Ambient Noise Tomography and Machine Learning Models to Reveal Geothermal Structure in the Taupo Volcanic Zone

doi:10.21203/rs.3.rs-3339951/v1

Download PDF

Article

Ambient Noise Tomography and Machine Learning Models to Reveal Geothermal Structure in the Taupo Volcanic Zone

https://doi.org/10.21203/rs.3.rs-3339951/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

To reveal potential geothermal fluid flows and temperature distribution in the Taupo Volcanic Zone, North Island of the New Zealand Hikurangi subduction zone, we analyzed seismometer and well log data. First, we extracted Rayleigh phase velocity dispersion curves from ambient noise cross-correlation analysis and then estimated S-wave velocity models through 3D surface wave tomography. Second, we constructed the 3D temperature model derived from our machine learning models using 3D velocity and temperature log data. Faulted/fractured zones, which can host fluids or magma, led to significant reductions in S-wave velocity within the subsurface. As a result, our S-wave velocities were lower in existing geothermal reservoirs and through flow pathways (i.e., active fault zones), particularly within the NE-SW directional structure toward Lake Taupo. The most suitable 3D temperature model (or others) was obtained based on the Gaussian process regression model, compared to other models in all machine learning algorithms. High-temperature areas ranging up to 300 °C or more are consistent with flow paths through the structure. Our approach could contribute to the unrevealed geothermal structures and fluid flow pathways in this region.

Earth and environmental sciences/Solid earth sciences

Physical sciences/Energy science and technology

Physical sciences/Engineering

Geothermal fields are predominant in regions where tectonism, magmatism, and volcanism are strongly interacted; therefore, they are commonly located in heterogeneous geological environments^1–3. There are a variety of geothermal environments linked to volcanic-magmatic sources in different tectonic environments. In the Taupo Volcanic Zone, New Zealand, geothermal-volcanic systems are closely linked to the volcanic rift zone related to the Hikurangi oblique subduction (Fig. 1). The central region of the Taupo Volcanic Zone is known as a complex caldera system including, for example, Whakamaru, Ohkuri, and Reporoa calderas, which possess many geothermal systems like Wairakei-Tauhara, Rotokawa, and Ngatamariki geothermal systems. Similarly, ⁴ studied the Silali – Paka volcanic area (20km apart), Kenya, where geothermal-volcanic systems are related to continental drift. Regarding a study area from Kuju volcano, Japan⁵, geothermal-volcanic systems are related to the graben-depression zone caused by tectonic lines around the center of Kyushu Island. Therefore, geophysical exploration in such heterogeneous tectonic-volcanic settings is required for geothermal reservoir characterization.

In the geothermal-volcanic environments, hot fluids in fracture or porous rock changed and altered host rocks, leading to clay mineral-dominated formations¹. In addition, tectonic dynamics resulted in magma- volcanic eruptions produced a wide range scale of fractures and different litho-stratigraphy. Therefore, in such environment, high contrast of seismic velocity or other physical properties (e.g., gravity, resistivity) between non-fractured and fractured rock or different types of rocks can be observed⁶. Seismic, gravity, and electromagnetic analysis among geothermal exploration techniques are used to map dominant fracture zones of accumulated hot fluids^5,7–9. The seismic approach can be used to investigate natural resources such as oil/gas reservoirs, geothermal energy reservoirs, or ore mineral deposits⁷. However, these resources are hosted by porous and fractured rock in different environments. Oil/gas hosted sediment velocities are mainly influenced by their fluid saturation/content rather than fracture because the sedimentary basin environment/formation (e.g., clastic sediments) is less complex, which is suitably investigated using the seismic reflection method^7,10. In geothermal environments, ground water has always existed, and the main contrast velocity is among fractured, non-fractured rock, altered or non-altered rock⁶ due to the dynamical deformation of tectonic-magmatic-volcanic mechanisms. For mineralized ore environments (i.e., hydrothermal/geothermal mineral deposits), the ore minerals as magmatic fluids filled into fractures as ore forming processes^11–13. These ore mineral-filled fractures tend to cause high velocity (or high gravity or magnetics) compared to the surrounding rock¹⁴. In this regard, this method/approach could also be suitable for application in mineral exploration.

There are more than 70% of countries on the world map that can possess and develop geothermal energy resources¹⁵. In the current situation, geothermal energy exploration and development still face challenges because geothermal energy resources are located in complex geological environments of active tectonic zones, leading to high exploration costs and associated risks. Therefore, in this study, we aim to address these issues by utilizing low-cost ambient noise data and enhance imaging resolution through advanced velocity analysis for geothermal energy exploration and development. Furthermore, we used only limited seismometer data for velocity analysis integrated with seven borehole data to investigate large-scale geothermal structures (e.g., 3D temperature), and we assess the feasibility of reservoir-scale exploration. First, we implemented ambient noise tomography then using a 3D velocity model and well-log data to estimate spatial physical rock properties (i.e., 3D temperature) via the machine learning approach.

Ambient noise tomography or surface wave tomography, depending on natural sources, is based on cross-correlation/seismic interferometry between two seismometer stations to retrieve the propagation of the surface wave. Surface wave dispersion curves are first estimated by surface wave analysis^16–20, and then the 3D S-wave velocity structure structures are derived through either conventional or direct surface wave tomography techniques^19,21,22. Conventional surface wave tomography is based on dispersion data inverted into 2D phase/group velocity profile by performing point-wise inversion of 1D S-wave at each grid, then combined to obtain a 3D S-wave velocity model. While a direct surface wave tomography, without a step of constructing phase/group velocity map, directly inverts surface wave dispersion traveltimes into 3D S-wave velocity structure²¹.

The machine learning or deep learning technique has two main parts (i.e., classification and regression model). There are many types of classification and regression models²³ and so far they have been applied to many fields of studies, e.g., for environmental forecasting (e.g., weather, fires, flood and water contamination, landslides)²³; for facial, voice recognition, or traffic classification^24,25; for predicting reservoir properties (e.g., saturation or pore pressure)^26,27. In this study, we considered using limited data (seismometer and well logs) to estimate the spatial geothermal structure (i.e., velocity and temperature). Additionally, we applied these integration methods and evaluated exploration scale strategies to choose the reservoir site. It is difficult to select reservoir sites for the volcanic-geothermal exploration stage. In this regard, we tried to solve this problem which we revealed and discussed in this paper. Additionally, we provided insights into understanding of shallow crustal structures and mechanisms.

Overview of study area. The Taupo Volcanic Zone lies within the North Island of New Zealand off the Hikurangi oblique subduction, where the Pacific tectonic plate is subducting beneath the Australian tectonic plate at a rate of ~4-5cm/year^28–31. Due to this oblique plate subduction (Fig. 1a), there are two major fault zones that produced abundant faults/fractures throughout this region³¹. The first fault zone is the Taupo Volcanic Zone, associated with the volcanic rift zone. The second fault zone is the strike-slip fault region that is apparently parallel with the Hikurangi subduction margin (North Island dextral fault system; NIDFS in Fig. 1). The Taupo Volcanic Zone’s formation could be attributed to past strike slip fault zone caused by oblique subduction, leading to the development of volcanic rift structure. Volcanic activity in the Taupo Volcanic Zone commenced ~2Ma. In the mid-crust, the mantle wedge melted from the plate boundary was expelled, contributing to numerous volcanic eruptions on the North Island of New Zealand over the past time^28,32,33. As a result, the present volcanic rift structures have been formed.

In the Taupo Volcanic Zone (e.g., central study area, Fig. 1a), the two major volcanic mechanisms have occurred. The older zone at 2 - 0.7 Ma had the main volcanism of andesitic magma, and the younger zone at 0.6 Ma - present has the major volcanism of rhyolitic magma, less contents of andesitic magma, then basaltic and dacitic magmas in minor volume²⁸. The younger period section has a rifting rate between 7 and 18mm/annum. The source of magmatism or volcanism is important due to the silica content, lithology, degree of temperature, and type of volcano. The present of sediment/volcanic successions for this shallow region are overlaid by clastic sediments, pyroclastic deposits, rhyolite, basalt, andesite, mixed/altered materials and cut through by many volcanic rifts/fluids^28,30. Due to such conditions and environments, many geothermal-volcanic systems have formed (Fig. 1). The Taupo Volcanic Zone is stretching over 200 km in length and up to 60 km in width, dimensionally due to vent positions and caldera structural geometry. The basement is graywacke, which form faulted-volcanic metasediment formations^11,34. The dominant magma-erupted type is Rhyolite. The abundance of magma fluids, fault-volcanic rift structure, high heat flux, and voluminous rhyolitic volcanism (a long rift arc) provide plentiful geothermal energy resources in this region. However, details of each geothermal volcanic system, geothermal exploration and development in this region can be found in previous studies^{3,13,28,30,32,35–37}.

We implemented two main approaches to estimate the 3D geothermal velocity and temperature structure in this region, in which we first applied a direct ambient noise tomography to all selected phase velocity dispersion data (Fig. 2), then we estimated a 3D temperature model derived from our 3D S-wave velocity model incorporating with borehole data (Fig. 3).

3D S-wave velocity model versus active fault distribution. We conducted a checkerboard test to check our 3D model inversion. The checkerboard test apparently recovered the 3D S-wave resolution of the anomaly grids (0.015^ox0.015^o) with a velocity perturbation of 20% as shown in Fig. 4, and geological features of the input anomaly can be apparently resolved in our study area (Fig. 4), in which the resolution is limited to depths shallower than 6 km below the surface (Fig. 2e). Our study reveals the shallow S wave velocity structure at different depths that reflects the geothermal structure in this region. The results (Fig. 5) show variations in velocity contrast that enable us to distinguish potential geothermal structure from surrounding rocks and understand how they are controlled through the dominant fracture systems or flow pathways (Figs. 5-6). Remarkably, we used 25 seismometer data to cover a large exploration area within the Taupo volcanic systems, known for hosting numerous geothermal systems^32,38. This region is characterized by complex caldera systems comprising of volcanic rift structures. These rift structure hold the potential for geothermal structure/reservoirs to host the fluids. The heat source was probably at a depth of > 5km from the surface^30,39, whose magmatic fluids mixed with shallow meteoric water can migrate convectively and conductively³⁰. The low S-wave velocity infers the fracture/fault zones that may serve as geothermal fluid pathway because the fracture of alter or non-altered rock caused a significant velocity reduction compared to non-fractured rock⁶.

According to our results (Fig. 5), we observed that at these various depths, the low velocity anomalies align with the geothermal systems/reservoirs along the flow pathways. Even with a large-scale study, we are still able to better understand the flow pathways and investigate new geothermal systems in the central study area (Fig. 5). This is important to locate a new reservoir and deploy dense seismometers in a future target area or for the reservoir scale. Therefore, identifying a reservoir scale from this study can contribute to reducing exploration cost and time, providing valuable geothermal exploration imaging/information for other geophysical studies/applications.

3D S-wave velocity structure versus seismicity distribution. Based on our results overlayed with the GNS seismicity catalog during 2000-2022 of depth from 0 to 8 km and known geothermal systems (Figs. 5-7), the seismicity distribution (≤ M_w 3.5) appeared to occur within areas of low velocity and high temperature and relevance to geothermal field development/production distributed over the central study area (gray shading areas in Fig. 5a; ⁴⁰). Such low velocity areas can indicate natural fluid flows and accumulations, which possibly act as a seismicity triggering mechanism. The Taupo region is a volcanically active zone consisting of seismogenic faults/fractures (Fig. 1; ⁴¹). Based on the study by³⁰, there is significant heat flux in this study area. Therefore, magmatic/geothermal fluids can be widely distributed within this shallow region. These fluid migrations or intrusions can change pore pressure and then generate crustal seismicity over this study area^42–44.

Additionally, the past supervolcano eruption (e.g., results as the big Lake Taupo) last occurred in 70Ka³⁹. This phenomenon can be caused by the enriched magmatic intrusions, which were supplied vertically and possibly horizontally by geothermal fluid flows. The migration of magmatic fluids cannot be rapidly moved because of the viscosity and structure of crystal mushes⁴⁵. However, fluid flows (magma mixed with groundwater through fracture systems) can horizontally move at ease toward Lake Taupo over a geological time (Figs. 5-8), which could contribute to fluid supply mechanisms related to past eruptions in this lake area ³⁹. In this regard, predominant magmatic intrusion/natural fluid flows as fluid influx and/or exsolved gases can accumulate high pore fluid pressure, which can influence fault slip triggering processes or incipient fractures/cracks⁴² in this region.

3D geothermal structure (e.g., 3D temperature distribution). We revealed flow pathways and low S-wave geothermal systems over this study area (Figs. 5-8). Another crucial factor is detecting the dominant fracture zone and evaluating other main reservoir rock properties such as spatial temperature, permeability, heat flux, etc. In this study, we tested/used the machine learning model based on seven local borehole data; therefore, we assessed the exploration scale rather than the detailed reservoir scale.

To predict the 3D temperature, we used various models as we described in Method Section (Fig. 3). The best suitable model is GPR, which we trained/validated using velocity (our 1D S-wave), depth, and temperature from seven well data⁴⁶ (Figs. 3a-b). As a result, we only obtained ~84% which can prevent our GPR model from being underfitted and overfitted, compared to other models in Table 1.

Table 1 The selected regression model and metrics for evaluating and predicting the 3D temperature model.

Models	RMSE	MSE	R-Square	MAE
GPR	27.5313	757.9723	0.8356	19.8406
Decision Tree	35.7006	1274.5370	0.7235	27.2537
Support Vector Machine	33.1585	1099.4886	0.7614	23.7207
Ensemble of Trees	32.1347	1032.6421	0.7759	24.9376
Neural Network	30.7417	945.0564	0.7949	23.4099
Kernel Regression Model	37.9056	1436.8341	0.6882	29.3366

Based on the results in Figs. 7a-b, we can obtain 3D temperature, in which potential fluid pathways correspond to high-temperature fluids. Temperature has also appeared to be rising in flow pathways (e.g., ~320^oC at 2km depth), compared to non-fluid flow pathways. However, we used limited seismometers and well logs due to our access to these data, future research with more comprehensive data coverage, including reservoir-scale studies and distributed exploration boreholes, could further improve accuracy. In the previous study⁵, we used dense seismometers to reveal a high-resolution geothermal structure on the reservoir scale of the Kuju volcano, Japan. The 3D high-resolution S-wave model can thus enhance predictions of precise 3D heat flux, permeability/fluid flows, temperature, and relevance to reservoir performance and monitoring.

By a cross-correlation analysis of ambient noise and machine learning models, we obtained 3D S-wave velocity and 3D temperature models, which is the first application in this volcanic rift region. We discussed various mechanisms through which geothermic fluids were controlled by dominant volcanic rifts as flow pathways, particularly toward Lake Taupo. The key findings are summarized as follows.

We revealed existing geothermal systems and potential reservoirs along the flow pathways (i.e., active fault map) within complex caldera systems.
Frequent seismicity occurrences seem to be triggered by natural fluid flows and/or geothermal fluid extraction.
Hot fluid flows correspond to flow pathways, and supercritical ones ~320^oC or more can be at a depth just greater than 2 km from the surface.
We investigated known geothermal systems using limited seismometer and well data in a large-scale exploration area. Our approach is capable of applying and locating new geothermal systems for reservoir-scale studies and borehole exploration.

The scale of exploration is important. Our approach here can utilize limited resources/data and enhance seismic imaging resolution, which can lower the risk, time, and cost in too many geothermal fields and provide crucial information for understanding shallow crustal structures as well as a novel approach for geothermal exploration/development in other active tectonic zones (e.g., a lengthy arc African rift).

Seismic Data in the Taupo Volcanic Zone. We used continuous seismic data recorded by broadband seismometers during August - December 2010 by⁴⁷. Only 25 seismometers for ambient noise data were used, considering the quality of data. The seismometers were distributed over many volcanic geothermal systems (Fig. 1a). The sampling time interval is 100 ms. The distances between the seismometers were spaced between 5 and 45 km (Fig. 1b), and the area of coverage is approximately 45 km × 45 km. We should note that geothermal reservoirs and magmatic heat sources can last more than a couple of thousand years^30,48; therefore, velocity (derived from four-month data) and temperature models (including exploited areas) may not change significantly over time because of continuous heat sources.

Analysis of surface wave dispersion. Our first step involved estimating phase velocity dispersion curves between station pairs from ambient noise data. We have pre-processed the ambient noise data by dividing the vertical component of the continuous ambient noise data into 30-min segments and applied the bandpass filtering at 0.2–5Hz with instrument response correction. By cross-correlating data and linearly stacking all segments for all station pairs, we could observe surface wave propagation on the stacked cross-correlation (CC) functions in the 0.1–1 Hz frequency band (Fig. 2a). Then we applied a whitening and velocity filter to cross-correlation functions ¹⁷, and extracted phase velocity dispersion curves for each pair of stations through the zero-crossing method^17,49 (Figs. 2b-c). Finally, we performed surface wave tomography to obtain a 3D S-wave velocity model^19,21.

The cross-spectrum between two seismometers was calculated using the following equation.

$${\rho }_{z}\left(f\right)= \frac{{u}_{i}\left(f\right){u}_{j}^{*}\left(f\right)}{\left|{u}_{i}\left(f\right)\right|\left|{u}_{j}\left(f\right)\right|};\left(1\right)$$

where ${\rho }_{z}$ is the vertical cross-spectrum, $f$ is the frequency, $u$ is the seismic record in the frequency domain from station i or j, and the asterisk denotes the complex conjugate. Cross-spectrum was used to obtain the phase velocity by the zero-crossing method.

The zero-crossing method⁴⁹ estimates phase velocities between station pairs by fitting zeros of the observed cross-spectrum to zeros of the theoretical cross-spectrum in the frequency domain. The phase velocity C obtained by the zero-crossing method is expressed as follows.

Here, we described other important points for our velocity analysis. We used only the daily cross-correlation waveforms that resembled others and rejected bad station/data. We determined the zone of phase velocity picking within one wavelength and six wavelength criteria based on the phase velocity and frequency diagram of its station pair (between red lines in Figs. 2b-c). The unique zero-cross dispersion curve (phase velocity or group velocity) is within the zone of the acceptable range of phase velocities (Fig. 2c). Noticeably, Rayleigh group velocities were determined from the stacked cross-correlation functions using continuous wavelet transform analysis from⁵⁰.

Zero-cross dispersion curve analysis is different from surface wave dispersion curved that was analyzed based on time-frequency representation of the data (i.e., multiple filter technique or continuous wavelet transform analysis in which the dispersion curves were picked based on high energy semblance)⁵¹. We should note that the surface wave arrival time is thus related to the propagation of the wave energy, defined as an energy maximum, which can be difficult to identify due to the presence of other seismic waves through scattering, multipathing, wave type conversions, and other unwanted signals and noises⁵¹. The velocity filter and the zero-crossing method¹⁷ thus minimized these issues and provided stable dispersion curves in geothermal velocity analysis. A single zero-cross dispersion curve can be picked/selected, which is not changed based on a variation of a high maximum energy semblance velocity map.

Ambient noise tomography for S-wave inversion. For this study, we employed a direct 3D S-wave inversion analysis using DSurfTomo²¹ based on the LSQR inversion approach⁵² in Eq. (3). After selecting all phase velocity dispersion curves (Fig. 2), 3D surface wave tomography was applied to invert these phase velocity pairs into a 3D S-wave velocity model²¹. In this inversion, one-dimensional S-wave velocity models (i.e., averaging from 1.1 times phase velocity pairs; black circles in Fig. 2e) were used as initial velocity models. We used 10 iterations for the grid size (0.015^ox0.015^o) and depth direction (< 6km) with a smoothing constraint for the speed change and a damping value in the DSurfTomo program based on several trials and errors. The outcome results were not corrected for topographic effects due to minimal elevation difference (Fig. 1b).

The main equation for the S-wave velocity inversion is as follows:

$$\delta {t}_{i}f=\sum _{k=1}^{K}-\frac{{v}_{ik}}{{C}_{k}^{2}}\sum _{j=1}^{J}\left[{R}_{\alpha }{Z}_{j}\frac{\partial {C}_{k}}{\partial {Vp}_{k}{Z}_{j}}+ {R}_{\rho }{Z}_{j}\frac{\partial {C}_{k}}{\partial {\rho }_{k}{Z}_{j}}+ \frac{\partial {C}_{k}}{\partial {Vs}_{k}{Z}_{j}}\right]{|}_{{\theta }_{k}} \delta {Vs}_{k}{Z}_{j}=\sum _{l=1}^{M}{G}_{il}{m}_{jl} ;\left(3\right)$$

where ${t}_{i}f$ is the calculated travel times from a reference model, which can be updated in the inversion, the ${v}_{ik}$ denotes the bilinear interpolation coefficients of the raypaths associated with the ith traveltime datum, and ${C}_{k}$ represents the phase velocity at the kth 2D surface grid point at frequency f, the scaling factors ${R}_{\alpha }{Z}_{j}$ and ${R}_{\rho }{Z}_{j}$ were derived based on the empirical relations ⁵³; ${\theta }_{k}$ represents the 1D reference model and the kth surface grid point on the surface; ${Vp}_{k}{Z}_{j}$, ${Vs}_{k}{Z}_{j}$ and ${\rho }_{k}{Z}_{j}$ are the P-wave velocity, S-wave velocity, and mass density at the jth node in the depth direction, respectively. The total number of grid points for the 3D model is M = KJ. The direct surface wave tomography described in the equation above can be solved for model $m$ (i.e., S-wave velocity) by further formulated in matrix inverse form of the travel time residual vector for all raypaths at different frequencies, and $G$ is the data sensitivity matrix. This is solved by the LSQR algorithm of⁵².

Analysis of machine learning models. To predict the spatial variation of temperature, we used a 3D S-wave velocity model derived from ambient noise tomography and seven borehole logs⁴⁶ in Ngatamiri geothermal reservoir (Fig. 3). Well data and our 1D S-wave velocity were used to train the training model (Fig. 3). We used machine learning model algorithms in MATLAB, which is a statistical technique for building the linear or non-linear regression model between predictor and response variables. There are three main steps, regardless of any machine learning model: training data, test data, and validation data. The predictor variables or independent variables were used as inputs or features to train the regression model, then we optimized the model using the polynomial fit function, and finally we can automatically obtain the best polynomial degree for the model that is not under-fitted or overfitted based on the validation data. It is dependent on which we split the training data and validation set to predict the output by the cross-validation or holdout technique in MATLAB.

Through our machine learning approach, we built various regression models and predicted the results. We automatically performed all trained models and compared all of them (i.e., nonlinear regression models such as Decision/regression Trees, Support Vector Machines, Gaussian Progress Regression models, Kernel Approximation Regression Models, Ensembles of Trees, and Neural Networks) presented in Table 1. Then, we can find the best model that is not under/over fitted^23,54. Our predictors/input parameters are 7 temperature - depth profiles, 1D velocity of each well location as a training model and performed validation model based on cross-validation fold technique (Fig. 3); as a result, the best-performing model is Gaussian Process Regression (GPR) model. In this study, only data from local wells that we could access were used. If we use more well data and their distribution over a large area, the accuracy can be most optimized.

The most suitable model in our study area is based on the GPR model, which we briefly described below. A non-linear fit function of the parameters can be formulated as follows:

$${y}_{i} = f\left[x,\right]+ϵ ;\left(4\right)$$

where ${y}_{i}$ is the output/response feature, $f$ is function to evaluate the input/predictor feature ${x}_{i}$, and  is the regression coefficients and $ϵ$ represents the error terms.

The GPR is the nonparametric kernel-based probabilistic model. Response $y$ (e.g., temperature) can be modeled from:

$$P\left({y}_{i}|f\left({x}_{i}\right),{x}_{i}\right) \sim N\left({y}_{i}|{h\left({x}_{i}\right)}^{T} + f\left({x}_{i}\right),{\sigma }^{2}\right) ;\left(5\right)$$

where $N\left(0,{\sigma }^{2}\right) \sim ϵ. {\sigma }^{2}$is error variance, $h$is the explicit basis function. Based on this GPR model, the fit model will be formed by obtaining the basis function coefficients, , the noise variance, ${\sigma }^{2}$, and the hyperparameters of the kernel function, derived from the input data set while training the GPR model. The other set of fit functions relevant to many other models (e.g., Decision Trees, Support Vector Machines, Neural Networks), and regression metrics to evaluate/validate the model performance based on residual ${r}_{i }= {y}_{i}-\widehat{{y}_{i}}$ in which $\widehat{{y}_{i}}$ refer to the predicted response valued from a regression model as follows:

Mean Absolute Error (MAE) $=$$\frac{1}{n}{\sum }_{i=1}^{n}\left|{y}_{i}-\widehat{{y}_{i}}\right|;$

Mean Square Error (MSE) $=$ $\frac{1}{n}{\sum }_{i=1}^{n}{\left({y}_{i}-\widehat{{y}_{i}}\right)}^{2}$;

Root Mean Square Error (RMSE)$= \sqrt{MSE}$

and R-square $=$ $\frac{ \frac{1}{n}{\sum }_{i=1}^{n}{\left({y}_{i}-\stackrel{-}{{y}_{i}}\right)}^{2} - \frac{1}{n}{\sum }_{i=1}^{n}{\left({y}_{i}-\widehat{{y}_{i}}\right)}^{2}}{ \frac{1}{n}{\sum }_{i=1}^{n}{\left({y}_{i}-\stackrel{-}{{y}_{i}}\right)}^{2}}$ in which $\stackrel{-}{{y}_{i}}$ is the mean response. The R-square is valued between 0 and 1. The perfect fitting model is when R-square is close to 1.

Fundings

This study was supported by the Leading Enhanced Notable Geothermal Optimization (LENGO) of the Science and Technology Research Partnership for Sustainable Development (SATREPS) and partially supported by the Japan Society for the Promotion of Science (KAKENHI grant JP20H01997, JP21H05202, and JP22H05108).

Acknowledgements

The authors are deeply indebted to Yasuhiro Fujimitsu (Kyushu Univ.) for the leadership and support of the LENGO-SATREPS projects. The authors thank Tatsunori Ikeda (Kyushu University) for technical discussions. We thank Ahmad Bahaa and Sherief Hashima (Kyushu University) for introducing us Machine Learning.

Author contributions

C.C. executed the data analysis and drafted the initial manuscript. C.C. and T.T. reviewed and revised the manuscript. All authors read and approve the final manuscript.

Data availability

The data that support the findings of this study are open-source; for example, seismometer data network from⁴⁷ or from the IRIS portal (http://fdsn.adc1.iris.edu/networks/detail/Z8_2009/); active fault database and seismicity data catalog from GNS science (Institute of Geological and Nuclear Science; https://www.gns.cri.nz/).

Competing interests

The authors declare no competing interests.

Dunkley, P. N., Smith, M., Allen, D. J. & Darling, W. G. The geothermal activity and geology of the northern sector of the Kenya Rift Valley. (1993).
Richards, J. P. & Mumin, A. H. Magmatic-hydrothermal processes within an evolving Earth: Iron oxide-copper-gold and porphyry Cu ± Mo ± Au deposits. Geology 41, 767–770 (2013).
Ratouis, T. M. P. & Zarrouk, S. J. Factors controlling large-scale hydrodynamic convection in the Taupo Volcanic Zone (TVZ), New Zealand. Geothermics 59, 236–251 (2016).
Chhun, C., Tsuji, T. & Ikeda, T. Potential Geothermal Reservoir Systems in the Kenyan Great Rift Valley and Volcanic Region Assessed by Ambient Noise Analysis. Available SSRN 4465700 (2023).
Chhun, C., Tsuji, T. & Ikeda, T. Potential Accumulations of Fluid Flows and Geothermal Structure in Kuju Revealed by Azimuthal Anisotropic Ambient Noise Tomography. Available SSRN 4470075 (2023).
Rempe, M. et al. Damage and seismic velocity structure of pulverized rocks near the San Andreas Fault. J. Geophys. Res. Solid Earth 118, 2813–2831 (2013).
Kana, J. D., Djongyang, N., Raïdandi, D., Nouck, P. N. & Dadjé, A. A review of geophysical methods for geothermal exploration. Renew. Sustain. Energy Rev. 44, 87–95 (2015).
Kanda, I., Fujimitsu, Y. & Nishijima, J. Geological structures controlling the placement and geometry of heat sources within the Menengai geothermal field, Kenya as evidenced by gravity study. Geothermics 79, 67–81 (2019).
Omollo, P., Nishijima, J., Fujimitsu, Y. & Sawayama, K. Resistivity structural imaging of the Olkaria Domes geothermal field in Kenya using 2D and 3D MT Data inversion. Geothermics 103, 102414 (2022).
Chhun, C., Kioka, A., Jia, J. & Tsuji, T. Characterization of hydrate and gas reservoirs in plate convergent margin by applying rock physics to high-resolution seismic velocity model. Mar. Pet. Geol. 92, 719–732 (2018).
Rowland, J. V & Simmons, S. F. Hydrologic, magmatic, and tectonic controls on hydrothermal flow, Taupo Volcanic Zone, New Zealand: Implications for the formation of epithermal vein deposits. Econ. Geol. 107, 427–457 (2012).
Payne, C. E. & Peters, K. J. Geochemistry in prospectivity modelling: investigating gold mineralization in the Taupo Volcanic Zone, New Zealand. Geochemistry Explor. Environ. Anal. 15, 193–204 (2015).
Simmons, S. F., Brown, K. L., Browne, P. R. L. & Rowland, J. V. Gold and silver resources in Taupo Volcanic Zone geothermal systems. Geothermics 59, 205–214 (2016).
Liu, Q. et al. Metallic mineral exploration by using ambient noise tomography in Ashele copper mine, Xinjiang, China. Geophysics 87, B221–B231 (2022).
Geothermal Energy Association. World map of geothermal power countries installed and developing. (2015).
Ekström, G. Time domain analysis of Earth’s long‐period background seismic radiation. J. Geophys. Res. Solid Earth 106, 26483–26493 (2001).
Sadeghisorkhani, H., Gudmundsson, Ó. & Tryggvason, A. GSpecDisp: A matlab GUI package for phase-velocity dispersion measurements from ambient-noise correlations. Comput. Geosci. 110, 41–53 (2018).
Suemoto, Y., Ikeda, T., Tsuji, T. & Iio, Y. Identification of a nascent tectonic boundary in the San-in area, southwest Japan, using a 3D S-wave velocity structure obtained by ambient noise surface wave tomography. Earth, Planets Sp. 72, 1–13 (2020).
Nimiya, H., Ikeda, T. & Tsuji, T. Three‐Dimensional S Wave Velocity Structure of Central Japan Estimated by Surface‐Wave Tomography Using Ambient Noise. J. Geophys. Res. Solid Earth 125, e2019JB019043 (2020).
Nthaba, B., Ikeda, T., Nimiya, H., Tsuji, T. & Iio, Y. Ambient noise tomography for a high-resolution 3D S-wave velocity model of the Kinki Region, Southwestern Japan, using dense seismic array data. Earth, Planets Sp. 74, 96 (2022).
Fang, H., Yao, H., Zhang, H., Huang, Y.-C. & van der Hilst, R. D. Direct inversion of surface wave dispersion for three-dimensional shallow crustal structure based on ray tracing: methodology and application. Geophys. J. Int. 201, 1251–1263 (2015).
Lehujeur, M. et al. Reservoir imaging using ambient noise correlation from a dense seismic network. J. Geophys. Res. Solid Earth 123, 6671–6686 (2018).
Nadler, D. W. Decision support: using machine learning through MATLAB to analyze environmental data. J. Environ. Stud. Sci. 9, 419–428 (2019).
Ahmad, A. B., Saibi, H., Belkacem, A. N. & Tsuji, T. Vehicle Auto-Classification Using Machine Learning Algorithms Based on Seismic Fingerprinting. Computers 11, 148 (2022).
Ahmad, A. B. & Tsuji, T. Traffic monitoring system based on deep learning and seismometer data. Appl. Sci. 11, 4590 (2021).
Singha, D. K., Chatterjee, R., Sen, M. K. & Sain, K. Pore pressure prediction in gas-hydrate bearing sediments of Krishna–Godavari basin, India. Mar. Geol. 357, 1–11 (2014).
Keynejad, S., Sbar, M. L. & Johnson, R. A. Comparison of model-based generalized regression neural network and prestack inversion in predicting Poisson’s ratio in Heidrun Field, North Sea. Lead. Edge 36, 938–946 (2017).
Wilson, C. J. N. et al. Volcanic and structural evolution of Taupo Volcanic Zone, New Zealand: a review. J. Volcanol. Geotherm. Res. 68, 1–28 (1995).
Downs, D. T. et al. Evolution of the intra-arc Taupo-Reporoa basin within the Taupo Volcanic Zone of New Zealand. Geosphere 10, 185–206 (2014).
Wilson, C. J. N. & Rowland, J. V. The volcanic, magmatic and tectonic setting of the Taupo Volcanic Zone, New Zealand, reviewed from a geothermal perspective. Geothermics 59, 168–187 (2016).
Delano, J. E., Stahl, T., Howell, A. & Clark, K. 3D coseismic surface displacements from historical aerial photographs of the 1987 Edgecumbe earthquake, New Zealand. Authorea Prepr. (2022).
Hurst, T., Heise, W., Hreinsdottir, S. & Hamling, I. Geophysics of the Taupo Volcanic Zone: A review of recent developments. Geothermics 59, 188–204 (2016).
Chambefort, I. et al. Ngatamariki geothermal system: Magmatic to epithermal transition in the Taupo volcanic zone, New Zealand. Econ. Geol. 112, 319–346 (2017).
Davey, F. J., Henrys, S. A. & Lodolo, E. Asymmetric rifting in a continental back-arc environment, North Island, New Zealand. J. Volcanol. Geotherm. Res. 68, 209–238 (1995).
Bibby, H. M., Caldwell, T. G., Davey, F. J. & Webb, T. H. Geophysical evidence on the structure of the Taupo Volcanic Zone and its hydrothermal circulation. J. Volcanol. Geotherm. Res. 68, 29–58 (1995).
Chambefort, I., Buscarlet, E., Wallis, I. C., Sewell, S. & Wilmarth, M. Ngatamariki geothermal field, New Zealand: Geology, geophysics, chemistry and conceptual model. Geothermics 59, 266–280 (2016).
White, B. R. & Chambefort, I. Geothermal development history of the Taupo Volcanic Zone. Geothermics 59, 148–167 (2016).
Sutton, A. N., Blake, S. & Wilson, C. J. N. An outline geochemistry of rhyolite eruptives from Taupo volcanic centre, New Zealand. J. Volcanol. Geotherm. Res. 68, 153–175 (1995).
Allan, A. S. R., Wilson, C. J. N., Millet, M.-A. & Wysoczanski, R. J. The invisible hand: Tectonic triggering and modulation of a rhyolitic supereruption. Geology 40, 563–566 (2012).
Climo, M., Milicich, S. D. & White, B. A history of geothermal direct use development in the Taupo Volcanic Zone, New Zealand. Geothermics 59, 215–224 (2016).
Sherburn, S., Bannister, S. & Bibby, H. Seismic velocity structure of the central Taupo Volcanic Zone, New Zealand, from local earthquake tomography. J. Volcanol. Geotherm. Res. 122, 69–88 (2003).
Brenguier, F. et al. Postseismic relaxation along the San Andreas fault at Parkfield from continuous seismological observations. Science (80-. ). 321, 1478–1481 (2008).
Sabra, K. G. Forecast from noise. Nat. Geosci. 1, 89–90 (2008).
Biggs, J., Anthony, E. Y. & Ebinger, C. J. Multiple inflation and deflation events at Kenyan volcanoes, East African Rift. Geology 37, 979–982 (2009).
Maguire, R. et al. Magma accumulation at depths of prior rhyolite storage beneath Yellowstone Caldera. Science (80-. ). 378, 1001–1004 (2022).
Bignall, G. Ngatamariki geothermal field geoscience overview. GNS Sci. Consult. Rep. 2009 94, 35 (2009).
Stephen Bannister. Deep Geothermal HADES seismic array. (2009). doi:10.7914/SN/Z8_2009
Watanabe, K., Hayashi, M. & Fujino, T. Fission track age of volcanoes in the Kuju volcanic region in relation to geothermal activity. J. Geotherm. Res. Soc. Japan 9, 207–217 (1987).
Ekström, G., Abers, G. A. & Webb, S. C. Determination of surface‐wave phase velocities across USArray from noise and Aki’s spectral formulation. Geophys. Res. Lett. 36, (2009).
Jiang, C. & Denolle, M. A. NoisePy: A new high‐performance python tool for ambient‐noise seismology. Seismol. Res. Lett. 91, 1853–1866 (2020).
Schimmel, M., Stutzmann, E. & Ventosa, S. Measuring group velocity in seismic noise correlation studies based on phase coherence and resampling strategies. IEEE Trans. Geosci. Remote Sens. 55, 1928–1935 (2017).
Paige, C. C. & Saunders, M. A. LSQR: An algorithm for sparse linear equations and sparse least squares. ACM Trans. Math. Softw. 8, 43–71 (1982).
Brocher, T. M. Empirical relations between elastic wavespeeds and density in the Earth’s crust. Bull. Seismol. Soc. Am. 95, 2081–2092 (2005).
Sunori, S. K. et al. Machine Learning Based Prediction of Soil pH. in 2021 5th International Conference on Electronics, Communication and Aerospace Technology (ICECA) 884–889 (IEEE, 2021).
Langridge, R. M. et al. The New Zealand active faults database. New Zeal. J. Geol. Geophys. 59, 86–96 (2016).
Wessel, P. et al. The generic mapping tools version 6. Geochemistry, Geophys. Geosystems 20, 5556–5564 (2019).

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

Ambient Noise Tomography and Machine Learning Models to Reveal Geothermal Structure in the Taupo Volcanic Zone

Status:

Version 1

Abstract

Figures

Introduction

Results and discussions

Conclusions

Data and Methods

Declarations

References

Additional Declarations

Status:

Version 1