Search

Despite their great success, neural networks still remain as black-boxes due to the lack of interpretability. Here we propose a new analyzing method, namely the weight pathway analysis (WPA), to make them transparent. We consider weights in pathways that link neurons longitudinally from input neurons to output neurons, or simply weight pathways, as the basic units for understanding a neural network, and decompose a neural network into a series of subnetworks of such weight pathways. A visualization scheme of the subnetworks is presented that gives longitudinal perspectives of the network like radiographs, making the internal structures of the network visible. Impacts of parameter adjustments or structural changes to the network can be visualized via such radiographs. Characteristic maps are established for subnetworks to characterize the enhancement or suppression of the influence of input samples on each output neuron. Using WPA, we discover that neural network store and utilize information in a holographic way, that is, subnetworks encode all training samples in a coherent structure and thus only by investigating the weight pathways can one explore samples stored in the network. Furthermore, with WPA, we reveal fundamental learning modes of a neural network: the linear learning mode and the nonlinear learning mode. The former extracts linearly separable features while the latter extracts linearly inseparable features. The hidden-layer neurons self-organize into different classes for establishing learning modes and for reaching the training goal. The finding of learning modes provides us the theoretical ground for understanding some of the fundamental problems of machine learning, such as the dynamics of learning process, the role of linear and nonlinear neurons, as well as the role of network width and depth., 16 pages, 10 figures

Additional file 1. PRISMA 2020 flow diagram.

Additional file 2. PRISMA 2020 checklist.

This paper presents a soft computing of integrating artificial neural networks (ANNs) and genetic algorithms (GAs) to back analyze the earth stress field based on hydraulic fracturing. In this method, the ANN model is employed to map the relationship between the earth stress parameters and hydraulic fracturing behavior instead of numerical computation, and the advantage of this work is that it can conveniently conduct the integration of ANN and optimization algorithm and effectively reduce the workload of numerical computation by using directly the field-measured information to build learning samples. In addition, this can also improve accuracy of earth stress determination from field test data sets for ANN model. The GA is applied to implement multi-objective earth stress parameters optimization on the basis of the objective function. The field monitoring information in a practical project of hydropower engineering is used to verify the proposed soft computing in this study. Investigation results demonstrate that the proposed methodology is capable and valuable in addressing geomechanical parameters determination in hydropower engineering.

This paper presents a displacement back analysis approach with back propagation neural network (BPNN) and genetic algorithm (GA) to characterize the deep fractured strata. In this method, BPNN is employed to perform the nonlinear relationship between the geotechnical properties and and the wellbore deformation behaviors, and to combine the field monitoring information to establish the fitness function for GA. The learning specimens for BPNN are from the numerical simulation and field measurements. GA is employed to perform multi-objective optimization for determination of geotechnical properties on the basis of the established fitness function in a large search space. Results of characterization of geotechnical properties show that the displacement back analysis approach based on the integration of BPNN and GA can effectively recognize the horizontal in–situ stresses and natural fracture properties from wellbore displacements during drilling. This is significant in deep rock mechanics investigation because accurate and low–cost information on geotechnical properties is critical in reducing well costs and optimizing drilling.

This paper presents an application of natural computation integrating artificial neural network and Genetic algorithm in geo-stress field analysis of hydropower engineering. The artificial neural network (ANN) model is used to implement simulation analysis of the hydraulic fracturing behavior of rock mass instead of numerical computation, which can well reduce work load of numerical modeling calculation and figure out the integration problem between numerical computation and optimization algorithm. The only limited field test dataset can be applied to drill the data-driven artificial neural network model, which is one of the main advantages of ANN. The genetic algorithm (GA) is applied to perform multi-objective optimization for identification of geo-mechanical parameters by means of its objective function. The developed modeling framework is verified with field measurements in a practical project of hydropower engineering. Validation results illustrate that the proposed approach is capable and valuable in settling geo-mechanical parameters identification of nonlinear behavior problem.

The partition behavior of hafnium among different phases in FGH97 PM(powder metallurgy) superalloy and its effects on the precipitation behaviors of MC carbide and 7' phase were studied by means of 3DAP,SEM,TEM and physiochemical phase analysis.The results showed that element Hf mainly exists in 7' phase and MC carbide,which makes 7' composition transform to(Ni, Co)3(Al,Ti,Nb,Hf),also makes MC transform to(Nb,Ti,Hf)C.With Hf addition increasing,the proportion of Hf in 7' maintains constant,but in MC carbide increases and in 7 decreases,which means that partition ratio(R\) between 7' phase and MC carbide is decreased,while partition ratio {R2) between 7' phase and 7 matrix is increased,the average partition ratio between 7' phase and MC carbide is about 1:0.1,and the average partition ratio between 7' phase and 7 matrix is about 1 '. 0.05.Hf is helpful to the precipitations of 7' phase and MC carbide,the morphology and size of 7' phase are influenced more by Hf than these of MC carbide.

A new technology of constructing ultra-thin antiseepage walls by polymer injection has been developing. The mechanism of polymer expanding and filling the slot formed by the direct pushing method is studied in this paper. The expansion-filling mechanism of the polymer in the preset slot in soil is simulated with the finite-volume method (FVM) and the volume-of-fluid (VOF) method. The construction technology of polymer injection is also investigated. The injecting effect has been verified through field experiment and the antiseepage effect of the formed wall has been theoretically analyzed by the finite-element method (FEM). The results show that the polymer-injection technology is suitable for forming a continuous anti-seepage wall in earth dams or dikes.

An alternative method of detection-loophole-free Bell test is proposed using local hidden variable (LHV) models with optimal detection efficiencies. A framework for constructing such optimal LHV models is presented. Optimal LHV models for maximally and non-maximally entangled twopartite states are constructed to reproduce the quantum correlations within the critical detection efficiencies. The LHV models are shown to be completely equivalent with the existing twopartite Bell inequalities in their optimized setups, and to have even lower critical efficiencies in the LHV modes' own optimized setups. Applications in Bell tests and in device-independent quantum information processing are discussed., 27 pages, 8 figures

The oxygen sensitivity of pure TiO2, nano-CeO2 coating micro-TiO2 and nano-CeO2 coating nano-TiO2 materials was investigated in the present work. The experiments showed that the oxygen sensitivity of the nano-CeO2 coating nano-TiO2 materials was much better than that of the pure TiO2 and CeO2 coating micro-TiO2 materials. Thermodynamic analyses and XPS experimental results indicated that CeO2 was easier to be reduced into non-stoichiometric compound than TiO2 under the same change of partial oxygen pressure. With large special surface area and capacity of storing and releasing oxygen while the atmosphere changes, nano-CeO2 is a very good catalyst for the oxygen sensitivity of TiO2.

This paper deals with the synthesis, microstructures and properties of γ-aluminum oxynitride (AlON). The thermodynamic properties of AlON were analyzed and the Gibbs energy of AlON with different compositions and temperatures were evaluated. Based on thermodynamic studies, AlON has been synthesized. The microstructures, mechanical properties and oxidation resistance of the synthetic AlON have been examined and discussed.

We have determined the Debye temperature of the MgCNi3 superconductor by using the Rietveld refinement method based on the powder x-ray diffraction data. MgCNi3 crystallizes in the cubic perovskite structure with space group Pm m and lattice constant a = 3.8089 A. The temperature factors of the atoms Mg, C and Ni are 0.52, 0.45 and 0.44, respectively. The Debye temperature of MgCNi3 is calculated to be ΘD = 440 K.

There are many disadvantages of the traditional methods in quality control of filled soil-stone compaction, such as the high cost, inefficiency, impact of construction and so on. So it is one of the important problems which should be settled urgently to develop some new detection technologies which bring about high precision, quick and non-destructive methods. The basic principle and application results of portable falling weight deflectometer (PFWD) are analyzed firstly. Based on dynamics eigenvalues(dynamic parameters) of PFWD and artificial intelligence technology, the density and moisture content of filled soil-stone is attempted to be back-calculated by back-propagation (BP) neural networks prediction model between dynamic parameters of PFWD and compactness of filled soil-stone. The application results show that back-calculation the density and moisture content of filled soil-stone is a quick and accurate method, which provides a successful solution to the key technological problems of earth-rock dam construction quality control.

The dendrite tip growth kinetic coefficients and Gauss distribution parameters for 3D-CAFE simulation are determined according to the main compositions of 430 ferrite stainless steel and the macrostructure forming under the experimental condition of slow cooling. Based on the repeated computation with different heat transfer coefficients, t heat transfer coefficient under slow condition is determined when the solidification structure by above simulation computation is basically the same as the experimental one under slow cooling. The temperature fields and flow fields of 430 ferrite stainless steel during the solidification process under the conditions of slow cooling, air cooling and water cooling were analyzed by use of 3D-CAFE method, respectively. The results show that the temperature field of solidification process under slow cooling condition is the most uniform and the solid-liquid region is the widest, followed by air cooling condition, while the temperature field