Wavelet holds an essential role in seismic data processing and characterization, for examples deconvolution and seismic inversion. Unfortunately, wavelet is an unknown data. Several existing methods attempt to estimate and extract the wavelet from seismic data. However, the methods give only a single wavelet from one seismic trace. When seismic data are non-stationer, single wavelet usage will cause a problem, that is raising the error. This paper proposes a time-varying wavelet estimation method to accommodate this problem. It uses matrix diagonalization to estimate a set of wavelets. Next, the time-varying wavelet is applied to deconvolution and seismic inversion. The experiment shows that time-varying wavelet improves the results in both deconvolution and seismic inversion. The errors decreased and spectrum bandwidth broadened.

A relatively straightforward methodology is presented for extending seismic bandwidth, and hence enhancing the seismic resolution by performing time-variant deconvolution. The generalized S-transform (GST) approach is used in order to properly compute the time-frequency components of the seismic reflection trace. In estimating the time-variant wavelet, a spectral modeling method is proposed, named multi-Ricker spectral approximation (MRA). After obtaining the estimated wavelet spectrum at each time sample, a deconvolution filter can then be built and applied in the S-transform domain. This proposed time-variant seismic enhancement method needs neither information on subsurface attenuation model nor assumption that the subsurface reflectivity is random. It is a data-driven methodology which is based on the seismic data only. This proposed method is validated on a noise-free and noisy synthetics and also applied to a field data. Results show that, after enhancement, overall seismic bandwidth can be extended resulting in a higher vertical resolution. Correlation with VSP corridor stack at well location ensures that the generated reflection details after enhancement are geologically plausible.

Estimating copula parameters remains a challenge when dealing with multiple correlated variables. Focused studies on the application of uncommon copula functions are also still scarce. Asymmetric dependence is necessary to be taken into account as symmetric dependence may not always be sufficient to model real data dependence. Asymmetric copulas were constructed using the Archimedean family as the basis copula. Linear inversion, random search, and Particle Swarm Optimization (PSO) were used to compare the estimations of copula parameters. Python was used as the main programming software to apply the proposed methods in this paper. From the comparison, linear inversion resulted in 1% of average absolute relative error while PSO and random search resulted in 4% and 19%, respectively. A different result was shown using a real data set. Real data often deal with local extreme values while performing the simulation. PSO was more stable than others when real data were used. It was concluded that PSO is the wisest method for real data cases and asymmetric copula parameter estimation.

No abstract

Data kecepatan gelombang S (shear) sangat diperlukan untuk karakterisasi reservoar dalam menentukan zona reservoar. Namun data kecepatan gelombang S sangat terbatas dan tersedia pada sumur tertentu saja. Penelitian ini dilakukan untuk memprediksi nilai kecepatan gelombang S dengan menggunakan metode supervised machine learning pada sumur S-1 lapangan migas di cekungan Sumatra Tengah. Simulasi algoritma machine learning dilakukan melalui tahapan sebelum dan setelah tuning pada algoritma library Scikit learn dan algoritma artificial neural network (ANN). Selain itu, parameter dan jumlah data yang digunakan dalam memprediksi nilai kecepatan gelombang akan menentukan nilai error dan akurasi. Hasil analisis menunjukkan bahwa algoritma yang digunakan untuk memperoleh akurasi terbaik pertama dalam memprediksi kecepatan gelombang S, yaitu random forest dengan nilai parameter n_estimator terbaik 10 dan algoritma kedua yang terbaik yaitu k-nearest neighbor dengan nilai parameter n_neighbor terbaik 5.