The presence of gas in the rock’s or sediment’s pore space significantly affects its seismic amplitude response and modifies its subsurface signature. Gas hydrates in the subsurface are often difficult to image with reflection seismic data if the seismic data lack a strong bottom simulating reflector (BSR). High-amplitude BSRs are caused by a sharp decrease in acoustic impedance in the rocks as the hydrates transition from their solid form, to a free gas form due to changing pressure and temperature conditions with depth beneath the seafloor. Two key reasons for weak BSRs include 1) insufficient free gas below the hydrate to create the needed impedance contrast, and 2) stratigraphy-parallel BSRs that are subtle and can only be identified with advanced seismic analysis. In these cases, the imaging and detection of hydrates becomes difficult, as traditional detection methods rely heavily on BSRs, gas chimneys, or pockmarks on the seafloor and other contextual components. Additionally, differentiating between low-saturation gas (LSG) and high-saturation gas (HSG) reservoirs remains another subsurface imaging issue, as they will have similar seismic amplitude and rock physics responses.

To address and understand these challenging imaging problems, I employ an unsupervised machine learning multi-attribute analysis to reduce their significant uncertainty in hydrocarbon exploration. The hydrate study looks at two 2D seismic datasets in the Pegasus Basin, offshore New Zealand, where BSRs are not continuously or clearly imaged. The first LSG investigates the amplitude responses within the King Kong/Lisa Anne and Ursa producing fields within the deepwater Gulf of Mexico and within the Scarborough gas field, offshore Australia. These analyses use principal component analysis (PCA) methods applied to a selected set of seismic attributes to identify meaningful combinations of attributes which provide insight into the seismic data. I then use self-organizing maps (SOMs) to help visualize and interpret the multi-dimensional PCA results. Several SOM hyperparameters were tested such as neuron count and the number of epochs (iterations) to create an optimized SOM that is computationally efficient, and effectively identifies the features of interest.

Optimized SOM results, which use a combination of attributes sensitive to attenuation, frequency, and small amplitude anomalies, help better resolve subtle hydrate response and can differentiate between LSG and HGS reservoirs. However, this method only proved to be successful within the Gulf of Mexico data volume and yielded disappointing results with the Carnarvon Basin. This difference is most likely due to the Carnarvon Basin having a different amplitude response resulting from a different burial history and fluid saturations than that of the Gulf of Mexico. Therefore, this method is non-transferrable, and a different combination of attributes may be needed in other LSG-prone basins.

Individually, some of these attributes have minimal success in identifying the seismically invisible hydrates and differentiating between low and high gas saturation reservoirs. However, when used together by employing multi-attribute analyses such as PCA and SOMs, these same attributes provide clearer insight into the subsurface responses needed to better identify the presence of hydrates and to distinguish between both reservoirs.