Search NASA⌕ Search

SEARCH · Search NASA

Results for “Direct numerical simulation”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 523 records · Page 29

Mean-velocity profile of smooth channel flow explained by a cospectral budget model with wall-blockage

A series of recent studies has shown that a model of the turbulent vertical velocity variance spectrum (Fvv) combined with a simplified cospectral budget can reproduce many macroscopic flow properties of turbulent wall-bounded flows, including various features of the mean-velocity profile (MVP), i.e., the “law of the wall”. While the approach reasonably models the MVP’s logarithmic layer, the buffer layer displays insufficient curvature compared to measurements. The assumptions are re-examined here using a direct numerical simulation (DNS) dataset at moderate Reynolds number that includes all the requisite spectral and co-spectral information. Starting with several hypotheses for the cause of the “missing” curvature in the buffer layer, it is shown that the curvature deficit is mainly due to mismatches between (i) the modelled and DNS-observed pressure-strain terms in the cospectral budget and (ii) the DNS-observed Fvv and the idealized form used in previous models. By replacing the current parameterization for the pressure-strain term with an expansive version that directly accounts for wall-blocking effects, the modelled and DNS reported pressure-strain profiles match each other in the buffer and logarithmic layers. Forcing the new model with DNS-reported Fvv rather than the idealized form previously used reproduces the missing buffer layer curvature to high fidelity thereby confirming the “spectral link” between Fvv and the MVP across the full profile. A broad implication of this work is that much of the macroscopic properties of the flow (such as the MVP) may be derived from the energy distribution in turbulent eddies (i.e., Fvv) representing the microstate of the flow, provided the link between them accounts for wall-blocking.

54 ENVIRONMENTAL SCIENCES↗

High Resolution Stereoscopic PIV Measurements in a 5 x 5 Rod Bundle with Mixing Vane

Spacer grids are used in the pressurized water reactor (PWR) to fix, clamp the fuel rods, and enhance the turbulence mixing within the spaces of the rods and overall heat transfer performance. Proper understanding on the thermal-hydraulic characteristics of the turbulent mixing and heat transfer created by the space grids is important. However, spacer grids having complex design structure with dimple, mixing vane and spring have been challenges for computational fluid dynamics (CFD) turbulent models. High fidelity experimental data with spatial-temporal resolutions is essential to validate and assess the performances of CFD models. Baglietto and Ninokata applied an improved realizable anisotropic turbulence models via direct numerical simulation (DNS)-based near-wall formulation data to study the wall shear stresses and flow structures in tight bundle subchannels. These authors have shown that their models are able to capture the stress distribution and velocity field at an acceptable accuracy. The authors numerically studied the flow field characteristics at the region downstream of the spacer grids, and validated their results to particle imagery velocimetry (PIV) measurements. Dominguez and Hassan performed two-dimensional time resolved PIV (TR-PIV) measurements within a 5 x 5 PWR rod bundle with spacer-grids that utilized a matching index of refraction (MIR) approach. Their averaged data showed a rather uniform direction velocity profile after the spacer grid while the downstream illustrated a no slip wall profile. In this paper, we presents flow measurements in the MIR 5 x 5 fuel rod bundle with a spacer grid. Time resolved stereoscopic PIV (TR-SPIV) measurements were performed to obtain high spatial-temporal resolution two-dimensional three-component (2D3C) velocity vector fields in the subchannels between the rods. Flow statistics such as mean velocity, and Reynolds stresses were computed. The obtained experimental data can be used to improve the computational modeling and validate the numerical models that are currently applied to light and pressurized water reactors.

11 NUCLEAR FUEL CYCLE AND FUEL MATERIALS↗

A new class of finite element variational multiscale turbulence models for incompressible magnetohydrodynamics

New large eddy simulation (LES) turbulence models for incompressible magnetohydrodynamics (MHD) derived from the variational multiscale (VMS) formulation for finite element simulations are introduced. The new models include the variational multiscale formulation, a residual-based eddy viscosity model, and a mixed model that combines both of these component models. Each model contains terms that are proportional to the residual of the incompressible MHD equations and is therefore numerically consistent. Moreover, each model is also dynamic, in that its effect vanishes when this residual is small. The new models are tested on the decaying MHD Taylor Green vortex at low and high Reynolds numbers. The evaluation of the models is based on comparisons with available data from direct numerical simulations (DNS) of the time evolution of energies as well as energy spectra at various discrete times. Thus a numerical study, on a sequence of meshes, is presented that demonstrates that the large eddy simulation approaches the DNS solution for these quantities with spatial mesh refinement.

71 CLASSICAL AND QUANTUM MECHANICS, GENERAL PHYSIC↗

Test Problem: Tilted Rayleigh-Taylor for 2-D Mixing Studies

The 'tilted-rig' test problem originates from a series of experiments (Smeeton & Youngs, 1987, Youngs, 1989) performed at AWE in the late 1980's, that followed from the 'rocket-rig' experiments (Burrows et al., 1984; Read & Youngs, 1983), and exploratory experiments performed at Imperial College (Andrews, 1986; Andrews and Spalding, 1990). A schematic of the experiment is shown in Figure 1, and comprises a tank filled with light fluid above heavy, and then 'tilted' on one side of the apparatus, thus causing an 'angled interface' to the acceleration history due to rockets. Details of the configuration given in the next chapter include: fluids, dimensions, and other necessary details to simulate the experiment. Figure 2 shows results from two experiments, Case 110 (which is the source for this test problem) that has an Atwood number of 0.5, and Case 115 (a secondary source described in Appendix B), with Atwood of 0.9 Inspection of the photograph in Figure 2 (the main experimental diagnostic) for Case 110. reveals two main areas for mix development; 1) a large-scale overturning motion that produces a rising plume (spike) on the left, and falling plume (bubble) on the right, that are almost symmetric; and 2) a Rayleigh-Taylor driven mixing central mixing region that has a large-scale rotation associated with the rising and falling plumes, and also experiences lateral strain due to stretching of the interface by the plumes, and shear across the interface due to upper fluid moving downward and to the right, and lower fluid moving upward and to the left. Case 115 is similar but differs by a much larger Atwood of 0.9 that drives a strong asymmetry between a left side heavy spike penetration and a right side light bubble penetration. Case 110 is chosen as the source for the present test problem as the fluids have low surface tension (unlike Case 115) due the addition of a surfactant, the asymmetry small (no need to have fine grids for the spike), and there is extensive reasonable quality photographic data. The photographs in Figure 2 also reveal the appearance of a boundary layer at the left and right walls; this boundary layer has not been included in the test problem as preliminary calculations suggested it had a negligible effect on plume penetration and RT mixing. The significance of this test problem is that, unlike planar RT experiments such as the Rocket-Rig (Youngs, 1984), Linear Electric Motor - LEM (Dimonte, 1990), or the Water Tunnel (Andrews, 1992), the Tilted-Rig is a unique two-dimensional RT mixing experiment that has experimental data and now (in this TP) Direct Numerical Simulation data from Livescu and Wei. The availability of DNS data for the tilted-rig has made this TP viable as it provides detailed results for comparison purposes. The purpose of the test problem is to provide 3D simulation results, validated by comparison with experiment, which can be used for the development and validation of 2D RANS models. When such models are applied to 2D flows, various physics issues are raised such as double counting, combined buoyancy and shear, and 2-D strain, which have not yet been adequately addressed. The current objective of the test problem is to compare key results, which are needed for RANS model validation, obtained from high-Reynolds number DNS, high-resolution ILES or LES with explicit sub-grid-scale models. The experiment is incompressible and so is directly suitable for algorithms that are designed for incompressible flows (e.g. pressure correction algorithms with multi-grid); however, we have extended the TP so that compressible algorithms, run at low Mach number, may also be used if careful consideration is given to initial pressure fields. Thus, this TP serves as a useful tool for incompressible and compressible simulation codes, and mathematical models. In the remainder of this TP we provide a detailed specification; the next section provides the underlying assumptions for the TP, fluids, geometry details, boundary conditions (and alternative set-ups), initial conditions, and acceleration history (and ways to treat the acceleration ramp at the start of the experiment). This is followed by a section that defines data to be collected from the simulations, with results from the experiments and DNS from Livescu using the CFDNS code, and ILES simulations from Youngs using the compressible TURMOIL code and Andrews using the incompressible RTI3D code. We close the TP with concluding remarks, and Appendices that includes details of the sister Case 115, initial condition specifications for density and pressure fields. The Tilted-Rig Test Problem is intended to serve as a validation problem for RANS models, and as such we have provided ILES and DNS simulations in support of the test problem definition. The generally good agreement between experiment, ILES and DNS supports our assertion that the Tilted-Rig is useful, and the only 2-D TP that can be used to validate RANS models.

33 ADVANCED PROPULSION SYSTEMS↗

Multi-scale approach to the modeling of fission gas discharge during hypothetical loss-of-flow accident in gen-IV sodium fast reactor

The required technological and safety standards for future Gen IV Reactors can only be achieved if advanced simulation capabilities become available, which combine high performance computing with the necessary level of modeling detail and high accuracy of predictions. The purpose of this paper is to present new results of multi-scale three-dimensional (3D) simulations of the inter-related phenomena, which occur as a result of fuel element heat-up and cladding failure, including the injection of a jet of gaseous fission products into a partially blocked Sodium Fast Reactor (SFR) coolant channel, and gas/molten sodium transport along the coolant channels. The computational approach to the analysis of the overall accident scenario is based on using two different inter-communicating computational multiphase fluid dynamics (CMFD) codes: a CFD code, PHASTA, and a RANS code, NPHASE-CMFD. Using the geometry and time history of cladding failure and the gas injection rate, direct numerical simulations (DNS), combined with the Level Set method, of two-phase turbulent flow have been performed by the PHASTA code. The model allows one to track the evolution of gas/liquid interfaces at a centimeter scale. The simulated phenomena include the formation and breakup of the jet of fission products injected into the liquid sodium coolant. The PHASTA outflow has been averaged over time to obtain mean phasic velocities and volumetric concentrations, as well as the liquid turbulent kinetic energy and turbulence dissipation rate, all of which have served as the input to the core-scale simulations using the NPHASE-CMFD code. A sliding window time averaging has been used to capture mean flow parameters for transient cases. The results presented in the paper include testing and validation of the proposed models, as well the predictions of fission-gas/liquid-sodium transport along a multi-rod fuel assembly of SFR during a partial loss-of-flow accident. (authors)

21 SPECIFIC NUCLEAR REACTORS AND ASSOCIATED PLANTS↗

Physics-informed machine learning of the Lagrangian dynamics of velocity gradient tensor

Reduced models describing the Lagrangian dynamics of the velocity gradient tensor (VGT) in homogeneous isotropic turbulence (HIT) are developed under the physics-informed machine learning (PIML) framework. We consider the VGT at both Kolmogorov scale and coarse-grained scale within the inertial range of HIT. Building reduced models requires resolving the pressure Hessian and subfilter contributions, which is accomplished by constructing them using the integrity bases and invariants of the VGT. The developed models can be expressed using the extended tensor basis neural network (TBNN) introduced by Ling et al. [J. Fluid Mech. 807, 155 (2016)]. Physical constraints, such as Galilean invariance, rotational invariance, and incompressibility condition, are thus embedded in the models explicitly. Our PIML models are trained on the Lagrangian data from a high-Reynolds number direct numerical simulation (DNS). To validate the results, we perform a comprehensive out-of-sample test. We observe that the PIML model provides an improved representation for the magnitude and orientation of the small-scale pressure Hessian contributions. Statistics of the flow, as indicated by the joint PDF of second and third invariants of the VGT, show good agreement with the “ground-truth” DNS data. A number of other important features describing the structure of HIT are reproduced by the model successfully. We have also identified challenges in modeling inertial range dynamics, which indicates that a richer modeling strategy is required. This helps us identify important directions for future research, in particular towards including inertial range geometry into the TBNN.

71 CLASSICAL AND QUANTUM MECHANICS, GENERAL PHYSIC↗

Computational characterization of the secondary droplets formed during the impingement of a train of ethanol drops

This paper uniquely characterizes the secondary droplets formed during the impingement of a train of ethanol drops, using three-dimensional direct numerical simulations performed under conditions studied experimentally by Yarin and Weiss. Our numerical results have been previously validated against experimental data demonstrating the ability to accurately capture the splashing dynamics. In this work, the predictive ability of the model is leveraged to gain further insight into secondary droplet formation. We present a robust post-processing algorithm, which scrutinizes the liquid volume fraction field in the volume-of-fluid method and quantifies the number, volume and velocity of secondary droplets. The high-resolution computational simulations enable secondary droplet characterization within close proximity of the impingement point at small length and time scales, which is extremely challenging to achieve experimentally. By studying the temporal evolution of secondary droplet formation, direct connections are made between liquid structures seen in the simulation and the instantaneous distribution of secondary droplets, leading to detailed insight into the instability-driven breakup process of lamellae. Time-averaged secondary droplet characteristics are also studied to describe the global distribution of secondary droplets. Such analysis is vital to understanding fuel drop impingement in direct injection engines, facilitating the development of highly accurate spray–wall interaction models for use in Lagrangian solvers.

42 ENGINEERING↗

Dynamic Multiscale Averaging (DMA) of Turbulent Flow

A new approach called dynamic multiscale averaging (DMA) for computing the effects of turbulent flow is described. The new method encompasses multiple applications of temporal and spatial averaging, that is, multiscale operations. Initially, a direct numerical simulation (DNS) is performed for a relatively short time; it is envisioned that this short time should be long enough to capture several fluctuating time periods of the smallest scales. The flow field variables are subject to running time averaging during the DNS. After the relatively short time, the time-averaged variables are volume averaged onto a coarser grid. Both time and volume averaging of the describing equations generate correlations in the averaged equations. These correlations are computed from the flow field and added as source terms to the computation on the next coarser mesh. They represent coupling between the two adjacent scales. Since they are computed directly from first principles, there is no modeling involved. However, there is approximation involved in the coupling correlations as the flow field has been computed for only a relatively short time. After the time and spatial averaging operations are applied at a given stage, new computations are performed on the next coarser mesh using a larger time step. The process continues until the coarsest scale needed is reached. New correlations are created for each averaging procedure. The number of averaging operations needed is expected to be problem dependent. The new DMA approach is applied to a relatively low Reynolds number flow in a square duct segment. Time-averaged stream-wise velocity and vorticity contours from the DMA approach appear to be very similar to a full DNS for a similar flow reported in the literature. Expected symmetry for the final results is produced for the DMA method. The results obtained indicate that DMA holds significant potential in being able to accurately compute turbulent flow without modeling for practical engineering applications.

97 MATHEMATICS AND COMPUTING↗

Self-contained filtered density function

The filtered density function (FDF) closure is extended to a “self-contained” format to include the subgrid-scale (SGS) statistics of all of the hydro-thermo-chemical variables in turbulent flows. These are the thermodynamic pressure, the specific internal energy, the velocity vector, and the composition field. In this format, the model is comprehensive and facilitates large-eddy simulation (LES) of flows at both low and high compressibility levels. A transport equation is developed for the joint pressure-energy-velocity-composition filtered mass density function (PEVC-FMDF). In this equation, the effect of convection appears in closed form. The coupling of the hydrodynamics and thermochemistry is modeled via a set of stochastic differential equation for each of the transport variables. This yields a self-contained SGS closure. We demonstrated how LES is conducted of a turbulent shear flow with transport of a passive scalar. Finally, the consistency of the PEVC-FMDF formulation is established, and its overall predictive capability is appraised via comparison with direct numerical simulation (DNS) data.

71 CLASSICAL AND QUANTUM MECHANICS, GENERAL PHYSIC↗

Advanced Model Development for Large Eddy Simulation of Oxy-Combustion and Supercritical Carbon Dioxide Power Cycles

A joint experimental and numerical study is performed to observe the characteristics of a supercritical carbon dioxide turbulent mixing layer in the presence of strong nonlinearities in the thermodynamic and transport properties. A bespoke experimental setup is designed and employed for this purpose and provides insight into macroscopic mixing behavior. The mixing is experimentally observed using two techniques: shadowgraphy and spontaneous Raman scattering. Qualitative and quantitative intensity fields obtained via these techniques yield instantaneous and mean density data. Spanwise temperature data is also collected using analogue resistance temperature detectors. These measurements are used to quantify the level of mixed material within the field. The experimental data are supplemented by a companion high-fidelity numerical study. The numerical results are obtained through fully resolved, three-dimensional direct numerical simulation. The numerical dataset permits observation of the near-field mixing characteristics, which are difficult to measure experimentally due to the rapid dynamics and sharp thermophysical gradients in this area. Qualitative field visualizations are presented, followed by quantitative mixed material results and observations regarding thermodynamic property trends at select locations within the field. One-dimensional spectra of the turbulent kinetic energy and solenoidal dissipation are provided to observe the spectral characteristics of the flow. Reynolds stress anisotropy is analyzed graphically through anisotropy invariance maps (Lumley triangles). The mixing quantification, spectral data and anisotropy analysis of a flow at these thermodynamic conditions represent the main outcomes of the work.

20 FOSSIL-FUELED POWER PLANTS↗

Some Specific CASL Requirements for Advanced Multiphase Flow Simulation of Light Water Reactors

Because of the diversity of physical phenomena occuring in boiling, flashing, and bubble collapse, and of the length and time scales of LWR systems, it is imperative that the models have the following features: • Both vapor and liquid phases (and noncondensible phases, if present) must be treated as compressible. • Models must be mathematically and numerically well-posed. • The models methodology must be multi-scale. A fundamental derivation of the multiphase governing equation system, that should be used as a basis for advanced multiphase modeling in LWR coolant systems, is given in the Appendix using the ensemble averaging method. The remainder of this work focuses specifically on the compressible, well-posed, and multi-scale requirements of advanced simulation methods for these LWR coolant systems, because without these are the most fundamental aspects, without which widespread advancement cannot be claimed. Because of the expense of developing multiple special-purpose codes and the inherent inability to couple information from the multiple, separate length- and time-scales, efforts within CASL should be focused toward development of a multi-scale approaches to solve those multiphase flow problems relevant to LWR design and safety analysis. Efforts should be aimed at developing well-designed unified physical/mathematical and high-resolution numerical models for compressible, all-speed multiphase flows spanning: (1) Well-posed general mixture level (true multiphase) models for fast transient situations and safety analysis, (2) DNS (Direct Numerical Simulation)-like models to resolve interface level phenmena like flashing and boiling flows, and critical heat flux determination (necessarily including conjugate heat transfer), and (3) Multi-scale methods to resolve both (1) and (2) automatically, depending upon specified mesh resolution, and to couple different flow models (single-phase, multiphase with several velocities and pressures, multiphase with single velocity and pressure, etc.) A unified, multi-scale approach is advocated to extend the necessary foundations and build the capability to simultaneously solve the fluid dynamic interface problems (interface resolution) as well as multiphase mixtures (homogenization).

11 NUCLEAR FUEL CYCLE AND FUEL MATERIALS↗

Life cycle of streaks in the buffer layer of wall-bounded turbulence

In this work, streaks in the buffer layer of wall-bounded turbulence are tracked in time to study their life cycle. Spatially and temporally resolved direct numerical simulation data are used to analyze the strong wall-parallel movements conditioned to low-speed streamwise flow. The analysis of the streaks shows that there is a clear distinction between wall-attached and detached streaks, and that the wall-attached streaks can be further categorized into streaks that are contained in the buffer layer and the ones that reach the outer region. The results reveal that streaks are born in the buffer layer, coalescing with each other to create larger streaks that are still attached to the wall. Once the streak becomes large enough, it starts to meander due to the large streamwise-to-wall-normal aspect ratio, and consequently the elongation in the streamwise direction, which makes it more difficult for the streak to be oriented strictly in the streamwise direction. While the continuous interaction of the streaks allows the superstructure to span extremely long temporal and length scales, individual streak components are relatively small and short-lived. Tall-attached streaks eventually split into wall-attached and wall-detached components. These wall-detached streaks have a strong wall-normal velocity away from the wall, similar to ejections or bursts observed in the literature. Conditionally averaging the flow fields to these split events show that the detached streak has not only a larger wall-normal velocity compared to the wall-attached counterpart, it also has a larger (less negative) streamwise velocity, similar to the velocity field at the tip of a vortex cluster.

42 ENGINEERING↗

Some observations concerning "laminarization" in heated vertical tubes

For low "turbulent" Reynolds numbers in strongly-heated vertical gas flow, fundamental results of existing direct numerical simulation (DNS) databases have been examined to deduce differences between cases which "revert" to turbulent and those that yield integral parameters which correspond to laminar flow (henceforth laminarizing or laminarized). The objectives of this paper are (1) to examine the streamwise evolution of near-wall flow structures and (2) to determine which, if any, proposed laminarization parameters could be used to discriminate between turbulent and laminarizing flows resulting. Views of streamlines in r-θ cross sections showed evidence of a ring of irregular-shaped streamwise vortices near the wall at all locations. The transverse advection of these vortices is hypothesized to provide a path of least resistance to the transport of streamwise momentum in the wall-normal direction. It is demonstrated that a steady streamwise laminar vortex can provide apparent turbulent momentum transport (as a consequence of the Reynolds decomposition) such that a reasonable mean velocity profile is predicted near the wall. For the present cases, the values of the non-dimensional radius (y c,w + = Re τ,w ) and of the modified wall Reynolds number did discriminate whether turbulent or laminarizing flow resulted.

99 GENERAL AND MISCELLANEOUS↗

Strongly Heated Turbulent Flow in a Channel with Pin Fins

Large-eddy simulations (LES) were performed to study the turbulent flow in a channel of height H with a staggered array of pin fins with diameter D = H/2 as a function of heating loads that are relevant to the cooling of turbine blades and vanes. The following three heating loads were investigated—wall-to-coolant temperatures of Tw/Tc = 1.01, 2.0, and 4.0 - where the Reynolds number at the channel inlet was 10,000 and the back pressure at the channel outlet was 1 bar. For the LES, two different subgrid-scale models—the dynamic kinetic energy model (DKEM) and the wall-adapting local eddy-viscosity model (WALE)—were examined and compared. This study was validated by comparing with data from direct numerical simulation and experimental measurements. The results obtained show high heating loads to create wall jets next to all heated surfaces that significantly alter the structure of the turbulent flow. Results generated on effects of heat loads on the mean and fluctuating components of velocity and temperature, turbulent kinetic energy, the anisotropy of the Reynolds stresses, and velocity-temperature correlations can be used to improve existing RANS models.

42 ENGINEERING↗

Effects of equivalence ratio variations on turbulent flame speed in lean methane/air mixtures under lean-burn natural gas engine operating conditions

Direct numerical simulations of turbulent premixed air flames were carried out using an inflow-outflow configuration to study the effects of equivalence ratio on the turbulent flame speed in lean mixtures. The inflow velocity was dynamically adjusted at a run-time to stabilize the flame brush location within the computational domain. Linear forcing was applied to maintain the turbulent intensities at desired levels. Numerous equivalence ratios near the lean limit were selected for the same turbulence properties and the normalized turbulent flame speed was shown to be a function of the equivalence ratio. Simulations were performed for over 80 eddy turnover times with the turbulent flame speed obtained by averaging the inflow velocity. Results revealed that the equivalence ratio does not have an explicit effect on the normalized turbulent flame speed over the lean limit. Analysis of flame surface area showed that the surface wrinkling produced by eddies of varying scales was not influenced by the change in equivalence ratios when the Karlovitz and Damkohler numbers are fixed. Lastly based on the flame surface statistics flame surface normal is preferentially parallel to the most compressive strain rate direction for all equivalence ratios.

03 NATURAL GAS↗

Automated and efficient local adaptive regression for principal component-based reduced-order modeling of turbulent reacting flows

Principal Component Analysis can be used to reduce the cost of Computational Fluid Dynamics simulations of turbulent reacting flows by reducing the dimensionality of the transported variables through projection of the thermochemical state onto a lower-dimensional manifold. However, because of the nonlinearity of the principal component source terms, nonlinear regression techniques must be utilized for the source terms in terms of the principal components. Unfortunately, widely available and utilized nonlinear regression techniques can have prohibitive computational requirements and/or accuracy that is highly dependent on user experience in ad hoc tuning of model architecture and hyperparameters. Here, in this work, a new nonlinear regression approach is proposed that is both computationally efficient and automated so does not require any user input. The approach is evaluated through a priori prediction of principal component source terms using data from a Direct Numerical Simulation of a turbulent nonpremixed n-heptane/air jet flame. In particular, the proposed framework consists of local regressions whose complexity is adapted according to the local nonlinearity of the data: local linear regression when accurate enough and local Artificial Neural Networks when nonlinear regression is required. The number of local clusters for local regression is determined automatically using the Davies-Bouldin index. In addition, Bayesian optimization is utilized for model training (i.e., to select the best architectures and hyperparameters of the nonlinear regressions in an unsupervised fashion), eliminating ad hoc hand-tuning and/or expensive grid searches. Overall, compared to a single, global neural network, the new local adaptive regression approach is shown to have comparable accuracy but 69% less training time due to the utilization of local linear regression and faster training of local neural networks.

42 ENGINEERING↗