Method, device and storage medium for reconstructing physical field of three-dimensional interstellar medium

Abstract

The application relates to a three-dimensional interstellar medium physical field reconstruction method, device and storage medium, wherein the three-dimensional interstellar medium physical field reconstruction method comprises the following steps: acquiring three-dimensional interstellar medium physical field simulation data, inputting the three-dimensional interstellar medium physical field simulation data into a three-dimensional photochemical radiation model, generating a synthetic observation data set and a true value three-dimensional interstellar medium physical field data set, and constructing a training sample; inputting the training sample into an initial deep learning architecture to obtain a three-dimensional interstellar medium physical field prediction intermediate result; calculating a loss based on the difference between the result and the true value three-dimensional interstellar medium physical field data set, iteratively training the initial deep learning architecture through gradient back propagation, and obtaining a final inversion deep learning architecture; and inputting real observation data into the deep learning architecture to obtain a three-dimensional interstellar medium physical field reconstruction result. Through the application, the problem that in the related art, it is difficult to accurately restore observation data to the complex physical state of a real three-dimensional interstellar medium is solved.

Description

Technical Field

[0001] This application relates to the field of astrophysical data processing, and in particular to methods, apparatus and storage media for reconstructing the physical fields of three-dimensional interstellar media. Background Technology

[0002] The density, temperature, and chemical structure of the interstellar medium are crucial for understanding molecular cloud evolution and star formation mechanisms, and its three-dimensional structure reconstruction is of paramount importance to astrophysical research. However, current techniques that use simple empirical models to fit data struggle to accurately reconstruct the complex physical states of the real three-dimensional interstellar medium from observational data.

[0003] Currently, no effective solution has been proposed to address the problem of accurately reconstructing observational data to the complex physical state of the real three-dimensional interstellar medium in related technologies. Summary of the Invention

[0004] This application provides a method, apparatus, and storage medium for reconstructing the physical field of a three-dimensional interstellar medium, in order to at least solve the problem in related technologies that it is difficult to accurately restore observation data to the complex physical state of the real three-dimensional interstellar medium.

[0005] In a first aspect, embodiments of this application provide a method for reconstructing the physical field of a three-dimensional interstellar medium, the method comprising:

[0006] Acquire three-dimensional interstellar medium physical field simulation data;

[0007] The simulated data of the three-dimensional interstellar medium physical field is input into a preset three-dimensional photochemical radiation model to generate a synthetic observation dataset and a true three-dimensional interstellar medium physical field dataset. The true three-dimensional interstellar medium physical field dataset is formed by further calculation and supplementation based on the simulated data of the three-dimensional interstellar medium physical field through the three-dimensional photochemical radiation model.

[0008] Training samples are constructed based on the synthetic observation dataset and the true three-dimensional interstellar medium physical field dataset;

[0009] The training samples are input into the initial three-dimensional interstellar medium physics field inversion deep learning architecture to obtain intermediate prediction results of the three-dimensional interstellar medium physics field; the loss result is calculated based on the difference between the intermediate prediction results of the three-dimensional interstellar medium physics field and the ground truth three-dimensional interstellar medium physics field dataset.

[0010] The gradient of the loss result is backpropagated to the initial three-dimensional interstellar medium physics field inversion deep learning architecture for iterative training, to obtain the three-dimensional interstellar medium physics field inversion deep learning architecture.

[0011] The acquired real observation data is input into the three-dimensional interstellar medium physics field inversion deep learning architecture to obtain the three-dimensional interstellar medium physics field reconstruction results.

[0012] In some embodiments, the calculation of the loss result based on the difference between the intermediate predicted results of the three-dimensional interstellar medium physics field and the ground truth three-dimensional interstellar medium physics field dataset includes:

[0013] Based on the difference between the intermediate results of the predicted three-dimensional interstellar medium physical field and the true three-dimensional interstellar medium physical field dataset, a first loss is constructed.

[0014] The intermediate results of the three-dimensional interstellar medium physical field prediction are input into a preset fast rendering module to obtain the synthetic observation prediction intermediate results; the fast rendering module is the approximate calculation module of the three-dimensional photochemical radiation model.

[0015] A second loss is constructed based on the difference between the intermediate results of the synthetic observation prediction and the synthetic observation dataset.

[0016] The loss result is obtained based on the first loss and the second loss.

[0017] In some embodiments, obtaining the loss result based on the first loss and the second loss includes:

[0018] Based on the intermediate results of the three-dimensional interstellar medium physical field prediction and the preset physical constraints, a third loss is constructed;

[0019] The loss result is obtained based on the first loss, the second loss, and the third loss.

[0020] In some embodiments, acquiring three-dimensional interstellar medium physical field simulation data includes:

[0021] Obtain raw three-dimensional interstellar medium physical field simulation data;

[0022] The original three-dimensional interstellar medium physical field simulation data is resampled to a regular cubic mesh to obtain the three-dimensional interstellar medium physical field simulation data.

[0023] In some embodiments, constructing training samples based on the synthetic observation dataset and the ground truth three-dimensional interstellar medium physics dataset includes:

[0024] Based on the synthetic observation dataset and the true three-dimensional interstellar medium physical field dataset, original training samples are constructed;

[0025] The original training samples are cropped and augmented from multiple perspectives, and then stored in a preset format to obtain the training samples.

[0026] In some embodiments, after obtaining the three-dimensional interstellar medium physical field reconstruction result, the method further includes:

[0027] The reconstructed physical field of the three-dimensional interstellar medium is input into a preset radiation transfer module to obtain the synthetic observation results;

[0028] The prediction effect is verified based on the synthetic observation results and the real observation data.

[0029] In some embodiments, verifying the prediction effect based on the synthetic observation results and the real observation data includes:

[0030] The synthetic observation results and the real observation data are quantitatively compared to obtain the prediction effect; the quantitative comparison includes line intensity error, spectral line type comparison, image structure comparison and statistical consistency.

[0031] In some embodiments, the true three-dimensional interstellar medium physics dataset includes at least a three-dimensional density distribution.

[0032] Secondly, embodiments of this application provide a three-dimensional interstellar medium physical field reconstruction device, the device comprising:

[0033] The three-dimensional data acquisition module is used to acquire three-dimensional interstellar medium physical field simulation data;

[0034] The dataset generation module is used to input the three-dimensional interstellar medium physical field simulation data into a preset three-dimensional photochemical radiation model to generate a synthetic observation dataset and a true three-dimensional interstellar medium physical field dataset; the true three-dimensional interstellar medium physical field dataset is formed based on the three-dimensional interstellar medium physical field simulation data and further calculated and supplemented by the three-dimensional photochemical radiation model.

[0035] The training sample construction module is used to construct training samples based on the synthetic observation dataset and the true three-dimensional interstellar medium physical field dataset;

[0036] The loss calculation module is used to input the training samples into the initial three-dimensional interstellar medium physics field inversion deep learning architecture to obtain intermediate prediction results of the three-dimensional interstellar medium physics field; and to calculate the loss result based on the difference between the intermediate prediction results of the three-dimensional interstellar medium physics field and the ground truth three-dimensional interstellar medium physics field dataset.

[0037] The model training module is used to backpropagate the gradient of the loss result to the initial three-dimensional interstellar medium physics field inversion deep learning architecture for iterative training, so as to obtain the three-dimensional interstellar medium physics field inversion deep learning architecture.

[0038] The model inference module is used to input the acquired real observation data into the three-dimensional interstellar medium physical field inversion deep learning architecture to obtain the three-dimensional interstellar medium physical field reconstruction result.

[0039] Thirdly, embodiments of this application provide a storage medium storing a computer program that, when executed by a processor, implements the three-dimensional interstellar medium physics field reconstruction method as described in the first aspect above.

[0040] Compared to related technologies, the three-dimensional interstellar medium physics field reconstruction method, apparatus, and storage medium provided in this application reconstruct the physics field of the interstellar medium by acquiring three-dimensional interstellar medium physics field simulation data; inputting the three-dimensional interstellar medium physics field simulation data into a preset three-dimensional photochemical radiation model to generate a synthetic observation dataset and a true three-dimensional interstellar medium physics field dataset; the true three-dimensional interstellar medium physics field dataset is formed by further calculation and supplementation based on the three-dimensional interstellar medium physics field simulation data and the three-dimensional photochemical radiation model; training samples are constructed based on the synthetic observation dataset and the true three-dimensional interstellar medium physics field dataset; and the training samples are input into the initial three-dimensional interstellar medium physics field model. A deep learning architecture for field inversion is used to obtain intermediate prediction results of the physical field of the three-dimensional interstellar medium. Based on the difference between the intermediate prediction results and the true three-dimensional interstellar medium physical field dataset, a loss result is calculated. The gradient of the loss result is backpropagated to the initial deep learning architecture for three-dimensional interstellar medium physical field inversion for iterative training to obtain the deep learning architecture for three-dimensional interstellar medium physical field inversion. The acquired real observation data is input into the deep learning architecture for three-dimensional interstellar medium physical field inversion to obtain the reconstruction result of the physical field of the three-dimensional interstellar medium. This solves the problem in related technologies that it is difficult to accurately restore the observation data to the complex physical state of the real three-dimensional interstellar medium.

[0041] Details of one or more embodiments of this application are set forth in the following drawings and description to make other features, objects and advantages of this application more readily apparent. Attached Figure Description

[0042] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0043] Figure 1 This is a hardware structure block diagram of a terminal for a three-dimensional interstellar medium physical field reconstruction method according to an embodiment of this application;

[0044] Figure 2 This is a flowchart of a three-dimensional interstellar medium physical field reconstruction method according to an embodiment of this application;

[0045] Figure 3 This is a schematic diagram of the framework of a three-dimensional interstellar medium physical field reconstruction system according to an embodiment of this application;

[0046] Figure 4 This is a structural block diagram of a three-dimensional interstellar medium physics field reconstruction device according to an embodiment of this application. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this application clearer, the application is described and illustrated below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the application. All other embodiments obtained by those skilled in the art based on the embodiments provided in this application without inventive effort are within the scope of protection of this application. Furthermore, it is understood that although the efforts made in such a development process may be complex and lengthy, for those skilled in the art related to the content disclosed in this application, modifications to design, manufacturing, or production based on the technical content disclosed in this application are merely conventional technical means and should not be construed as insufficient disclosure of the content of this application.

[0048] In this application, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment that is mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this application may be combined with other embodiments without conflict.

[0049] Unless otherwise defined, the technical or scientific terms used in this application shall have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms “a,” “an,” “an,” “the,” and similar words used in this application do not indicate quantity limitation and may indicate singular or plural. The terms “comprising,” “including,” “having,” and any variations thereof used in this application are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or modules (units) is not limited to the listed steps or units, but may also include steps or units not listed, or may include other steps or units inherent to these processes, methods, products, or devices. The terms “connected,” “linked,” “coupled,” and similar words used in this application are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. “Multiple” used in this application means two or more. “And / or” describes the relationship between related objects, indicating that three relationships may exist; for example, “A and / or B” can represent: A alone, A and B simultaneously, and B alone. The terms “first,” “second,” “third,” etc., used in this application are merely to distinguish similar objects and do not represent a specific ordering of the objects.

[0050] The method embodiments provided in this example can be executed on a terminal, computer, or similar computing device. Taking running on a terminal as an example, Figure 1 This is a hardware structure block diagram of a terminal for a three-dimensional interstellar medium physical field reconstruction method according to an embodiment of this application. Figure 1 As shown, a terminal may include one or more ( Figure 1 Only one is shown in the diagram. A processor 102 (which may include, but is not limited to, a microprocessor MCU or a programmable logic device FPGA, etc.) and a memory 104 for storing data are also shown. Optionally, the terminal may further include a transmission device 106 for communication functions and an input / output device 108. Those skilled in the art will understand that... Figure 1 The structure shown is for illustrative purposes only and does not limit the structure of the terminal described above. For example, the terminal may also include components that are larger than... Figure 1 The more or fewer components shown, or having the same Figure 1 The different configurations shown.

[0051] The memory 104 can be used to store computer programs, such as application software programs and modules, like the computer program corresponding to the three-dimensional interstellar medium physical field reconstruction method in this embodiment. The processor 102 executes various functional applications and data processing by running the computer program stored in the memory 104, thereby implementing the above-described method. The memory 104 may include high-speed random access memory and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include memory remotely located relative to the processor 102, and these remote memories can be connected to the terminal via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0052] The transmission device 106 is used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the terminal's communication provider. In one example, the transmission device 106 includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device 106 may be a Radio Frequency (RF) module used for wireless communication with the Internet.

[0053] This embodiment provides a method for reconstructing the physical field of a three-dimensional interstellar medium. Figure 2 This is a flowchart of a three-dimensional interstellar medium physics field reconstruction method according to an embodiment of this application, such as... Figure 2 As shown, the process includes the following steps:

[0054] Step S201: Obtain three-dimensional interstellar medium physical field simulation data.

[0055] Specifically, obtaining simulation data of the three-dimensional interstellar medium physics field can be achieved by acquiring simulation data from the STARFORGE project, a multi-institutional joint project aimed at conducting high-resolution three-dimensional radiation-magnetohydrodynamic (MHD) star formation simulations. STARFORGE simulates nebula evolution processes including gravity, magnetic fields, and stellar feedback (jet streams, stellar winds, radiation, and supernova explosions), generating three-dimensional physical fields such as gas density, velocity, temperature, star formation, and cluster assembly.

[0056] Step S202: Input the three-dimensional interstellar medium physical field simulation data into the preset three-dimensional photochemical radiation model to generate a synthetic observation dataset and a true three-dimensional interstellar medium physical field dataset; the true three-dimensional interstellar medium physical field dataset is formed by further calculation and supplementation based on the three-dimensional interstellar medium physical field simulation data and the three-dimensional photochemical radiation model.

[0057] Among them, the three-dimensional photochemical radiation model can be the 3D-PDR model, which is a chemical-thermal equilibrium-radiative transfer code suitable for three-dimensional photolysis regions. It can handle three-dimensional clouds with arbitrary density distributions and uses HEALPix ray tracing to calculate far-ultraviolet (FUV) radiation attenuation, thermal equilibrium, and chemical reactions.

[0058] By inputting three-dimensional interstellar medium physical field simulation data into a three-dimensional photochemical radiation model, on the one hand, a true three-dimensional interstellar medium physical field dataset containing key physical quantities such as chemical abundance distribution (e.g., H2, C, CO) and gas temperature distribution is generated; on the other hand, two-dimensional or spectral line cube images related to multi-band spectral line emission (e.g., CO, [C II], [O I] cooling lines) are generated through the radiation transfer process, forming a synthetic observation dataset, which provides core data support with strong physical consistency and data dimension matching for subsequent training sample construction and model training.

[0059] It should be noted that the above-mentioned true three-dimensional interstellar medium physical field dataset is not constructed from scratch, but is derived from the original three-dimensional interstellar medium physical field simulation data and further supplemented by the three-dimensional photochemical radiation model. It covers at least the three-dimensional density distribution and may also include extended physical quantities such as temperature distribution, chemical abundance distribution and far-ultraviolet radiation field.

[0060] Step S203: Construct training samples based on the synthetic observation dataset and the ground truth three-dimensional interstellar medium physical field dataset.

[0061] In the process of constructing training samples based on the synthetic observation dataset and the ground truth three-dimensional interstellar medium physical field dataset generated in step S202, the synthetic observation data is first used as input features (covering two-dimensional images, spectral line cubes, etc., covering multiple bands or spectral lines, such as CO, [C II], [O I] and other cooling line related data), and the corresponding ground truth three-dimensional physical field is used as a label (containing at least three-dimensional density distribution, and may also include temperature distribution, chemical abundance distribution, far-ultraviolet radiation field and other sets of physical quantities), forming a training sample pair corresponding to "input-label".

[0062] Step S204: Input the training samples into the initial three-dimensional interstellar medium physics field inversion deep learning architecture to obtain intermediate prediction results of the three-dimensional interstellar medium physics field; calculate the loss result based on the difference between the intermediate prediction results of the three-dimensional interstellar medium physics field and the true three-dimensional interstellar medium physics field dataset.

[0063] Specifically, the training samples are input into a pre-constructed initial three-dimensional interstellar medium physical field inversion deep learning architecture. This architecture can use deep neural network structures such as 3D U-Net, V-Net, Encoder–Decoder + Skip Connections, Diffusion Model, or Schrödinger Bridge inversion model, which can realize the mapping from two-dimensional / 2.5D low-dimensional observation data to high-dimensional three-dimensional physical fields. Through feature extraction, dimensionality upscaling, and mapping learning of the network, intermediate prediction results of the three-dimensional interstellar medium physical field, including core physical quantities such as three-dimensional density distribution, temperature distribution, and chemical abundance distribution, are output.

[0064] Subsequently, a loss function (i.e. supervised inversion loss, such as L1 or L2 error) is constructed based on the numerical difference between the predicted intermediate results and the true three-dimensional interstellar medium physics field dataset (which is derived from three-dimensional numerical simulation and 3D-PDR photochemical radiation model, and contains multiple physical quantities such as density, temperature, and chemical abundance). Finally, the loss result is obtained.

[0065] Step S205: The gradient of the loss result is backpropagated to the initial three-dimensional interstellar medium physics field inversion deep learning architecture for iterative training, resulting in the three-dimensional interstellar medium physics field inversion deep learning architecture.

[0066] After calculating the loss result, the gradient of the loss is backpropagated along the network layers of the initial three-dimensional interstellar medium physics inversion deep learning architecture. Gradient descent optimization algorithms (such as Adam, SGD, etc.) are used to iteratively update the learnable parameters of the network, such as the convolution kernel parameters and the weights of the fully connected layers. In each iteration, the deviation between the model prediction result and the true three-dimensional interstellar medium physics dataset is continuously reduced.

[0067] In this process, the convergence of the model is determined by monitoring the changes in loss between the training and validation sets, the model prediction accuracy, and the degree to which physical constraints are satisfied. When the loss value tends to stabilize and no longer decreases significantly, and the numerical error of the prediction results reaches the preset standard, iterative training is stopped. Finally, a deep learning architecture for inverting the physical field of the three-dimensional interstellar medium is obtained, which has the ability to accurately invert the physical field of the three-dimensional interstellar medium from two-dimensional / 2.5D observation data. This architecture can stably output prediction results containing at least three-dimensional density distribution, as well as optional core physical quantities such as temperature distribution and chemical abundance distribution, and supports uncertainty estimation to improve the interpretability and credibility of the results.

[0068] It should be noted that uncertainty assessment refers to the model simultaneously providing reliability and uncertainty information for the reconstruction results at each spatial location while outputting the 3D reconstruction results of the interstellar medium. The uncertainty mainly stems from three categories: first, the unavoidable physical uncertainty caused by incomplete observational information, i.e., 2D projection cannot uniquely determine the 3D structure (ill-conditioned inverse problem), and different 3D structures may exhibit similar projection effects; second, the uncertainty brought about by observational noise and instrument effects, which makes 100% certainty difficult even with a perfect model; and third, the uncertainty arising from the approximation between the model and the physical process, due to the approximation in the solution process and the fact that machine learning models do not provide analytical solutions. This uncertainty estimation makes the output no longer a single "unique solution," but a "credible solution space," which can be probabilistically verified against actual observations. Larger observational biases are allowed in areas of high uncertainty, while strict consistency is required in areas of low uncertainty. This enhances the interpretability and credibility of the reconstruction results, clearly distinguishes between uncertainties caused by insufficient observational data and model predictions, avoids over-interpretation of the 3D structure, and provides crucial evidence for reliability assessment of reconstruction results in different spatial regions during the inference phase.

[0069] Step S206: Input the acquired real observation data into the three-dimensional interstellar medium physical field inversion deep learning architecture to obtain the three-dimensional interstellar medium physical field reconstruction result.

[0070] This step requires preprocessing the acquired real observation data (from multi-band spectral lines and two-dimensional projection images captured by the telescope, such as CO, [C II], [OI] and other cooling line related data). The data format should be standardized, scaled, and noise should be preliminarily processed according to the training sample format standard to ensure that it is consistent with the characteristics of the input data during model training.

[0071] Subsequently, the preprocessed real observation data is input into a deep learning architecture for inverting the three-dimensional interstellar medium physics field, which has been optimized through multiple rounds of iterative training. This architecture has been trained with a loss function and has a stable ability to map low-dimensional observation data to high-dimensional three-dimensional physics fields. Based on the input real observation information, it can accurately invert and output the reconstructed three-dimensional interstellar medium physics field. The result includes at least the core physical quantity of three-dimensional density distribution, and can also output key physical parameters such as temperature distribution and chemical abundance distribution according to actual needs. At the same time, it supports the output of uncertainty estimates for each spatial location, clarifying the reliability of the reconstruction results in different regions, and providing high-precision and interpretable three-dimensional physics field data support for subsequent observation verification and astronomical research.

[0072] Steps S201 to S206 above involve acquiring simulated data of the three-dimensional interstellar medium physical field, inputting it into a preset three-dimensional photochemical radiation model to generate a synthetic observation dataset and a true three-dimensional interstellar medium physical field dataset based on the simulated data and further calculated by the model. Training samples are then constructed based on the two datasets, and the training samples are input into the initial three-dimensional interstellar medium physical field inversion deep learning architecture to obtain intermediate prediction results and calculate the loss results. Subsequently, the gradient of the loss results is backpropagated to the initial architecture for iterative training to obtain an optimized inversion deep learning architecture. Finally, real observation data is input to obtain the three-dimensional interstellar medium physical field reconstruction results. This achieves a deep integration of real physical simulation, photochemical radiation calculation, and machine learning inverse problem modeling, breaking through the limitations of traditional methods that rely on idealized assumptions or simple empirical fitting. It solves the technical problem that observation data is difficult to accurately reproduce the complex physical state of the real three-dimensional interstellar medium, achieving a high-precision, physically self-consistent, and interpretable three-dimensional interstellar medium physical field reconstruction effect. At the same time, it has good generalization ability and scalability, and can adapt to observation data of different bands and different instruments, providing strong technical support for astrophysical research such as molecular cloud evolution and star formation mechanisms.

[0073] In some embodiments, the calculation of the loss result based on the difference between the intermediate predicted results of the three-dimensional interstellar medium physics field and the ground truth three-dimensional interstellar medium physics field dataset includes:

[0074] Based on the difference between the intermediate results of the predicted three-dimensional interstellar medium physical field and the true three-dimensional interstellar medium physical field dataset, a first loss is constructed.

[0075] The intermediate results of the three-dimensional interstellar medium physical field prediction are input into a preset fast rendering module to obtain the synthetic observation prediction intermediate results; the fast rendering module is the approximate calculation module of the three-dimensional photochemical radiation model.

[0076] A second loss is constructed based on the difference between the intermediate results of the synthetic observation prediction and the synthetic observation dataset.

[0077] The loss result is obtained based on the first loss and the second loss.

[0078] First, the first loss (i.e., supervised inversion loss) is constructed. The core function of the first loss is to enable the model to quickly learn the fundamental mapping relationship from observational data to the three-dimensional physical field, ensuring that the predicted results are numerically close to the actual physical state. Specifically, after inputting the training samples into the initial three-dimensional interstellar medium physical field inversion deep learning architecture, intermediate prediction results containing core physical quantities such as three-dimensional density distribution and possible temperature and chemical abundance distributions are obtained. This intermediate result is then compared pixel-by-pixel or voxel-by-voxel with the ground truth three-dimensional interstellar medium physical field dataset (based on three-dimensional numerical simulation and further calculated and supplemented by the 3D-PDR photochemical radiation model, covering a set of true values ​​for multiple physical quantities such as density, temperature, and chemical abundance). The differences between the two are quantified using loss calculation methods such as L1 error and L2 error, thus constructing the first loss and providing a basic supervisory signal for model training.

[0079] Secondly, a second loss (i.e., observation consistency constraint loss) is constructed. To prevent the model from pursuing only numerical fit and deviating from the actual observation scenario, an observation consistency constraint needs to be introduced. This process relies on a pre-defined fast rendering module. This fast rendering module is an approximation calculation module for three-dimensional photochemical radiation models (such as 3D-PDR). Its core advantage is that it significantly reduces computational complexity while ensuring the main physical consistency. Specifically, this can be achieved by performing weighted integration along the line of sight on the predicted three-dimensional physical field, approximate emission calculation based on physical empirical formulas, lookup table or interpolation-based radiation calculation, or approximating the three-dimensional photochemical radiation model using a pre-trained surrogate model. After inputting the above-mentioned intermediate results of the three-dimensional interstellar medium physical field prediction into this fast rendering module, a synthetic observation prediction intermediate result is generated, which simulates a two-dimensional projection image or spectral cube data under the real observation scenario. Subsequently, the intermediate results of the synthetic observation predictions are compared with the corresponding synthetic observation dataset (standard observation data generated from the three-dimensional true physical field through a complete radiative transfer process). The differences between the two in key observation features such as spectral line intensity, image structure, and spectral line type are quantified, and then a second loss is constructed to ensure that the three-dimensional physical field predicted by the model can be reversed to the results that conform to the actual observation laws.

[0080] Finally, the final loss result is obtained by weighted fusion of the first and second losses. By configuring reasonable adjustable weights for the first and second losses, the importance of numerical fitting accuracy and observation consistency constraints can be balanced according to training needs. For example, in the early stages of training, the weight of the first loss can be appropriately increased to allow the model to quickly grasp the basic structural features of the three-dimensional physical field; in the later stages of training, the weight of the second loss can be increased to strengthen the matching degree between the model's prediction results and the observed data. The loss result formed by the weighted sum of the two includes both the error constraints at the numerical level and the physical consistency requirements at the observation level, providing a comprehensive and reliable guiding signal for subsequent iterative optimization of model parameters.

[0081] The above steps construct a first loss based on the difference between the intermediate predicted results of the 3D interstellar medium physical field and the true 3D interstellar medium physical field dataset. This ensures that the model predictions numerically match the real physical state. The intermediate predicted results are then input into the fast rendering module of the 3D photochemical radiation model approximation calculation module to obtain the synthetic observation prediction intermediate results. A second loss is constructed based on the difference between this result and the synthetic observation dataset. Finally, the first and second losses are combined to obtain the final loss result. This achieves a dual constraint on the model prediction, ensuring the numerical accuracy of the 3D physical field prediction while preventing the model from pursuing only numerical fitting and deviating from actual observational patterns through observation consistency constraints. This allows the model to learn the mapping relationship from 2D / 2.5D observation data to the 3D physical field, which not only conforms to the numerical truth but also corresponds to the spectral lines, images, and other features in the real observation scenario. This significantly improves the reliability and physical consistency of the model inversion, laying a solid foundation for the subsequent accurate reconstruction of the 3D interstellar medium physical field.

[0082] In some embodiments, obtaining the loss result based on the first loss and the second loss includes:

[0083] Based on the intermediate results of the three-dimensional interstellar medium physical field prediction and the preset physical constraints, a third loss is constructed;

[0084] The loss result is obtained based on the first loss, the second loss, and the third loss.

[0085] The process of obtaining the loss result based on the first and second losses further introduces a regularization mechanism based on physical constraints. By constructing a third loss and integrating it with the first two, a comprehensive loss result that combines numerical accuracy, observational consistency, and physical rationality is formed. Specifically, firstly, the third loss (i.e., physical constraint regularization loss) is constructed. The core purpose is to prevent the model from outputting three-dimensional physical field results that violate basic physical laws in the process of pursuing numerical fitting and observation matching, thus ensuring the physical self-consistency of the reconstruction results. The preset physical constraints are formulated based on the inherent physical properties of the interstellar medium and cover several key rules: density non-negativity constraint, since the density of matter does not have a negative value, it is necessary to restrict the density value of all spatial locations in the intermediate prediction results to be greater than or equal to 0; temperature range constraint, based on the common sense of astrophysics, to limit the gas temperature to a reasonable range, avoiding extreme temperature values ​​far exceeding the actual interstellar environment; chemical abundance normalization constraint, to ensure that the abundance ratio of various chemical species (such as H2, C, CO, etc.) conforms to the stoichiometric law and the sum is within a reasonable range; in addition, constraints such as energy conservation and mass conservation may also be included to prevent the model output from showing situations that violate the physical essence, such as energy being generated out of thin air or mass not being conserved. When constructing the third loss, a special physical constraint penalty function Φ is designed to quantify and penalize the part of the prediction intermediate results that violates the above constraints. The more severe the constraint, the larger the penalty value, and finally the third loss is formed.

[0086] Subsequently, the final loss result is obtained by weighted fusion of the first, second, and third losses. By configuring adjustable weights α, β, and γ (α>0, β>0, γ>0) for the three loss terms, the importance of each constraint dimension can be flexibly balanced according to the actual needs of model training. For example, in the early stages of training, the weight of α can be appropriately increased to allow the model to grasp the basic numerical characteristics of the three-dimensional physical field first; as training progresses, the weight of β can be increased to strengthen observation consistency matching; at the same time, a certain weight of γ is retained to continuously constrain physical rationality and avoid physical violations in the later stages of the model. The total loss result is obtained by weighted summation of the three losses. (in, The first loss, This is the second loss. (As the third loss), this result includes not only the numerical difference between the predicted value and the true value, but also incorporates the consistency requirements at the observation level and the rule constraints at the physical level. It provides comprehensive and scientific guidance signals for the iterative optimization of model parameters, ensuring that the trained model can accurately invert the three-dimensional physical field and meet the physical law requirements of real astronomical scenes.

[0087] The above steps involve constructing a third loss based on intermediate predictions of the 3D interstellar medium physical field and pre-defined physical constraints (such as non-negative density, reasonable temperature range, normalized chemical abundance, and mass conservation, which conform to the inherent physical properties of the interstellar medium). This third loss is then weighted and fused with a first loss to ensure numerical fitting accuracy and a second loss to enhance observational consistency, resulting in the final loss. This achieves a triple synergistic constraint on model training. The first loss ensures that the predicted results numerically match the true 3D physical field; the second loss allows the model output to correspond inversely to the spectral lines and image features of real observations; and the third loss prevents the model from producing unreasonable results that violate fundamental physical laws. By integrating domain physics knowledge into the model optimization process in the form of soft constraints, this significantly improves the physical consistency, prediction reliability, and interpretability of the 3D interstellar medium physical field inversion. It also enhances the model's robustness in scenarios with sparse or noisy observational data, providing a comprehensive and scientific optimization guide for subsequent accurate and reasonable 3D interstellar medium structure reconstruction.

[0088] In some embodiments, acquiring three-dimensional interstellar medium physical field simulation data includes:

[0089] Obtain raw three-dimensional interstellar medium physical field simulation data;

[0090] The original three-dimensional interstellar medium physical field simulation data is resampled to a regular cubic mesh to obtain the three-dimensional interstellar medium physical field simulation data.

[0091] First, we acquire the original three-dimensional interstellar medium physical field simulation data. This data comes from the high-resolution three-dimensional radiation-magnetohydrodynamic star formation simulation project jointly carried out by the STARFORGE project and other institutions. It covers complex physical processes such as gravity, magnetic field, and stellar feedback (jet, stellar wind, radiation, supernova explosion). It includes key three-dimensional physical information such as gas density field, velocity field, temperature distribution, chemical abundance, and star formation and cluster assembly. Its original output format may be unstructured grid or adaptive mesh refinement (AMR) format, and the grid resolution is dynamically adjusted according to the intensity of the physical process.

[0092] Subsequently, the original simulation data is resampled to map the irregularly distributed original data to a preset regular cubic grid (e.g., 128³). At the same time, the conservation of physical quantities and data integrity are maintained during the resampling process, and the interference caused by grid inconsistency in the original data is eliminated. Finally, three-dimensional interstellar medium physical field simulation data with regular structure and uniform dimensions is obtained, which can be directly input into the subsequent three-dimensional photochemical radiation model.

[0093] Through the above steps, the original simulation data was standardized and normalized, eliminating the interference of subsequent processing caused by the inconsistency of the mesh in the original data. This enabled the processed data to adapt to the input requirements of the three-dimensional photochemical radiation model, providing a unified, dimensionally standardized and physically complete basic data support for subsequent photochemical calculations, training sample construction and model training, ensuring the smooth progress and data consistency of the entire three-dimensional reconstruction process.

[0094] In some embodiments, constructing training samples based on the synthetic observation dataset and the ground truth three-dimensional interstellar medium physics dataset includes:

[0095] Based on the synthetic observation dataset and the true three-dimensional interstellar medium physical field dataset, original training samples are constructed;

[0096] The original training samples are cropped and augmented from multiple perspectives, and then stored in a preset format to obtain the training samples.

[0097] First, synthetic observation datasets (including multi-band spectral line images or spectral line cubes such as CO, [C II], and [OI]) are used as input features, and ground truth three-dimensional interstellar medium physical field datasets (including core physical quantities such as density, temperature, and chemical abundance) are used as corresponding labels to construct original training samples with a one-to-one correspondence between "input" and "label", ensuring the physical consistency and mapping correlation of the samples.

[0098] Subsequently, multi-view cropping was performed on the original training samples, with projection cropping along different directions of the x, y, and z axes to generate sub-samples of multiple scales and perspectives. This enriched the perspective diversity of the samples and enhanced the model's ability to understand features at different scales. At the same time, targeted data augmentation operations were carried out, including geometric transformations (rotation, mirroring, flipping), observation simulation (adding common astronomical observation noises such as photon shot noise, dark noise, and readout noise), resolution perturbation, and numerical scaling. These effectively improved the model's generalization ability and robustness, and reduced the risk of overfitting.

[0099] Finally, the cropped and enhanced samples are stored uniformly in preset standard formats such as HDF5. This format can efficiently carry massive amounts of complex physical data and retain complete metadata. During storage, key metadata such as the viewpoint parameters, band type, simulation time, and physical variable fields corresponding to the samples are attached simultaneously, resulting in training samples that are structurally regular, sufficiently diverse, physically complete, and adapted to the model training input requirements.

[0100] The above steps achieve standardized, diversified, and high-quality construction of training samples, ensuring the physical consistency and mapping correlation of the sample "input-label" relationship. They also broaden the learning boundary of the model through multi-view expansion and data augmentation, reducing the risk of overfitting. At the same time, the unified format and complete meta-information ensure the adaptability of the samples to the subsequent machine learning model training input, providing solid data support for the model to efficiently learn the stable mapping relationship from two-dimensional / 2.5D observation data to three-dimensional physical fields, significantly improving the efficiency of model training and the reliability of the final inversion.

[0101] In some embodiments, after obtaining the three-dimensional interstellar medium physical field reconstruction result, the method further includes:

[0102] The reconstructed physical field of the three-dimensional interstellar medium is input into a preset radiation transfer module to obtain the synthetic observation results;

[0103] The prediction effect is verified based on the synthetic observation results and the real observation data.

[0104] Specifically, the reconstruction results, including core physical quantities such as three-dimensional density distribution, temperature distribution, and chemical abundance distribution, are first input into the preset radiation transfer module. This module solves the radiation transfer equation by ray tracing along the preset line of sight, Monte Carlo method, LVG approximation or escape probability model, etc., fully considering absorption, emission and velocity Doppler effects, and generating synthetic observation results with the same dimension and format as the real observation data.

[0105] Subsequently, a comprehensive quantitative comparison was conducted between the synthetic observation results and real observation data (derived from telescope multi-band spectral lines, two-dimensional projection images, etc.). Through various methods such as line intensity error analysis, spectral line type feature comparison, image structure consistency verification, and statistical measure (such as mean, variance, correlation coefficient, etc.) matching degree evaluation, the reliability and physical consistency of the reconstruction results were systematically verified. If the verification results meet the preset accuracy standards, the three-dimensional reconstruction effect is confirmed to be effective. If there are deviations, they can be fed back to the model training stage for parameter adjustment, forming a closed loop of "reconstruction-verification-optimization" to further ensure the accuracy and credibility of the three-dimensional interstellar medium physical field reconstruction results.

[0106] The above steps realize a closed-loop verification mechanism of "reconstruction-verification", which can not only objectively and accurately evaluate the reliability and physical consistency of the three-dimensional reconstruction results, but also promptly detect the deviation between the reconstruction results and the actual observations, providing feedback for model parameter optimization or data preprocessing improvement. This effectively avoids the problem of unreliable reconstruction results caused by lack of verification, further ensuring the accuracy and scientific nature of the reconstruction of the three-dimensional interstellar medium physical field, making the reconstruction results more consistent with the actual scenarios and physical laws of astronomical observation, and providing reliable data support for astrophysical research such as molecular cloud evolution and star formation mechanisms.

[0107] In some embodiments, verifying the prediction effect based on the synthetic observation results and the real observation data includes:

[0108] The synthetic observation results and the real observation data are quantitatively compared to obtain the prediction effect; the quantitative comparison includes line intensity error, spectral line type comparison, image structure comparison and statistical consistency.

[0109] The above-mentioned verification of prediction results based on synthetic observation results and real observation data specifically includes a comprehensive quantitative comparison between synthetic observation results obtained from the reconstruction results of the three-dimensional interstellar medium physical field through the radiation transfer module (generated by solving the radiation transfer equation through ray tracing, Monte Carlo methods, etc.) and real observation data from telescope multi-band spectral lines and two-dimensional projection images, and then a comprehensive judgment on the prediction effect.

[0110] The quantitative comparison encompasses the verification of core observational features across multiple dimensions: line intensity error analysis, which quantifies the matching degree of radiation intensity by calculating the absolute and relative errors of the two intensities of key spectral lines (such as CO, [C II], [O I], etc.); spectral line shape comparison, which compares key parameters such as the contour shape, peak position, and half-width at half-maximum of spectral lines one by one to verify the accuracy of the reconstruction of physical processes such as the velocity Doppler effect; image structure comparison, which analyzes the consistency between synthetic and real observations in spatial structure and morphological distribution through edge detection, feature extraction, and other methods to ensure the reliability of the spatial features of the 3D reconstruction; and statistical consistency assessment, which calculates statistical parameters such as the mean, variance, and correlation coefficient of the two to verify the stability and rationality of the prediction results from the perspective of the overall data distribution. Through the comprehensive consideration of multi-dimensional quantitative indicators, the accuracy and physical self-consistency of the 3D interstellar medium physical field reconstruction results are fully and objectively verified.

[0111] The above steps enable multi-dimensional and objective verification of the 3D reconstruction results. This not only allows for precise evaluation of the consistency between the reconstruction results and real observations in terms of numerical characteristics, physical process reproduction, and spatial structure presentation, but also comprehensively explores the sources of deviation between the two. This provides clear feedback for model parameter optimization and data preprocessing improvement, effectively avoiding the problem of unreliable reconstruction results caused by lack of verification. It further enhances the accuracy, physical consistency, and scientific reliability of the 3D interstellar medium physical field reconstruction, providing reliable data support for molecular cloud evolution and star formation mechanism analysis in astrophysical research.

[0112] In some embodiments, the true three-dimensional interstellar medium physics dataset includes at least a three-dimensional density distribution.

[0113] Among them, the true three-dimensional interstellar medium physics field dataset includes at least the core physical quantity of three-dimensional density distribution. This data is based on the original density field data generated by high-resolution three-dimensional radiation-magnetohydrodynamic star formation simulation such as STARFORGE, and is further calculated and supplemented by the 3D-PDR photochemical radiation model. It can accurately reflect the material distribution characteristics of the interstellar medium in three-dimensional space. At the same time, this dataset can also be further included with multi-dimensional physical quantities such as far-ultraviolet radiation field, three-dimensional temperature distribution, and chemical abundance distribution according to actual application needs, forming a complete set of multi-physical quantities. This design uses three-dimensional density distribution as a fundamental physical quantity, ensuring the core supporting role of the ground truth dataset and providing the most critical inversion target benchmark for machine learning models. Meanwhile, the scalable multi-physical quantities enrich the label dimensions of model training, enabling the model to learn more comprehensive correlations of interstellar medium physical properties. At the same time, this dataset combines the physical realism of the original simulation with the supplementary completeness of photochemical calculations, providing high-precision, physically consistent label data for training sample construction. This ensures the accuracy of model inversion from the source and lays a solid foundation for the subsequent collaborative inversion of the three-dimensional field of multi-physical quantities, making the final three-dimensional interstellar medium physical field reconstruction result more consistent with the complex physical state of real astronomical scenarios.

[0114] Figure 3 This is a schematic diagram of the framework of a three-dimensional interstellar medium physical field reconstruction system according to an embodiment of this application. The diagram illustrates the complete technical process: First, simulated three-dimensional interstellar medium physical field data is input into a three-dimensional photochemical radiation model to generate a true three-dimensional interstellar medium physical field dataset and a synthetic observation dataset. The two are combined to construct training samples. Subsequently, the training samples are used to train a deep learning architecture for three-dimensional interstellar medium physical field inversion. After the architecture is trained, real observation data is input into the architecture, and the three-dimensional interstellar medium physical field reconstruction result is finally obtained through the inference process. The entire system realizes the end-to-end physical field reconstruction from simulation data generation and model training to real observation inversion.

[0115] The training dataset constructed in this application is derived from real physical processes (gas dynamics, radiation feedback, chemical evolution), thus the learned model possesses stronger physical robustness, rather than relying solely on empirical fitting. Utilizing machine learning models, it achieves the ability to directly output three-dimensional structures from two-dimensional observations, greatly improving the efficiency of three-dimensional reconstruction and making it suitable for large-scale nebulae, star clusters, or galaxy environments. The introduction of a forward rendering verification mechanism automatically assesses the consistency between the reconstructed results and real observations, improving the reliability of the results and enabling applications in scientific research and engineering. This system is scalable to observational data from different wavebands and instruments, possessing good applicability and versatility, providing a powerful tool for interstellar structure research, star formation mechanism investigation, and astronomical observation interpretation. Compared to traditional methods, this application's embodiments simultaneously possess three major advantages: "realistic physical simulation," "efficient machine learning inversion," and "verification through real observations," filling the gap in existing technologies for mapping observations to three-dimensional structures.

[0116] The ability to scale to observation data from different bands and instruments means that this method decouples physical modeling from observation representation, allowing the same 3D inversion framework to flexibly adapt to different band characteristics and instrument responses without requiring a redesign of the model structure for specific observation equipment. Specifically, different band observations correspond to different physical processes, and the 3D-PDR model itself can output multiple spectral lines and radiation intensities in different bands, giving the synthesized observation data a natural multi-band attribute. Furthermore, considering the differences in spatial resolution, spectral resolution, observation noise, and band coverage among different instruments, the synthesis observation generation stage can be adapted by adding different beam convolutions, different channel sampling, and different noise models. This allows the same 3D physical field to be "rendered" as observation data in different styles, such as ALMA, JWST, and single-antenna telescopes. Ultimately, the model learns the invariant physical relationship of "observation → 3D structure," rather than the specific characteristics of a particular instrument, thus achieving unified input and efficient processing of observation data from different telescopes and different bands.

[0117] This embodiment also provides a three-dimensional interstellar medium physical field reconstruction device, which is used to implement the above embodiments and preferred embodiments, and will not be repeated as already described. As used below, the terms "module," "unit," "subunit," etc., can refer to a combination of software and / or hardware that performs a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0118] Figure 4 This is a structural block diagram of a three-dimensional interstellar medium physics field reconstruction device according to an embodiment of this application, such as... Figure 4 As shown, the device includes:

[0119] The three-dimensional data acquisition module 41 is used to acquire three-dimensional interstellar medium physical field simulation data;

[0120] The dataset generation module 42 is used to input the three-dimensional interstellar medium physical field simulation data into the preset three-dimensional photochemical radiation model to generate a synthetic observation dataset and a true three-dimensional interstellar medium physical field dataset; the true three-dimensional interstellar medium physical field dataset is formed by further calculation and supplementation based on the three-dimensional interstellar medium physical field simulation data through the three-dimensional photochemical radiation model.

[0121] Training sample construction module 43 is used to construct training samples based on synthetic observation datasets and ground truth three-dimensional interstellar medium physical field datasets;

[0122] The loss calculation module 44 is used to input training samples into the initial three-dimensional interstellar medium physics field inversion deep learning architecture to obtain intermediate prediction results of the three-dimensional interstellar medium physics field; and to calculate the loss result based on the difference between the intermediate prediction results of the three-dimensional interstellar medium physics field and the true three-dimensional interstellar medium physics field dataset.

[0123] Model training module 45 is used to backpropagate the gradient of the loss result to the initial three-dimensional interstellar medium physical field inversion deep learning architecture for iterative training, so as to obtain the three-dimensional interstellar medium physical field inversion deep learning architecture.

[0124] The model inference module 46 is used to input the acquired real observation data into the three-dimensional interstellar medium physical field inversion deep learning architecture to obtain the three-dimensional interstellar medium physical field reconstruction results.

[0125] It should be noted that the above modules can be functional modules or program modules, and can be implemented by software or hardware. For modules implemented by hardware, the above modules can reside in the same processor; or the above modules can be located in different processors in any combination. Specific examples in this embodiment can be found in the examples described in the above embodiments and optional implementations, and will not be repeated in this embodiment.

[0126] Furthermore, in conjunction with the three-dimensional interstellar medium physical field reconstruction method in the above embodiments, this application embodiment can provide a storage medium for implementation. This storage medium stores a computer program; when executed by a processor, the computer program implements any of the three-dimensional interstellar medium physical field reconstruction methods in the above embodiments.

[0127] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0128] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0129] Those skilled in the art should understand that the technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments have been described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0130] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for reconstructing the physical field of a three-dimensional interstellar medium, characterized in that, include: Acquire three-dimensional interstellar medium physical field simulation data; The simulated data of the three-dimensional interstellar medium physical field is input into a preset three-dimensional photochemical radiation model to generate a synthetic observation dataset and a true three-dimensional interstellar medium physical field dataset. The true three-dimensional interstellar medium physical field dataset is formed by further calculation and supplementation based on the simulated data of the three-dimensional interstellar medium physical field through the three-dimensional photochemical radiation model. Training samples are constructed based on the synthetic observation dataset and the true three-dimensional interstellar medium physical field dataset; The training samples are input into the initial three-dimensional interstellar medium physics field inversion deep learning architecture to obtain intermediate results of three-dimensional interstellar medium physics field prediction. Based on the difference between the intermediate predicted results of the three-dimensional interstellar medium physics field and the true three-dimensional interstellar medium physics field dataset, the loss result is calculated, including: Based on the difference between the intermediate results of the predicted three-dimensional interstellar medium physical field and the true three-dimensional interstellar medium physical field dataset, a first loss is constructed. The intermediate results of the three-dimensional interstellar medium physical field prediction are input into a preset fast rendering module to obtain the synthetic observation prediction intermediate results; the fast rendering module is the approximate calculation module of the three-dimensional photochemical radiation model. A second loss is constructed based on the difference between the intermediate results of the synthetic observation prediction and the synthetic observation dataset. The loss result is obtained based on the first loss and the second loss; The gradient of the loss result is backpropagated to the initial three-dimensional interstellar medium physics field inversion deep learning architecture for iterative training, to obtain the three-dimensional interstellar medium physics field inversion deep learning architecture. The acquired real observation data is input into the three-dimensional interstellar medium physics field inversion deep learning architecture to obtain the three-dimensional interstellar medium physics field reconstruction results.

2. The method for reconstructing the physical field of a three-dimensional interstellar medium according to claim 1, characterized in that, The process of obtaining the loss result based on the first loss and the second loss includes: Based on the intermediate results of the three-dimensional interstellar medium physical field prediction and the preset physical constraints, a third loss is constructed; The loss result is obtained based on the first loss, the second loss, and the third loss.

3. The method for reconstructing the physical field of a three-dimensional interstellar medium according to claim 1, characterized in that, The acquisition of three-dimensional interstellar medium physical field simulation data includes: Obtain raw three-dimensional interstellar medium physical field simulation data; The original three-dimensional interstellar medium physical field simulation data is resampled to a regular cubic mesh to obtain the three-dimensional interstellar medium physical field simulation data.

4. The method for reconstructing the physical field of a three-dimensional interstellar medium according to claim 1, characterized in that, The construction of training samples based on the synthetic observation dataset and the ground truth three-dimensional interstellar medium physical field dataset includes: Based on the synthetic observation dataset and the true three-dimensional interstellar medium physical field dataset, original training samples are constructed; The original training samples are cropped and augmented from multiple perspectives, and then stored in a preset format to obtain the training samples.

5. The method for reconstructing the physical field of a three-dimensional interstellar medium according to claim 1, characterized in that, After obtaining the three-dimensional interstellar medium physical field reconstruction results, the following is also included: The reconstructed physical field of the three-dimensional interstellar medium is input into a preset radiation transfer module to obtain the synthetic observation results; The prediction effect is verified based on the synthetic observation results and the real observation data.

6. The method for reconstructing the physical field of a three-dimensional interstellar medium according to claim 5, characterized in that, The verification of the prediction effect based on the synthetic observation results and the real observation data includes: The synthetic observation results and the real observation data are quantitatively compared to obtain the prediction effect; the quantitative comparison includes line intensity error, spectral line type comparison, image structure comparison and statistical consistency.

7. The method for reconstructing the physical field of a three-dimensional interstellar medium according to any one of claims 1 to 6, characterized in that, The true three-dimensional interstellar medium physics dataset includes at least a three-dimensional density distribution.

8. A three-dimensional interstellar medium physical field reconstruction device, characterized in that, The device includes: The three-dimensional data acquisition module is used to acquire three-dimensional interstellar medium physical field simulation data; The dataset generation module is used to input the three-dimensional interstellar medium physical field simulation data into a preset three-dimensional photochemical radiation model to generate a synthetic observation dataset and a true three-dimensional interstellar medium physical field dataset; the true three-dimensional interstellar medium physical field dataset is formed based on the three-dimensional interstellar medium physical field simulation data and further calculated and supplemented by the three-dimensional photochemical radiation model. The training sample construction module is used to construct training samples based on the synthetic observation dataset and the true three-dimensional interstellar medium physical field dataset; The loss calculation module is used to input the training samples into the initial three-dimensional interstellar medium physics field inversion deep learning architecture to obtain intermediate prediction results of the three-dimensional interstellar medium physics field; and to calculate the loss result based on the difference between the intermediate prediction results of the three-dimensional interstellar medium physics field and the ground truth three-dimensional interstellar medium physics field dataset. The loss calculation module is further configured to: construct a first loss based on the difference between the intermediate prediction result of the three-dimensional interstellar medium physical field and the true three-dimensional interstellar medium physical field dataset; input the intermediate prediction result of the three-dimensional interstellar medium physical field to a preset fast rendering module to obtain a synthetic observation prediction intermediate result; the fast rendering module is an approximate calculation module for the three-dimensional photochemical radiation model; construct a second loss based on the difference between the synthetic observation prediction intermediate result and the synthetic observation dataset; and obtain the loss result based on the first loss and the second loss. The model training module is used to backpropagate the gradient of the loss result to the initial three-dimensional interstellar medium physics field inversion deep learning architecture for iterative training, so as to obtain the three-dimensional interstellar medium physics field inversion deep learning architecture. The model inference module is used to input the acquired real observation data into the three-dimensional interstellar medium physical field inversion deep learning architecture to obtain the three-dimensional interstellar medium physical field reconstruction result.

9. A storage medium, characterized in that, The storage medium stores a computer program, wherein the computer program is configured to execute the three-dimensional interstellar medium physics field reconstruction method according to any one of claims 1 to 7 when it is run.

Images