Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “Inverse Optimization”

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 1,009 records · Page 56

Automation and Intensity Modulated Radiation Therapy for Individualized High-Quality Tangent Breast Treatment Plans

Purpose: To demonstrate the large-scale clinical implementation and performance of an automated treatment planning methodology for tangential breast intensity modulated radiation therapy (IMRT). Methods and Materials: Automated planning was used to prospectively plan tangential breast IMRT treatment for 1661 patients between June 2009 and November 2012. The automated planning method emulates the manual steps performed by the user during treatment planning, including anatomical segmentation, beam placement, optimization, dose calculation, and plan documentation. The user specifies clinical requirements of the plan to be generated through a user interface embedded in the planning system. The automated method uses heuristic algorithms to define and simplify the technical aspects of the treatment planning process. Results: Automated planning was used in 1661 of 1708 patients receiving tangential breast IMRT during the time interval studied. Therefore, automated planning was applicable in greater than 97% of cases. The time for treatment planning using the automated process is routinely 5 to 6 minutes on standard commercially available planning hardware. We have shown a consistent reduction in plan rejections from plan reviews through the standard quality control process or weekly quality review multidisciplinary breast rounds as we have automated the planning process for tangential breast IMRT. Clinical plan acceptance increased from 97.3% using our previous semiautomated inverse method to 98.9% using the fully automated method. Conclusions: Automation has become the routine standard method for treatment planning of tangential breast IMRT at our institution and is clinically feasible on a large scale. The method has wide clinical applicability and can add tremendous efficiency, standardization, and quality to the current treatment planning process. The use of automated methods can allow centers to more rapidly adopt IMRT and enhance access to the documented improvements in care for breast cancer patients, using technologies that are widely available and already in clinical use.

62 RADIOLOGY AND NUCLEAR MEDICINE↗

Inverse design of a pyrochlore lattice of DNA origami through model-driven experiments

Sophisticated statistical mechanics approaches and human intuition have demonstrated the possibility of self-assembling complex lattices or finite-size constructs. However, attempts so far have mostly only been successful in silico and often fail in experiment because of unpredicted traps associated with kinetic slowing down (gelation, glass transition) and competing ordered structures. Theoretical predictions also face the difficulty of encoding the desired interparticle interaction potential with the experimentally available nano- and micrometer-sized particles. To overcome these issues, we combine SAT assembly (a patchy-particle interaction design algorithm based on constrained optimization) with coarse-grained simulations of DNA nanotechnology to experimentally realize trap-free self-assembly pathways. In this paper, we use this approach to assemble a pyrochlore three-dimensional lattice, coveted for its promise in the construction of optical metamaterials, and characterize it with small-angle x-ray scattering and scanning electron microscopy visualization.

77 NANOSCIENCE AND NANOTECHNOLOGY↗

Numerical schemes for dynamically orthogonal equations of stochastic fluid and ocean flows

The quantification of uncertainties is critical when systems are nonlinear and have uncertain terms in their governing equations or are constrained by limited knowledge of initial and boundary conditions. Such situations are common in multiscale, intermittent and non-homogeneous fluid and ocean flows. The dynamically orthogonal (DO) field equations provide an adaptive methodology to predict the probability density functions of such flows. The present work derives efficient computational schemes for the DO methodology applied to unsteady stochastic Navier-Stokes and Boussinesq equations, and illustrates and studies the numerical aspects of these schemes. Semi-implicit projection methods are developed for the mean and for the DO modes, and time-marching schemes of first to fourth order are used for the stochastic coefficients. Conservative second-order finite-volumes are employed in physical space with new advection schemes based on total variation diminishing methods. Other results include: (i) the definition of pseudo-stochastic pressures to obtain a number of pressure equations that is linear in the subspace size instead of quadratic; (ii) symmetric advection schemes for the stochastic velocities; (iii) the use of generalized inversion to deal with singular subspace covariances or deterministic modes; and (iv) schemes to maintain orthonormal modes at the numerical level. To verify our implementation and study the properties of our schemes and their variations, a set of stochastic flow benchmarks are defined including asymmetric Dirac and symmetric lock-exchange flows, lid-driven cavity flows, and flows past objects in a confined channel. Different Reynolds number and Grashof number regimes are employed to illustrate robustness. Optimal convergence under both time and space refinements is shown as well as the convergence of the probability density functions with the number of stochastic realizations.

71 CLASSICAL AND QUANTUM MECHANICS, GENERAL PHYSIC↗

Adaptive Activation Functions Accelerate Convergence in Deep and Physics-informed Neural Networks

We employ adaptive activation functions for regression in deep and physics-informed neural networks (PINNs) to approximate smooth and discontinuous functions as well as solutions of linear and nonlinear partial differential equations. In particular, we solve the nonlinear Klein-Gordon equation, which has smooth solutions, the nonlinear Burgers equation, which can admit high gradient solutions, and the Helmholtz equation. We introduce a scalable hyper-parameter in the activation function, which can be optimized to achieve best performance of the network as it changes dynamically the topology of the loss function involved in the optimization process. The adaptive activation function has better learning capabilities than the traditional one (fixed activation) as it improves greatly the convergence rate, especially at early training, as well as the solution accuracy. To better understand the learning process, we plot the neural network solution in the frequency domain to examine how the network captures successively different frequency bands present in the solution. We consider both forward problems, where the approximate solutions are obtained, as well as inverse problems, where parameters involved in the governing equation are identified. Our simulation results show that the proposed method is a very simple and effective approach to increase the efficiency, robustness and accuracy of the neural network approximation of nonlinear functions as well as solutions of partial differential equations, especially for forward problems. We theoretically prove that in the proposed method, gradient descent algorithms are not attracted to suboptimal critical points or local minima.

machine leaning, Bad minima, Inverse problems, Phy↗

Bimetallic-Derived Catalytic Structures for CO 2 -Assisted Ethane Activation

In recent years, the simultaneous upgrading of CO 2 and ethane has emerged as a promising approach for generating valuable gaseous (CO, H 2 , and ethylene) and liquid (aromatics and C 3 oxygenates) chemicals from the greenhouse gas CO 2 and large-reserved shale gas. The key challenges for controlling product selectivity lie in the selective C-H and C-C bond cleavage of ethane with the assistance of CO 2 . Bimetallic-derived catalysts likely undergo alloying or oxygen-induced segregation under reaction conditions, thus providing diverse types of interfacial sites, e.g., metal/support (M/M'O x ) interface and metal oxide/metal (M'O x /M) inverse interface, that are beneficial for selective CO 2 -assisted ethane upgrading. The alloying extent can be initially predicted by cohesive energy and atomic radius (or Wigner-Seitz radius), while the preference for segregation to form the on-top suboxide can be approximated using the work function, electronegativity, and binding strength of adsorbed oxygen. Furthermore, bimetallic-derived catalysts are typically supported on high surface area oxides. Modifying the reducibility and acidity/basicity of the oxide supports and introducing surface defects facilitate CO 2 activation and oxygen supplies for ethane activation. Using in situ synchrotron characterization and density functional theory (DFT) calculations, we found that the electronic properties of oxygen species influence the selective cleavage of C-H/C-C bonds in ethane, with electron-deficient oxygen over the metal (or alloy) surface promoting nonselective bond scission to produce syngas and electron-enriched oxygen over the metal oxide/metal interface enhancing selective C-H scission to yield ethylene. Here, we further demonstrate that the preferred structures of the catalyst surfaces, either alloy surfaces or metal oxide/metal inverse interfaces, can be controlled through the appropriate choice of metal combinations and their atomic ratios. Through a comprehensive comparison of experimental results and DFT calculations, the selectivity of C-C/C-H bond scission is correlated with the thermodynamically favorable bimetallic-derived structures (i.e., alloy surfaces or metal oxide/metal inverse interfaces) under reaction conditions over a wide range of bimetallic catalysts. These findings not only offer structural and mechanistic insights into bimetallic-derived catalysts but also provide design principles for selective catalysts for CO 2 -assisted activation of ethane and other light alkanes. This Account concludes by discussing challenges and opportunities in designing advanced bimetallic-derived catalysts, incorporating new reaction chemistries for other products, employing precise synthesis strategies for well-defined structures with optimized site densities, and leveraging time/spatial/energy-resolved in situ spectroscopy/scattering/microscopy techniques for comprehensive structural analysis. The research methodologies established here are helpful for the investigation of dynamic alloy and interfacial structures and should inspire more efforts toward the simultaneous upgrading of CO 2 and shale gas.

03 NATURAL GAS↗

Evaluation of a commercial biologically based IMRT treatment planning system

A new inverse treatment planning system (TPS) for external beam radiation therapy with high energy photons is commercially available that utilizes both dose-volume-based cost functions and a selection of cost functions which are based on biological models. The purpose of this work is to evaluate quality of intensity-modulated radiation therapy (IMRT) plans resulting from the use of biological cost functions in comparison to plans designed using a traditional TPS employing dose-volume-based optimization. Treatment planning was performed independently at two institutions. For six cancer patients, including head and neck (one case from each institution), prostate, brain, liver, and rectal cases, segmental multileaf collimator IMRT plans were designed using biological cost functions and compared with clinically used dose-based plans for the same patients. Dose-volume histograms and dosimetric indices, such as minimum, maximum, and mean dose, were extracted and compared between the two types of treatment plans. Comparisons of the generalized equivalent uniform dose (EUD), a previously proposed plan quality index (fEUD), target conformity and heterogeneity indices, and the number of segments and monitor units were also performed. The most prominent feature of the biologically based plans was better sparing of organs at risk (OARs). When all plans from both institutions were combined, the biologically based plans resulted in smaller EUD values for 26 out of 33 OARs by an average of 5.6 Gy (range 0.24 to 15 Gy). Owing to more efficient beam segmentation and leaf sequencing tools implemented in the biologically based TPS compared to the dose-based TPS, an estimated treatment delivery time was shorter in most (five out of six) cases with some plans showing up to 50% reduction. The biologically based plans were generally characterized by a smaller conformity index, but greater heterogeneity index compared to the dose-based plans. Overall, compared to plans based on dose-volume optimization, plans with equivalent target coverage obtained using the biologically based TPS demonstrate improved dose distributions for the majority of normal structures.

62 RADIOLOGY AND NUCLEAR MEDICINE↗

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↗

Calculations of single crystal elastic constants for yttria partially stabilised zirconia from powder diffraction data

Yttria Stabilised Zirconia (YSZ) is a tough, phase-transforming ceramic that finds use in a wide range of commercial applications from dental prostheses to thermal barrier coatings. Micromechanical modelling of phase transformation can deliver reliable predictions in terms of the influence of temperature and stress. However, models must rely on the accurate knowledge of single crystal elastic stiffness constants. Some techniques for elastic stiffness determination are well-established. The most popular of these involve exploiting frequency shifts and phase velocities of acoustic waves. However, the application of these techniques to YSZ can be problematic due to the micro-twinning observed in larger crystals. Here, we propose an alternative approach based on selective elastic strain sampling (e.g., by diffraction) of grain ensembles sharing certain orientation, and the prediction of the same quantities by polycrystalline modelling, for example, the Reuss or Voigt average. The inverse problem arises consisting of adjusting the single crystal stiffness matrix to match the polycrystal predictions to observations. In the present model-matching study, we sought to determine the single crystal stiffness matrix of tetragonal YSZ using the results of time-of-flight neutron diffraction obtained from an in situ compression experiment and Finite Element modelling of the deformation of polycrystalline tetragonal YSZ. The best match between the model predictions and observations was obtained for the optimized stiffness values of C11 = 451, C33 = 302, C44 = 39, C66 = 82, C12 = 240, and C13 = 50 (units: GPa). Considering the significant amount of scatter in the published literature data, our result appears reasonably consistent.

36 MATERIALS SCIENCE↗

Electron Cyclotron Emission–Based Separatrix Identification in ITER with OMFIT Synthetic Modeling

Accurate determination of the separatrix location is essential for understanding edge plasma behavior and optimizing confinement in tokamaks, especially in next-generation devices such as ITER. In this study, a synthetic microwave diagnostics module was developed and implemented in the OMFIT framework to assess the feasibility of an electron cyclotron emission–based separatrix detection method in ITER plasmas. Simulations were carried out using ITER H-mode equilibrium scenarios with different plasma density profiles and different pedestal widths. Here, the results show that the electron emission temperature profiles consistently exhibit an inversion pattern near the edge, with a well-defined minimum point that could serve as a proxy for the separatrix location. However, unlike in the DIII-D, the minimum point in ITER is systematically offset by approximately 2 cm into the scrape-off layer, independent of density or pedestal width, which is within the radial resolution range (2 to 5 cm) determined by the 500-MHz channel spacing. While the method does not provide the exact separatrix location, it offers a reliable indicator of the boundary region and has potential applications for real-time boundary monitoring in ITER and other future fusion devices.

Electron cyclotron emission↗

Impact of the HIV-1 genetic background and HIV-1 population size on the evolution of raltegravir resistance

Background Emergence of resistance against integrase inhibitor raltegravir in human immunodeficiency virus type 1 (HIV-1) patients is typically associated with selection of one of three signature mutations: Y143C/R, Q148K/H/R or N155H, representing three distinct resistance pathways. The mechanisms that drive selection of a specific pathway are still poorly understood. We reviewed the effect of the HIV-1 genetic background and population dynamics on the emergence of raltegravir resistance. Using deep sequencing we analyzed the integrase coding sequence (CDS) in longitudinal samples from five patients who initiated raltegravir plus optimized background therapy at viral loads > 5000 copies/ml. To investigate the role of the HIV-1 genetic background we created recombinant viruses containing the viral integrase coding region from pre-raltegravir samples from two patients in whom raltegravir resistance developed through different pathways. The in vitro selections performed with these recombinant viruses were designed to mimic natural population bottlenecks. Results Deep sequencing analysis of the viral integrase CDS revealed that the virological response to raltegravir containing therapy inversely correlated with the relative amount of unique sequence variants that emerged suggesting diversifying selection during drug pressure. In 4/5 patients multiple signature mutations representing different resistance pathways were observed. Interestingly, the resistant population can consist of a single resistant variant that completely dominates the population but also of multiple variants from different resistance pathways that coexist in the viral population. We also found evidence for increased diversification after stronger bottlenecks. In vitro selections with low viral titers, mimicking population bottlenecks, revealed that both recombinant viruses and HXB2 reference virus were able to select mutations from different resistance pathways, although typically only one resistance pathway emerged in each individual culture. Conclusions The generation of a specific raltegravir resistant variant is not predisposed in the genetic background of the viral integrase CDS. Typically, in the early phases of therapy failure the sequence space is explored and multiple resistance pathways emerge and then compete for dominance which frequently results in a switch of the dominant population over time towards the fittest variant or even multiple variants of similar fitness that can coexist in the viral population.

60 APPLIED LIFE SCIENCES↗

Detecting and quantifying materials in containers utilizing an inverse algorithm with adaptive regularization

A method, system and computer program product for automatically detecting and quantifying materials in containers (e.g., luggage, storage containers, equipment or componentry for processing or handling). An X-ray radiograph of a container is generated. A set of materials of interest (e.g., plutonium, steel) is selected to determine if the object comprises or contains materials of interest. An estimate of the areal densities or thicknesses of each of the selected materials of interest is obtained by minimizing an objective function with respect to the areal densities or thicknesses for each of the selected materials of interest. Adaptive regularization is implemented in the objective function to improve optimization results by adding a constraint to the objective function, where the constraint penalizes the objective function for solutions that do not line up with a prior belief about the solution form (e.g., the solution should not be negative or especially noisy).

97 MATHEMATICS AND COMPUTING↗

Theoretical study of optical properties of anti phase domains in GaP

III-V/Si heterostructures are currently investigated for silicon photonics and solar energy conversion. In particular, dilute nitride alloy GaAsPN grown on a GaP/Si platform exhibits lattice match with Si and an optimal band gap configuration for tandem solar cell devices. However, monolithic “coherent” growth of the GaP thin layer on Si suffers from the nucleation of extended structural defects, which can hamper device operation as well as the GaP/Si interface level and through their propagation inside the overall heterostructure. However, the effect of such structural defects on optical and transport properties is actually not well understood in details. In this letter, we investigate the anti phase domains defect (also called inversion domains) by means of ab initio calculations giving insights into the alteration of optical and transport properties of GaP due to the defective GaP/Si interface.

75 CONDENSED MATTER PHYSICS, SUPERCONDUCTIVITY AND↗

Efficient 3-D velocity model building using joint inline and crossline plane-wave wave-equation migration velocity analyses

SUMMARY Wave-equation migration velocity analysis (WEMVA) is an image-domain inversion method for velocity model building. Automatic plane-wave WEMVA (PWEMVA) calculates the moveouts of plane-wave common-image gathers (CIGs) by searching a best-fitting parabola with semblance analysis and backprojects residual CIG moveouts into wavefield wave paths with a reflection tomographic kernel. However, 3-D PWEMVA is very computationally expensive because 3-D reflection tomographic inversion requires at least five 3-D reverse-time migrations per iteration and stores two types of source wavefields at model boundaries. We develop a joint inline and crossline PWEMVA method for efficient 3-D velocity model building. We alternatively implement the inline and crossline PWEMVAs with a constraint for each other, in which we iteratively construct the 3-D velocity model update through 1-D spline interpolation of 2-D gradients. The inline and crossline joint inversion is practical since PWEMVA only inverts for low-wavenumber velocity perturbations along wave paths, and the method can take less than 1 per cent of the computational cost of full 3-D PWEMVA. To construct unaliased plane waves for our joint inline and crossline PWEMVA, we develop a 3-D data interpolation method in the frequency–wavenumber (FK) domain to recover regularly and randomly missing traces. The method minimizes the misfit on sufficiently localized data subsets with iterative optimal step lengths and a gradient preconditioner that iteratively selects dominant dips along different azimuths. In numerical experiments, we use a 3-D synthetic seismic data set and a land 3-D field seismic data set acquired at the Farnsworth CO2-EOR (enhanced oil recovery) field to demonstrate the efficacy of our velocity model building and data interpolation methods.

Liu, Xuejian↗

Pillars of the Mantle: Imaging the Interior of the Earth with Adjoint Tomography. In: PEARC17 Proceedings of the Practice and Experience in Advanced Research Computing 2017 on Sustainability, Success and Impact, Article No. 75

In this work, we investigate global seismic tomographic models obtained by spectral-element simulations of seismic wave propagation and adjoint methods. Global crustal and mantle models are obtained based on an iterative conjugate-gradient type of optimization scheme. Forward and adjoint seismic wave propagation simulations, which result in synthetic seismic data to make measurements and data sensitivity kernels to compute gradient for model updates, respectively, are performed by the SPECFEM3D_GLOBE package [1] [2] at the Oak Ridge Leadership Computing Facility (OLCF) to study the structure of the Earth at unprecedented levels. Using advances in solver techniques that run on the GPUs on Titan at the OLCF, scientists are able to perform large-scale seismic inverse modeling and imaging. Using seismic data from global and regional networks from global CMT earthquakes, scientists are using SPECFEM3D_GLOBE to understand the structure of the mantle layer of the Earth. Visualization of the generated data sets provide an effective way to understand the computed wave perturbations which define the structure of mantle in the Earth.

Pugmire, David↗

Sparse measurement medical CT reconstruction using multi-fused block matching denoising priors

A major challenge for medical X-ray CT imaging is reducing the number of X-ray projections to lower radiation dosage and reduce scan times without compromising image quality. However these under-determined inverse imaging problems rely on the formulation of an expressive prior model to constrain the solution space while remaining computationally tractable. Traditional analytical reconstruction methods like Filtered Back Projection (FBP) often fail with sparse measurements, producing artifacts due to their reliance on the Shannon-Nyquist Sampling Theorem. Consensus Equilibrium, which is a generalization of Plug and Play, is a recent advancement in Model-Based Iterative Reconstruction (MBIR), has facilitated the use of multiple denoisers are prior models in an optimization free framework to capture complex, non-linear prior information. However, 3D prior modelling in a Plug and Play approach for volumetric image reconstruction requires long processing time due to high computing requirement. Instead of directly using a 3D prior, this work proposes a BM3D Multi Slice Fusion (BM3D-MSF) prior that uses multiple 2D image denoisers fused to act as a fully 3D prior model in Plug and Play reconstruction approach. Our approach does not require training and are thus able to circumvent ethical issues related with patient training data and are readily deployable in varying noise and measurement sparsity levels. In addition, reconstruction with the BM3D-MSF prior achieves similar reconstruction image quality as fully 3D image priors, but with significantly reduced computational complexity. We test our method on clinical CT data and demonstrate that our approach improves reconstructed image quality.

Hossain, Maliha [ORNL]↗

Smoothing tool design and performance during subaperture glass polishing

During subaperture tool grinding and polishing, overlaps of the tool influence function can result in undesirable mid-spatial frequency (MSF) errors in the form of surface ripples, which are often corrected using a smoothing polishing step. Here, in this study, flat multi-layer smoothing polishing tools are designed and tested to simultaneously (1) reduce or remove MSF errors, (2) minimize surface figure degradation, and (3) maximize the material removal rate. A time-dependent convergence model in which spatial material removal varies with a workpiece-tool height mismatch, combined with a finite element mechanical analysis to determine the interface contact pressure distribution, was developed to evaluate various smoothing tool designs as a function of tool material properties, thicknesses, pad textures, and displacements. An improvement in smoothing tool performance is achieved when the gap pressure constant, $\overline{h}$ (which describes the inverse rate at which the pressure drops with a workpiece-tool height mismatch), is minimized for smaller spatial scale length surface features (namely, MSF errors) and maximized for large spatial scale length features (i.e., surface figure). Five specific smoothing tool designs were experimentally evaluated. A two-layer smoothing tool using a thin, grooved IC1000 polyurethane pad (with a high elastic modulus, E pad =360 MPa ), thicker blue foam (with an intermediate modulus, E foam =5.3 MPa ) underlayer, and an optimized displacement (d t =1 mm ) provided the best overall performance (namely, high MSF error convergence, minimal surface figure degradation, and high material removal rate).

47 OTHER INSTRUMENTATION↗

Integrating Natural Language Processing and Machine Learning Algorithms to Categorize Oncologic Response in Radiology Reports

A significant volume of medical data remains unstructured. Natural language processing (NLP) and machine learning (ML) techniques have shown to successfully extract insights from radiology reports. However, the codependent effects of NLP and ML in this context have not been well-studied. Between April 1, 2015 and November 1, 2016, 9418 cross-sectional abdomen/pelvis CT and MR examinations containing our internal structured reporting element for cancer were separated into four categories: Progression, Stable Disease, Improvement, or No Cancer. We combined each of three NLP techniques with five ML algorithms to predict the assigned label using the unstructured report text and compared the performance of each combination. The three NLP algorithms included term frequency-inverse document frequency (TF-IDF), term frequency weighting (TF), and 16-bit feature hashing. The ML algorithms included logistic regression (LR), random decision forest (RDF), one-vs-all support vector machine (SVM), one-vs-all Bayes point machine (BPM), and fully connected neural network (NN). The best-performing NLP model consisted of tokenized unigrams and bigrams with TF-IDF. Increasing N-gram length yielded little to no added benefit for most ML algorithms. With all parameters optimized, SVM had the best performance on the test dataset, with 90.6 average accuracy and F score of 0.813. The interplay between ML and NLP algorithms and their effect on interpretation accuracy is complex. The best accuracy is achieved when both algorithms are optimized concurrently.

62 RADIOLOGY AND NUCLEAR MEDICINE↗

Robust scalable initialization for Bayesian variational inference with multi-modal Laplace approximations

Predictive modeling typically relies on Bayesian model calibration to provide uncertainty quantification. Variational inference utilizing fully independent (“mean-field”) Gaussian distributions are often used as approximate probability density functions. This simplification is attractive since the number of variational parameters grows only linearly with the number of unknown model parameters. However, the resulting diagonal covariance structure and unimodal behavior can be too restrictive to provide useful approximations of intractable Bayesian posteriors that exhibit highly non-Gaussian behavior, including multimodality. High-fidelity surrogate posteriors for these problems can be obtained by considering the family of Gaussian mixtures. Gaussian mixtures are capable of capturing multiple modes and approximating any distribution to an arbitrary degree of accuracy, while maintaining some analytical tractability. Unfortunately, variational inference using Gaussian mixtures with full-covariance structures suffers from a quadratic growth in variational parameters with the number of model parameters. The existence of multiple local minima due to strong nonconvex trends in the loss functions often associated with variational inference present additional complications, These challenges motivate the need for robust initialization procedures to improve the performance and computational scalability of variational inference with mixture models. In this work, we propose a method for constructing an initial Gaussian mixture model approximation that can be used to warm-start the iterative solvers for variational inference. The procedure begins with a global optimization stage in model parameter space. In this step, local gradient-based optimization, globalized through multistart, is used to determine a set of local maxima, which we take to approximate the mixture component centers. Around each mode, a local Gaussian approximation is constructed via the Laplace approximation. Finally, the mixture weights are determined through constrained least squares regression. The robustness and scalability of the proposed methodology is demonstrated through application to an ensemble of synthetic tests using high-dimensional, multimodal probability density functions. Here, the practical aspects of the approach are demonstrated with inversion problems in structural dynamics.

97 MATHEMATICS AND COMPUTING↗