Analysis and Development Enlightenment on Source Term Inversion Technology of Nuclear and Chemical Hazards

Source Term Inversion (STI) is of great significance to mitigate and contain the sources of nuclear and chemical hazards, accurately predict the spatiotemporal transmission and diffusion of nuclear and chemical hazards, assist combat operations and support decision-making. This paper summarizes and analyzes the key algorithms and application platform of the current nuclear and chemical hazard source inversion technology, and puts forward the enlightenment and suggestions for its development, which has a certain theoretical reference and reference value.


Introduction
Usually, nuclear and chemical attacks are accompanied by sudden and huge lethality. It is an important task to determine the information related to the nuclear, chemical and biological safety and national defense security of the country, which can help to alleviate and contain pollution sources, and more accurately predict the spatiotemporal transmission and diffusion of pollutants, providing a decision for the army's combat protection, combat decontamination and combat operations Policy basis [1] . Therefore, based on the inversion technology of nuclear and chemical hazards source term, this paper summarizes the relevant methods, technologies and platforms, analyzes the focus of next research.

STI based on Euler method
Most source term inversion algorithms use Euler method to obtain source term information by minimizing the difference function, such as optimal interpolation method, genetic algorithm, Kalman filter, ensemble Kalman filter, four-dimensional variation, set four-dimensional variation, etc. [2][3][4][5][6] . Euler methods usually either rely on statistical methods for selecting test solutions or use adjoint models to calculate backward from observation time to release time. Both methods may include iterative optimization solution [7][8][9] .
In Bayes formula and Kalman filter, random sampling is used to generate source information, which is then used as the input of AT&D model [10][11][12] . The difference between the subsequent concentration output and the concentration observation can be used to determine the possibility of these initial estimates. All probability density functions of source information generated by all estimation likelihood calculations can be used to optimize the program to obtain iterative improved estimates of source locations [13][14][15][16] . In addition, Bourne Joseph R et al. [17] , of the Automatic Machine Control Laboratory, Department of Mechanical Engineering, University of Utah, USA, proposed a new algorithm for plume source term estimation and source search motion planning based on nonparametric Bayes, which uses the coordination and plume estimation model among multiple robots to realize faster and more robust pollution source location determination and source intensity estimation. The algorithm has been used in the design of mobile robots equipped with gas concentration sensors. Many Euler STI contain adjoint models. On the basis of adjoint model, variational assimilation method is gradually applied to source term inversion, and a large number of experiments and applications have been carried out in the research process. The most representative is the CEREA research team headed by Marc Bocquet [18] , which has successively verified the effectiveness of variational data assimilation in the inversion of nuclear accident source terms by using numerical simulation experiments, wind tunnel experimental data and European atmospheric diffusion experimental data, and then estimated and studied the release source terms of Chernobyl and Fukushima nuclear accidents.
In addition, there are other Euler methods that do not require adjoint model or Bayesian reasoning. By iteratively adjusting the AT&D prediction driven by source information, these methods directly obtain the source information from the decentralized model to match the observed pollution concentration values.

STI based on Lagrange method
The second type of STI algorithm is Lagrange method, but it is not widely studied and applied as Euler algorithm. Lagrange methods belong to the category of entity backtracking, that is, they trace the state of an entity to its original state. This method is similar to the target tracking problem proposed by Hall and McMullen [19] (2004). The traditional Lagrangian backtracking method is to trace the source by analyzing the time of a single fluid package [20][21][22] . The requirements of wind field and pollutant concentration data are often too high to be used for source term estimation. In addition, the reverse flow must converge in order to accurately trace the fluid back to its source, or the fluid must evolve in time so that multiple trajectories intersect at the source [23] .
Based on the traditional Lagrangian package backtracking method, some scholars have extended and improved it from the perspective of multi-source inversion and multi-scale inversion, and proposed Lagrange entity backtracking, Lagrange particle model and other methods. Andrew J. Annunzio, Sue Ellen Haupt and others [24][25][26][27] (2007, 2012) , of Pennsylvania State University, have carried out a series of studies in the field of pollution transmission and diffusion. In view of the complex problem of the number of uncertain pollution sources and the overlapping of emission, they proposed a multi entity field approximation (MEFA) method based on Lagrange state estimation to locate the location of multiple instantaneous or continuous release pollution sources by integrating the effects of turbulence and solid diffusion on multi-source concentration field. However, the influence of complex underlying surfaces such as buildings and terrain on pollutant diffusion is not considered in the literature.
Stohl A of Greece used the Lagrange particle model [28] in RODOS to conduct STI research based on data of numerical simulation experiment and Mol Belgium site tracer experiment, which preliminarily realized the inversion of the release source term of nuclear accident using dose rate data. Subsequently, in order to strengthen the ability of RODOS to predict and evaluate the consequences of nuclear accidents, they launched the project of "data assimilation of off-site nuclear accident emergency" based on Kalman filter [29] . Unfortunately, due to the complexity of the actual nuclear accident, the experimental results are quite different from the actual situation, so it has not been carried out all the time [30] . Mahmoud Bady [31] (2017) carried out Lagrangian particle inversion modeling according to the basic principles of computational fluid dynamics, aiming to determine the location of urban air pollution sources by using direct inversion technology.

STI based on deep learning
Deep learning provides a new idea for source term inversion. It does not need to explore the mathematical relationship between the source term and the observed value, but only needs to learn enough training data to quickly predict and retrieve. The mapping relationship of deep learning organization is shown in Figure 2. Ilias Bougoudis et al. [32] (2015) proposed a triple intelligent integrated system (HISYCOL) based on hybrid machine learning. Through cluster data sets and unsupervised machine learning, the system realizes clustering tracking of data vectors and hidden knowledge mining, and has significant effect in processing correlation analysis and source term inversion under high pollutant concentration.
Jan Kleine, University of Twente, the Netherlands Deters et al. [33] (2017) proposed a machine learning inversion method based on six-year meteorological data and PM2.5 pollution data to solve the problem of urban particulate matter pollution diffusion. The test and analysis show that the inversion accuracy of this method is better than that of stable weather under strong wind or high precipitation, because under stable weather, the influence of underlying surface on PM2.5 pollution diffusion is far greater than that under extreme environment.
Julien Brajard et al. [34] (2020) of Sorbonne University in France proposed a hybrid data assimilation method based on the combination of Kalman filter and neural network to solve the problems of large noise impact in observation data and low accuracy of pollution prediction model, and carried out numerical experimental analysis through Lorenz 96 model. Compared with Kalman filter, the hybrid method can not only shorten the calculation time by two times, but also ensure the accuracy of inversion and prediction decline steadily with the increase of observation noise.

Application Platform of Nuclear and Chemical Hazards STI
In Europe and the United States, the inversion of nuclear and chemical hazard source terms developed to the peak in the 1970s and 1980s, and the technical research system was basically formed [35] . Since the 1990s, with the development of relevant basic disciplines, inversion technology has been integrated with numerical analysis of micro meteorological field, pollutant diffusion of complex terrain, virtual environment simulation, etc. many countries, such as the United States, the United Kingdom, the Netherlands, Sweden and other countries have successively established comprehensive emergency response systems including leakage source model, wind field model and diffusion model, and have a variety of evaluation software systems, which have been widely used in the international business, and have reached a mature degree of industrialization. The most representative systems are HG, NARAC and SAFER system. HG system [36] mainly includes thermodynamic model, escape model, plume jet model, heavy gas diffusion model, long-distance diffusion model and so on, which can evaluate the diffusion of gas and liquid or the two-phase release of multi-component mixture. It is used for gas escape, flash evaporation, evaporation tank, heavy gas diffusion, pure diffusion and UF6 gas diffusion with chemical reaction. NARAC system [37] can simulate complex flow field, detailed particle diffusion, dry and wet deposition process on various spatial scales, including local and regional meteorological prediction, diffusion model and nuclear explosion settlement model, and can simulate and analyze the leakage and release of nuclear and chemical hazards in complex environment. SAFER system [38] can handle various types of releases, including instantaneous, continuous, transient flows, ground level, uplift releases, and low or high momentum jets.
In addition, the British NAME system [39] can retrieve and predict the instantaneous or continuous time air concentration, including the concentration, deposition and dose of radioisotopes, by tracking the three-dimensional trajectory of fluid particles and calculating the air concentration by Monte Carlo method, simulating the medium and long-term transport and deposition of pollutants. GASMAL [40] , a decision support system developed by the Netherlands Institute of Applied Sciences (TNO), combines calculation speed, graphic display and database information to enhance the fast decision-making, which is crucial for chemical accident emergency, and ensures the timeliness of emergency response.

Table1.
Nuclear and chemical hazards emergency response system

Enlightenment to my development
To sum up, the research on the inversion of source terms of nuclear and chemical hazards in foreign countries has been quite large. From the perspective of application platform, some countries such as the United States, the United Kingdom, the Netherlands and other countries have developed relatively mature inversion and diffusion prediction platforms; from the perspective of inversion algorithm, whether it is Euler principle, Lagrange principle or deep learning, each method gradually presents the trend of integration and integration.

Deep Learning Will Become a New Research Hotspot
Compared with the optimal interpolation method, genetic algorithm, Kalman filter and other methods, deep learning has obvious advantages in improving the accuracy and speed of hazard inversion, which brings new research ideas for source term inversion. However, deep learning requires high training samples and computational performance, and it is difficult to learn complex inversion samples, which is still in the primary research stage. The inversion algorithm based on deep learning, fourdimensional variation and ensemble Kalman filter is bound to become an important means to solve the inversion problem in complex environment.

Inversion of Multi Pollution Sources Will Become a New Research Focus
According to the research status, domestic and foreign scholars mainly focus on single point source continuous or instantaneous release inversion, and there are few studies on multi pollution sources term inversion. Some research institutions in the United States have carried out experimental research, and the domestic is still in the stage of mathematical derivation. However, in the actual combat environment, there is no single point source pollution. The inversion research of continuous, instantaneous and even mixed release of multiple pollution sources is more important for hazard prediction and protection decisionmaking. Therefore, the source term inversion of multiple pollution sources will be the focus of practical application in the next step.

Qualitative and Quantitative Mixed Inversion Will Become a Research Difficulty
At present, the information obtained by all kinds of reconnaissance and monitoring equipment in the battlefield environment is diverse and in different formats. Most of the monitored data are qualitative data based on sensitivity alarm, and the collection and acquisition methods of quantitative monitoring value data are relatively less. At the same time, there are many uncertainties in the acquisition of information such as pollution species, hazard source intensity, attack mode, meteorological field and so on, which is difficult to guarantee the error of the result. Therefore, it is a difficult problem to put forward the qualitative and quantitative hybrid inversion technology suitable for battlefield environment.
In addition, the scientific layout of monitoring equipment is of great significance to reduce redundant interference and improve inversion accuracy, but there are relatively few studies that can be used for reference, which is also a key link that needs to be strengthened.