Deep learning with cross-shape deep Boltzmann machine for pre-stack inversion problem
Son Dang Thai Phan, Mrinal Kanti Sen · 2019
For AVA inversion, we design a cross-shape deep Boltzmann machine structure for deep learning purpose by connecting four different restricted Boltzmann machines located at the vertices via a hidden neuron layer at the center of the cross. The network training process is performed by minimizing an energy function that is similar to the least square solution of an inverse problem. Unlike common network designs which have training datasets fed into visible layers located at the two ends of the structure, the cross-shape design allows the network to populate changes from input data at any visible layer toward the remaining ones. We then use the network to perform pre-stack seismic inversion to predict the impedances by training it to learn the non-linear relationship between the rock properties and seismic amplitudes. The inversion algorithm requires low frequency starting models to normalize the inputs before training the network, and to convert the results into absolute values after network applications. The resulting models show that the algorithm is capable of capturing all features in the training dataset while accurately reconstructing the input logs at the well locations and producing a geologically plausible impedance section. Furthermore, the network has the potential to be used for solving any similar non-linear inverse problem, provided sufficient training data is available. Presentation Date: Tuesday, September 17, 2019 Session Start Time: 8:30 AM Presentation Time: 10:10 AM Location: 221D Presentation Type: Oral