A method for plasma disruption attribution based on a large model in the field of nuclear fusion

Abstract

The specification discloses a plasma disruption attribution method based on a nuclear fusion field large model, determines a disruption precursor event, a time of occurrence of the disruption precursor event and an occurrence confidence through a multi-modal discriminator according to experimental data of a controllable nuclear fusion experiment, determines a connection relationship between each disruption precursor event based on the disruption precursor event, the time of occurrence of the disruption precursor event and the occurrence confidence through a nuclear fusion large model, and determines a corresponding disruption event chain of the controllable nuclear fusion experiment according to the connection relationship, so as to determine a reason for plasma disruption according to the disruption event chain. It can be seen that, through the above scheme, the nuclear fusion large model and the multi-modal discriminator fully utilize the multi-modal and fragmented experimental data and expert experience of the controllable nuclear fusion, realize analysis on the reason for plasma disruption, determine a control decision of the next experiment based on the reason for plasma disruption, and thus improve the iteration efficiency of the controllable nuclear fusion experiment.

Description

Technical Field

[0001] This specification relates to the field of computer technology, and in particular to a plasma breakup attribution method based on a large-scale model in the field of nuclear fusion. Background Technology

[0002] Controlled nuclear fusion is a crucial technology for effectively addressing the energy dilemma. With technological advancements and increased international cooperation, magnetic confinement fusion research is poised to enter the era of combustion plasma. However, the high resource and time costs associated with experimental data collection have slowed progress in studying fusion physics and control strategies. In recent years, with the rapid development of artificial intelligence (AI) technology, both domestic and international researchers have begun exploring its use to address bottlenecks in fusion experimental research.

[0003] Tokamaks are currently one of the most promising methods for achieving magnetically confined controlled nuclear fusion. In the process of fusion control using tokamaks, plasma travels at extremely high speeds within the device. By controlling the poloidal and circumferential magnetic coils around the device, magnetic confinement of the plasma can be achieved, with the goal of ensuring stable, high-parameter operation of the plasma within the tokamak. A crucial control objective is to prevent plasma fragmentation during high-speed operation. Fragmentation not only leads to discharge failure but also causes extensive damage to the device due to large-scale fragmentation.

[0004] In existing technologies, deep learning-based methods are used to predict the occurrence of plasma breakup.

[0005] However, from the perspective of fusion control, simply predicting the occurrence of plasma breakup is still insufficient to guide the avoidance of plasma breakup in the next experiment. It is also necessary to pinpoint the cause of plasma breakup in order to provide control decisions for the next experiment.

[0006] Based on this, this specification provides a plasma breakup attribution method based on a large-scale model in the field of nuclear fusion. Summary of the Invention

[0007] This specification provides a plasma breakup attribution method based on a large-scale model in the field of nuclear fusion, in order to partially solve the aforementioned problems existing in the prior art.

[0008] The following technical solution is adopted in this specification:

[0009] This specification provides a plasma breakup attribution method based on a large-scale model in the field of nuclear fusion, including:

[0010] Acquire experimental data from controlled nuclear fusion experiments, including experimental logs, target image sequences, magnetic probe timing signals, and timing signals of the plasma geometric center location;

[0011] The experimental data is input into a pre-trained multimodal discriminator to obtain multiple rupture precursor events corresponding to the controlled nuclear fusion experiment, the occurrence confidence of each rupture precursor event, and the occurrence time of each rupture precursor event, as output by the multimodal discriminator.

[0012] The precursory events of the controlled nuclear fusion experiment, the occurrence confidence levels of each precursory event, and the occurrence times of each precursory event are used as inputs to a pre-trained nuclear fusion large-scale model. The large-scale model determines the connections between the precursory events and obtains the chain of events corresponding to the controlled nuclear fusion experiment based on the connections. The large-scale model is based on corpus related to the controlled nuclear fusion experiment and a pre-trained natural language model.

[0013] The cause of plasma rupture in the controlled nuclear fusion experiment is determined based on the rupture event chain.

[0014] This specification provides a plasma breakup attribution device based on a large-scale model in the field of nuclear fusion, including:

[0015] The experimental data acquisition module is used to acquire experimental data from controlled nuclear fusion experiments. The experimental data includes experimental logs, target image sequences, magnetic probe timing signals, and timing signals of the plasma geometric center position.

[0016] The multimodal discrimination module is used to input the experimental data into a pre-trained multimodal discriminator to obtain multiple rupture precursor events corresponding to the controlled nuclear fusion experiment, the occurrence confidence of each rupture precursor event, and the occurrence time of each rupture precursor event, as output by the multimodal discriminator.

[0017] The rupture event chain determination module is used to input each rupture precursor event corresponding to the controlled nuclear fusion experiment, the occurrence confidence of each rupture precursor event, and the occurrence time of each rupture precursor event as inputs into a pre-trained nuclear fusion large model, determine the connection relationship between each rupture precursor event through the nuclear fusion large model, and obtain the rupture event chain corresponding to the controlled nuclear fusion experiment based on the connection relationship;

[0018] The rupture cause determination module is used to determine the cause of plasma rupture in the controlled nuclear fusion experiment based on the rupture event chain.

[0019] This specification provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned plasma breakup attribution method based on a large-scale model in the field of nuclear fusion.

[0020] This specification provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the aforementioned plasma breakup attribution method based on a large-scale model in the field of nuclear fusion.

[0021] The above-mentioned technical solutions adopted in this specification can achieve the following beneficial effects:

[0022] The plasma breakup attribution method based on a large-scale model in the field of nuclear fusion provided in this specification determines the precursor events, their occurrence times, and confidence levels using a multimodal discriminator based on experimental data from controlled nuclear fusion experiments. Then, using a large-scale nuclear fusion model, the connection relationships between these precursor events are determined based on their occurrence times and confidence levels, thus establishing the breakup event chain corresponding to the controlled nuclear fusion experiment. Finally, the cause of plasma breakup is determined based on this event chain. Therefore, this approach, by fully utilizing the multimodal and fragmented experimental data and expert experience from controlled nuclear fusion experiments through a large-scale nuclear fusion model and a multimodal discriminator, enables the analysis of plasma breakup causes. This allows for the determination of control decisions for the next experiment based on the cause of the plasma breakup, thereby improving the iterative efficiency of controlled nuclear fusion experiments. Attached Figure Description

[0023] The accompanying drawings, which are included to provide a further understanding of this specification and form part of this specification, illustrate exemplary embodiments and are used to explain this specification, but do not constitute an undue limitation thereof. In the drawings:

[0024] Figure 1 This is a flowchart illustrating a plasma breakup attribution method based on a large-scale model in the field of nuclear fusion, as described in this specification.

[0025] Figure 2 This is a schematic diagram of a rupture event chain in this specification;

[0026] Figure 3 This is a schematic diagram of a multimodal discriminator described in this specification;

[0027] Figure 4 This is a flowchart illustrating a plasma breakup attribution method based on a large-scale model in the field of nuclear fusion, as described in this specification.

[0028] Figure 5 This is a flowchart illustrating a plasma breakup attribution method based on a large-scale model in the field of nuclear fusion, as described in this specification.

[0029] Figure 6 A schematic diagram of a plasma breakup attribution device based on a large model in the field of nuclear fusion, provided for this specification;

[0030] Figure 7 The corresponding information provided in this specification Figure 1 A schematic diagram of an electronic device. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this specification clearer, the technical solutions of this specification will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this specification, and not all of them. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this specification.

[0032] Additionally, it should be noted that all actions involving the acquisition of signals, information, or data in this manual are performed in accordance with the relevant data protection regulations and policies of the locality and with authorization from the owner of the relevant device.

[0033] It should be noted that, unless otherwise specified, the features in the following embodiments and implementation methods can be combined with each other.

[0034] Controlled nuclear fusion is a crucial technology for effectively addressing the energy dilemma. With technological advancements and increased international cooperation, magnetic confinement fusion research is poised to enter the era of combustion plasma. However, the high resource and time costs associated with experimental data collection have slowed progress in studying fusion physics and control strategies. In recent years, with the rapid development of artificial intelligence (AI) technology, both domestic and international researchers have begun exploring its use to address bottlenecks in fusion experimental research.

[0035] Tokamaks are one of the most promising methods for achieving magnetically confined controlled nuclear fusion. A tokamak is a toroidal vacuum chamber with magnetic coils. The basic principle of fusion control using tokamaks is to inject hydrogen isotopes deuterium and tritium into a vacuum device, and then excite them through heating and ionization, allowing electrons and ions to gain sufficient energy and velocity to form plasma. When certain temperature, density, and energy confinement time are reached, deuterium-deuterium atoms or deuterium-tritium atoms can eventually overcome Coulomb repulsion and form a fusion reaction; that is, the product of the above three factors can meet the requirements of the Lawson criterion. The macroscopic expectation is that the fusion output power will exceed the input power. During this period, the plasma will move at extremely high speeds within the device. By controlling the poloidal and toroidal magnetic coils around the device, magnetic confinement of the plasma can be achieved, with the goal of allowing the plasma to operate stably within the tokamak device in a high-parameter manner.

[0036] Therefore, an important goal of control is to prevent plasma from breaking up during high-speed operation. Breaking up not only leads to discharge failure, but large-scale breaking up can also cause extensive damage to the device.

[0037] Regarding plasma breakup, if breakup could be predicted during discharge based on certain precursor events, and the plasma state adjusted in real time to achieve breakup, it would not only help reduce the risk of breakup in current experiments but also accumulate experience for future experiments. However, during the precursor development stage, the causes of breakup, such as plasma instability, often evolve and accumulate gradually. The data for this process are highly diverse, with different triggers and precursor events mutually inducing and coupling to form complex event chains. The duration of precursor development also varies depending on the cause of breakup, ranging from a few milliseconds to several seconds. Current research cannot yet provide a clear and comprehensive classification of the causes and precursors of breakup.

[0038] Currently, deep learning-based predictions of plasma breakup actually predict whether a breakup will occur, but do not address the cause of the breakup. Therefore, how to accurately determine the cause of breakup in controlled nuclear fusion experiments and thus prevent plasma breakup through certain controls is an urgent problem to be solved.

[0039] Based on this, this specification provides a plasma breakup attribution method based on a large model in the field of nuclear fusion. By fully utilizing the large nuclear fusion model and multimodal discriminator, it makes full use of multimodal and fragmented experimental data and expert experience in controlled nuclear fusion to analyze the causes of plasma breakup. Then, based on the causes of plasma breakup, it determines the control decisions for the next experiment, thereby improving the iterative efficiency of controlled nuclear fusion experiments.

[0040] The technical solutions provided in the various embodiments of this specification are described in detail below with reference to the accompanying drawings.

[0041] Figure 1 This document presents a flowchart illustrating a plasma breakup attribution method based on a large-scale model in the field of nuclear fusion.

[0042] S100: Acquire experimental data from the controlled nuclear fusion experiment, including experimental logs, target image sequences, magnetic probe timing signals, and timing signals of the plasma geometric center position.

[0043] This specification provides a plasma breakup attribution method based on a large-scale model in the field of nuclear fusion. The execution of this method can be performed by electronic devices such as servers used for plasma breakup attribution. Furthermore, during the execution of this method, the trained multimodal discriminator and the large-scale nuclear fusion model involved, as well as the electronic devices used for model training and execution of this method, can be the same or different; this specification does not impose any limitations on this.

[0044] In controlled nuclear fusion experiments, multimodal experimental data are generated, including experimental logs, target image sequences, magnetic probe timing signals, and timing signals of the plasma geometric center location.

[0045] The experimental log includes the research objectives and expected results of the experiment, experimental equipment and parameters, experimental operation procedures (including equipment startup, heating or cooling stages, experimental duration, etc.), experimental results, and any problems, abnormal situations, or unclear observations encountered during the experiment. The experimental log can be recorded by the operator in natural language; this manual does not limit the type of natural language used in the information in the experimental log.

[0046] The target image sequence comprises multiple charge-coupled device (CCD) images of the plasma acquired by designated image acquisition equipment during controlled nuclear fusion experiments. In the sequence, the CCD images are arranged sequentially according to the time of acquisition. CCD images are used to observe and record the plasma discharge process and can be used to study the morphology, structure, and dynamic characteristics of the plasma. The image acquisition equipment employs photoelectric conversion technology to convert light signals into electrical signals, which are then digitally processed and recorded to ultimately form CCD images. By analyzing these CCD images, the characteristics of the plasma in controlled nuclear fusion experiments can be clearly identified, and the potential causes of plasma disruption during the experiment can be analyzed.

[0047] Magnetic probe timing signals are essentially signals obtained in controlled nuclear fusion experiments by measuring the distribution and properties of the vertical and horizontal magnetic fields at the boundary using magnetic probes. The measured values ​​obtained from the magnetic probes are arranged sequentially according to their acquisition times to form the magnetic probe timing signal. In this specification, the magnetic probe can be a Mirnov Diode Based Probe (MDBP). Magnetic probe timing signals can characterize the spatial distribution information of the magnetic field in the plasma, including its magnitude, shape, and variations. This data is crucial for analyzing plasma stability, motion characteristics, magnetic field topology, and plasma-magnetic field interactions. Therefore, based on the magnetic probe timing signals, the potential causes of plasma disruption during experiments can be analyzed.

[0048] The time-series signal of the geometric center position of the plasma is actually the result of multiple displacements of the plasma in the vertical and horizontal directions during a controlled nuclear fusion experiment. The time-series signal is obtained by arranging the plasma according to the acquisition time. It is used to characterize the position and motion state of the plasma. Detecting and recording the changes in the geometric center position of the plasma can determine the stability, morphological evolution and other dynamic information of the plasma, thereby analyzing the potential causes of plasma breakup in the experiment.

[0049] In practical applications, different types of data acquisition devices can be deployed in controlled nuclear fusion experiments to collect experimental data from different modes. The sampling rates of these different acquisition devices can vary; for example, the sampling rate for CCD images can be 2 ms, for magnetic probes it can be 50 ms, and for plasma geometric center location it can be 1 ms. The specific sampling rate can be flexibly adjusted according to the actual application scenario, and this manual does not impose any limitations on it.

[0050] S102: Input the experimental data into a pre-trained multimodal discriminator to obtain multiple rupture precursor events corresponding to the controlled nuclear fusion experiment, the occurrence confidence of each rupture precursor event, and the occurrence time of each rupture precursor event, as output by the multimodal discriminator.

[0051] In this specification, in order to achieve the comprehensive analysis of the causes of plasma rupture by combining experimental data from multiple modes, a pre-trained multimodal discriminator is used to extract features of different modes from the experimental data of different modes. The features of different modes are then fused. The fused features combine the information of the current controlled nuclear fusion experiment represented by the experimental data of each mode. Thus, the multimodal discriminator can further predict each rupture precursor event, the occurrence time of each rupture precursor event, and the occurrence confidence level based on the fused features.

[0052] Among them, rupture precursor events are early warning signals indicating impending plasma rupture, and precursory signs or phenomena of plasma instability and rupture. When a plasma loses stability and cannot maintain its expected shape and motion, it may undergo a rupture process. Rupture precursor events are not the direct cause of plasma rupture. Rupture precursor events can include plasma configuration-related physical events, current distribution-related physical events, temperature distribution-transport-related physical events, pressure distribution-related physical events, energy storage-related physical events, thermal radiation-related physical events, magnetohydrodynamic (MHD)-related physical events, and motor-power-coil control events, etc. In addition, rupture precursor events can be divided into primary events and secondary events. Primary events contain secondary events. For example, a horizontal position control anomaly during the plasma shaping stage is a primary event, while a horizontal position deviation during the plasma shaping stage is a secondary event. It can be seen that a secondary event is one type of primary event. In addition, the event attributes of the rupture precursor event and the type of observation data corresponding to the event can be determined at the same time. For example, the event attributes of the plasma forming stage horizontal position inward event include plasma horizontal position, preset horizontal position, and plasma inner boundary position, and its observation data includes CCD image and plasma current.

[0053] The multimodal discriminator can also output the occurrence confidence level corresponding to each rupture precursor event. This occurrence confidence level can indicate the probability of the corresponding rupture precursor event occurring in a controlled nuclear fusion experiment. The higher the occurrence confidence level, the greater the probability that the corresponding rupture precursor event will occur in a controlled nuclear fusion experiment. Conversely, the lower the occurrence confidence level, the less likely the corresponding rupture precursor event will occur in a controlled nuclear fusion experiment.

[0054] Additionally, it should be noted that the occurrence time corresponding to each rupture precursor event output by the multimodal discriminator does not necessarily refer to the exact time of the rupture precursor event, but rather to an approximate time. That is, the occurrence time can be a specific moment or a time interval. For example, the occurrence time of the impurity level rise event can be 191ms or 191 to 225ms. In other words, the multimodal discriminator determines that the impurity level rise event may occur at 191ms or within the time interval of 191 to 225ms.

[0055] S104: The precursor events of the controlled nuclear fusion experiment, the occurrence confidence levels of each precursor event, and the occurrence times of each precursor event are input into a pre-trained nuclear fusion model. The model determines the connections between the precursor events and obtains the chain of events corresponding to the controlled nuclear fusion experiment based on these connections. The nuclear fusion model is based on corpus related to the controlled nuclear fusion experiment and a pre-trained natural language model.

[0056] Furthermore, this specification employs a pre-trained large-scale nuclear fusion model. This model is derived from a pre-trained natural language model by incorporating knowledge and expert experience from the field of nuclear fusion and adjusting the model parameters of the pre-trained natural language model. The large-scale nuclear fusion model effectively integrates fundamental physics, fusion physics, and tokamak control-related knowledge. It can provide corresponding physical mechanism analyses based on experimental phenomena. Therefore, based on the input of controlled nuclear fusion experiments and their corresponding prodromal events, confidence levels, and occurrence times, the large-scale nuclear fusion model can analyze the causes of plasma breakup in controlled nuclear fusion experiments.

[0057] It is understood that the pre-trained nuclear fusion large-scale model can be obtained through supervised fine-tuning based on a pre-trained natural language model, or it can be trained through any existing transfer learning method such as Reinforcement Learning from Human Feedback (RLHF). This specification does not limit the specific training method of the nuclear fusion large-scale model. However, since the base model of the nuclear fusion large-scale model is a pre-trained natural language model, the occurrence confidence of each rupture precursor event and the occurrence time of each rupture precursor event, which are inputs to the nuclear fusion large-scale model, are all in natural language. For example, the input of the nuclear fusion large-scale model can be "plasma sweep weak field side aperture barrier event, occurrence confidence is 0.7, occurrence time 191-225ms". However, the connection relationship between the rupture precursor events output by the nuclear fusion large-scale model can be in the form of natural language or in the form of a graph, which is not limited in this specification.

[0058] In this specification, the large-scale nuclear fusion model can analyze the development direction of each rupture precursor event based on the plasma's occurrence at approximately the time of occurrence or within the occurrence period, along with their occurrence confidence levels, and the physics knowledge and expert experience acquired during the training phase. This analysis yields the connectivity relationships between the rupture precursor events. These connectivity relationships characterize the correlation between two interconnected rupture precursor events, and can also indicate the connection direction between two interconnected rupture precursor events, which in turn indicates the development direction of the event between them.

[0059] Furthermore, based on the connectivity between the various rupture precursor events, the rupture precursor events are sequentially connected according to the direction of event development to form a rupture event chain.

[0060] It should be noted that the development direction of each rupture precursor event cannot be analyzed solely based on the temporal dimension to determine the sequence of events; it can also be analyzed based on the physical knowledge learned by the large model. That is, in the rupture event chain, a rupture precursor event that occurs earlier is not necessarily ordered before a rupture precursor event that occurs later.

[0061] For example, the inputs to the large-scale nuclear fusion model are: Event A: Plasma sweeping weak field side aperture barrier, occurrence period 191–225 ms, occurrence confidence 0.7; Event B: Impurity level rise, occurrence period 191–225 ms, occurrence confidence 0.85; Event C: Plasma edge cooling or current peaking small break, occurrence time 225 ms, occurrence confidence 0.5; Event D: Plasma horizontal outward movement small break, occurrence time 45 ms, occurrence confidence 0.6; Event E: Plasma horizontal outward movement small break, occurrence time 267 ms, occurrence confidence 0.67; Event F: Plasma horizontal outward movement small break, occurrence time 304 ms, occurrence confidence 0.88; Event G: CS volt-second depletion, occurrence time 340 ms, occurrence confidence 0.67. The aforementioned events A through G are all precursor events of rupture. Using events A through G as inputs to the large-scale nuclear fusion model, the connection relationships between events A through G are obtained from the model's output. Furthermore, a rupture event chain is constructed based on these connection relationships, such as... Figure 2 As shown, Figure 2 This characterizes the development direction of events A through G and the correlation between each event. Figure 2 The rupture event chain ultimately leads to the plasma rupture event, namely event H: plasma horizontal inward rupture.

[0062] S106: Determine the cause of plasma rupture in the controlled nuclear fusion experiment based on the rupture event chain.

[0063] Therefore, based on the rupture event chain finally derived by S104, the operator can quickly locate and analyze the cause of plasma rupture, and make full use of fragmented knowledge involving many disciplines such as physics, materials, control, and diagnostics, thereby improving the efficiency and accuracy of locating and analyzing the cause of plasma rupture, and providing a highly accurate basis for adjusting the control strategy of the next controlled nuclear fusion experiment.

[0064] The plasma breakup attribution method based on a large-scale model in the field of nuclear fusion provided in this specification determines the breakup precursor events, their occurrence times, and their confidence levels using a multimodal discriminator based on experimental data from controlled nuclear fusion experiments. Then, the connection relationships between each breakup precursor event are determined using a large-scale nuclear fusion model based on the breakup precursor events, their occurrence times, and their confidence levels. This determines the breakup event chain corresponding to the controlled nuclear fusion experiment, and the cause of plasma breakup is determined based on the breakup event chain.

[0065] It is evident that by employing the above scheme, and by fully utilizing the multimodal and fragmented experimental data and expert experience of controlled nuclear fusion through the large nuclear fusion model and multimodal discriminator, the causes of plasma breakup can be analyzed. Based on the causes of plasma breakup, control decisions for the next experiment can be determined, thereby improving the iterative efficiency of controlled nuclear fusion experiments.

[0066] In one or more embodiments described herein, the pre-trained multimodal discriminator used in S102 can process multimodal experimental data. The multimodal discriminator can quantize and encode experimental data from different modalities separately, and fuse the encoded vectors of experimental data from different modalities to utilize the rich information obtained from the fusion to determine information such as precursory events leading to plasma rupture in controlled nuclear fusion experiments. Therefore, the multimodal discriminator includes feature extraction modules for experimental data from multiple different modalities, and prediction modules that predict precursory events based on the features of different modalities output by the multiple feature extraction modules. Figure 3 As shown. Based on the structure of the multimodal discriminator described above, S102 can specifically be:

[0067] Step 1: Input each character contained in the experimental log in the experimental data into the character feature extraction module to obtain the character features of the experimental log.

[0068] In this specification, the experimental logs are natural language data, containing various types of character data. Therefore, the characters contained in the experimental logs can be input into the character feature extraction module of the multimodal discriminator to obtain the character features of the experimental logs. The character feature extraction module can be constructed using any existing structure for character feature extraction, such as word embeddings or the BERT model; this specification does not impose any limitations on it.

[0069] Step 2: Take each target image contained in the target image sequence in the experimental data as input and input it into the image feature extraction module to obtain the image features of the target image sequence.

[0070] Specifically, the target image sequence contains target images (such as CCD images) acquired at different acquisition times. Since the target images are two-dimensional, each target image in the target image sequence is used as input to the image feature extraction module within the multimodal discriminator to obtain image features. The image feature extraction module can be constructed using any existing model structure for image feature extraction, such as convolutional neural networks or feature pyramids; this specification does not limit its use.

[0071] Step 3: Input the magnetic probe timing signal from the experimental data into the magnetic probe timing signal feature extraction module to obtain the first feature corresponding to the magnetic probe timing signal.

[0072] Specifically, the magnetic probe time-series signal feature extraction module can extract features from the magnetic probe time-series signal that describe the spatial distribution changes of the magnetic field in the plasma at different acquisition times. These features are the first features corresponding to the magnetic probe time-series signal. The magnetic probe time-series signal feature extraction module can be constructed using any existing model structure for feature extraction of time-series signals, such as long short-term memory networks or temporal convolutional networks. This specification does not limit its use in this regard.

[0073] Step 4: Take the time-series signal of the plasma geometric center position in the experimental data as input and input it into the plasma geometric center position feature extraction module to obtain the second feature corresponding to the time-series signal of the plasma geometric center position.

[0074] Specifically, the plasma geometric center location feature extraction module can extract features from the time-series signal of the plasma geometric center location to describe the position and motion state of the plasma at different acquisition times. These features are the second features corresponding to the time-series signal of the plasma geometric center location. The plasma geometric center location time-series signal feature extraction module can be constructed using any existing model structure for feature extraction of time-series signals, such as Long Short-Term Memory Networks or Temporal Convolutional Networks. This specification does not limit its use in this regard.

[0075] Additionally, it should be noted that the aforementioned character feature extraction module, image feature extraction module, magnetic probe time-series signal feature extraction module, and plasma geometric center position feature extraction module can be pre-trained separately or jointly trained with the prediction module; this specification does not impose any limitations on this.

[0076] Step 5: Fuse the character features, the image features, the first feature, and the second feature to obtain the fused features.

[0077] The features output by the feature extraction modules of different modalities are fused to improve feature representation capability. The feature fusion method can be any existing type of feature fusion method, such as concatenation, addition, multiplication, weighted summation, etc., and this specification does not limit it.

[0078] Step 6: Input the fusion features into the prediction module to obtain the multiple rupture precursor events corresponding to the controlled nuclear fusion experiment, the occurrence confidence of each rupture precursor event, and the occurrence time of each rupture precursor event, as output by the prediction module.

[0079] In one or more embodiments of this specification, the large-scale nuclear fusion model used in S104 can be pre-trained using the following scheme, such as... Figure 4 As shown, specifically:

[0080] S200: Pre-acquire all question-answer pairs related to the controlled nuclear fusion experiment as the first target corpus, and adjust the model parameters of the pre-trained natural language model according to the first target corpus to obtain the intermediate model.

[0081] In this specification, the RLHF training method can be used to train a large-scale nuclear fusion model adapted to the field of nuclear fusion based on a pre-trained natural language model. The RLHF training method includes: supervised fine-tuning of the pre-trained natural language model, obtaining the reward model after training, and retraining the fine-tuned model using reinforcement learning based on the reward model and the fine-tuned model.

[0082] Based on the training concept of the RLHF training method described above, in this step, the pre-trained natural language model is first trained with supervised fine-tuning using the first target corpus as training samples. The first target corpus includes question-answer pairs related to controlled nuclear fusion experiments. These question-answer pairs are stored in a question-and-answer format, meaning there is a binding relationship between a question and an answer.

[0083] Specifically, the questions contained in each question-answer pair in the first target corpus are used as training samples, and the answers contained in each question-answer pair are used as annotations for the training samples.

[0084] Based on the training samples and their annotations, the model parameters of the pre-trained natural language model are adjusted to obtain an intermediate model. The training method is supervised training; that is, training samples are input into the pre-trained natural language model to obtain the predicted answers corresponding to the training samples output by the natural language model. Then, the model parameters are adjusted with the optimization direction of minimizing the difference between the predicted answers and the annotations. After multiple iterations, a fully trained intermediate model is obtained.

[0085] S202: Obtain the completed training reward model.

[0086] Specifically, the reward model can be viewed as a scoring model used to score and rank the answers output by intermediate models during subsequent reinforcement learning. The reward model can be obtained based on pre-set rules, or it can be trained in a supervised manner based on pre-constructed training samples and their corresponding scores.

[0087] S204: Obtain the target question related to the controlled nuclear fusion experiment and the target answer corresponding to the target question, and concatenate the target question and the target answer to obtain the target text.

[0088] The target questions related to controlled nuclear fusion experiments can be obtained from relevant reference documents and operational logs. These can be obtained through rule-based text extraction, word segmentation, length-based text segmentation, or manual annotation, etc., and this manual does not limit the specific methods used. The target answers to the target questions can be obtained through human answers, extraction and generation from reference documents using a pre-trained natural language model, or retrieval from a knowledge base in the field of nuclear fusion, etc., and this manual does not limit the specific methods used either.

[0089] In this specification, the target texts involved in this step may contain different or the same target questions, but generally, the target answers contained in different target texts are usually different.

[0090] S206: Input the target text into the reward model to obtain the reward value corresponding to the target text output by the reward model.

[0091] The input to the trained reward model can be target text obtained by concatenating a target question and a target answer related to controlled nuclear fusion experiments. The output can be the target answer contained in the target text, and its accuracy and credibility relative to the target question contained in the target text. Therefore, the input to the reward model is text in natural language form, and the output is a numerical value between [m, n]. The values ​​of m and n can be adjusted according to the specific application scenario, and this specification does not impose any restrictions on them.

[0092] S208: Based on the target text and the reward value corresponding to the target text, the model parameters of the intermediate model are adjusted using a reinforcement learning training method to obtain the trained nuclear fusion large model.

[0093] The training method using RLHF actually defines the state space based on the target question in the target text, and the action space based on the target answer in the target text. The reward value comes from the reward value corresponding to the target text output by the reward model in S206. This reward value is actually used to characterize the accuracy of the target answer in the target text for the target question.

[0094] The intermediate model can be trained using conventional reinforcement learning methods to adjust its parameters, or it can be trained using the Proximal Policy Optimization (PPO) algorithm. This manual does not specify which method is preferred.

[0095] In addition, in an optional embodiment of this specification, to further improve the performance of the large-scale nuclear fusion model, the training process of RLHF can be performed iteratively, as follows:

[0096] In the first iteration of training, a pre-trained natural language model is used as the base model. An intermediate model is obtained based on the first target corpus and the base model through the above S200. The reward value corresponding to the target text is obtained according to the trained reward model. Then, based on the target text and the reward value corresponding to the target text, the model parameters of the intermediate model are adjusted by the reinforcement learning training method to obtain the nuclear fusion large model.

[0097] In subsequent iterative training, it is first determined whether the nuclear fusion large model obtained from the previous iteration meets the preset conditions. These preset conditions may include, for example, that the test results obtained from the test samples meet a preset threshold. If the nuclear fusion large model obtained from the previous iteration does not meet the preset conditions, then the nuclear fusion large model obtained from the previous iteration is used as the base model. An intermediate model is then obtained based on the first target corpus and the base model. The reward value corresponding to the target text is then obtained again based on the reward model. Finally, based on the target text and the reward value corresponding to the target text, the model parameters of the intermediate model are adjusted using reinforcement learning training methods to obtain the nuclear fusion large model. This process continues until the nuclear fusion large model obtained from the iterative training meets the preset conditions.

[0098] In practical applications, after the large-scale nuclear fusion model is launched, it can be fine-tuned again based on online feedback so that it can adapt to specific application scenarios.

[0099] In an optional embodiment of this specification, to further enhance the understanding of nuclear fusion field expertise by the large-scale nuclear fusion model, the intermediate model obtained by S200 based on the first target corpus may also be based on the following scheme:

[0100] First, a second target corpus related to controlled nuclear fusion experiments is acquired, and the model parameters of a pre-trained natural language model are adjusted based on this second target corpus to obtain candidate models. The second target corpus includes at least natural language data and formula / code data.

[0101] In this specification, candidate models are obtained by fine-tuning the model parameters of a pre-trained natural language model based on a second target corpus. The purpose is to adapt the model to be fine-tuned to the nuclear fusion field before supervised fine-tuning training, thereby improving the efficiency and performance of supervised fine-tuning training.

[0102] The difference between the second and first target corpora lies in the fact that the second target corpus is a high-quality declarative knowledge corpus in the field of nuclear fusion, used to enrich the vertical domain knowledge of the candidate models obtained during training. The first target corpus, on the other hand, consists of question-and-answer pairs related to nuclear fusion experiments.

[0103] Furthermore, because nuclear fusion is an interdisciplinary field involving physics, materials science, control, diagnostics, and many other disciplines, it suffers from fragmented professional knowledge stored in various reference documents and near-reliable operational logs. Additionally, the extracted corpus contains formulas. If formulas are simply treated as natural language, the underlying physical knowledge and mechanisms they embody cannot be fully identified. Therefore, when formulas related to controlled nuclear fusion experiments are obtained, a pre-configured formula code conversion tool can be used to convert the identified formulas into formula code in a specified programming language (such as Mathematica). This formula code can then represent the physical knowledge and mechanisms underlying the formulas.

[0104] Therefore, the second target corpus includes natural language corpus and formula code corpus, and of course there are other types of corpus, such as algorithm code corpus.

[0105] In this step, the pre-trained natural language model is trained and its parameters are adjusted using an autoregressive approach, using the first target corpus as a sample, to obtain a candidate model.

[0106] Optionally, obtaining the second target corpus related to controlled nuclear fusion experiments can specifically include:

[0107] Step 1: Obtain reference documents related to the controlled nuclear fusion experiment and the reference operation log of the controlled nuclear fusion experiment in advance.

[0108] Reference documents can be text files such as textbooks and journal articles related to the field of nuclear fusion. These include fundamental knowledge of nuclear fusion (basic principles, reaction pathways, energy release, etc.), plasma physics (properties, stability, magnetic confinement, etc. of plasma), fusion device technology (design, operating parameters, magnetic field configuration, etc. of different types of fusion devices (such as tokamak, spherical inertial confinement fusion, etc.), and plasma diagnostics (introduction to various diagnostic methods and instruments used to measure plasma properties, temperature, density, fluid dynamics, etc.). The reference operation log for controlled nuclear fusion experiments can be the operation log of multiple completed controlled nuclear fusion experiments. The operation log includes the operations performed by the operator during the controlled nuclear fusion experiment and the parameters of the experiment. For example: at 50ms, reduce the voltage VPF5 of the No. 5 superconducting magnet (Poloidal Field, PF) coil, and set V... LIMIT =600; Increase VPF5 at 30ms and decrease VPF7 at 50ms. I P The ascent went well, and the discharge was successful.

[0109] Step 2: Convert the reference file and the reference operation log into candidate characters.

[0110] Text conversion technology is used to convert the text in reference documents and reference operation logs into candidate characters. Text conversion can be done using Optical Character Recognition (OCR) technology.

[0111] Step 3: Delete the candidate characters that contain characters of a specified type, and segment the remaining candidate characters according to the specified character length to obtain multiple character fragments as the second target corpus related to controlled nuclear fusion experiments.

[0112] In this step, the extracted candidate Zefu data undergoes data cleaning and filtering. The step of deleting candidates containing characters of a specified type includes characters unrelated to the field of nuclear fusion, such as those appearing in the table of contents, title, author, preface, etc. Filtering includes deleting sentences shorter than a specified length, or characters containing only numbers and punctuation (typically formula numbers).

[0113] Additionally, the remaining candidate characters need to be segmented into multiple character fragments according to a specified character length. This is because the input to the pre-trained natural language model is limited by character length, and segmentation according to the specified length of the pre-trained natural language model is necessary to ensure the semantic integrity and coherence of the sample fragments. For example, using Qwen-7B-chat as the base, the maximum length is 8k, so the length of each character fragment after segmentation is less than 8k, and it can be divided according to chapters (textbooks) or natural paragraphs (papers). For the Mathematica code of the formulas, it is also necessary to ensure that the natural language context of the formulas exists.

[0114] Therefore, the model parameters of the candidate model are adjusted according to the first target corpus to obtain an intermediate model.

[0115] In an optional embodiment of this specification, S202 may be implemented according to the following scheme, such as... Figure 5 As shown:

[0116] S300: Obtain a reference question related to the controlled nuclear fusion experiment, and obtain multiple candidate answers corresponding to the reference question based on the reference question and the intermediate model.

[0117] In this specification, the reward model is obtained through supervised training. The training samples used to train the reward model include reference problems related to controlled nuclear fusion experiments. The method for obtaining the reference problems is similar to that for obtaining the aforementioned target problems. Alternatively, the aforementioned target problems can be directly used as reference problems. This specification does not restrict whether the target problems and reference problems are the same.

[0118] Since the reward model evaluates whether an answer is accurate, it's necessary to obtain candidate answers for the reference question during training. In this specification, to expand the sample size, a scheme is adopted to obtain multiple candidate answers for each reference question. One way to obtain candidate answers is to input the reference question multiple times into the trained intermediate model, obtaining multiple candidate answers output by the intermediate model each time.

[0119] S302: Sort the candidate answers corresponding to the reference question, and determine the reference indicators corresponding to each candidate answer according to the order of the candidate answers.

[0120] In S300, multiple candidate answers were obtained for each reference question. Although nuclear fusion is an interdisciplinary field, multiple candidate answers to the same reference question from different perspectives may all be correct, but their accuracy and reliability vary. Therefore, in this step, the candidate answers are ranked based on their accuracy in answering the reference question. A score is then assigned to each candidate answer based on its order; this score serves as the reference index. A higher reference index indicates a more accurate and reliable answer, while a lower reference index indicates a less accurate and questionable answer.

[0121] S304: The reference question and each candidate answer are used as inputs to the reward model to be trained, and the prediction index corresponding to each candidate answer is output by the reward model.

[0122] In this specification, the training sample of the reward model is a concatenation of the reference question and the candidate answer to the reference question. That is, the reference question and one of the multiple candidate answers are concatenated to obtain a concatenated text as a training sample. The annotation of this training sample is the reference index corresponding to the candidate answer obtained in S302 above.

[0123] The reward model to be trained can predict the accuracy of a candidate answer in answering the reference question based on the input reference question and the candidate answers to the reference question, and give a prediction index corresponding to the candidate answer. This prediction index is used to indicate the accuracy of the prediction of the candidate answer in answering the reference question by the reward model.

[0124] S306: Based on the prediction index corresponding to each candidate answer and the reference index corresponding to each candidate answer, train the reward model to be trained to obtain the trained reward model.

[0125] The reward model is trained with the goal of minimizing the difference between the predicted and reference metrics of the candidate answers. The trained reward model is obtained after iterative training.

[0126] The above are one or more embodiments of the plasma breakage attribution method based on a large-scale model in the field of nuclear fusion provided in this specification. Based on the same idea, this specification also provides a corresponding plasma breakage attribution device based on a large-scale model in the field of nuclear fusion, such as... Figure 6 As shown.

[0127] Figure 6 This specification provides a schematic diagram of a plasma fracture attribution device based on a large-scale model in the field of nuclear fusion, specifically including:

[0128] The experimental data acquisition module 400 is used to acquire experimental data from controlled nuclear fusion experiments. The experimental data includes experimental logs, target image sequences, magnetic probe timing signals, and timing signals of the plasma geometric center position.

[0129] The multimodal discrimination module 402 is used to input the experimental data into a pre-trained multimodal discriminator to obtain multiple rupture precursor events corresponding to the controlled nuclear fusion experiment, the occurrence confidence of each rupture precursor event, and the occurrence time of each rupture precursor event, as output by the multimodal discriminator.

[0130] The rupture event chain determination module 404 is used to input each rupture precursor event corresponding to the controlled nuclear fusion experiment, the occurrence confidence of each rupture precursor event, and the occurrence time of each rupture precursor event as inputs into a pre-trained nuclear fusion large model, determine the connection relationship between each rupture precursor event through the nuclear fusion large model, and obtain the rupture event chain corresponding to the controlled nuclear fusion experiment based on the connection relationship;

[0131] The rupture cause determination module 406 is used to determine the cause of plasma rupture in the controlled nuclear fusion experiment based on the rupture event chain.

[0132] Optionally, the pre-trained multimodal discriminator includes a character feature extraction module, an image feature extraction module, a magnetic probe time-series signal feature extraction module, a plasma geometric center position feature extraction module, and a prediction module;

[0133] Optionally, the multimodal discrimination module 402 is specifically used to: input each character contained in the experimental log in the experimental data into the character feature extraction module to obtain the character features of the experimental log; input each target image contained in the target image sequence in the experimental data into the image feature extraction module to obtain the image features of the target image sequence; input the magnetic probe timing signal in the experimental data into the magnetic probe timing signal feature extraction module to obtain the first feature corresponding to the magnetic probe timing signal; input the timing signal of the plasma geometric center position in the experimental data into the plasma geometric center position feature extraction module to obtain the second feature corresponding to the timing signal of the plasma geometric center position; fuse the character features, the image features, the first feature, and the second feature to obtain a fused feature; input the fused feature into the prediction module to obtain the multiple rupture precursor events corresponding to the controlled nuclear fusion experiment, the occurrence confidence of each rupture precursor event, and the occurrence time of each rupture precursor event, as output by the prediction module.

[0134] Optionally, the device further includes:

[0135] The training module 408 is specifically used to pre-acquire question-answer pairs related to controlled nuclear fusion experiments as the first target corpus, and adjust the model parameters of the pre-trained natural language model according to the first target corpus to obtain an intermediate model; acquire the trained reward model; acquire the target questions related to controlled nuclear fusion experiments and the target answers corresponding to the target questions, concatenate the target questions and the target answers to obtain the target text; input the target text into the reward model to obtain the reward value corresponding to the target text output by the reward model; and adjust the model parameters of the intermediate model using a reinforcement learning training method according to the target text and the reward value corresponding to the target text to obtain the trained nuclear fusion large model.

[0136] Optionally, the training module 408 is further configured to acquire a second target corpus related to controlled nuclear fusion experiments, and adjust the model parameters of the pre-trained natural language model according to the second target corpus to obtain a candidate model; wherein the second target corpus includes at least natural language type corpus and formula code type corpus;

[0137] Optionally, the training module 408 is specifically used to adjust the model parameters of the candidate model according to the first target corpus to obtain an intermediate model.

[0138] Optionally, the training module 408 is specifically used to: pre-acquire reference documents related to the controlled nuclear fusion experiment and reference operation logs of the controlled nuclear fusion experiment; convert the reference documents and reference operation logs into candidate characters; delete the candidate characters containing characters of a specified type, and segment the remaining candidate characters according to a specified character length to obtain multiple character fragments as a second target corpus related to the controlled nuclear fusion experiment.

[0139] Optionally, the training module 408 is specifically used to: use the questions contained in each question-answer pair in the first target corpus as training samples, and use the answers contained in each question-answer pair as annotations of the training samples; adjust the model parameters of the pre-trained natural language model according to the training samples and the annotations of the training samples to obtain an intermediate model.

[0140] Optionally, the training module 408 is specifically configured to: acquire a reference question related to controlled nuclear fusion experiments; obtain multiple candidate answers corresponding to the reference question based on the reference question and the intermediate model; sort the candidate answers corresponding to the reference question and determine the reference indicators corresponding to each candidate answer based on the order of the candidate answers; input the reference question and the candidate answers as inputs into the reward model to be trained to obtain the predicted indicators corresponding to each candidate answer output by the reward model; and train the reward model to be trained based on the predicted indicators corresponding to each candidate answer and the reference indicators corresponding to each candidate answer to obtain the trained reward model.

[0141] This specification also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described... Figure 1 The plasma breakup attribution method shown is based on a large model in the field of nuclear fusion.

[0142] This instruction manual also provides Figure 7 The diagram shows a schematic structural representation of the electronic device. Figure 7 At the hardware level, the electronic device includes a processor, internal bus, network interface, memory, and non-volatile memory, and may also include other hardware required for the business operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then runs it to achieve the above-mentioned functions. Figure 1 The method shown is a plasma breakup attribution method based on a large-scale model in the field of nuclear fusion. Of course, in addition to software implementation, this specification does not exclude other implementation methods, such as logic devices or a combination of hardware and software, etc. That is to say, the execution subject of the following processing flow is not limited to individual logic units, but can also be hardware or logic devices.

[0143] In the 1990s, improvements to a technology could be clearly distinguished as either hardware improvements (e.g., improvements to the circuit structure of diodes, transistors, switches, etc.) or software improvements (improvements to the methodology). However, with technological advancements, many methodological improvements today can be considered direct improvements to the hardware circuit structure. Designers almost always obtain the corresponding hardware circuit structure by programming the improved methodology into the hardware circuit. Therefore, it cannot be said that a methodological improvement cannot be implemented using hardware physical modules. For example, a Programmable Logic Device (PLD) (such as a Field Programmable Gate Array (FPGA)) is such an integrated circuit whose logic function is determined by the user programming the device. Designers can program and "integrate" a digital system onto a PLD themselves, without needing chip manufacturers to design and manufacture dedicated integrated circuit chips. Furthermore, nowadays, instead of manually manufacturing integrated circuit chips, this programming is mostly implemented using "logic compiler" software. Similar to the software compiler used in program development, the original code before compilation must be written in a specific programming language, called a Hardware Description Language (HDL). There are many HDLs, such as ABEL (Advanced Boolean Expression Language), AHDL (Altera Hardware Description Language), Confluence, CUPL (Cornell University Programming Language), HDCal, JHDL (Java Hardware Description Language), Lava, Lola, MyHDL, PALASM, and RHDL (Ruby Hardware Description Language). Currently, the most commonly used are VHDL (Very-High-Speed ​​Integrated Circuit Hardware Description Language) and Verilog. Those skilled in the art should understand that by simply performing some logic programming on the method flow using one of these hardware description languages ​​and programming it into an integrated circuit, the hardware circuit implementing the logical method flow can be easily obtained.

[0144] The controller can be implemented in any suitable manner. For example, it can take the form of a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) executable by the (micro)processor, logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers, and embedded microcontrollers. Examples of controllers include, but are not limited to, the following microcontrollers: ARC 625D, Atmel AT91SAM, Microchip PIC18F26K20, and Silicon Labs C8051F320. A memory controller can also be implemented as part of the control logic of the memory. Those skilled in the art will also recognize that, in addition to implementing the controller in purely computer-readable program code form, the same functionality can be achieved by logically programming the method steps to make the controller take the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, such a controller can be considered a hardware component, and the means included therein for implementing various functions can also be considered as structures within the hardware component. Alternatively, the means for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0145] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, a computer can be, for example, a personal computer, laptop computer, cellular phone, camera phone, smartphone, personal digital assistant, media player, navigation device, email device, game console, tablet computer, wearable device, or any combination of these devices.

[0146] For ease of description, the above devices are described in terms of function, divided into various units. Of course, in implementing this specification, the functions of each unit can be implemented in one or more software and / or hardware components.

[0147] Those skilled in the art will understand that embodiments of this specification can be provided as methods, systems, or computer program products. Therefore, this specification may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this specification may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0148] This specification is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this specification. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0149] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0150] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0151] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0152] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0153] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0154] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0155] Those skilled in the art will understand that the embodiments of this specification can be provided as methods, systems, or computer program products. Therefore, this specification may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this specification may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0156] This specification can be described in the general context of computer-executable instructions that are executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This specification can also be practiced in distributed computing environments, where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0157] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0158] The above description is merely an embodiment of this specification and is not intended to limit this specification. Various modifications and variations can be made to this specification by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this specification should be included within the scope of the claims of this specification.

Claims

1. A plasma breakup attribution method based on a large-scale model in the field of nuclear fusion, characterized in that, include: The experimental data of the controlled nuclear fusion experiment are obtained, including experimental logs, target image sequences, magnetic probe timing signals, and timing signals of the plasma geometric center position. The experimental data is input into a pre-trained multimodal discriminator, which, based on the character features of the experimental log, the image features of the target image sequence, the first feature of the spatial distribution of the magnetic field in the plasma as a function of time, characterized by the time-series signal of the magnetic probe, and the second feature of the position and motion state of the plasma as a function of time, characterized by the time-series signal of the geometric center of the plasma, obtains multiple rupture precursor events corresponding to the controlled nuclear fusion experiment, the occurrence confidence of each rupture precursor event, and the occurrence time of each rupture precursor event. The precursor events of the controlled nuclear fusion experiment, the occurrence confidence levels of each precursor event, and the occurrence times of each precursor event are input into a pre-trained large-scale nuclear fusion model. The large-scale nuclear fusion model determines the connection relationships between the precursor events and obtains the chain of precursor events corresponding to the controlled nuclear fusion experiment based on the connection relationships. The connection relationships are used to characterize the correlation between two interconnected precursor events, and the connection relationships include a connection direction to indicate the event development direction between the two precursor events. The cause of plasma rupture in the controlled nuclear fusion experiment was determined based on the rupture event chain. The large-scale nuclear fusion model is based on corpus related to controlled nuclear fusion experiments and a pre-trained natural language model, and is obtained through the following training process: Acquire declarative knowledge corpus related to controlled nuclear fusion experiments, and fine-tune the model parameters of a pre-trained natural language model based on the declarative knowledge corpus to obtain candidate models adapted to the field of nuclear fusion. Acquire all question-answer pairs related to controlled nuclear fusion experiments as the first target corpus, and perform supervised training on the candidate model based on the first target corpus to obtain an intermediate model; Obtain target text related to controlled nuclear fusion experiments and the corresponding reward value of the target text, and perform reinforcement learning training on the intermediate model based on the target text and the corresponding reward value of the target text to obtain a trained nuclear fusion large model. The target text includes target questions related to controlled nuclear fusion experiments and target answers corresponding to the target questions. The reward value corresponding to the target text is used to characterize the accuracy of the target answer in the target text for the target question.

2. The method as described in claim 1, characterized in that, The pre-trained multimodal discriminator includes a character feature extraction module, an image feature extraction module, a magnetic probe time-series signal feature extraction module, a plasma geometric center position feature extraction module, and a prediction module; The experimental data is input into a pre-trained multimodal discriminator, which, based on the character features of the experimental log, the image features of the target image sequence, the first feature of the spatial distribution of the magnetic field in the plasma over time as represented by the magnetic probe time-series signal, and the second feature of the position and motion state of the plasma over time as represented by the time-series signal of the plasma geometric center position, obtains multiple prodromal events corresponding to the controlled nuclear fusion experiment, the occurrence confidence of each prodromal event, and the occurrence time of each prodromal event. Specifically, this includes: Input each character contained in the experimental log in the experimental data into the character feature extraction module to obtain the character features of the experimental log; The target images contained in the target image sequence in the experimental data are used as inputs to the image feature extraction module to obtain the image features of the target image sequence. The magnetic probe timing signal in the experimental data is input into the magnetic probe timing signal feature extraction module to obtain the first feature corresponding to the magnetic probe timing signal. The time-series signal of the plasma geometric center position in the experimental data is used as input and input into the plasma geometric center position feature extraction module to obtain the second feature corresponding to the time-series signal of the plasma geometric center position; The character features, the image features, the first feature, and the second feature are fused to obtain the fused feature; The fusion features are input into the prediction module to obtain multiple rupture precursor events corresponding to the controlled nuclear fusion experiment, the occurrence confidence of each rupture precursor event, and the occurrence time of each rupture precursor event, as output by the prediction module.

3. The method as described in claim 1, characterized in that, The acquisition of target text related to controlled nuclear fusion experiments and the corresponding reward value for the target text specifically includes: Obtain the target questions related to controlled nuclear fusion experiments and the corresponding target answers, and concatenate the target questions and the answers to obtain the target text; The target text is input into the reward model to obtain the reward value corresponding to the target text output by the reward model.

4. The method as described in claim 1, characterized in that, The declarative knowledge corpus includes at least natural language corpus and formula code corpus.

5. The method as described in claim 1, characterized in that, The acquisition of declarative knowledge corpus related to controlled nuclear fusion experiments specifically includes: Pre-acquire reference documents related to the controlled nuclear fusion experiment and the reference operation log of the controlled nuclear fusion experiment; Convert the reference file and the reference operation log into candidate characters; The candidate characters are deleted if they contain characters of a specified type, and the remaining candidate characters are segmented according to a specified character length to obtain multiple character fragments as declarative knowledge corpus related to controlled nuclear fusion experiments.

6. The method as described in claim 1, characterized in that, The step of supervising the candidate model based on the first target corpus to obtain an intermediate model specifically includes: The questions contained in each question-answer pair in the first target corpus are used as training samples, and the answers contained in each question-answer pair are used as annotations for the training samples. Based on the training samples and their annotations, the model parameters of the candidate model are adjusted to obtain an intermediate model.

7. The method as described in claim 3, characterized in that, The reward model was obtained through training as follows: To obtain reference questions related to controlled nuclear fusion experiments, Based on the reference question and the intermediate model, multiple candidate answers corresponding to the reference question are obtained; Sort the candidate answers to the reference questions and determine the reference indicators corresponding to each candidate answer according to the order of the candidate answers. The reference question and each candidate answer are used as inputs to the reward model to be trained, and the prediction index corresponding to each candidate answer is obtained by the reward model output. The reward model to be trained is trained based on the prediction index corresponding to each candidate answer and the reference index corresponding to each candidate answer, so as to obtain the trained reward model.

8. A plasma fracture attribution device based on a large-scale model in the field of nuclear fusion, characterized in that, include: The experimental data acquisition module is used to acquire experimental data from controlled nuclear fusion experiments. The experimental data includes experimental logs, target image sequences, magnetic probe timing signals, and timing signals of the plasma geometric center position. A multimodal discrimination module is used to input the experimental data into a pre-trained multimodal discriminator, so that the multimodal discriminator can obtain multiple rupture precursor events corresponding to the controlled nuclear fusion experiment, the occurrence confidence of each rupture precursor event, and the occurrence time of each rupture precursor event based on the character features of the experimental log, the image features of the target image sequence, the first feature of the spatial distribution of the magnetic field in the plasma as a function of time represented by the magnetic probe time series signal, and the second feature of the position and motion state of the plasma as a function of time represented by the time series signal of the geometric center position of the plasma. The rupture event chain determination module is used to input the rupture precursor events corresponding to the controlled nuclear fusion experiment, the occurrence confidence levels of each rupture precursor event, and the occurrence times of each rupture precursor event as inputs into a pre-trained large-scale nuclear fusion model. The large-scale nuclear fusion model determines the connection relationships between the rupture precursor events and obtains the rupture event chain corresponding to the controlled nuclear fusion experiment based on these connection relationships. The connection relationships characterize the correlation between two interconnected rupture precursor events, and include a connection direction to indicate the event development direction between the two rupture precursor events. The rupture cause determination module is used to determine the cause of plasma rupture in the controlled nuclear fusion experiment based on the rupture event chain. The large-scale nuclear fusion model is based on corpus related to controlled nuclear fusion experiments and a pre-trained natural language model, and is obtained through the following training process: Acquire declarative knowledge corpus related to controlled nuclear fusion experiments, and fine-tune the model parameters of a pre-trained natural language model based on the declarative knowledge corpus to obtain candidate models adapted to the field of nuclear fusion. Acquire all question-answer pairs related to controlled nuclear fusion experiments as the first target corpus, and perform supervised training on the candidate model based on the first target corpus to obtain an intermediate model; Obtain target text related to controlled nuclear fusion experiments and the corresponding reward value of the target text, and perform reinforcement learning training on the intermediate model based on the target text and the corresponding reward value of the target text to obtain a trained nuclear fusion large model. The target text includes target questions related to controlled nuclear fusion experiments and target answers corresponding to the target questions. The reward value corresponding to the target text is used to characterize the accuracy of the target answer in the target text for the target question.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the method described in any one of claims 1 to 7.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method described in any one of claims 1 to 7.

Images