Search NASA⌕ Search

SEARCH · Search NASA

Results for “CASE METHOD”

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 1,711 records · Page 95

NSFnets (Navier-Stokes flow nets): Physics-informed neural networks for the incompressible Navier-Stokes equations

In the last 50 years there has been a tremendous progress in solving numerically the Navier-Stokes equations using finite differences, finite elements, spectral, and even meshless methods. Yet, in many real cases, we still cannot incorporate seamlessly (multi-fidelity) data into existing algorithms, and for industrial-complexity applications the mesh generation is time consuming and still an art. Moreover, solving ill-posed problems (e.g., lacking boundary conditions) or inverse problems is often prohibitively expensive and requires different formulations and new computer codes. Here, we employ physics-informed neural networks (PINNs), encoding the governing equations directly into the deep neural network via automatic differentiation, to overcome some of the aforementioned limitations for simulating incompressible laminar and turbulent flows. We develop the Navier-Stokes flow nets (NSFnets) by considering two different mathematical formulations of the Navier-Stokes equations: the velocity-pressure (VP) formulation and the vorticity-velocity (VV) formulation. Since this is a new approach, we first select some standard benchmark problems to assess the accuracy, convergence rate, computational cost and flexibility of NSFnets; analytical solutions and direct numerical simulation (DNS) databases provide proper initial and boundary conditions for the NSFnet simulations. The spatial and temporal coordinates are the inputs of the NSFnets, while the instantaneous velocity and pressure fields are the outputs for the VP-NSFnet, and the instantaneous velocity and vorticity fields are the outputs for the VV-NSFnet. This is unsupervised learning and, hence, no labeled data are required beyond boundary and initial conditions and the fluid properties. The residuals of the VP or VV governing equations, together with the initial and boundary conditions, are embedded into the loss function of the NSFnets. No data is provided for the pressure to the VP-NSFnet, which is a hidden state and is obtained via the incompressibility constraint without extra computational cost. Unlike the traditional numerical methods, NSFnets inherit the properties of neural networks (NNs), hence the total error is composed of the approximation, the optimization, and the generalization errors. Here, we empirically attempt to quantify these errors by varying the sampling (“residual”) points, the iterative solvers, and the size of the NN architecture. For the laminar flow solutions, we show that both the VP and the VV formulations are comparable in accuracy but their best performance corresponds to different NN architectures. The initial convergence rate is fast but the error eventually saturates to a plateau due to the dominance of the optimization error. For the turbulent channel flow, we show that NSFnets can sustain turbulence at , but due to expensive training we only consider part of the channel domain and enforce velocity boundary conditions on the subdomain boundaries provided by the DNS data base. We also perform a systematic study on the weights used in the loss function for balancing the data and physics components, and investigate a new way of computing the weights dynamically to accelerate training and enhance accuracy. In the last part, we demonstrate how NSFnets should be used in practice, namely for ill-posed problems with incomplete or noisy boundary conditions as well as for inverse problems. We obtain reasonably accurate solutions for such cases as well without the need to change the NSFnets and at the same computational cost as in the forward well-posed problems. As a result, we also present a simple example of transfer learning that will aid in accelerating the training of NSFnets for different parameter settings.

97 MATHEMATICS AND COMPUTING↗

Wall-Resolved Large-Eddy Simulation of Flow Separation Over NASA Wall-Mounted Hump

This paper reports the findings from a study that applies wall-resolved large-eddy simulation to investigate flow separation over the NASA wall-mounted hump geometry. Despite its conceptually simple flow configuration, this benchmark problem has proven to be a challenging test case for various turbulence simulation methods that have attempted to predict flow separation arising from the adverse pressure gradient on the aft region of the hump. The momentum-thickness Reynolds number of the incoming boundary layer has a value that is near the upper limit achieved by recent direct numerical simulation and large-eddy simulation of incompressible turbulent boundary layers. The high Reynolds number of the problem necessitates a significant number of grid points for wall-resolved calculations. The present simulations show a significant improvement in the separation-bubble length prediction compared to Reynolds-Averaged Navier-Stokes calculations. The current simulations also provide good overall prediction of the skin-friction distribution, including the relaminarization observed over the front portion of the hump due to the strong favorable pressure gradient. We discuss a number of problems that were encountered during the course of this work and present possible solutions. A systematic study regarding the effect of domain span, subgrid-scale model, tunnel back pressure, upstream boundary layer conditions and grid refinement is performed. The predicted separation-bubble length is found to be sensitive to the span of the domain. Despite the large number of grid points used in the simulations, some differences between the predictions and experimental observations still exist (particularly for Reynolds stresses) in the case of the wide-span simulation, suggesting that additional grid resolution may be required.

Uzun, Ali↗

Explicit physics-informed neural networks for nonlinear closure: The case of transport in tissues

In upscaling methods, closures for nonlinear problems present a well-known challenge. While a number of theoretical methods have been proposed for handling such closures, nonlinearities still remain a significant obstacle for many problems. In this work, we use a combination of formal upscaling and data-driven machine learning for explicitly closing a nonlinear transport and reaction process in multiscale tissues. The classical effectiveness factor model is used to formulate the macroscale reaction kinetics. We train a multilayer perceptron network using training data generated by direct numerical simulations over microscale examples. Once trained, the network is used in an algorithm for numerically solving the upscaled (coarse-grained) differential equation describing mass transport and reaction in two example tissues. The network is described as being explicit in the sense that the network is trained using macroscale concentrations and gradients of concentration as components of the feature space rather than incorporating them as part of a constraint in the optimization process. Network training and solutions to the macroscale transport equations were computed for two different tissues. The two tissue types (brain and liver) exhibit markedly different geometrical complexity and spatial scale (cell size and sample size). The upscaled solutions for the average concentration are compared with numerical solutions derived from the microscale concentration fields by a posteriori averaging. There are three outcomes of this work of particular note. 1) Our overall approach results in an upscaled nonlinear PDE. The PDE is closed using a neural network, and our approach results in the definition of the classical effectiveness factor for effecting closure. 2) We identify particular source terms for the closure problem that are important for representing the structure of the closure. These source terms involve macroscale concentrations and their gradients. We adopt these source terms to use as explicit features in the learning algorithm. We find the trained networks that include the macroscale source terms generate models that are able to predict the correction factor with increased fidelity over those that do not. 3) We find that the trained network exhibits good generalizability, and it is able to predict the effectiveness factor with high fidelity for realistically-structured tissues despite the significantly different scale and geometrical complexity of the two example tissue types. This latter result emphasizes our purposeful connection between conventional averaging methods with the use of machine learning for closure; this contrasts with some machine learning methods for upscaling where the exact form of the macroscale equation remains unknown.

97 MATHEMATICS AND COMPUTING↗

Flat-plate collector performance evaluation. The case for a solar simulation approach

A method is proposed for determining the performance of flat-plate solar collector using a simulated sun. Collector test variables that will help establish the basis for the indoor test facility at the Lewis Research Center are discussed. The use of the indoor testing should permit a standard test for the convenient and accurate determination of collector performance. Preliminary test results are reported as an example of the type of collector performance data to be expected from the simulation approach.

Simon, F. F.↗

Improved solution for potential flow about arbitrary axisymmetric bodies by the use of a higher-order surface source method

An investigation is conducted of a case of axisymmetric bodies in which the application of main interest is an inlet, possibly with centerbody and ring vanes. The technique employed makes use of curved surface elements and a source density which varies over the element. Such an approach is designated a higher-order implementation. Questions of surface element geometry are discussed along with the computation of the induced velocity matrices and the organization of the calculation. The calculated results are compared with analytic solutions.

Hess, J. L.↗

Biological consequences of MLC calibration errors in IMRT delivery and QA

Purpose: The purpose of this work is threefold: (1) to explore biological consequences of the multileaf collimator (MLC) calibration errors in intensity modulated radiotherapy (IMRT) of prostate and head and neck cancers, (2) to determine levels of planning target volume (PTV) and normal tissue under- or overdose flagged with clinically used QA action limits, and (3) to provide biologically based input for MLC QA and IMRT QA action limits. Methods: Ten consecutive prostate IMRT cases and ten consecutive head and neck IMRT cases were used. Systematic MLC offsets (i.e., calibration error) were introduced for each control point of the plan separately for X1 and X2 leaf banks. Offsets were from - 2 to 2 mm with a 0.5 mm increment. The modified files were imported into the planning system for forward dose recalculation. The original plan served as the reference. The generalized equivalent uniform dose (gEUD) was used as the biological index for the targets, rectum, parotid glands, brainstem, and spinal cord. Each plan was recalculated on a CT scan of a 27 cm diameter cylindrical phantom with a contoured 0.6 cc ion chamber. Dose to ion chamber and 3D gamma analysis were compared to the reference plan. QA pass criteria: (1) at least 95% of voxels with a dose cutoff of 50% of maximum dose have to pass at 3 mm/3% and (2) dose to chamber within 2% of the reference dose. Results: For prostate cases, differences in PTV and rectum gEUD greater than 2% were identified. However, a larger proportion of plans leading to greater than 2% difference in prostate PTV gEUD passed the ion chamber QA but not 3D gamma QA. A similar trend was found for the rectum gEUD. For head and neck IMRT, the QA pass criteria flagged plans leading to greater than 4% differences in PTV gEUD and greater than 5% differences in the maximum dose to brainstem. If pass criteria were relaxed to 90% for gamma and 3% for ion chamber QA, plans leading to a 5% difference in PTV gEUD and a 5%-8% difference in brainstem maximum dose would likely pass IMRT QA. A larger proportion of head and neck plans with greater than 2% PTV gEUD difference passed 3D gamma QA compared to ion chamber QA. Conclusions: For low modulation plans, there is a better chance to catch MLC calibration errors with 3D gamma QA rather than ion chamber QA. Conversely, for high modulation plans, there is a better chance to catch MLC calibration errors with ion chamber QA rather than with 3D gamma QA. Ion chamber and 3D gamma analysis IMRT QA can detect greater than 2% change in gEUD for PTVs and critical structures for low modulation treatment plans. For high modulation treatment plans, ion chamber and 3D gamma analysis can detect greater than 2% change in gEUD for PTVs and a 5% change in critical structure gEUD since either QA methods passes the QA criteria. For gEUD changes less than those listed above, either QA method has the same proportion of passing rate.

62 RADIOLOGY AND NUCLEAR MEDICINE↗

Breakup of radioactive nuclear beams at intermediate energies as indirect method for nuclear astrophysics

We discuss the use of one-nucleon removal reactions of loosely bound nuclei at intermediate energies as an indirect method in nuclear astrophysics. These breakup reactions are good spectroscopic tools and can be used to study a large number of loosely bound proton- or neutron-rich nuclei over a wide range of beam energies. They are peripheral processes that can be used to extract asymptotic normalization coefficients (ANC) from which direct capture proton reaction rates of astrophysical interest can be calculated parameter free. We emphasize the importance of reaction model calculations and of exclusive measurements to check them. We review several cases: the breakup of {sup 8}B, {sup 9}C, {sup 15}C and {sup 23}Al. Firrst we review how we have used the data for the breakup of {sup 8}B at energies from 30 to 1000 MeV/nucleon on light and heavy targets to extract the astrophysical factor S{sub 17}(0) = 18.7{+-}1.9 eV{center_dot}b for the key reaction for solar neutrino production. Glauber model calculations in the eikonal approximation and in the optical limit using different effective NN interactions were found to give consistent, though slightly different results, for both {sup 8}B and {sup 9}C cases, enabling also us to evaluate the precision of the method. The third case is that of a neutron-rich nucleus and the ANC alone does not lead to unambiguous estimates of the associated (n,{gamma}) capture reaction. The {sup 23}Al case is that of an sd-shell nucleus, suspected of halo properties, for which our recent GANIL experiment revealed a complex configuration mixing in its ground state. The method outlined here has the big advantage that it can be used for beams of low quality, such as cocktail beams, and intensities as low as a few pps. As such it is very appropriate for the existing and future radioactive beam facilities. The breakup reactions are therefore complementary or a very good alternative to the use of proton and neutron transfer reactions (the ANC method) which require radioactive nuclear beams of much better purity and intensity.

73 NUCLEAR PHYSICS AND RADIATION PHYSICS↗

SU-E-T-187: Feasibility Study of Stereotactic Liver Radiation Therapy Using Multiple Divided Partial Arcs in Volumetric Modulated Arc Therapy

Purpose: To verify volumetric modulated arc therapy (VMAT) using flattening filter free (FFF) mode with jaw tracking (JT) feature for single breath hold as long as 15 s per arc in liver stereotactic body radiation therapy (SBRT) against intensity modulated radiation therapy (IMRT) FFF-JT. Methods: Ten hepatocellular carcinoma (HCC) cases were planned with 10 MV FFF using Pinnacle3 treatment planning system which delivered by TrueBeam to administer 48 Gy/ 4 fractions. Eight non-coplanar beams were assigned to IMRT using step-and-shoot technique. For VMAT, two or three non-coplanar partial arcs (up to 180 degrees) were further divided into subarcs with gantry rotation less than 80 degrees to limit delivery time within 15 s. Dose distributions were verified using OCTAVIUS II system and pass rates were evaluated using gamma analysis with criteria of 3%/3 mm at threshold of 5% to the maximum dose. The actual irradiation time was measured. Results: The VMAT-FFF-JT of partial-arcs with sub-divided arcs was able to produce a highly conformal plan as well as IMRT-FFF-JT. Isodose lines and DVH showed slight improvement in dosimetry when JT was employed for both IMRT and VMAT. Consequently, VMAT-FFF-JT was superior in reducing the dose to liver minus gross tumor volume. VMAT-FFF-JT has shorter total treatment time compared with 3D conformal radiation therapy (3D-CRT) FFF because the gantry was rotated simultaneously with the beam delivery in VMAT. Moreover, due to the small and regular shape of HCC, VMAT-FFF-JT offered less multileaf collimator motion, thus the interplay effect is expected to be reduced. The patient specific QA of IMRT and VMAT acquired the pass rates higher than 90%. Conclusion: VMAT-FFF-JT could be a promising technique for liver SBRT as the sub-divided arcs method was able to accommodate a single breath hold irradiation time of less than 15 s without deterioration of the dose distribution compared with IMRT-FFF-JT.

60 APPLIED LIFE SCIENCES↗

Accelerating one-dimensional searches.

In the search technique presented, the number of function evaluations is reduced by making use of the derivative of the function. The search method is combined with the variable metric algorithm reported by Davidon (1959) to solve a 17-parameter optimal atmospheric flight of a space shuttle vehicle. In the case of the problem, the proposed method requires one or two function evaluations per iteration less than the standard cubic fit-golden section method.

Johnson, I. L., Jr.↗

EnsembleGraph: Interactive Visual Analysis of Spatial-Temporal Behavior for Ensemble Simulation Data

We present a novel visualization framework—EnsembleGraph— for analyzing ensemble simulation data, in order to help scientists understand behavior similarities between ensemble members over space and time. A graph-based representation is used to visualize individual spatiotemporal regions with similar behaviors, which are extracted by hierarchical clustering algorithms. A user interface with multiple-linked views is provided, which enables users to explore, locate, and compare regions that have similar behaviors between and then users can investigate and analyze the selected regions in detail. The driving application of this paper is the studies on regional emission influences over tropospheric ozone, which is based on ensemble simulations conducted with different anthropogenic emission absences using the MOZART-4 (model of ozone and related tracers, version 4) model. We demonstrate the effectiveness of our method by visualizing the MOZART-4 ensemble simulation data and evaluating the relative regional emission influences on tropospheric ozone concentrations. Positive feedbacks from domain experts and two case studies prove efficiency of our method.

ensemble simulation↗

SU-E-T-290: Dosimetric Accuracy of Acuros XB and Analytical Anisotropic Algorithm in Stereotactic Ablative Radiotherapy of Small Lung Lesions

Purpose: The accuracy of dose calculation for lung stereotactic ablative radiotherapy (SABR) of small lesions critically depends on the proper modeling of the lateral scatter in heterogeneous media. In recent years, grid-based Boltzman solvers such as Acuros XB (AXB) have been introduced for enhanced modeling of radiation transport in heterogeneous media. The purpose of this study is to evaluate the dosimetric impact of dose calculation between AXB and convolution-superposition algorithms such as analytical anisotropic algorithm (AAA) for small lesion sizes and different beam energies. Methods: Five lung SABR VMAT cases with GTV ranged from 0.8cm to 2.5cm in diameter were studied. For each case, doses were calculated by AAA, AXB (V11031) and Monte Carlo simulation (MC) with the same plan parameters for 10MV and 6MV. The dose calculation accuracy were evaluated by comparing DVHs and dose distributions with MC as the benchmark. The accuracy of calculated dose was also validated by EBT3 film measurement with a field size of 3cmx3cm in a thorax phantom. Results: For 10MV and GTV less than 1cm, dose calculated by AXB agrees well with MC compared to AAA. Dose difference calculated with AXB and AAA could be up to 30%. For GTV greater than 2cm, calculation results of AXB and AAA agree within 5% in GTV. For 6MV, the difference between calculated doses by AXB and AAA is less 10% for GTV less than 1cm. Based on film measurements, lung dose was overestimated 10% and 20% by AAA for 6MV and 10MV. Conclusion: Lateral scatter and transport is modeled more accurately by AXB than AAA in heterogeneous media, especially for small field size and high energy beams. The accuracy depends on the assigned material in calculation. If grid-based Boltzman solvers or MC are not available for calculation, lower energy beams should be used for treatment.

60 APPLIED LIFE SCIENCES↗

Comparison of spherical harmonics method and discrete ordinates method for radiative transfer in a turbulent jet flame

Here, in this study, we systematically compared the accuracy and computational cost of two popular solution methods for the radiative transfer equation (RTE): the spherical harmonics method (P N ) and the discrete ordinates method (DOM). We first investigated convergence characteristics of different orders of P N and DOM in a series of 1D homogeneous configurations with varying optical thicknesses. Both solvers perform better for optically thicker cases. The accuracy of P N methods increases with its order, , but the gain in accuracy reduces with the increase in , i.e., improvement of P 7 over P 5 is less than that of P 3 over P 1 . This decreasing trend becomes more prominent as the optical thickness decreases. On the other hand, DOM’s accuracy increases almost linearly with the increase in the number of ordinates (or polar angles in this study) in all cases. While comparing the directional profile of radiative intensity, both solvers perform better when the radiative intensity is more isotropic. These solvers were then connected with a full spectrum k-distribution (FSK) spectral model and used to perform radiation-coupled simulations of a turbulent jet flame in an axi-symmetric cylindrical domain. Results obtained from P 1 to P 7 approximations for P N , and 2 x 4, 4 x 4, 4 x 8, 8 x 8 finite angles for DOM are compared with that from an optically thin model, and a reference solution from line-by-line (LBL) photon Monte Carlo (PMC) method. The choice of radiation solver shows a noticeable impact on the temperature distribution of the flame. The P N solvers lead to slightly higher radiant fractions and the DOM solvers lead to slightly lower radiant fractions than the PMC benchmark solution. Finally, the computational costs of each of these solvers are also reported and an intermittent evaluation / time blending scheme to improve the computational efficiency of radiation solvers in radiation-coupled simulations are also demonstrated.

42 ENGINEERING↗

Quantum dynamics study of fulvene double bond photoisomerization: The role of intramolecular vibrational energy redistribution and excitation energy

The double bond photoisomerization of fulvene has been studied with quantum dynamics calculations using the multi-configuration time-dependent Hartree method. Fulvene is a test case to develop optical control strategies based on the knowledge of the excited state decay mechanism. The decay takes place on a time scale of several hundred femtoseconds, and the potential energy surface is centered around a conical intersection seam between the ground and excited state. The competition between unreactive decay and photoisomerization depends on the region of the seam accessed during the decay. The dynamics are carried out on a four-dimensional model surface, parametrized from complete active space self-consistent field calculations, that captures the main features of the seam (energy and locus of the seam and associated branching space vectors). Wave packet propagations initiated by single laser pulses of 5-25 fs duration and 1.85-4 eV excitation energy show the principal characteristics of the first 150 fs of the photodynamics. Initially, the excitation energy is transferred to a bond stretching mode that leads the wave packet to the seam, inducing the regeneration of the reactant. The photoisomerization starts after the vibrational energy has flowed from the bond stretching to the torsional mode. In our propagations, intramolecular energy redistribution (IVR) is accelerated for higher excess energies along the bond stretch mode. Thus, the competition between unreactive decay and isomerization depends on the rate of IVR between the bond stretch and torsion coordinates, which in turn depends on the excitation energy. These results set the ground for the development of future optical control strategies.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗

Developing Fatigue Pre-crack Procedure to Evaluate Fracture Toughness of Pipeline Steels Using Spiral Notch Torsion Test

The spiral notch torsion test (SNTT) has been utilized to investigate the crack growth behavior of X52 steel base and welded materials used for hydrogen infrastructures. The X52 steel materials are received from a welded pipe using friction stir welding techniques. Finite element models were established to study the crack growth behavior of steel SNTT steel samples, which were assumed to be isotropic material. A series SNTT models were set up to cover various crack penetration cases, of which the ratios between crack depth to diameter (a/D ratio) ranging from 0.10 to 0.45. The evolution of compliance and energy release rates in the SNTT method have been investigated with different cases, including different geometries and materials. Indices of characteristic compliance and energy release rates have been proposed. Good agreement has been achieved between predictions from different cases in the same trend. These work shed lights on a successful protocol for SNTT application in wide range of structural materials. The further effort needed for compliance function development is to extend the current developed compliance function to the deep crack penetration arena, in the range of 0.55 to 0.85 to effectively determine fracture toughness for extremely tough materials.

fracture toughness↗

The method of complex characteristics for transonic airfoil design, with an application to compressors

The use of mathematical models to study physical problems of current interest to aeronautical engineers has been made possible by the development of numerical techniques to compute solutions of the differential equations of transonic aerodynamics. These advances have encouraged the improvement of supercritical wing technology. A method to determined steady, shockless flow of an inviscid, compressible fluid past a cascade of airfoils in the (x,y)-plane is considered, taking into account also the case of an isolated airfoil. The method of complex characteristics solves the equations in the hodograph plane by extending all variables into the complex domain, where the notion of type is no longer significant. Attention is given to the mathematical background, the method of complex characteristics, and numerical calculations.

Bledsoe, M.↗

SU-D-9A-07: Imaging Dose and Cancer Risk in Image-Guided Radiotherapy of Cancers

Purpose: To systematically evaluate the imaging doses and cancer risks associated with various imaging procedures involving ionizing radiation during image-guided radiotherapy of an increasingly large number of cancer patients. Methods: 141 patients (52 brain cases, 47 thoracic cases, 42 abdominal cases, aged 3 to 91 years old) treated between October 2009 and March 2010 were included in this IRB-approved retrospective study. During the whole radiotherapy course, each patient underwent at least one type of imaging procedures, i.e., kV portal, MV portal and kVCBCT, besides CT simulations. Based on Monte Carlo modeling and particle transport in human anatomy of various dimensions, the correlations between the radiation doses to the various organs-at-risk (OARs) at the head, the thoracic and the abdominal regions and one's weight, circumference, scan mAs and kVp have been obtained and used to estimate the radiation dose from a specific imaging procedure. The radiation-induced excess relative risk (ERR) was then estimated with BEIR VII formulism based on one's gender, age and radiation dose. 1+ ERR was reported in this study as relative cancer risk. Results: For the whole cohort of 141 patients, the mean imaging doses from various imaging procedures were 8.3 cGy to the brain, 10.5 cGy to the lungs and 19.2 cGy to the red bone marrow, respectively. Accordingly, the cancer risks were 1.140, 1.369 and 2.671, respectively. In comparison, MV portal deposited largest doses to the lungs while kVCBCT delivered the highest doses to the red bone marrow. Conclusion: The compiled imaging doses to a patient during his/her treatment course were patient-specific and site-dependent, varying from 1.2 to 263.5 cGy on average, which were clinically significant and should be included in the treatment planning and overall decision-making. Our results indicated the necessity of personalized imaging to maximize its clinical benefits while reducing the associated cancer risks. Sichuan University Scholarship.

60 APPLIED LIFE SCIENCES↗

A Workflow for Visual Diagnostics of Binary Classifiers using Instance-Level Explanations

In this paper, we propose a visual analytics workflow to help data scientists explore, diagnose, and understand the decisions made by a binary classifier. The approach leverages “instance-level explanations”, measures of local feature relevance that explain single instances and uses them to build a set of visual representations that guide the users in their investigation. The workflow is based on three main visual representations and steps: one based on aggregate statistics to see how data distributes across correct or incorrect decisions; one based on explanations to understand which features are used to make these decisions; and one based on raw data, to derive insights on potential root causes for the observed patterns. The work is validated through a long-term collaboration with a group of healthcare professionals who used our method to make sense of machine learning models they developed. The case studies from this collaboration demonstrate that the proposed method helps experts derive useful knowledge about the model and the phenomena it describes and also generates useful hypotheses on how a model can be improved.

Krause, Josua↗

Differential equation based method for accurate approximations in optimization

A method to efficiently and accurately approximate the effect of design changes on structural response is described. The key to this method is to interpret sensitivity equations as differential equations that may be solved explicitly for closed form approximations, hence, the method is denoted the Differential Equation Based (DEB) method. Approximations were developed for vibration frequencies, mode shapes and static displacements. The DEB approximation method was applied to a cantilever beam and results compared with the commonly-used linear Taylor series approximations and exact solutions. The test calculations involved perturbing the height, width, cross-sectional area, tip mass, and bending inertia of the beam. The DEB method proved to be very accurate, and in most cases, was more accurate than the linear Taylor series approximation. The method is applicable to simultaneous perturbation of several design variables. Also, the approximations may be used to calculate other system response quantities. For example, the approximations for displacements are used to approximate bending stresses.

Pritchard, Jocelyn I.↗