Detection of Dynamic Phenomena Associated with Underground Nuclear Explosion Using Multiple Seismic Surveys and Machine Learning

The application of an active seismic method for detecting the source location of an underground nuclear explosion (UNE) is an ongoing field of research. The objective of active seismic in On-Site Inspection (OSI) is to detect the static signatures such as the cavity created by the UNE. Along with characteristic static signatures, UNEs produce dynamic phenomena such as groundwater mounding, which gradually revert to pre-test conditions. These dynamic phenomena are observable for an extended period, even up to several decades. The magnitude of these phenomena is prominent near the source origin and results from the redistribution of residual energy, such as pressure, temperature, and saturation. These dynamic changes in sub-surface rock and fluid properties will affect the seismic property of the rock, resulting in changes of P-wave velocity. These changes can be detected by using an active seismic survey. This study highlights the potential of using time-lapse seismic to identify ground zero by monitoring post-explosion variation in the seismic signature. Time-lapse seismic, also known as 4D seismic, is a well-known technology, used in the oil and gas industry for several decades for petroleum production monitoring and management. It involves taking more than one 2D/3D survey at different calendar times over the same reservoir and studying the difference in seismic attributes. This study investigates the characteristic dynamic phenomena associated with the UNE and their impact on the emplacement rock’s seismic property. Groundwater mounding (GWM) is one of the phenomena with a high gradient of dissipation during the initial days immediately after the explosion. We look at the impact of GWM variation on seismic P-wave velocity and discuss the potential of using time-lapse seismic for its detection. The challenges of implementing time-lapse seismic, such as non-repeatability, seasonal variations and time constraints, are discussed. A frequent seismic monitoring survey method (time-lapse seismic) is proposed to monitor rock and fluid properties changes due to the post-UNE dynamic phenomena. Due to the time constraint for the OSI activity, conventional time-lapse seismic processing would not be suitable. Therefore, a machine learning-based 4D detection workflow is presented. The near-real-time 4D detection workflow using machine learning can be implemented during the OSI to identify the source location or ground zero.


Introduction
The United Nations set up the Preparatory Commission for the Comprehensive Nuclear-Test-Ban Treaty Organisation (CTBTO) in 1996 to build verification regimes and monitor activities related to nuclear weapon proliferation. When it comes into force, the Comprehensive Test Ban Treaty (CTBT) will prohibit the testing of nuclear weapons.
OSI is one of the verification regimes under the CTBT. After the Treaty comes into force, any member state can request an OSI to ascertain whether a nuclear explosion has occurred in violation of the Treaty. As the most invasive inspection, this is considered the ultimate deterrent for nuclear testing. Therefore, it adds confidence to the verification regime. It entails a thorough search of a designated inspection area in the Inspected State Party (ISP) territory, utilising the Treaty-defined inspection methodologies (CTBT, 2012). The requirements to investigate possible violation of the CTBT during an OSI are unique and depend on the time, the inspection area, and techniques used. The objective of an OSI is to identify the location of the triggering event and gather evidence of a possible Treaty violation by detecting the anomalous characteristic signatures associated with the nuclear explosion.

Characteristic Signatures of UNE
A UNE releases a tremendous amount of energy in a very short time. The resultant high pressure and temperatures produce characteristic static and dynamic signatures. Geological deformation, such as a cavity, rubble chimney, rock compaction, and fracturing, collectively known as the 'cavity system', is the characteristic static signature associated with UNE (Adushkin & Spivak, 2015;Germain & Kahn, 1968). The dimensions of these deformations depend on the explosive yield strength and geology of the surrounding rock. The cavity has an average radius of about 7-10 m/kt 1/3 , where 'kt' is the yield strength expressed as kiloton. The rock deformations are observed up to six times the cavity radius. The P-wave velocity of the cavity is about 50-80% lower than the initial background velocity (Adushkin & Spivak, 2015;Derlich, 1970).

The Objective
During an OSI campaign, reconnaissance surveys using scientific methods, such as visual observation, multi-spectral imaging, air-borne gamma spectroscopy, gamma radiation monitoring, and various other methods are carried out in a large area (about 1000 km 2 ) to isolate suspected areas for more detailed investigation. Once the likely ground zero zones are isolated, more specific and intrusive techniques such as active seismic surveys are brought in for a more thorough examination of the suspected area to confirm the location as ground zero by detecting potential UNE related phenomena. An active seismic method can be potentially employed in OSI to detect static phenomena such as cavity systems caused by the UNE. However, successful imaging of the cavity system may not be a conclusive indication of an underground nuclear test. There could be other naturally occurring features such as karst and cavity present in the area, which would produce a similar seismic signature. Therefore, it is essential to detect dynamic phenomena typically associated with UNE to confirm the source.
The scope of this study is to determine the feasibility of using time-lapse seismic during an OSI to identify ground zero by monitoring post-explosion dynamic phenomena. This study presents a 4D detection routine using unsupervised machine learning, which can be implemented on a non-crossequalized 4D seismic dataset. The suggested technique is envisaged for field deployment to locate the zone of 4D change in a seismic section.

Dynamic Phenomena of UNE
Similar to the characteristic mechanical changes of rock medium, there are certain characteristic dynamic phenomena associated with UNEs. These phenomena persist for an extended period, such as for a few months to even several decades. The magnitude of these dynamic phenomena decrease with distance from the cavity and time. They are most prominent near the source origin, while their spatial and temporal behaviour depends on the explosive yield strength and surrounding rock geology (Laczniak et al., 1996). Such changes in residual energy would affect the elastic and seismic properties of the rock layers, just as in producing oil and gas reservoir rock. The rock property changes could be monitored using time-lapse seismic (Johnston, 2013). The main 'dynamic phenomena' observed are temperature, pressure, and saturation variations.
The UNE generates enormous heat, which melts the host rock and device materials surrounding the test location and forms a mass of glass that slowly cools over time (Carle et al., 2003). The melt glass debris retains the majority of the chemical makeup of the host rock, but includes the heavier radionuclides produced by the test. If the test is conducted under or within 100 m of the water table, then as the cavity starts cooling, groundwater infills the cavity. Test heat is the main factor which maintains the groundwater flow within the cavity. It drives the convection cells that transport radionuclides upward into the rubble chimney above the collapsed cavity (Carle et al., 2003;Maxwell et al., 2000). Thermal conduction and groundwater flow are the main cause of residual thermal energy dissipation at the test site. Depending on the geology, the rate of test heat dissipation varies for different test sites. Due to the relatively slow thermal conduction rates of rock, the amount of time test heat remains in a system is mainly determined by the groundwater flow rate, which is a function of the region's permeability (Carle et al., 2003;Maxwell et al., 2000;Pawloski et al., 2001).
UNE causes an immediate rise in pressure within the cavity. It is attributed to the combination of explosion shock wave and sudden decrease in cavity volume caused by cavity collapse. The studies carried out in CHESHIRE, GREELEY, and ALMENDRO tests show that the high initial cavity pressure drops below hydrostatic pressure within minutes and dissipates completely by about 20 days (Carle et al., 2003). The cavity pressure then slowly rises as the groundwater accumulates in the cavity, due to drainage from surrounding rocks.
Explosion-induced groundwater mounding is another significant phenomenon associated with UNE. It has the most significant gradient of dissipation during the first 100 days period, which could produce detectable changes in the seismic properties (Knox et al., 1965).

Changes in Seismic Property Due to Groundwater Mounding
Underground explosions, particularly those that occur under or within 100 m of the water table, alter the hydrologic conditions in the surrounding area (Carle et al., 2003). The water level fluctuations are caused by changes in permeability, interstitial fluid pressures, mineral transformations, temperature, and water composition (Pohlmann et al., 1999).
Evidence indicates that rock compaction caused by shock waves may account for 30% of the cavity volumes following the explosion (Knox et al., 1965). This results in the compaction of pores, which were already saturated before the explosion, producing a stress field that drives the water from those pores into both pre-existing and explosion-induced fractures. This process creates a large groundwater mound (GWM). As the cavity collapses, forming a rubble chimney, a central depression in the groundwater mound is formed (Fig. 1). The central depression is attributed to the increase in permeability and porosity of the chimney region and very high test-heat from the glass melt zone, which vaporises the water. Knox et al., (1965); Thordarson (1987) Garber (1971); Burkhard and Rambo (1991), and many others have reported that, for tests carried out below the water table, post-shot water levels within the cavity and chimney were hundreds of feet lower than pre-test levels. This is because of water draining away through fractured rock and residual test heat in the cavity that vaporises the water. As the glass melt zone starts cooling, the water starts to infill the cavity, and slowly the water level rises within the cavity. The test heat drives convection and maintains the hydraulic gradient associated with the GWM, and results in a groundwater flow field that spreads over a large region (Carle et al., 2003;Knox et al., 1965;Pawloski et al., 2001).
Outside the chimney region, water levels have been hundreds of feet higher than preshot levels, and the rise varies according to distance from the explosion point and rock type. Measurements taken during and after a few tests indicate that a circular mound of elevated water levels appears immediately after explosion and then begins to disperse by drainage outward into adjacent rock and inward to the chimney region (Laczniak et al., 1996). The initial rise in water level beyond the rubble chimney has been attributed to the decrease in porosity as a result of inelastic compression of the rock (Burkhard & Rambo, 1991;Knox et al., 1965;Laczniak et al., 1996). Continued draining of water outward results in an increase in the water level, which migrates outward over time (Fig. 1). The size of the increase normally diminishes with increasing distance from the cavity and time. Field observations show that test effects may last a few months or may endure for several years before reestablishing equilibrium (Laczniak et al., 1996). Several post test observations indicate that the magnitude of GWM is highest near the test location and the rate of dissipation is as high as 60 cm/day during the initial 100 days following the explosion.

Spatial Extent of Groundwater Mounding
The observed spatial extent of the post-explosion groundwater mound is proportional to the yield strength and geology. Knox et al. (1965) reported that a 35 kt Aardvark test caused a 200-m water level rise in an observation well at a scaled distance of Vol. 180, (2023) Detection of Dynamic Phenomena Associated with Underground Nuclear Explosion 1289 94 m/kt 1/3 . Forty days after the explosion, the water level remained 17 m above the initial hydrostatic level. According to Chu (1986), the maximum impact limit of groundwater rise was a scaled distance of 900 m/kt 1/3 . Adushkin et al. (1993), reported that a 20 kt UNE caused a water level rise of 20, 12, 4, and 3 m at scaled distances of 295, 185, 370, and 580 m/kt 1/3 . Groundwater mounding has been documented for many of the underground nuclear detonations carried out in the US (Charlie et al., 1996). A linear regression analysis of the experimental underground explosion data, in which the maximum distances of groundwater table rise or residual pore water pressure increases were known, shows that an underground explosion can induce groundwater level to rise up to a scaled distance of 879 m/kt 1/3 (Charlie & Doehring, 2007;Charlie et al., 1996).
The test heat reportedly has a very high magnitude and persists for a long duration but is limited to the cavity and rubble chimney. Within this region, the pressure also varies dynamically-first, it rises to a very significant level and then suddenly drops to below hydrostatic values and then slowly increases as water starts flowing in. In addition, chemical reactions due to radioactive material from UNE are also reported. This shows that the cavity and rubble chimney region has a very dynamic condition and behaves like a chemical reactor (Carrigan et al., 2016). This creates a non-linear zone within the cavity and chimney region (Agnew, 2020). However, the spatial extent of the cavity and chimney, where the dynamic non-linear changes are observed, is much smaller, compared to the spatial extent of groundwater mounding, which is the prominent phenomenon observable during OSI (for 1 kt explosion, the approximate size of the cavity is 15 m in radius and GWM extent is 1000 m). Therefore, in this study the effects of non-linear conditions at the cavity were not considered.

Effect of GWM on Seismic Property
A numerical modelling study was carried out to understand the impact of seismic property variation as a result of saturation change due to the groundwater mounding, specifically, the P-wave velocity change. The initial rock properties for this study were derived from the well log data of the Nevada test site, published in Wagoner (2014). The depth of burial of the yield for this specific test site was 350 m, and the geology at the target is mixed alluvium. A test scenario, where the UNE is carried out above the groundwater table but within the saturation zone is considered for this study. This would create a groundwater mounding as observed for the Aardvark test and explained in the study published by Knox et al. (1965). The explosion yield assumed for this study is 1 kt, which would produce a groundwater level rise for a radial distance of about 1 km (Charlie & Doehring, 2007). The magnitude of static rock property changes, such as changes in P-wave velocity and porosity, which is a direct impact of the explosion on the emplacement rock, is incorporated in the initial model in accordance with studies published by Adushkin and Spivak (2004). The change in seismic velocity due to GWM depletion was computed using the Biot-Gassmann Eq. (1) (Batzle & Wang, 1992), where K sat is the fluid-saturated bulk modulus, K dry is the dry bulk modulus of the medium, K o is the mineral bulk modulus, and K fl is the bulk modulus of the fluid (air and water) in the pore space. / is the porosity. K dry and K o were determined from literature values for the materials assumed in the models. K fl was calculated by the Reuss average Eq.
where K l and K g are the bulk moduli of liquid and gas. This equation shows that the fluid with the lowest bulk modulus (air/gas) has a significant impact on the effective fluid modulus. Hence, a little gas/air can have a significant effect on the seismic property. So in the case of a cavity created by UNE, the heated chimney that has steam or gas during the initial days will significantly affect the seismic property. Seismic P-wave velocity (V p ) is calculated using Eq. (3): where l sat is the saturated shear modulus of the rock frame, which is assumed to be equal to the dry shear modulus of the rock frame (l dry ), and q is the mineral density. The effect of groundwater mound variation on the seismic property is predicted for two instances, where for the first instance, the groundwater mound is higher than in the second instance. This assumption is based on the observations from various test sites that the groundwater mound dissipates as the calendar time increases (Carle et al., 2003). Figure 2 shows the change in longitudinal wave (V p ) velocity due to groundwater mound as computed by Eq. (3). The analysis indicates that groundwater mound variation produces about 30-40% change in primary velocity. Publications by Pohlmann et al., (1999) and Vincent et al., (2011) show that the gradient of pressure and thermal energy dissipation during the initial 100 days is minimal compared to groundwater mound dissipation. Therefore, pressure and temperature are kept constant for this analysis. The UNE creates a heterogeneous fractured zone surrounding the cavity. This is called 'the Halo'. The study by Bonal and Desilets (2015) shows that the magnitude of velocity change is even higher in the halo area. However, to keep the model simple, this study has not considered the halo effect. Even with the assumption of homogeneous saturation, the velocity change observed is about 40%, which can be detected by time-lapse seismic (Johnston, 2013).

Time-Lapse Seismic
Seismic imaging of UNE static signature such as a cavity system with an active seismic method is very challenging. Even if the seismic imaging produces results similar to a cavity system, it will still not be conclusive evidence of sub-surface explosion. Any naturally occurring cavity/karst or gas plume may generate a seismic signature similar to the cavity system. Therefore, it is essential to detect dynamic phenomena typically associated with UNE to confirm the source.
Time-lapse seismic is an effective method for detecting the rock property changes due to dynamic phenomena. Time-lapse seismic, also called 4D seismic, has been successfully used in the oil and gas industry for petroleum production monitoring and reservoir management for the last few decades (Calvert, 2005;Campbell et al., 2015;Johnston, 2013). Due to production, pressure and fluid saturation, changes lead to changes in elastic properties (i.e., bulk modulus and density) and corresponding changes in P-and S-wave velocities at the reservoir level. These changes affect the seismic response and are detectable using 4D seismic (Nguyen et al., 2015;Sambo et al., 2020). Time-lapse seismic involves taking two or more 2D/3D seismic at different calendar times over the same reservoir and computing Vol. 180, (2023) Detection of Dynamic Phenomena Associated with Underground Nuclear Explosion 1291 the difference between the two seismic observations. The critical assumption is that acquisition parameters, including receiver locations and data processing parameters, are constant. Thus, the anomalies observed between the two seismic surveys are assumed to be attributable to changes in the reservoir. Time-lapse seismic has the potential to be an effective method to detect rock property changes due to UNE related dynamic phenomena. Here we investigate the prospect of applying time-lapse seismic during OSI for detecting sub-surface changes due to groundwater mounding. In an ideal condition, the difference between seismic images of surveys acquired with the same parameters at different calendar times will highlight changes at the sub-surface target. However, in the real field, several factors cause non-repeatability between surveys. Therefore, careful processing is required with bespoke parameters to highlight the 4D difference due to changes in sub-surface target. Repeatability between the base and monitor survey is key to successfully detecting the sub-surface changes. Repeatability is commonly measured as normalized root-mean-square (NRMS) difference. The NRMS is the RMS amplitude of the difference between the base and monitor normalized by the average RMS amplitudes of the base and monitor (Johnston, 2013).
where the RMS operator is defined as The summation is over N number of samples x i in time window. NRMS = 0 for perfectly repeatable data, and for an anticorrelated base and monitor seismic, the NRMS = 2. NRMS is sensitive to phase and amplitude differences, time shifts and noise.

Factors Affecting Repeatability in Land 4D Seismic
A seismic image is a function of the propagation of seismic waves. Several factors affect the propagation path between base and monitor surveys. The difference in propagation path affects the repeatability. The most significant repeatability challenges on land are lateral variations in near-surface velocity. Bakulin et al. (2012) showed that increasing receiver depth improved image quality and repeatability. In this 4D study, four 2D survey lines were recorded with receivers placed at depth 0, 10, 20 and 30 m, respectively, for each survey. The best-stacked section was obtained for receivers buried at 30 m depth. This is because, as the receivers are buried deeper, it creates less scatter from the near-surface and improves the imaging. Table 1 below shows the comparison of repeatability for different source-receiver depth configurations. A numerical study was carried out to understand the impact on repeatability for different source-receiver depth configurations. Three different source-receiver depth configurations were considered-(a) where both source and receiver are at the surface, (b) where source is at the surface and receivers are buried at 20 m depth, and (c) where both source and receivers are buried at 20 m depth. It is very evident from the results displayed in Fig. 3 that the repeatability is excellent when both source and receiver are buried. Burying the source and receivers helps overcome the non-repeatability caused by near-surface and acquisition geometry errors. However, for OSI, we are restricted to the surface laid source and receiver option, which provides relatively poor repeatability.
Even if sources and receivers are precisely relocated, the near-surface can change with time, thereby introducing non-repeatability (Bakulin et al., 2014;Johnston, 2013). Bakulin et al. (2014) studied the repeatability of 11 repeat 2D surveys. They observed that non-repeatability (NRMS) increases as the time between the monitor surveys and baseline survey increases. In contrast, non-repeatability values are small when consecutive surveys are compared, i.e. when the times between surveys are short. This shows that frequent seismic observation will have less impact due to seasonal variation and have good repeatability.

Character of GWM 4D Signature in a Heterogeneous Cavity System
In addition to the factors that affect the repeatability described above, the scattering due to the heterogeneous background (i.e., heterogeneous cavity system) impacts the character of the 4D signature. We carried out a modelling study to understand the nature of the 4D signature due to groundwater dissipation in a perfectly homogeneous background with and without a scattered cavity system (Fig. 4). All conditions, including the near-surface, were considered perfectly repeatable in this study. The groundwater mound dissipation rate used for this study is as presented by Knox et al. (1965) and shown in Fig. 5.
The comparison shows that as the magnitude of the 4D signal increases with time due to groundwater dissipation, the non-repeatability of the 4D signal in a scattered medium significantly increases. Therefore, this would lower the chance of detecting and correctly interpreting the 4D signature. This implies that it would be even more challenging to detect the 4D signature in the presence of a heterogeneous cavity system in real field conditions. Therefore, we would require a smarter detection method for OSI, such as using a machine-learning algorithm to detect the 4D signal in a heterogeneous background.

Methodology
To detect the time-lapse seismic signature due to dissipation of groundwater mounding during OSI, we propose an active seismic method that involves frequent seismic observation and 4D detection workflow using machine learning.

Frequent Seismic Monitoring
Frequent or continuous seismic monitoring is when the time-lapse seismic data is acquired either continuously or on a very frequent time interval, i.e., daily. This technology, known as 'SeisMovie TM ', was previously used for reservoir monitoring, where the receiver arrays were permanently buried, and piezoelectric vibrator sources were used (Cotton et al., 2013). Since the source and receivers are buried, and monitoring is frequent, this type of acquisition provides exceptional repeatability and a very high signal-to-noise ratio. This frequent seismic monitoring system routinely achieves a sensitivity of 0.05 ms  (Calvert, 2005). During the OSI, when the potential ground zero zone is isolated, the seismic monitoring should be carried out daily when the gradient of groundwater mound dissipation is very high, i.e., during the initial 100 days following the explosion, to capture the seismic property variation. Frequent 4D seismic monitoring would effectively detect the rock property variations due to the groundwater mound dissipation. Additionally, as discussed in Sect. 3, frequent observation will be less impacted by seasonal variation and therefore, it will have relatively better repeatability. The synthetic example shows the impact of a heterogeneous cavity system on the 4D signal. In this example, no changes in the near-surface condition are assumed. a 4D difference for a scenario without a cavity, and b 4D difference with a heterogeneous cavity. c Comparison of NRMS error computed at the target, indicating that the non-repeatability increases as the groundwater mound depletes. This shows that the scattered target medium is highly non-repeatable, and consequently, the 4D signature in a heterogeneous cavity system will be challenging to detect 1294 S. Mathew et al. Pure Appl. Geophys. Figure 5 shows the groundwater mound dissipation curve and the potential period of detection using time-lapse seismic. The groundwater mound dissipation curve is adapted from Knox et al. (1965).

4D Detection Workflow
Time-lapse seismic data has to be carefully processed using the same processing workflow to detect the 4D signature. Typical land 4D seismic data processing involves bespoke workflows and careful parameterization to highlight the 4D signature. These conventional 4D processing workflows often have a high turnaround time and, hence, will not be appropriate for the OSI where near real-time processing is required (Jervis et al., 2018). We present a workflow using self-organizing maps (SOM) for fast detecting 4D signatures on frequently acquired multiple monitor land 4D seismic data. SOM is a neural network algorithm based on unsupervised learning (Kohonen, 1990). It is widely used for pattern recognition, classification, and data analysis. It captures the information residing in (multiple) input attributes by reorganizing data samples based on their topological relation. SOM is an excellent facies analysis/classification tool, commonly used to address static problems such as seismic interpretation (Ramirez et al., 2012). The proposed workflow shows that SOM can address dynamic problems.
Researchers have shown that SOM can be used for detecting anomalies (Hormann & Fischer, 2019). SOM has been used to monitor the condition of mechanical and electronic systems where the minimum quantization error (MQE) of a test data observation to the SOM has been used as an indicator to evaluate the system's health. The proposed workflow (Fig. 6) consists of the following steps: (1) generate instantaneous attributes from multiple monitor seismic data, (2) compute 4D attribute of these instantaneous attributes, i.e. amplitude vs time (AVT) for these attributes and calculate intercept (A), gradient (B) and the product of intercept and gradient (A 9 B), (3) normalize and convert the attribute to training vectors, (4) train SOM and find best matching unit (BMU), which is the neuron closest to the training data observation, (5) compute the MQE and determine the threshold for the outlier. The non-linear dimensionality reduction property of SOM would help to map the progressively varying 4D change from the input 4D attributes provided. In the SOM output map, the progressively varying zone of 4D will be highlighted, while other regions will not be highlighted.
The SOM is trained iteratively until all the weight vectors of the map are grouped into clusters Figure 5 curve shows the rate of groundwater mound dissipation. The graph is adapted from the work of Knox et al., (1965). The seismic section shows the best period for the time-lapse seismic to be carried out which is during the initial 150 days (about 5 months) after an explosion. Day 0 is the time when UNE is conducted Vol. 180, (2023) Detection of Dynamic Phenomena Associated with Underground Nuclear Explosion according to their distance. After the training process, the MQE is computed, which is described as the distance between the input data observation U(t) and the BMU of the SOM. The MQE is calculated as in Eq. (6): where : j j j j 2 F is the Frobenius norm. MQE indicates that the input data observation belongs to a space further away from the BMU; therefore, it will be an outlier (Tian et al., 2014). In this workflow, the MQE highlights the region in the seismic section that is progressively changing, which is the time-lapse change at the target.

Numerical Results
The proposed workflow was tested on synthetic datasets. Four time-lapse seismic monitors were modelled, simulating frequent 4D seismic observation over depleting groundwater mound targets. Two sets of such synthetic data were generated-one with a heterogeneous cavity system and the other without one. Velocity models both with and without the cavity system, which incorporates velocity change due to groundwater mound dissipation, are shown in Fig. 7. Random noise was added in the first 20 m depth with a mean ranging from 35 to 85 m/s and standard deviation from 125 to 185 m/s to simulate non-repeatability of land seismic due to varying nearsurface conditions. At the target level, which is at a depth of 500 m, velocity change corresponding to the saturation change due to the groundwater mound was introduced. Figure 8 shows the velocity difference between each monitor and the base survey.
2D synthetic seismic shot gathers were generated using acoustic finite-difference modelling. The shot and receiver spacing for the 2D survey was 20 m and 10 m, respectively. Both shots and receivers were modelled at the surface. There were 121 shots generated for each monitor with a 25 Hz zero-phase Ricker wavelet as source signal. The generated shots had a record length for 2 s with a 2 ms sample interval and a maximum offset of 1.2 km. The shot gathers were then processed, migrated using a reverse time migration algorithm, and stacks were created. The 4D difference sections for three monitors are displayed in Fig. 9a.
Non-cross-equalized migrated stack sections of monitor surveys were given as input to the 4D detection workflow. Instantaneous attributes, such as quadrature amplitude, envelope, instantaneous phase, instantaneous frequency, and sweetness, were generated for each monitor stack. 4D attributesintercept (A) and gradient (B)-are determined by computing the attribute vs time for each instantaneous attribute. Normalized gradient (B) and product of intercept and gradient (A 9 B) are then given as input vectors for SOM training. After the SOM training, MQE is computed. The quantization error is used to detect an outlier in training data. Since the SOM is trained using time-lapse attributes, the outlier highlights regions of 4D changes. A threshold is applied to the MQE to highlight the outlier (Fig. 9b).
The result displayed in Fig. 9b shows that the SOM workflow successfully detects the zone of 4D change on the non-cross-equalised time-lapse dataset. The workflow could be applied iteratively, i.e. as and when a new monitor is acquired and thus confirm the zone of 4D change as each new monitor is added to the equation. The main advantage of the SOM workflow is that it does not require a cross-equalized dataset for 4D detection.

4D detection in a Heterogeneous Target
The presence of a heterogeneous target, such as a heavily fractured heterogeneous cavity system, poses an additional challenge in detecting underground explosion-related time-lapse change. The scattering from the heterogeneous cavity system alters the character of the 4D signal, as discussed in Sect. 3.2. The 4D detection workflow was tested on a synthetic 4D dataset with a heterogeneous cavity system as the target background.
Velocity models, seismic 4D difference section and workflow output for targets with and without heterogeneity are compared in Fig. 10. The result shows that the SOM workflow clearly detects the zone of 4D change in both instances. As expected, the presence of a heterogeneous cavity system creates a scattering of the seismic signal, which distorts the 4D signature. In this instance, the 4D detection workflow highlights the region of 4D change.

Discussions and Conclusions
In addition to the static phenomena, a UNE produces characteristic dynamic phenomena that are prominent near the source origin and observable for an extended period. These dynamic phenomena, such as groundwater mounding, have a very high gradient of dissipation during initial days and can produce changes to elastic properties of rock, which can be detected using active seismic. In this study, a potential workflow is presented employing time-lapse seismic survey for locating ground zero during OSI by detecting the UNE related dynamic changes. A frequent time-lapse seismic monitoring is proposed in Figure 7 Velocity models without heterogeneous cavity used for generating synthetic time-lapse seismic dataset. Base is the velocity model of first survey in which effects of groundwater mound is incorporated. Monitor 1, 2 and 3 are velocity models incorporating velocity change due to depleting groundwater mound at later calendar times. Colour bar shows P-wave velocity in m/s Figure 8 shows the difference in velocity models between base and corresponding monitors highlighting the velocity perturbation incorporated. Random velocity change is incorporated at the shallow to represent seasonal variations expected and gradual velocity change is included at the target due to groundwater mound depletion Vol. 180, (2023) Detection of Dynamic Phenomena Associated with Underground Nuclear Explosion 1297 which the sources and receivers for the acquisition will be surface laid. The rate of change of the UNErelated dynamic phenomena (especially groundwater mounding) is very high during the initial hundred days after the explosion. Therefore, the proposition is to conduct frequent (daily) seismic monitoring during this period. At least three monitor surveys are required to effectively confirm the 4D target. Detecting time-lapse change in land 4D data has been a long-standing problem due to poor repeatability and low signal-to-noise ratio. Due to the time constraint for OSI, it will not be possible to carry out conventional 4D seismic processing to cross-equalise multiple monitor surveys to enhance the 4D signal. Therefore, a fast and efficient 4D detection workflow is essential to detect 4D signatures from non-crossequalized input data. The proposed 4D detection workflow based on the Self-Organizing Map highlights the regions of dynamic changes in the timelapse seismic section. The technique is aimed at field deployment, such as during an On-Site Inspection. We have shown that the SOM workflow highlights the location of 4D changes, otherwise not possible from basic processing alone, by separating dynamic changes at the target from near-surface variation. The synthetic studies presented show that the proposed workflow can detect the zone of 4D change even when a scattering cavity system is present. During OSI, the region of interest would be a cavity system (heterogenous) where the workflow highlights 4D changes. The workflow is suitable for 4D detection, which sufficiently fulfils the objectives of an OSI campaign.
Survey parameters are site-specific and depend on the objective and target depth. Based on the general objective of OSI, receiver spacing of 5 m is desirable for denoising and achieving a good signal-to-noise ratio, and 25 m shot spacing would provide the fold coverage required for imaging the target. The logistical challenge in deploying active seismic survey equipment will be a factor when implementing this method for OSI. In order to deploy this technique, a five-member team would be required. Time-lapse seismic requires seismic sources with very good repeatable source signatures. Regarding receivers, the new nimble nodes or nodal receivers with wireless data transfer capability are very well suited for OSI. These receivers require far fewer resources to deploy 4D difference (base -monitor) of processed seismic data is shown in a. The 4D difference for monitors 1, 2 and 3, shows that it is challenging to infer 4D signal due to the non-repeatability caused by near-surface variations. b The 4D detection SOM workflow output overlayed on 4D difference section. The result highlights the zone of 4D (arrow marked), whilst differences elsewhere due to non-repeatability are not highlighted 1298 S. Mathew et al. Pure Appl. Geophys. than conventional receivers and can be used in any terrain. They can continuously record seismic data for an extended period and therefore are helpful for the time-lapse seismic monitoring as proposed in this research.