Avoiding pathologies in very deep networks

Choosing appropriate architectures and regularization strategies of deep networks is crucial to good predictive performance. To shed light on this problem, we analyze the analogous problem of constructing useful priors on compositions of functions. Specifically, we study the deep Gaussian process, a type of infinitely-wide, deep neural network. We show that in standard architectures, the representational capacity of the network tends to capture fewer degrees of freedom as the number of layers increases, retaining only a single degree of freedom in the limit. We propose an alternate network architecture which does not suffer from this pathology. We also examine deep covariance functions, obtained by composing infinitely many feature transforms. Lastly, we characterize the class of models obtained by performing dropout on Gaussian processes.

Thermo-hydro-geomechanical simulations of methane gas production from deep sea methane hydrate formations

A new constitutive model called Methane Hydrate Critical State (MHCS) model was conducted to investigate the geomechanical response of the gas-hydrate-bearing sediments at the Nankai Trough during the wellbore construction process. The strength and dilatancy of gas-hydrate-bearing soil would gradually disappear when the bonds are destroyed because of excessively shearing, which are often observed in dense soils and also in bonded soils such as cemented soil and unsaturated soil. In this study, the MHCS model, which presents such softening features, would be incorporated into a staged-finite-element model in ABAQUS, which mainly considered the loading history of soils and the interaction between cement-casing-formation. This model shows the influence of gas-hydrate-bearing soil to the deformation and stability of a wellbore and the surrounding sediments during wellbore construction. At the same time, the conventional Mohr-Coulomb model was used in the model to show the advantages of MHCS model by comparing the results of the two models.

Fibre optic monitoring of a deep circular excavation

This paper describes part of the monitoring undertaken at Abbey Mills shaft F, one of the main shafts of Thames Water's Lee tunnel project in London, UK. This shaft, with an external diameter of 30 m and 73 m deep, is one of the largest ever constructed in the UK and consequently penetrates layered and challenging ground conditions (Terrace Gravel, London Clay, Lambeth Group, Thanet Sand Formation, Chalk Formation). Three out of the twenty 1-2 m thick and 84 m deep diaphragm wall panels were equipped with fibre optic instrumentation. Bending and circumferential hoop strains were measured using Brillouin optical time-domain reflectometry and analysis technologies. These measurements showed that the overall radial movement of the wall was very small. Prior to excavation during a dewatering trial, the shaft may have experienced three-dimensional deformation due to differential water pressures. During excavation, the measured hoop and bending strains of the wall in the chalk exceeded the predictions. This appears to be related to the verticality tolerances of the diaphragm wall and lower circumferential hoop stiffness of the diaphragm walls at deep depths. The findings from this case study provide valuable information for future deep shafts in London. © ICE Publishing: All rights reserved.