Dual domain network architecture for non-linear ultrasound transmission tomography reconstruction

Author(s):  
Yuling Fan ◽  
Hongjian Wang ◽  
Hartmut Gemmeke ◽  
Torsten Hopp ◽  
Juergen Hesser
2002 ◽  
Vol 23 (1) ◽  
pp. 95-104 ◽  
Author(s):  
Marc Molinari ◽  
Simon J Cox ◽  
Barry H Blott ◽  
Geoffrey J Daniell

2021 ◽  
Author(s):  
Lysander Bresinsky ◽  
Jannes Kordilla ◽  
Irina Engelhardt ◽  
Martin Sauter

<p><span>Present methods to quantify recharge in karst aquifers in many cases rely on spatially and temporally aggregated precipitation values, neglecting the highly erratic, non-linear nature of infiltration dynamics that give rise to a dual-domain behavior with a slow diffuse and fast focused recharge component. Here, we demonstrate the applicability of integrated surface-subsurface flow models to simulate diffuse and preferential infiltration within the large scale Western-Mountain-Aquifer (WMA) in Israel and the Palestinian territories. A semi-arid climate region with a highly pronounced seasonality of precipitation and intense short-duration rainfalls, such as the Mediterranean region, emphasizes the importance of understanding and accounting for the complex dynamics of dual-domain infiltration and partitioning of the precipitation input signal via spatially discretized overland flow processes.</span></p><p><span>We apply HydroGeoSphere as a dual-continuum flow simulator for transient variably-saturated water flows, discretizing the rock matrix and secondary porosity (i.e., conduits and fractures) as separate overlapping continua. Flow is respectively computed via the Richards' equation with volume-averaged van Genuchten parameters, assuming that the Richards' equation is valid for both domains. The presented model accounts for surface flow via the two-dimensional Saint-Vénant equations under nonexistent inertial forces. We apply precipitation directly to the overland flow continuum and naturally account for the partitioning into Horton overland flow and percolating water. However, modeling of unsaturated flow through the conduit/fracture continuum with the van Genuchten parameterization is often limited, as the Richards' equation describes flow solely in terms of capillary forces, leading to high matric suction in the matrix continuum as a result of the smaller pore spaces (and hence constant exchange from the fracture continuum to the matrix system). In a natural system, non-linear transfer processes govern the transfer between fracture/conduit and matrix flow, such as inertia-driven infiltration (i.e., droplet, rivulet, and film flow) that initially retains itself from equilibration of capillary pressure heads and avoids instant matrix imbibition. This study demonstrates parametrization strategies to allow for infiltration through the fracture/conduit continuum using small-scale process-based simulations. Further, we offer procedures that help to achieve convergence of complex catchment-scale variably-saturated simulations.</span></p>


1993 ◽  
Vol 1 (1) ◽  
pp. 1-11 ◽  
Author(s):  
Tormod Næs ◽  
Knut Kvaal ◽  
Tomas Isaksson ◽  
Charles Miller

This paper is about the use of artificial neural networks for multivariate calibration. We discuss network architecture and estimation as well as the relationship between neural networks and related linear and non-linear techniques. A feed-forward network is tested on two applications of near infrared spectroscopy, both of which have been treated previously and which have indicated non-linear features. In both cases, the network gives more precise prediction results than the linear calibration method of PCR.


2014 ◽  
Vol 20 (1) ◽  
pp. 116-132 ◽  
Author(s):  
Gheorghe Ruxanda ◽  
Laura Maria Badea

Making accurate predictions for stock market values with advanced non-linear methods creates opportunities for business practitioners, especially nowadays, with highly volatile stock market evolutions. Well suited for approaching non-linear problems, Artificial Neural Networks provide a number of features which make possible reasonably accurate forecasts. But, like the old Latin saying “Primus inter pares”, not all Artificial Neural Networks perform the same, end results depending very much on the network architecture and, more specifically, on the chosen training algorithm. This paper provides suggestions on how to configure Artificial Neural Networks for performing stock market predictions, with an application on the Romanian BET index. Final results are confirmed by testing the trained networks on the Croatian Stock Market data. End remarks entitle Broyden-Fletcher-Goldfarb-Shanno training algorithm as a good choice in terms of model convergence and generalization capacity.


Sign in / Sign up

Export Citation Format

Share Document