1 Introduction

India ranks third globally in oil/gas consumption, around 5 million barrels per day, with a yearly demand growth of 3–4%. Hydrocarbon exploration in India was initiated in the middle of the nineteenth century and geared up with commercial oil production in Assam in 1889. Since then, several fields in northeast India with proven oil/gas reserves have been explored. Previously, various geoscientists have applied many conventional geophysical methods to identify and delineate the responsible factors for hydrocarbon accumulations (Darijani et al. 2019; Narayan et al. 2022a, b, 2023b). However, these methods provide limited insights regarding the characteristics of the different litho-facies at depth (Biswas et al. 2014). In the recent past, several studies have demonstrated the successful application of inversion and machine learning algorithms integrating seismic and wire-line log data to minimize the risk involved in hydrocarbon exploration (Kumar et al. 2016; Gogoi et al. 2018; Narayan et al. 2022a, b). Tulamura anticline is a linear, NNW–SSE trending, doubly plunging, asymmetric anticline, about 40 km long and 18 km wide at its widest part (figures 1 and 2). The study area has experienced intense folding, faulting, and thrusting during post-collision orogeny (Murty 1983; Naik et al. 2001). In general, the thick clastic sequence of Tertiary sediments represents the stratigraphy of the study area. Based on the relative abundance of coarse and finer clastics presence, the Bhuban Formation (Early to Middle Miocene) was subdivided into three members such as Lower Bhuban, Middle Bhuban and Upper Bhuban (Evans 1932).

Figure 1
figure 1

Tectonic map of northeast India. The study area (Tulamura field) is shown by the red rectangle. MH: Mikir Hill, SP: Shillong Plateau, DF: Dauki Fault and DuF: Dupitila Fault.

Full size image
Figure 2
figure 2

The DEM map shows the topographic variation in the study area. Red represents topographic high, and blue represents topographic low.

Full size image

In the present study, we applied various cutting-edge geophysical and machine learning tools on high-resolution 3D seismic and well-log data to delineate and characterize the possible dispersal pattern of the hydrocarbon-containing sandstone facies in the study area. Spectrally decomposed seismic data was initially analyzed to bring out possible channel geometry. Then, the seismic inversion method was used to calculate the rock parameters to identify and characterize the litho-facies. The acoustic impedance and Vp/Vs ratio were estimated by performing pre-stack seismic inversion (Latimer et al. 2000). In general, acoustic impedance is a good indicator of lithology. In contrast, the Vp/Vs ratio indicates the presence of fluid types within the reservoir rock.

Moreover, the Vp/Vs ratio also discriminates between sand and shale facies based on their distinct magnitude response. Generally, channel features consist of porous sands that create more accommodation space for the concentration of hydrocarbon and mineralization fluids (Bhattacharjee et al. 2017). Therefore, porosity estimation is crucial to assess the sand quality and the fluids' holding capacity. The neural network-based tool allows additional inputs from the seismic and well-log data to generate the target outcomes (Hami-Eddine et al. 2009; Singh et al. 2016; Pandey et al. 2020; Dixit et al. 2020; Narayan et al. 2023a, b). The primary objective of the present study is to delineate the possible channel geometry within the Bhuban Formation in the Tulamura field and its characterization based on petrophysical properties derived using joint inversion and machine learning methods. This study in Tulamura anticline, Tripura, unveils a promising lead for hydrocarbon exploration. The flowchart for delineation and reservoir characterization of the Bhuban Formation in Tulamura anticline, Assam Basin, is shown in figure 3(a).

Figure 3a
figure 3

Methodology flowchart for the delineation and reservoir characterization of Middle Bhuban in Tulamura anticline, Assam Basin.

Full size image
Figure 3b
figure 4

Stratigraphic chart of the Tripura–Cachar Fold Belt.

Full size image

2 Geology of the study area

2.1 Regional tectonics

The Assam and Assam Arakan Basin evolved due to rifting and drifting of the Indian plate towards the north and northeast after the Gondwanaland break up during the Cretaceous (Murty 1983; Naik et al. 2001; Narayan et al. 2023b, d). Indian subcontinent drifting caused Indian plate subduction below the Burmese plate. Oblique subduction of the Indian plate initiated the closing of the Neo-Tethys Ocean on the northeast and then gradually progressed southward. The westward migration of the prism complex extended into the Tripura–Cachar Fold Belt (TCFB) as the Neogene Accretionary Prism (Dasgupta and Nandy 1995). Initiation of structuration is considered to have occurred during the Mio-Pliocene, probably after the deposition of the Tipam Formation (Mohan and Pandey 1973). Detachment folding is the predominant mechanism of fold formation in the Tripura area. Several major anticlinal structures have been identified in Tripura through remote sensing studies and systematic geological mapping, and Tulamura anticline is among them.

2.2 Stratigraphy

The generalized litho-stratigraphy of the Tripura–Cachar area is shown in figure 3(b). A thick succession of argillaceous and arenaceous sequences represents the Neogene sediments of the Miocene Surma Group (Nandy et al. 1983). Surma Group is unconformably overlain by dominantly arenaceous Tipam Formation deposited in a fluvial environment. The Surma Group of sediments is further subdivided into Bhuban and Bokabil Formations. Bhuban Formation was further subdivided into three members: Lower Bhuban, Middle Bhuban, and Upper Bhuban, depending upon the relative abundance of coarser and finer clastics (Evans 1932). Bokabil Formation of Late Miocene to Early Pliocene age overlies the Bhuban Formations. Lower and Upper Bhuban Formations are more arenaceous than the Middle Bhuban, which is dominantly argillaceous, comprising mainly mudstone, shale, shale-siltstone alternation, and shale-sandstone alternation. The Surma Group of sedimentary succession lacks age-diagnostic fauna and marker horizons; hence, biostratigraphy control is minimal. During the Middle Eocene to Early Miocene, the collision of the Indian plate with the Burmese plate and Tibetan plate resulted in a major switch in sedimentation pattern over the Bengal Basin (Pal et al. 2016; Narayan et al. 2021, 2022a, b). The TCFB has undergone several stages of sedimentation. Surma Group sediments were deposited in a deltaic setting (Holtrop and Keizer 1970; Johnson and Nur Alam 1991) with occasional marine transgressions (Seshavataram et al. 1998). Most of the sand facies in Tripura are heterogeneous and lenticular. The sedimentological studies indicate that the Middle Bhuban sediments were deposited in the delta front environment. In contrast, the Upper Bhuban sediments in the area are believed to be deposited under marginal marine in the mouth bar to distal bar regime (Dutta 1993) with fluctuating sea levels. Sand facies consisting of channels, distributary mouth bars, tidal bars and channel lobes are expected to occur under this depositional setup.

3 Data used

The 3D survey (south Tulamura anticline) covers an area of around 50 km2 of pre-stack seismic data. The seismic data used for the study is the processed pre-stack time migration (PSTM) gathered in the time domain. It is found that the event is dipping at the far offset (figure 4a), which warrants gathering seismic data conditioning. To improve the quality of input seismic gathers, data conditioning through radon filtering was performed to remove the random noise. Subsequently, mild trim statics were applied for the event flattening. Figure 4(b) shows the seismic gather data quality improvement after conditioning. Further, well-to-seismic calibration was carried out to tie the formation tops marked at well with seismic reflectors (figure 4c). Three horizons (Upper, Middle and Lower Bhuban) have been mapped using the stacked seismic data. Figure 5(a) shows a representative seismic section showing the structural and stratigraphic configuration of the Tulamura anticline. Frequency bandwidth between 8 and 48 Hz was found in the zone of interest (figure 5b). The time structure map at the top of Middle Bhuban is depicted in figure 5(c). This two-way time structure map demonstrates the distinct anticlinal high signature bounded by the two longitudinal NNW–SSE trending thrust faults (figure 5c). The time range for the Middle Bhuban horizon varies from 1936 ms at anticline flanks to 569 ms at anticline crest. Faults are considered conduits and play a key role in fluid movement. Major faults with a substantial throw also restrict fluid movement and provide suitable entrapment conditions.

Figure 4
figure 5

Seismic gather data. (a) Vintage processed seismic gather, (b) conditioned seismic gather and (c) noise. (d) Well to seismic tie using pre-stack seismic data for the zone of interest (marked yellow dotted box). A good correlation was achieved between the synthetic and the actual seismic.

Full size image
Figure 5
figure 6

(a) The inverse seismic section passing through well W-1 demonstrates the formation tops and major and minor faults. (b) Frequency bandwidth between 8–48 Hz was found in the zone of interest. (c) The structure map at the Middle Bhuban top shows the Tulamura anticline and structure-bounding thrust faults.

Full size image

The only well (W-1) in the study area has a variety of recorded logs, viz., gamma-ray (GR), P- and S-waves (sonic wave velocity (DT) and shear wave velocity  (DTSM)), density (RHOB), resistivity and porosity throughout the Bhuban Formation. Depth matching between logs of a particular well is the first step for arriving at an accurate petrophysical evaluation. The well-log panel, comprising gamma-ray, density, resistivity, neutron porosity, P-impedance, Vp/Vs ratio, and effective porosity (PHIE), were studied to recognize the thickness of the sandstone facies from Middle Bhuban (2230–2340 ms). It is found that the identified sandstone facies were deposited in a fining upward sequence (figure 6a).

Figure 6
figure 7

(a) Well-log panel showing various log signatures. Gamma-ray log interpretation indicates thick sand facies deposited in a fining upward sequence at the Middle Bhuban base. (b) Well-based cross-plot between P-impedance vs. Vp/Vs ratio and (c) P-impedance vs. effective porosity (PHIE), indicating the characteristics of the sand (yellow), siltstone (brown) and shale facies (green).

Full size image

Well-based reservoir characterization was initially performed by cross-plotting the P-impedance vs. Vp/Vs ratio and P-impedance vs. effective porosity, colour-coded with litho-facies logs (figure 6b, c). It is found that the targeted sand facies correspond to the relatively low Vp/Vs ratio (1.6–1.8) and high effective porosity (10–20%) response with respect to shale and siltstone facies. On the other hand, it was also found that the sand, siltstone, and shale facies overlap in P-impedance response. Taking the lead from the above analysis, we attempted to delineate and characterize the sand facies using seismic-derived elastic and petrophysical properties.

4 Methods

4.1 Spectral decomposition and RGB blending

Spectral decomposition is also one of the most effective techniques for delineating channel geometry and its continuity (Chopra and Marfurt 2016). The wavelet transform technique known as the continuous wavelet transform (CWT) is extensively useful in the spectral decomposition of seismic data. Mathematically, the CWT technique is expressed as the inner product of the wavelet family ψσ,τ(t) with the signal u(t).

$${S}_{w}(\sigma ,\tau )=\frac{1}{\sqrt{\sigma }}\underset{-\infty }{\overset{\infty }{\int }}u(t)\overline{\psi } \left(\frac{t-\tau }{\sigma }\right)dt$$
(1)

where \(\overline{\psi }\) is the complex conjugate of ψ, and Sw is the time scale map used to extract the instantaneous frequency. τ is the time shift applied to the mother wavelet, and σ is the scale.

Spectral decomposition is the most effective technique for delineating channel geometry and its continuity (Chopra and Marfurt 2016). The PSTM stacked seismic data contains a frequency bandwidth of 8–48 Hz with a dominant frequency of 26 Hz (figure 5a, b). Considering this, we have generated iso-frequency volumes of 12, 20, 28 and 36 Hz. RGB (red-green-blue) blending of the spectrally decomposed (12, 20 and 28 Hz) seismic amplitude illuminates a southward trending channel geometry (figure 12a). Moreover, the identified channel feature must be characterized using advanced seismic attributes to ascertain the presence of good sand facies within the channel. The shale facies present in the surroundings of the channel geometry will provide the lateral seal to restrict the fluid flow that may contain hydrocarbons.

4.2 Pre-stack seismic inversion and P-impedance and Vp/Vs ratio estimation

Several authors have discussed pre-stack seismic inversion methods to obtain reliable estimates of P-impedance, S-impedance, density and Vp/Vs ratio (Simmons and Backus 1996; Buland and Omre 2003; Hampson et al. 2005; Russel et al. 2006; Zhang et al. 2011; Xiao et al. 2020; Sun et al. 2022). This research used model-based inversion to estimate the acoustic impedance and Vp/Vs ratio volumes. The discussion about the detailed methodology for estimating the P-impedance and Vp/Vs ratio is beyond the scope of this paper. This study found a suitable wavelet time length of 160 ms, requiring a seismic window of 500 to 600 ms for reliable inversion. Including shallower seismic responses in wavelet estimation captures the influence of lithologies encountered by seismic rays. However, readers seeking detailed information regarding the pre-stack inversion process may refer to the research articles by Fatti et al. (1994), Hampson et al. (2005) and Russel et al. (2006). We adopted the workflow to invert the 3D pre-stack seismic data shown in figure 7.

Figure 7
figure 8

Workflow adopted for the inversion of pre-stack seismic data.

Full size image

This study carried out pre-stack seismic inversion to generate P-impedance and Vp/Vs ratio volumes. The accuracy of the inverted results was evaluated through cross-plot analysis and comparison with log signatures at the well point (figure 8a–c). A linear relationship with good correlation (P-impedance: 87.8% and Vp/Vs ratio: 93.1%) was found between the inverted and actual logs with errors of nearly 337 m/s * gm/cm3 and 0.23, respectively. The estimated P-impedance and Vp/Vs ratio inline sections passing through well (W-1) overlaid with log P-impedance validate the inverted outcomes (figure 9a, b). The target sand facies are characterized by relatively moderate P-impedance (9800–10600 m/s * gm/cm3) and low Vp/Vs ratio (1.68–1.76), as shown in figure 9(a, b). It is also found that the Vp/Vs ratio and P-impedance attributes efficiently discriminate between the sand and shale facies.

Figure 8
figure 9

(a) Inversion analysis using single trace seismic attribute showing good correlation between original and inverted impedance/(Vp/Vs) ratio and lower synthetic error at well W-1. The wavelet used in this inversion process is shown in the inset. (b, c) Cross-plot results indicate a linear relationship between the inverted P-impedance/(Vp/Vs) ratio and the actual P-impedance/(Vp/Vs) ratio.

Full size image
Figure 9
figure 10

Inline passing through well W-1 extracted from (a) P-impedance and (b) Vp/Vs ratio are shown here. The log impedance and Vp/Vs ratio strips overlaid on the corresponding inverted impedance and Vp/Vs ratio show an excellent match.

Full size image

4.3 Multilayer perceptron (MLP) modelling and effective porosity prediction

The perceptron is a widely used neural network approach for binary problem-solving. It utilizes activation functions that exhibit monotonically increasing behaviours (Van der Baan and Jutten 2000; Dixit and Mandal 2020; Narayan et al. 2023a, c). As a fundamental model, the perceptron serves as a mathematical representation of how the human mind operates. It receives input information from the input layer (seismic attributes), applies weights to the inputs, calculates their sum, passes the result through an activation function, and generates an output in the output layer. The network’s performance is assessed based on its number of hidden layers. Insufficient hidden layers can lead to poor performance. In contrast, excessive hidden layers may cause the network to merely memorize the training data and fail when presented with unknown datasets (McCormack 1991).

To create an improved multilayer perceptron (MLP) model, adjustments are required to the number of neurons in the hidden layers, in addition to appropriately updating the weights based on the complexity of the problem at hand (Van der Baan and Jutten 2000; Hami-Eddine et al. 2009). Assuming the input vectors as x1, x2, ..., xn, and the corresponding weights as w1, w2, ..., wn, the output of a perceptron can be represented by the following equation:

$$y = \sum\limits_{i = 1}^{n} {w_{i} } x_{i} + bias.$$
(2)

It can also be written as,

$$y = w^{t} x + bias.$$
(3)

The difference between the expected and actual output is minimized through iterative back-propagation learning (Singh et al. 2016; Dixit and Mandal 2020). Seismic attributes (listed in table 1) involve diverse mathematical transformations of seismic trace data, enhancing predictive capabilities and pattern recognition through the nonlinear responses of attributes during MLP model training (Hampson et al. 2005). Before the combination, all input data were randomized and standardized to eliminate any potential bias. The model is trained using 75% of the input data, while the remaining 25% is left out for validation (testing) purposes. Sensitivity analyses identified vital features in effective porosity prediction and found that each attribute significantly contributes to property prediction. Among all input attributes, we found that the P-imp attribute has more feature importance than the other attributes (figure 10a). The correlation coefficient between initial and predicted PHIE values was 94.1% (figure 10b). The trained model achieved the lower mean absolute error (MAE) (~0.015) and average loss (~0.005) on the test data (figure 10c and d). MAE measures how close the predicted values are to the actual values on average. A lower MAE indicates better accuracy, representing a smaller average deviation from the actual values. The high correlation coefficient and reduced errors demonstrate the model’s efficacy in predicting PHIE volume. After getting an error below an acceptable threshold, the predicted PHIE was accepted, and the PHIE volume was generated. The predicted PHIE values closely matched the overlaid log PHIE strip. The PHIE measures the interconnected pore spaces in the sedimentary rock. The PHIE efficiently discriminates between the sand facies (relatively high PHIE) and shale facies (very low PHIE) and signifies the quality of sand facies. We found a high-PHIE (8–13%) response for the target sand zone (figure 11).

Table 1 Post-stack seismic derived attributes used for neural network training and validation.
Full size table
Figure 10
figure 11

(a) Relative feature importance found for different attributes used in machine learning modelling. (b) Cross-plot result indicating a linear relationship between the actual and predicted effective porosity with high correlation. (c) Lower training/validation mean absolute error. (d) Lower training/validation loss values obtained at 500 iterations indicate that the porosity volume was estimated with high accuracy.

Full size image
Figure 11
figure 12

Inline passing through well W-1 extracted from the effective porosity is shown here. The log porosity strip overlaid on the predicted porosity shows a good match.

Full size image

5 Results and discussions

We identified thick sandstone facies (~60 m) deposited in channel geometry at the Middle Bhuban base encountered in the well W-1. State-of-the-art geophysical and machine learning techniques were utilized to pursue this lead in the Tulamura field, combining well-log and seismic data. High-resolution 3D seismic data provides broader coverage with enhanced continuous subsurface imaging capability. Subsequently, it can be best utilized in mapping various litho-stratigraphic layers and in subsurface facies characterization from the surrounding rocks.

In the Tripura–Cachar Fold Belt, India, most channel features have been observed along the NNE–SSW direction with SSW flow (Kumar et al. 2016; Bhattacharjee et al. 2017). The sandstone facies within the channel geometry were mainly coarse to medium-grained, arkosic, and reducing. The studies reveal that the sandstone facies were deposited in fan, braided rivers of typical proximal facies with fining upward sequences (Kumar et al. 2016). We studied 3D seismic and well-log data to delineate the structural architecture and characterize the stratigraphic sequences in the Tulamura field. Gamma-ray log study suggests that the porous sand facies were deposited in a fining upward sequence (figure 6a). The overlying shale layer (impermeable) may act as the top seal for the hydrocarbon accumulations. The correlation between the sand facies and the elastic and petrophysical properties (P-impedance, Vp/Vs ratio and effective porosity) was established through cross-plot analysis at the well point (figure 6b, c). The results suggest that the ore-containing sandstone facies pose relatively moderate P-impedance, low-Vp/Vs ratio and highly effective porosity signatures with respect to surroundings. The structural mapping shows two NNW–SSE trending thrust faults bound the Tulamura anticline and may provide suitable structural entrapment conditions (particularly eastward fault) for fluid accumulation (figure 5b).

Seismic-derived attributes were analyzed to delineate and characterize the possible dispersal pattern of the identified sandstone facies in the study area. RGB blend spectrally decomposed seismic response indicates an almost southward trending channel geometry almost parallel to the NNE–SSW trending major fault (figure 12a). The varied tuning frequencies observed across different segments of the channel geometry signify the presence of lithological heterogeneities within its structure. The type and nature of the facies from the channel geometry were further characterized based on P-impedance, Vp/Vs ratio and effective porosity responses (figure 12b–d) (Wood et al. 2012). The zones within the channel geometry indicate relatively moderate P-impedance (figure 12b), low Vp/Vs ratio (figure 12c) and high effective porosity (figure 12d), indicating good porous sands (average effective porosity ~9%) for hydrocarbon accumulations in Tulamura field, Tripura, India. Our analysis reveals that the predominant sedimentary fill within the channel geometry consists of sand facies.

Figure 12
figure 13

Attribute maps were generated for the (a) RGB blend spectral decomposed seismic, (b) P-impedance, (c) Vp/Vs ratio, and (d) effective porosity.

Full size image

Nonetheless, transitions in litho-facies, potentially from clean sand to silt or shaly sand, are apparent, corroborating the findings derived from RGB blending. Dasgupta and Nandy (1995) suggested that the Bhuban Formation sediments were deposited in the delta front environment to the under marginal marine in the mouth bar to distal bar regime with fluctuating sea level. The levee part is argillaceous. These shale facies between channel geometry and the NNE–SSW trending major fault may provide good lateral entrapment conditions for fluid accumulations. The present study thoroughly discusses the structural and stratigraphic aspects of hydrocarbon exploration in the study area. Some studies have been conducted in the Penobscot field; Nova Scotia poses complex geological structures, limited well log data and moderate seismic data quality. These studies provide an efficient and reliable geological model in petroleum exploration (Narayan et al. 2023a, c).

6 Conclusions

In this study, we applied machine learning and state-of-the-art geophysical tools on 3D seismic data to delineate and characterize the geologic features containing hydrocarbon in the Tripura–Cachar Fold Belt, India. The spectral decomposition result efficiently delineates the almost southward trending channel geometry. The petrophysical properties (P-impedance, Vp/Vs ratio and effective porosity) estimated with high precision have been utilized to delineate and characterize the channel in-fill litho-facies. Based on relatively moderate P-impedance (9800–10600 m/s * gm/cm3), low Vp/Vs ratio (1.68–1.76) and high effective porosity (8–13%), it is found that the channel geometry is mainly filled with porous sand facies deposited in fining upward sequences. These porous sand facies could be a favourable reservoir for hydrocarbon exploration from Middle Bhuban deposits in the Tulamura field in Tripura. The impermeable shale facies at the sand top and between the channel geometry and NNE–SSW trending fault will provide a suitable stratigraphic entrapment for hydrocarbon fluids. The present study estimates the extension, orientation, and spatial occurrences of porous sand reservoir facies, and the sand facies with good porosity are the favourable zones possibly containing hydrocarbon. The present study also highlights the importance of machine learning and the utilization of geophysical tools in hydrocarbon exploration to get a clearer image of the subsurface. However, seismic data quality and limited well data are the main problems in the Tulamura field that need to be addressed.