishing it from purely data-driven methods. This mechanism plays a role in the model training phase. The total loss function during training consists of several parts. The first part is data loss, such as using mean squared error to calculate the difference between the geometry predicted by the network and the control point label data, ensuring that the model can learn from the experience of existing excellent designs. The second part is physical residual loss, which is crucial for achieving physical consistency. During the forward propagation of the training process, not only is the data loss calculated, but the "guessed" geometry output by the network in the current iteration, along with the input beam parameters, material parameters, etc., are also substituted into the physical control equations governing the operation of the target system for calculation. These physical equations include at least the beam dynamics equations describing the beam trajectory and the thermodynamic equations describing heat conduction. For single-loop optimization scenarios, the beam dynamics equations can be used to calculate the coverage integrity of the beam under a given geometry; the heat conduction equations are used to calculate the corresponding temperature field distribution. The theoretical values ​​calculated by the equations are compared with the values ​​required by real physical laws (such as the requirement that the heat source equals the heat dissipation due to energy conservation), and the difference constitutes the physical residual. This residual is added to the total loss function as a penalty term. The total loss function is therefore a weighted sum of the data loss term and the physical loss term, with hyperparameters used to balance their importance. Through backpropagation, gradients from the physical residuals force the network to adjust its internal parameters, making its output geometry increasingly conform to fundamental physical laws while still satisfying the training data. This means that even in areas not covered by the training data, physical constraints can guide the network to generate physically feasible designs, greatly enhancing the model's generalization ability and reliability. When the model's loss on the independent validation set no longer decreases significantly, training is complete, and a fixed set of network weights can be obtained.

[0024] Once the model training is complete, it can enter the deployment and application phase. Users can intuitively configure its design instructions through a graphical user interface by selecting modes, dragging sliders, and inputting values. For example, a user can select the "single-cycle optimization" mode, set "minimizing peak-to-average power ratio" as the primary objective, and simultaneously set an upper limit threshold for the maximum temperature. The system backend automatically encodes these interactive operations into standardized input vectors consistent with the training format. Subsequently, this vector is input into the pre-trained physical constraint neural network. The network performs a forward propagation operation, directly outputting optimized target surface geometric parameterization data within seconds or even milliseconds. Then, the system's built-in CAD kernel reads this data, automatically constructs the corresponding 3D geometric model, and encapsulates it into industry-standard format files such as STEP or IGES, which can be directly imported into manufacturing software or used for further detailed analysis.

[0025] To provide users with immediate feedback, the system also integrates a lightweight performance proxy model. This proxy model can be another pre-trained fast prediction network or an analytical model based on simplified physical formulas. It can quickly evaluate the performance of the generated design, predict its peak-to-average power ratio, maximum temperature, and other indicators, and present them to the user in the form of visual charts. If the user is not satisfied with the evaluation results, they can immediately return to the interface to fine-tune the preference weights or other parameters, and then trigger real-time generation again, forming a "design-evaluation-fine-tuning" intelligent closed loop, completely avoiding the time-consuming simulation loop in traditional methods.

[0026] The above is a complete description of the basic embodiment. To demonstrate the versatility of the invention, its specific implementation in the more challenging "multi-loop cooperative design" scenario will be further described below.

[0027] In actual operation of cyclotrons, the beam often operates in a multi-cycle mode, meaning that particle beams of different energies bombard the same physical target surface sequentially along different trajectories. This requires that the target surface geometry not only be optimized for a single cycle but also comprehensively consider the conflicting physical targets of multiple cycles. For example, a deeply recessed surface optimized for high-yield, high-energy cycles may result in incomplete beam coverage of low-energy cycles or the generation of local hot spots. Traditional single-cycle design methods are ineffective in addressing this, while optimization for multiple cycles is even more complex and time-consuming.

[0028] When applying the method of this invention, the user first sets the mode identifier to "multi-coil collaborative design" in the interactive interface. The corresponding standardized input vector also needs to be expanded; in addition to the general beam parameters and material parameters, it needs to include the specific beam parameters of all relevant coils, and assign a system-level preference weight to each coil. For example, if the design involves the... Three concentric circles, and the first If a circle contributes the most to the final output, it can be assigned a higher preference weight, while other circles have lower weights, thus reflecting the priority focus in the design.

[0029] Furthermore, to handle such complex multi-source inputs and system-level tradeoffs, the structure of physically constrained neural networks also needs adaptive enhancement. A preferred implementation is an encoder-decoder structure with an attention mechanism. Specifically, an independent encoding subnetwork is set up for each layer to extract the input features specific to that layer. Subsequently, a feature fusion module receives the encoded features from all layers and dynamically calculates the importance weights of different layer features in generating the final unified geometry using the attention mechanism. The attention mechanism allows the network to "focus" on the currently more critical layer information. Finally, the decoder generates a single, coordinated target geometry parameterization based on the fused global features.

[0030] Similarly, when training such a multi-layered collaborative network, the design of the loss function must also reflect the idea of ​​system-level optimization. In addition to including the data loss between the final generated unified geometry and the labeled geometry, the total loss should also include auxiliary performance prediction losses for each layer, such as predicting the uniformity of heat distribution under the generated geometry for each layer. These layer-specific loss terms are weighted and summed according to previously set preference weights, thereby guiding the network to learn during training how to make the best trade-offs between the competing needs of different layers.

[0031] The model trained in this way can generate an overall optimal target surface when faced with multi-coil collaborative design commands. This surface may have a larger curvature in coil regions where high yield is required to better match the beam, while it may be more gentle in coil regions where overheating needs to be avoided to facilitate heat dissipation.

[0032] Besides multi-coil coordination, practical engineering often faces various uncertainties, such as random drift in beam bombardment position due to equipment fluctuations. A design under ideal conditions may experience a sharp performance drop or even failure in actual operation due to parameter drift. Therefore, the method of this invention can also be specifically used for "robust design".

[0033] In this implementation, the user sets the pattern identifier to "robust design." To characterize uncertainty in the model, random variables need to be introduced into the beam parameters of the input vector. For example, the x-coordinate of the beam center can be defined as a random variable following a Gaussian distribution with nominal mean and variance. To train the network to generate robust geometry insensitive to input fluctuations, special processing needs to be introduced into the training strategy. An effective method is to actively inject noise into the network's feature space. For example, a random noise vector can be added to the transmitted feature vector between the encoder and decoder. This forces the network to learn that it should output a stable and smooth geometric response even with small perturbations in the input features, thereby enhancing the model's robustness. This randomness also needs to be adapted in the physical constraint embedding stage.

[0034] Traditional physical residual calculations are performed for a fixed set of inputs and outputs. Robust training, however, requires a Monte Carlo sampling method. For each training sample, multiple different beam position perturbation values ​​are randomly sampled from the aforementioned Gaussian distribution. For each sampled value, the physical equation residuals (e.g., calculating the temperature field uniformity under that perturbation) are calculated based on the geometry output by the network. Finally, the mean and variance of these sampled residuals are taken as the final physical loss term. The mean residual guides the network to optimize average performance, while the variance penalizes the performance's sensitivity to perturbations, encouraging the network to output a geometry that is stable under various possible perturbations. The model trained in this way can generate target profiles with smooth transitions and insensitivity to beam position drift when the user inputs robust design instructions containing uncertain parameters, thereby improving the reliability and lifespan of the design in real-world operating environments.

[0035] To further illustrate the effectiveness of the method of this invention in specific engineering problems, a detailed implementation case is given below, using the internal target design for producing the medical radioisotope At-211 in a TR-Alpha cyclotron accelerator as an example. The production of At-211 places extremely stringent requirements on the thermal management of the target surface; localized overheating can lead to target material volatilization (selective emission), severely affecting product purity and system safety. Traditional initial designs often start with simple planar targets, improving them through complex parameter scanning and simulation iterations. However, after applying the technical solution of this invention, several publicly available At-211 target design cases are first collected as seeds. These seed geometries are then mutated using a parametric model, systematically changing parameters such as tilt angle and radius of curvature to generate an initial geometric sample library covering a certain design space. For each geometric variant in the library, a rigorous dual simulation is performed: first, simulation is conducted using Geant4 or SRIM software. The stopping power and three-dimensional energy deposition distribution of particles in the bismuth target were analyzed. This energy deposition map was then used as a heat source and imported into multiphysics simulation software such as ANSYS or COMSOL. Combined with specific cooling channel parameters (flow rate, inlet temperature), conjugate heat transfer simulations were performed to obtain accurate steady-state temperature field and thermal stress distribution. The design with the best performance in the simulation results (e.g., the design with the lowest peak temperature while meeting interception efficiency requirements) and its corresponding input commands (set to "thermal management priority" mode, including specific beam and cooling parameters) were used as high-quality training samples.

[0036] Based on these samples, a physically constrained neural network for At-211 production scenarios was constructed and trained. The network's input vector has a dimension of 12, covering modes, beam energy and intensity, bismuth layer thickness, cooling water parameters, etc. The output is a 16-dimensional vector, defining 8 control points for the target surface profile. In the physical constraint embedding, thermal constraints are emphasized. The residual of Fourier's law of heat conduction is used as the core part of the physical loss, forcing the geometry of the network output to satisfy basic energy conservation and heat conduction laws. After training, designers can input the command "thermal management priority" through the interface, and the model can generate a completely new target surface design. High-fidelity simulation verification shows that compared with the initial planar baseline design, the peak temperature is reduced by about 15%, while the beam interception efficiency remains above 85%, effectively mitigating the risk of selective firing. In terms of efficiency, traditional parameter scanning optimization methods require approximately 140 GPU hours to complete a design exploration of similar depth, while the method of this invention requires approximately 100 GPU hours for a one-time model training. Most importantly, after training, the generated scheme takes only a few seconds for each new design instruction, and can flexibly adapt to changes in beam parameters or cooling conditions, demonstrating the efficiency advantage and adaptability of the technical solution of this invention.

[0037] In summary, this invention proposes a novel internal target design method based on physically constrained neural networks. This method combines the powerful representation and exploration capabilities of deep learning with the first-principles constraints of physical systems, creating a completely new internal target design approach. This method not only achieves an order-of-magnitude improvement in design efficiency but also ensures the reliability of the design through embedded physical laws. Furthermore, it enables unified processing of complex, multi-scenario, multi-objective, and uncertain design tasks through a programmable instruction system.

[0038] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0039] Furthermore, in this invention, descriptions involving terms such as "first," "second," and "a" are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0040] In this invention, unless otherwise explicitly specified and limited, the terms "connection," "fixed," etc., should be interpreted broadly. For example, "fixed" can mean a fixed connection, a detachable connection, or an integral part; it can mean a mechanical connection or an electrical connection; it can mean a direct connection or an indirect connection through an intermediate medium; it can mean the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0041] Furthermore, the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are feasible for those skilled in the art. If the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

Claims

1. An internal target design method based on a physically constrained neural network, characterized in that, Including the following steps: S1: Define a standardized input vector for receiving programmable design instructions and construct a neural network model that outputs target surface geometric parameterized data, thereby constructing and training a physically constrained neural network model; S2: During the training process of the physical constraint neural network model, the control equations of one or more physical fields that govern the operation of the target system are used as constraints. The physical equation residuals corresponding to the geometric data output by the physical constraint neural network model are calculated and incorporated into the loss function for embedding. S3: Encode the user-configured design instructions into standardized input vectors and input them into the trained physical constraint neural network model; S4: Through a single forward propagation operation of the physical constraint neural network model, the corresponding optimized target surface geometry data is inferred in real time based on the standardized input vector; S5: Convert and package the output optimized target surface geometry data into a 3D model file.

2. The internal target design method based on a physically constrained neural network as described in claim 1, characterized in that, In step S1, the standardized input vector includes: a pattern identifier for identifying the type of design task, beam parameters, material and cooling parameters, and preference weights for weighing conflicting physical objectives.

3. The internal target design method based on a physically constrained neural network as described in claim 2, characterized in that, The pattern identifier is used to refer to at least one design task type among single-cycle optimization design, multi-cycle collaborative design, and robust design.

4. The internal target design method based on a physically constrained neural network as described in claim 3, characterized in that, When the pattern identifier indicates a multi-loop collaborative design, the standardized input vector also includes beam parameters of multiple loops and loop-specific preference weights; the physical constraint neural network model adopts an encoder-decoder structure with an attention mechanism to generate a target geometry that uniformly coordinates the physical targets of multiple loops.

5. The internal target design method based on a physically constrained neural network as described in claim 3, characterized in that, When the mode identifier indicates a robust design, the beam parameters include random variables characterizing uncertainty; in step S2, target geometry insensitive to parameter fluctuations is generated by injecting noise into the network and using Monte Carlo sampling to estimate random biases in the physical equations.

6. The internal target design method based on a physically constrained neural network as described in claim 1, characterized in that, In step S1, the physical constraint neural network model adopts a multi-task learning architecture. The output layer of this multi-task learning architecture includes a main output head for generating target surface geometric parameterization data and an auxiliary output head for predicting one or more performance indicators.

7. The internal target design method based on a physically constrained neural network as described in claim 1, characterized in that, In step S2, the physical field that governs the operation of the target system includes at least one of the following: beam dynamics field and thermal conduction field.

8. The internal target design method based on a physically constrained neural network as described in claim 1, characterized in that, The S2 step also includes: designing activation functions, output layer transformations, or projection operations to make the geometric data output by the network automatically meet preset geometric boundary conditions.

9. The internal target design method based on a physically constrained neural network as described in claim 1, characterized in that, The step S4 is followed by the following step: using a lightweight proxy model to evaluate the performance of the output optimized target surface geometry data. If the evaluation result does not meet the preset performance index, the design instruction adjustment process is triggered and the process returns to step S3. If the evaluation result meets the preset performance index, the process proceeds to step S5.

10. The internal target design method based on a physically constrained neural network as described in claim 9, characterized in that, The preset performance indicators include at least one of the following: target output, maximum allowable temperature, peak-to-average power ratio, and beam interception efficiency.

Images