Bulletin of the American Physical Society
APS March Meeting 2023
Volume 68, Number 3
Las Vegas, Nevada (March 5-10)
Virtual (March 20-22); Time Zone: Pacific Time
Session M53: Computer VisionFocus
|
Hide Abstracts |
Sponsoring Units: GDS Chair: William Ratcliff, National Institute of Standards and Technology Room: Room 307 |
Wednesday, March 8, 2023 8:00AM - 8:36AM |
M53.00001: Machine Learning application for astrophysics: A case study for black hole images and strong gravitational lensing Invited Speaker: Joshua Yao-Yu Y Lin Modern astronomy research has been thriving due to new observations. To handle large and novel datasets, deep learning tools for computer vision (e.g. Convolutional Neural Networks and Vision Transformers) provide a new way to tackle the challenges in data analysis. In this talk, I will cover some of the recent development of deep learning on several astrophysics projects including supermassive black holes (SMBH), and strong gravitational lensing. |
Wednesday, March 8, 2023 8:36AM - 8:48AM |
M53.00002: Fusion for Reducing Domain Specificity in Computer Vision Models Laura E Brandt, Nicholas Roy The dream of computer vision is to enable autonomous visual processing of everything that can be seen in the world. Toward this end, the research community has focused recently on creating single, monolithic neurally-inspired models (given some task) which on average outperform the competition on some set of benchmarks, regardless of visual domain. The conventional wisdom has been that to make a neural algorithm perform simultaneously well in several domains, a single neural network should be trained on data from all domains of interest. More recently, synthetic datasets like FlyingThings3D, which contains random everyday objects from many domains flying along random trajectories through space, have attempted to reduce domain specificity by doing away entirely with scene structure. In this work we propose that the answer to the challenge of creating general perception systems is to recognise that different models will have different domains in which they perform well, and to fuse the estimates produced by separate perception models that are each "experts" in their own domains. We present a design paradigm for general model-fusion systems, and evaluate both quantitative and qualitative performance of such systems on image classification and segmentation. |
Wednesday, March 8, 2023 8:48AM - 9:00AM |
M53.00003: Physics-constrained 3D convolutional neural networks for relativistic electrodynamics Alexander Scheinker, Reeju Pokharel We present a physics-constrained neural network (PCNN) approach to calculating the electromagnetic fields of intense relativistic charged particle beams via 3D convolutional neural networks. Unlike the popular physics-informed neural networks (PINNs) approach, in which soft physics constraints are added as part of the network training cost function, our PCNNs respect hard physics constraints, such as ∇·B=0, by construction. Our 3D convolutional PCNNs map entire large (256x256x256 pixel) 3D volumes of time-varying current and charge densities to their associated electromagnetic fields. We demonstrate the method on space charge dominated, relativistic (5 MeV), short (hundreds of fs), high charge (2 nC) electron beams, such as those in the injector sections of modern free electron laser and plasma wakefield accelerators. We show that the method is accurate, respects physics constraints, and that the trained 3D convolutional PCNNs perform electromagnetic calculations orders of magnitude faster than traditional solvers which require a O(N2) process for calculating the space charge fields of intense charged particle beams. |
Wednesday, March 8, 2023 9:00AM - 9:12AM |
M53.00004: Image Classification Via Reversible Analog Superconducting Dynamics Ian Christie Landauer’s principle holds that logically irreversible digital operations are linked to physically irreversible state transitions, implying that irreversible operations must produce a finite amount of entropy. Dissipating this entropy to a thermal bath at finite frequency requires a minimum amount of power. To break this fundamental limit to power efficiency of digital devices, physically reversible hardware have been proposed. Digital devices are also being challenged by the rise of machine learning architectures which are often analog and inspired from analog biological neural networks (NN 1). This talk discusses the reversible limit of analog computing, presenting simulations of image classification using a network of superconducting-based analog flux parametrons (AFP 2). Utilizing Hamiltonian dynamics, these simulations explore realistic reversible and near-reversible hardware. We discuss how information chaos arises within phase-encoded AFP dynamics and the methods used to mitigate it. We show that even when mitigating for chaotic effects, energy dissipation per operation falls below Landauer's limit. Finally, we compare the success of various configurations of AFP-based NN implementing image classification benchmarks to the performance of related conventional artificial NN, finding similar accuracy. |
Wednesday, March 8, 2023 9:12AM - 9:24AM |
M53.00005: Machine Learning-Based Classification of Irregular Shape Defects in Metal Additive Manufacturing Elaine Jutamulia, Victoria A Ankel, Wei-Ying Chen, Alexander Heifetz We are developing algorithms for recognition of internal material defects in metal additive manufacturing from images obtained with pulsed thermal tomography (PTT). In prior work, we developed a convolutional neural network (CNN) which, having been trained on simulated 2D PTT images of subsurface elliptical defects, was able to classify the semi-major radii, semi-minor radii, and angular orientation of the best-fit ellipses in previously unseen PTT images. Training the CNN on irregular defect shapes, such as shapes imported from scanning electron microscopy (SEM) images of metallic laser powder bed fusion (LPBF)-printed specimens, would make the resulting classifications more descriptive of actual defect shapes. However, this requires a much higher volume of SEM images of material defects, which are difficult to obtain because of random occurrence of defects in LPBF. In one approach, we have developed a generative adversarial network (GAN) to augment the existing dataset of SEM defect images. In another approach, we are generating synthetic irregular-shaped defects, which have similar metrics (fractal dimension and area) as SEM images. Preliminary results demonstrate viability of these approaches. |
Wednesday, March 8, 2023 9:24AM - 9:36AM |
M53.00006: Deep Learning based Super-resolution models for Accelerating Multiphysics Simulations of Laser Powder Bed Fusion Francis Ogoke, Quanliang Liu, Olabode Ajenifujah, Amir Barati Farimani Laser Powder Bed Fusion is a method of additive manufacturing, where parts are constructed by iteratively fusing metal alloy powder, building complex 3D structures through laser melting. However, defects can form during the manufacturing process, where the meso-scale dynamics of the molten alloy near the laser, known as the melt pool, can directly contribute to the formation of undesirable porosity in the final part. Multiphysics simulation methods can describe the three-dimensional dynamics of the melt pool, but are computationally expensive at the resolutions required for accurate predictions. Therefore, in this work, we develop deep learning based super-resolution models to map low-resolution simulations of the melt pool temperature field to high-resolution simulations of the temperature field, avoiding the computational expense of performing multiple high-resolution simulations for analysis. To do so, we implement a 2-D diffusion model to upscale low-resolution cross-sections of the simulated melt pool to their corresponding high-resolution targets, by predicting the residual between the low-resolution and high-resolution images. We also implement a 3-D residual convolutional network super-resolution model to capture the full morphology of the melt pool. We demonstrate the preservation of key metrics of the melt pool physics between the ground truth simulation data and the super-resolution model output, such as the thermal field, the melt pool dimensions and the depth of the cavity formed by metal vaporization. |
Wednesday, March 8, 2023 9:36AM - 10:12AM |
M53.00007: Sujoy Ganguly Invited Speaker: Sujoy Ganguly
|
Wednesday, March 8, 2023 10:12AM - 10:24AM |
M53.00008: Jacobians in Deep Neural Networks : Criticality and beyond Darshil H Doshi, Tianyu He, Andrey Gromov Good parameter-Initialization is crucial for training Deep Neural Networks. Correct initialization ensures that the network function and gradients are well-behaved with depth. The conditions for such an initialization, known as “criticality”, help us select hyperparameters of the network. |
Wednesday, March 8, 2023 10:24AM - 10:36AM |
M53.00009: AutoInit: Automatic Initialization via Jacobian Tuning Tianyu He, Darshil H Doshi, Andrey Gromov Good initialization is essential for training Deep Neural Networks (DNNs). Oftentimes such initialization is found through a trial and error approach, which has to be applied anew every time an architecture is substantially modified, or inherited from smaller size networks leading to sub-optimal initialization. We will introduce a new and cheap algorithm, that allows one to find a good initialization automatically for general architectures. The algorithm utilizes the Jacobian between adjacent network basic blocks to tune the network hyperparameters to criticality. We will show the dynamics of the algorithm for fully connected networks with ReLU and derive conditions for its convergence. Then we will show that our method can find the automatic one-shot initialization for a variety of modern architectures with normalization layers and residual connections, where the initialization found by our method shows good performance on vision tasks. |
Wednesday, March 8, 2023 10:36AM - 10:48AM |
M53.00010: A Machine Learning Framework to Analyze and Optimize the Print Parameters of Direct Ink Writing (DIW) Systems Aldair Gongora, Deirdre Newton, Timothy D Yee, Zachary Doorenbos, Brian Giera, Thomas Yong-Jin Han, Kyle T Sullivan, Jennifer N Rodriguez Direct ink writing (DIW) is an extrusion based additive manufacturing technique that has gained significant popularity in a number of application areas due to its versatility in printable materials and part designs. However, analyzing and optimizing the print parameters of DIW systems remains a laborious, tedious, and resource intensive process. To further compound this challenge, the printability regime is intricate, complex, and depends on both the printing parameters and material/ink formulation. As a result, the criterion for assessing printability of DIW materials remains through extensive experimentation. Here, we present a data-driven machine learning (ML) framework for rapidly analyzing and optimizing the print parameters to meet a designer's desired printability metrics. The data-driven ML framework leverages techniques in image analysis, transfer learning, and probabilistic modeling for generating printability maps and uses Gaussian process regression to predict printability in the parameter space. This work represents a major advancement in accelerating the materials development and design cycle for DIW and offers transferable lessons for other manufacturing technologies. Prepared by LLNL under Contract DE-AC52-07NA27344. |
Wednesday, March 8, 2023 10:48AM - 11:00AM |
M53.00011: Odor discrimination and identification by graphene-based electronic nose system Gianaurelio Cuniberti, Shirong Huang, Alexander Croy, Bergoi Ibarlucea Olfaction is an evolutionary old sensory system, which provides sophisticated access to information about our surroundings. Inspired by the biological example, gas sensors in combination with efficient machine learning techniques aim to achieve similar performance and thus to digitize the sense of smell, which is termed as electronic nose (e-nose). Despite the significant progress of e-noses system, their compactness still remains challenging due to the complex layout design of sensor arrays with a multitude of receptor types or sensor materials, and the high working temperature. In this talk, we will present the odor discrimination and identification performance of graphene nanosensor-based electronic olfaction system. The developed prototype exhibits excellent odor discrimination and identification performance at room temperature, maximizing the obtained results from a single nanosensor. The underling adsorption mechanism between analyte gas molecules and functionalized graphene materials is elucidated by molecular dynamic simulations and density functional theory (DFT) calculations. This work may facilitate miniaturization of e-noses, digitization of odors, and distinction of volatile organic compounds (VOCs) in various emerging applications. |
Follow Us |
Engage
Become an APS Member |
My APS
Renew Membership |
Information for |
About APSThe American Physical Society (APS) is a non-profit membership organization working to advance the knowledge of physics. |
© 2024 American Physical Society
| All rights reserved | Terms of Use
| Contact Us
Headquarters
1 Physics Ellipse, College Park, MD 20740-3844
(301) 209-3200
Editorial Office
100 Motor Pkwy, Suite 110, Hauppauge, NY 11788
(631) 591-4000
Office of Public Affairs
529 14th St NW, Suite 1050, Washington, D.C. 20045-2001
(202) 662-8700