Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “Regularization”

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 289 records · Page 16
All sourcesNASA NTRSDOE OSTI
NASA NTRS2007

Coronas-F Orbit Monitoring and Re-Entry Prediction

Russian scientific satellite CORONAS-F was launched on July, 31, 2001. The object was inserted in near-circular orbit with the inclination 82.5deg and a mean altitude approx. 520 km. Due to the upper atmosphere drag CORONAS-F was permanently descended and as a result on December, 6, 2005 it has finished the earth-orbital flight, having lifetime in space approx. 4.5 years. The satellite structural features and its flight attitude control led to the significant variations of its ballistic coefficient during the flight. It was a cause of some specific difficulties in the fulfillment of the ballistic and navigation support of this space vehicle flight. Besides the main mission objective CORONAS-F also has been selected by the Inter-Agency Space Debris Coordination Committee (IADC) as a target object for the next regular international re-entry test campaign on a program of surveillance and re-entry prediction for the hazard space objects within their de-orbiting phases. Spacecraft (S/C) CORONAS-F kept its working state right up to the end of the flight - down to the atmosphere entry. This fact enabled to realization of the additional research experiments, concerning with an estimation of the atmospheric density within the low earth orbits (LEO) of the artificial satellites, and made possible to continue track the S/C during final phase of its flight by means of Russian regular command & tracking system, used for it control. Thus there appeared a unique possibility of using for tracking S/C at its de-orbiting phase not only passive radar facilities, belonging to the space surveillance systems and traditionally used for support of the IADC re-entry test campaigns, but also more precise active trajectory radio-tracking facilities from the ground control complex (GCC) applied for this object. Under the corresponding decision of the Russian side such capability of additional high-precise tracking control of the CORONAS-F flight in this period of time has been implemented. The organizing of the CORONAS-F ballistic and navigational support (BNS) and solving its main tasks (such as S/C orbit determination (OD) and its motion prediction and connected with them) both for regular mission stage and for additional flight program were realized by the group of specialists from the Mission Control Center (MCC). MCC was also assigned as a principal organization from the Russian side for participation in the 7th IADC re-entry test campaign on CORONAS-F. The CORONAS-F flight features and space environments circumstances during its flight as well as a methodology and technology of spacecraft ballistic and navigational support are given below. The BNS results for different phases of S/C flight, including the results of its re-entry predictions, obtained during the realization of the 7th IADC test campaign are submitted. The accuracy of space vehicle re-entry prediction and its dependence on various factors are analyzed in more details.

Ivanov, N. M.↗
NASA NTRS2006

High-sensitivity Cryogenic Temperature Sensors using Pressurized Fiber Bragg Gratings

Cryogenic temperature sensing was studied using a pressurized fiber Bragg grating (PFBG). The PFBG was obtained by simply applying a small diametric load to a regular fiber Bragg grating (FBG), which was coated with polyimide of a thickness of 11 micrometers. The Bragg wavelength of the PFBG was measured at temperatures from 295 to 4.2 K. A pressure-induced transition occurred at 200 K during the cooling cycle. As a result the temperature sensitivity of the PFBG was found to be nonlinear but reach 24 pm/K below 200 K, more than three times the regular FBG. For the temperature change from 80 K to 10 K, the PFBG has a total Bragg wavelength shift of about 470 pm, 10 times more than the regular FBG. From room temperature to liquid helium temperature the PFBG gives a total wavelength shift of 3.78 nm, compared to the FBG of 1.51 nm. The effect of the coating thickness on the temperature sensitivity of the gratings is also discussed.

Wu, Meng-Chou↗
NASA NTRS2009

LDPC Codes with Minimum Distance Proportional to Block Size

Low-density parity-check (LDPC) codes characterized by minimum Hamming distances proportional to block sizes have been demonstrated. Like the codes mentioned in the immediately preceding article, the present codes are error-correcting codes suitable for use in a variety of wireless data-communication systems that include noisy channels. The previously mentioned codes have low decoding thresholds and reasonably low error floors. However, the minimum Hamming distances of those codes do not grow linearly with code-block sizes. Codes that have this minimum-distance property exhibit very low error floors. Examples of such codes include regular LDPC codes with variable degrees of at least 3. Unfortunately, the decoding thresholds of regular LDPC codes are high. Hence, there is a need for LDPC codes characterized by both low decoding thresholds and, in order to obtain acceptably low error floors, minimum Hamming distances that are proportional to code-block sizes. The present codes were developed to satisfy this need. The minimum Hamming distances of the present codes have been shown, through consideration of ensemble-average weight enumerators, to be proportional to code block sizes. As in the cases of irregular ensembles, the properties of these codes are sensitive to the proportion of degree-2 variable nodes. A code having too few such nodes tends to have an iterative decoding threshold that is far from the capacity threshold. A code having too many such nodes tends not to exhibit a minimum distance that is proportional to block size. Results of computational simulations have shown that the decoding thresholds of codes of the present type are lower than those of regular LDPC codes. Included in the simulations were a few examples from a family of codes characterized by rates ranging from low to high and by thresholds that adhere closely to their respective channel capacity thresholds; the simulation results from these examples showed that the codes in question have low error floors as well as low decoding thresholds. As an example, the illustration shows the protograph (which represents the blueprint for overall construction) of one proposed code family for code rates greater than or equal to 1.2. Any size LDPC code can be obtained by copying the protograph structure N times, then permuting the edges. The illustration also provides Field Programmable Gate Array (FPGA) hardware performance simulations for this code family. In addition, the illustration provides minimum signal-to-noise ratios (Eb/No) in decibels (decoding thresholds) to achieve zero error rates as the code block size goes to infinity for various code rates. In comparison with the codes mentioned in the preceding article, these codes have slightly higher decoding thresholds.

Divsalar, Dariush↗
NASA NTRS2009

Simplified Approach to Predicting Rough Surface Transition

Turbine vane heat transfer predictions are given for smooth and rough vanes where the experimental data show transition moving forward on the vane as the surface roughness physical height increases. Consistent with smooth vane heat transfer, the transition moves forward for a fixed roughness height as the Reynolds number increases. Comparisons are presented with published experimental data. Some of the data are for a regular roughness geometry with a range of roughness heights, Reynolds numbers, and inlet turbulence intensities. The approach taken in this analysis is to treat the roughness in a statistical sense, consistent with what would be obtained from blades measured after exposure to actual engine environments. An approach is given to determine the equivalent sand grain roughness from the statistics of the regular geometry. This approach is guided by the experimental data. A roughness transition criterion is developed, and comparisons are made with experimental data over the entire range of experimental test conditions. Additional comparisons are made with experimental heat transfer data, where the roughness geometries are both regular and statistical. Using the developed analysis, heat transfer calculations are presented for the second stage vane of a high pressure turbine at hypothetical engine conditions.

Boyle, R. J.↗
NASA NTRS2007

Process for High-Rate Fabrication of Alumina Nanotemplates

An anodizing process, at an early stage of development at the time of reporting the information for this article, has shown promise as a means of fabricating alumina nanotemplates integrated with silicon wafers. Alumina nanotemplates are basically layers of alumina, typically several microns thick, in which are formed approximately regular hexagonal arrays of holes having typical diameters of the order of 10 to 100 nm. Interest in alumina nanotemplates has grown in recent years because they have been found to be useful as templates in the fabrication of nanoscale magnetic, electronic, optoelectronic, and other devices. The present anodizing process is attractive for the fabrication of alumina nanotemplates integrated with silicon wafers in two respects: (1) the process involves self-ordering of the holes; that is, the holes as formed by the process are spontaneously arranged in approximately regular hexagonal arrays; and (2) the rates of growth (that is, elongation) of the holes are high enough to make the process compatible with other processes used in the mass production of integrated circuits. In preparation for fabrication of alumina nanotemplates in this process, one first uses electron-beam evaporation to deposit thin films of titanium, followed by thin films of aluminum, on silicon wafers. Then the alumina nanotemplates are formed by anodizing the aluminum layers, as described below. In experiments in which the process was partially developed, the titanium films were 200 A thick and the aluminum films were 5 m thick. The aluminum films were oxidized to alumina, and the arrays of holes were formed by anodizing the aluminum in aqueous solutions of sulfuric and/or oxalic acid at room temperature (see figure). The diameters, spacings, and rates of growth of the holes were found to depend, variously, on the composition of the anodizing solution, the applied current, or the applied potential, as follows: In galvanostatically controlled anodizing, regardless of the chemical composition of the solution, relatively high current densities (50 to 100 mA/cm2) resulted in arrays of holes that were more nearly regular than were those formed at lower current densities. . The rates of elongation of the holes were found to depend linearly on the applied current density: the observed factor of proportionality was 1.2 (m/h)/(mA/cm2). For a given fixed current density and room temperature, the hole diameters were found to depend mainly on the chemical compositions of the anodizing solutions. The holes produced in sulfuric acid solutions were smaller than those produced in oxalic acid solutions. The arrays of holes produced in sulfuric acid were more ordered than were those produced in oxalic acid. . The breakdown voltage was found to decrease logarithmically with increasing concentration of sulfuric acid. The breakdown voltage was also found to decrease with temperature and to be accompanied by a decrease in hole diameter. The hole diameter was found to vary linearly with applied potential, with a slope of 2.1 nm/V. This slope differs from slopes (2.2 and 2.77 nm/V) reported for similar prior measurements on nanotemplates made from bulk aluminum. The differences among these slopes may be attributable to differences among impurities and defects in bulk and electron-beam-evaporated aluminum specimens.

Myung, Nosang↗
NASA NTRS2010

Uterine Contraction Modeling and Simulation

Building a training system for medical personnel to properly interpret fetal heart rate tracing requires developing accurate models that can relate various signal patterns to certain pathologies. In addition to modeling the fetal heart rate signal itself, the change of uterine pressure that bears strong relation to fetal heart rate and provides indications of maternal and fetal status should also be considered. In this work, we have developed a group of parametric models to simulate uterine contractions during labor and delivery. Through analysis of real patient records, we propose to model uterine contraction signals by three major components: regular contractions, impulsive noise caused by fetal movements, and low amplitude noise invoked by maternal breathing and measuring apparatus. The regular contractions are modeled by an asymmetric generalized Gaussian function and least squares estimation is used to compute the parameter values of the asymmetric generalized Gaussian function based on uterine contractions of real patients. Regular contractions are detected based on thresholding and derivative analysis of uterine contractions. Impulsive noise caused by fetal movements and low amplitude noise by maternal breathing and measuring apparatus are modeled by rational polynomial functions and Perlin noise, respectively. Experiment results show the synthesized uterine contractions can mimic the real uterine contractions realistically, demonstrating the effectiveness of the proposed algorithm.

Liu, Miao↗
NASA NTRS2011

Analysis of the Radiative Transfer Equation with Highly Asymmetric Phase Function

This paper considers a scalar radiative transfer problem with high scattering anisotropy, Two computational methods are presented based on decomposition of the diffuse light field into a regular and anisotropic part. The first algorithm (DOMAS) singles out the anisotropic radiance in the forward scattering peak using the Small-Angle Modification of RTE. The second algorithm (DOM2+) separates the single scattering radiance as an anisotropic part, which largely defines the fine detail of the total radiance in the backscattering directions. In both cases, the anisotropic part is represented analytically. With anisotropy subtraction, the regular part of the signal. which requires a numerical solution, is essentially smoothed as a function of angles. Further, the transport equation is obtained for the regular part that contains an additional source function from the anisotropic part of the signal. This equation is solved with the discrete ordinates method. A conducted numerical analysis of this work showed that algorithm DOMAS has a strong advantage as compared to the standard discrete ordinates method for simulation of the radiance transmission, and DOM2 + is the best of the three for the reflection computations. Both algorithms offer at least a factor of three acceleration of convergence of the azimuthal series for highly anisotropic phase functions.

Korkin, Sergey V.↗
NASA NTRS2014

A Technique for Mapping Characteristic Lengths to Preserve Energy Dissipated via Strain Softening in a Multiscale Analysis

It is often advantageous to account for the microstructure of the material directly using multiscale modeling. For computational tractability, an idealized repeating unit cell (RUC) is used to capture all of the pertinent features of the microstructure. Typically, the RUC is dimensionless and depends only on the relative volume fractions of the different phases in the material. This works well for non-linear and inelastic behavior exhibiting a positive-definite constitutive response. Although, once the material exhibits strain softening, or localization, a mesh objective failure theories, such as smeared fracture theories, nodal and element enrichment theories (XFEM), cohesive elements or virtual crack closure technique (VCCT), can be utilized at the microscale, but the dimensions of the RUC must then be defined. One major challenge in multiscale progressive damage modeling is relating the characteristic lengths across the scales in order to preserve the energy that is dissipated via localization at the microscale. If there is no effort to relate the size of the macroscale element to the microscale RUC, then the energy that is dissipated will remain mesh dependent at the macroscale, even if it is regularized at the microscale. Here, a technique for mapping characteristic lengths across the scales is proposed. The RUC will be modeled using the generalized method of cells (GMC) micromechanics theory, and local failure in the matrix constituent subcells will be modeled using the crack band theory. The subcell characteristic lengths used in the crack band calculations will be mapped to the macroscale finite element in order to regularize the local energy in a manner consistent with the global length scale. Examples will be provided with and without the regularization, and they will be compared to a baseline case where the size and shape of the element and RUC are coincident (ensuring energy is preserved across the scales).

softening↗
NASA NTRS2018

Outcomes of 7 Years of Airborne Trace Gas Measurements over California and Nevada: The Alpha Jet Atmospheric eXperiment (AJAX)

The Alpha Jet Atmospheric eXperiment (AJAX) has been flying a scientific payload since January 2011 measuring ozone, carbon dioxide, methane, formaldehyde and meteorological parameters up to 9 kilometers. AJAX is located and operated from the San Francisco Bay Area and has flown a total of 229 flights, on a regular basis (approximately 3 per month) over all seasons cataloguing a long-term record of trace gas concentrations over California and Nevada. The AJAX project focuses on science questions which benefit from routine, frequent observations with flexible scheduling. This presentation will provide an overview of AJAX activities including a discussion of airborne measurements for: Long-Range Transport (LRT) and Stratosphere-to-Troposphere Transport (STT). Regular sampling by AJAX has aided identification of LRT and evidence of STT, which during spring and summer months are visible as elevated O3 laminae within airborne profiles. Some laminae have the ability to impact surface level air quality; Satellite validation - Regular AJAX missions include flights to Railroad Valley, NV in coordination with GOSAT (Greenhouse Gases Observing Satellite) and OCO-2 (Carbon Observatory-2) observations, and more recently to provide coincident measurements under TROPOMI (TROPOspheric Monitoring Instrument); The AJAX project is uniquely flexible to incorporate specialized flights with limited planning notice, such as sampling emissions from California wildfires. Nine wildfires have been sampled, with some more than once allowing to observe emission changes as the fire progresses; Pandora validation - Future work will include development of flight strategies for validation of ground based Pandora spectrometers.

Yates, Emma L.↗
NASA NTRS2020

Lunar Magnetic Field Models From Lunar Prospector and SELENE/Kaguya Along-Track Magnetic Field Gradients

We use L1-norm model regularization of |Br| component at the surface on magnetic monopoles bases and along-track magnetic field differences alone (without vector observations) to derive high quality global magnetic field models at the surface of the Moon. The practical advantages to this strategy are the following: monopoles are more stable at closer spacing in comparison to dipoles, improving spatial resolution; L1-norm model regularization leads to sparse models which may be appropriate for the Moon which has regions of localized magnetic field features; and along-track differences reduce the need for ad-hoc external field noise reduction strategies. We examine also the use of Lunar Prospector and SELENE/Kaguya magnetometer data, combined and separately, and find that the Lunar Prospector along-track vector field differences lead to surface field models that require weaker regularization and, hence, result in higher spatial resolution. Significantly higher spatial resolution (wavelengths of roughly 25–30 km) and higher amplitude surface magnetic fields can be derived over localized regions of high amplitude anomalies (due to their higher signal-to-noise ratio). These high-resolution field models are also compared with the results of Surface Vector Mapping approach of Tsunakawa et al. (2015, https://doi.org/10.1002/2014JE004785). Finally, the monopoles- as well as dipoles-based patterns of the Serenitatis high amplitude magnetic feature have characteristic textbook patterns of Br and Bθ component fields from a nearly vertically downwardly magnetized source region and it implies that the principal source of the anomaly was formed when the region was much closer to the north magnetic pole of the Moon.

D. Ravat↗
NASA NTRS2022

Quantifying and Reducing Uncertainty in Microwave Vegetation Optical Depth and Soil Moisture Retrievals

Soil moisture and vegetation optical depth (VOD; related to vegetation water content) retrieved from SMAP and SMOS satellites are widely used for a range of hydrosphere and biosphere applications. However, while soil moisture has been globally well-validated, VOD validation has been sparse. Furthermore, simultaneously retrieval of these parameters results in uncertainties both individually in soil moisture and VOD retrievals as well as in compensation between the parameters. Here, we show global locations where soil moisture and VOD retrievals will have lower uncertainty, based on complementary brightness temperature information content and signal-to-noise ratio metrics. In these same locations, we show that error still propagates more into VOD. However, using VOD regularization algorithms, this error is greatly reduced, especially at sub-weekly timescales where algorithmic error can be most apparent. Despite these regularization approaches that reduce errors, there are yet vast differences in available global regularized retrievals originating from different algorithmic choices.

Passive microwave radiometry↗
DOE OSTI2025

An FFT-based micromechanical model for gradient enhanced brittle fracture

Damage models incorporated within FFT-based micromechanical methods have received much attention recently because of the need to better understand and predict brittle and ductile fracture. An important aspect of a damage model is non-local regularization, which removes the mesh dependence of the predictions that otherwise become physically unacceptable upon grid refinement. In this work, the Helmholtz-type equation for non-local gradient regularization of a damage model on a distorted grid is solved using an FFT-based approach. Further, the resulting system of equations is solved using the Jacobi iterative method. The model is applied to simulate brittle fracture of an intermetallic. The influence of the time and space discretization, the length-scale parameter, and intermetallic crystallographic orientation on crack evolution is studied.

36 MATERIALS SCIENCE↗
DOE OSTI2026

PDE-constrained high-order mesh optimization

Here, we present a novel framework for PDE-constrained r-adaptivity of high-order meshes. The proposed method formulates mesh movement as an optimization problem, with an objective function defined as a convex combination of a mesh quality metric and a measure of the accuracy of the PDE solution obtained via finite element discretization. The proposed formulation achieves optimized, well-defined high-order meshes by integrating mesh quality control, PDE solution accuracy, and robust gradient regularization. We adopt the Target-Matrix Optimization Paradigm to control geometric properties across the mesh, independent of the PDE of interest. To incorporate the accuracy of the PDE solution, we introduce error measures that control the finite element discretization error. The implicit dependence of these error measures on the mesh nodal positions is accurately captured by adjoint sensitivity analysis. Additionally, a convolution-based gradient regularization strategy is used to ensure stable and effective adaptation of high-order meshes. We demonstrate that the proposed framework can improve mesh quality and reduce the error by up to 10 times for the solution of Poisson and linear elasto-static problems. The approach is general with respect to the dimensionality, the order of the mesh, the types of mesh elements, and can be applied to any PDE that admits well-defined adjoint operators.

Computer science↗
DOE OSTI2025

Bulk nanocrystalline Al–Mg–Y alloys with amorphous grain boundary complexions display high strength and compressive plasticity

Although nanocrystalline alloys regularly exhibit high strengths, their use in structural applications often face challenges due to sample size limitations, unstable microstructures, and the limited ability to plastically deform. The incorporation of amorphous grain boundary complexions has been proposed to address these issues, by simultaneously stabilizing nanocrystalline grain structures for scale-up processing and improving alloy toughness. In the present study, the mechanical behavior of bulk nanocrystalline Al–Mg–Y is examined with macroscale compression testing, probing a length scale that is relevant to real-world structural applications. Bulk samples were fabricated via a simple powder metallurgy approach, with different hot-pressing temperatures and durations employed for consolidation in order to investigate microstructural and property evolution. All of the specimens contained primary face-centered cubic Al and secondary Al 4 C 3 and Al 3 Y phases, with the Al 3 Y particles exhibiting two populations of small equiaxed and larger elongated particles. Appreciable plasticity was measured along with high ultimate stresses over 800 MPa due to the presence of amorphous grain boundary complexions. Microstructural characterization of fracture surfaces revealed that the area fraction of dimpled regions increased with longer hot-pressing time. Most importantly, the elongated Al 3 Y particles formed regular cellular patterns with increasing hot-pressing time, delaying shear localization and significantly enhancing plasticity. The hierarchy present in the microstructure of the Al–Mg–Y alloy, from amorphous grain boundary complexions to secondary phases, gives rise to excellent bulk mechanical properties, which are attractive for structural applications.

Bulk nanocrystalline alloy↗
DOE OSTI2024

Clustering and Cliques in Preferential Attachment Random Graphs with Edge Insertion

In this paper, we investigate the global clustering coefficient (a.k.a transitivity) and clique number of graphs generated by a preferential attachment random graph model with an additional feature of allowing edge connections between existing vertices. Specifically, at each time step t, either a new vertex is added with probability f(t), or an edge is added between two existing vertices with probability 1 – f(t). We establish concentration inequalities for the global clustering and clique number of the resulting graphs under the assumption that f(t) is a regularly varying function at infinity with index of regular variation –$\gamma$, where $\gamma$ $\in$ [0, 1). Finally, we also demonstrate an inverse relation between these two statistics: the clique number is essentially the reciprocal of the global clustering coefficient.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2026

Memory-efficient nonsmooth dynamic optimization using adaptive randomized compression

Dynamic optimization problems arise in many applications including flow control, full waveform inversion, and medical imaging. These problems are plagued by significant computational challenges. One such challenge — and the focus of this work — is the memory limitation induced by the size of the underlying dynamical system. In particular, the entire dynamic trajectory is required for derivative computation and therefore must be stored or recomputed using, e.g., checkpointing. Although recent work demonstrated the use of adaptive randomized sketching to overcome the memory challenge, that work only applies to smooth unconstrained problems, prohibiting its use for nonsmooth regularized and constrained problems. The inclusion of nonsmooth regularizers and constraints is critical as they often arise in an attempt to preserve certain physical properties or to promote sparsity. To solve these problems, we introduce a trust-region algorithm for minimizing the sum of a smooth nonconvex function and a nonsmooth convex function that leverages randomized sketching to compress the dynamical system trajectories and adaptively adjust the sketch rank to satisfy a gradient inexactness condition. We prove convergence of this algorithm and demonstrate that it achieves substantial memory reduction on three discretized PDE-constrained optimization applications.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2025

Graph decomposition techniques for solving combinatorial optimization problems with variational quantum algorithms

The quantum approximate optimization algorithm (QAOA) has the potential to approximately solve complex combinatorial optimization problems in polynomial time. However, current noisy quantum devices cannot solve large problems due to hardware constraints. In this work, we develop an algorithm that decomposes the QAOA input problem graph into a smaller problem and solves MaxCut using QAOA on the reduced graph. The algorithm requires a subroutine that can be classical or quantum—in this work, we implement the algorithm twice on each graph. One implementation uses the classical solver Gurobi in the subroutine and the other uses QAOA. We solve these reduced problems with QAOA. On average, the reduced problems require only approximately 1/10 of the number of vertices than the original MaxCut instances. Furthermore, the average approximation ratio of the original MaxCut problems is 0.75, while the approximation ratios of the decomposed graphs are on average of 0.96 for both Gurobi and QAOA. With this decomposition, we are able to measure optimal solutions for ten 100-vertex graphs by running single-layer QAOA circuits on the Quantinuum trapped-ion quantum computer H1-1, sampling each circuit only 500 times. This approach is best suited for sparse, particularly k-regular graphs, as k-regular graphs on n vertices can be decomposed into a graph with at most $\frac{nk}{k+1}$ vertices in polynomial time. Further reductions can be obtained with a potential trade-off in computational time. In conclusion, while this paper applies the decomposition method to the MaxCut problem, it can be applied to more general classes of combinatorial optimization problems.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2024

Surrogate construction via weight parameterization of residual neural networks

Surrogate model development is a critical step for uncertainty quantification or other sample-intensive tasks for complex computational models. Here, in this work, we develop a multi-output surrogate form using a class of neural networks (NNs) that employ shortcut connections, namely Residual NNs (ResNets). ResNets are known to regularize the surrogate learning problem and improve the efficiency and accuracy of the resulting surrogate. Inspired by the continuous, Neural ODE analogy, we augment ResNets with weight parameterization strategy with respect to ResNet depth. Weight-parameterized ResNets regularize the NN surrogate learning problem and allow better generalization with a drastically reduced number of learnable parameters. We demonstrate that weight-parameterized ResNets are more accurate and efficient than conventional feed-forward multi-layer perceptron networks. We also compare various options for parameterization of the weights as functions of ResNet depth. We demonstrate the results on both synthetic examples and a large scale earth system model of interest.

97 MATHEMATICS AND COMPUTING↗