Bulletin of the American Physical Society
APS March Meeting 2020
Volume 65, Number 1
Monday–Friday, March 2–6, 2020; Denver, Colorado
Session R20: Data Science III: Deep LearningFocus
|
Hide Abstracts |
Sponsoring Units: GDS Chair: Emine Kucukbenli, Harvard University Room: 301 |
Thursday, March 5, 2020 8:00AM - 8:36AM |
R20.00001: Deep Learning-enabled Computational Microscopy and Sensing Invited Speaker: Aydogan Ozcan Deep learning is a class of machine learning techniques that uses multi-layered artificial neural networks for automated analysis of signals or data. The name comes from the general structure of deep neural networks, which consist of several layers of artificial neurons, each performing a nonlinear operation, stacked over each other. Beyond its main stream applications such as the recognition and labeling of specific features in images, deep learning holds numerous opportunities for revolutionizing image formation, reconstruction and sensing fields. In fact, deep learning is mysteriously powerful and has been surprising optics researchers in what it can achieve for advancing optical microscopy, and introducing new image reconstruction and transformation methods. From physics-inspired optical designs and devices, we are moving toward data-driven designs that will holistically change both optical hardware and software of next generation microscopy and sensing, blending the two in new ways. Today, we sample an image and then act on it using a computer. Powered by deep learning, next generation optical microscopes and sensors will understand a scene or an object and accordingly decide on how and what to sample based on a given task – this will require a perfect marriage of deep learning with new optical microscopy hardware that is designed based on data. For such a thinking microscope, unsupervised learning would be the key to scale up its impact on various areas of science and engineering, where access to labeled image data might not be immediately available or very costly, difficult to acquire. In this presentation, I will provide an overview of some of our recent work on the use of deep neural networks in advancing computational microscopy and sensing systems, also covering their biomedical applications. |
Thursday, March 5, 2020 8:36AM - 8:48AM |
R20.00002: Exploring Organic Ferroelectrics Using Data-driven Approaches Ayana Ghosh, Nicholas Lubbers, Serge M Nakhmanson, Jian-Xin Zhu Recent advances in the synthesis of polar molecular materials have produced practical alternatives to ferroelectric ceramics, opening up exciting new avenues for the incorporation of such compounds into modern electronic devices. However, in order to fully realize the prospects of polar polymer and molecular crystals for modern technological applications, it is paramount to acquire diverse datasets of potential organic ferroelectrics such that the mechanisms governing the emergence of ferroelectricity can be studied. Here we propose to use data-driven approaches to judiciously shortlist candidates from a wide range of chemical space with ferroelectric functionalities. First, this investigation will be governed by identification of chemical similarities between existing molecular compounds exhibiting similar ferroelectric behavior. Second, we investigate machine learning (ML) and deep neural network models for estimating charge transfer effects in organic chemistry. The dipole moment and ferroelectric properties estimated by ML can then be used to supplement the data-driven screening of possible organic ferroelectrics. |
Thursday, March 5, 2020 8:48AM - 9:00AM |
R20.00003: Deep Learning Model for Finding New Superconductors Tomohiko Konno, Hodaka Kurokawa, Fuyuki Nabeshima, Yuki Sakishita, Ryo Ogawa, Iwao Hosako, Atsutaka Maeda It is very difficult for both theories and computational methods to predict the superconducting transition temperatures Tc of superconductors for strongly correlated systems, in which high-temperature superconductivity emerges. Exploration of new superconductors still relies on the experience and intuition of experts, and is largely a process of experimental trial and error. In one study, only 3% of the candidate materials showed superconductivity. Here we report an interdisciplinary attempt for finding new superconductors based on deep learning. We represented the periodic table in a way that allows a deep learning model to learn it. Although we used only the chemical composition of materials as information, we obtained an R2 value of 0.92 for predicting Tc for materials in a database of superconductors. We obtained three remarkable results. The deep learning method can predict superconductivity for a material with a precision of 62%, which shows the usefulness of the model; it found the recently discovered superconductor CaBi2, which is not in the superconductor database; and it found Fe-based high-temperature superconductors (discovered in 2008) from the training data before 2008. These results open the way for the discovery of new high-temperature superconductor families. |
Thursday, March 5, 2020 9:00AM - 9:12AM |
R20.00004: Deep Learning for Energetic Materials: Predicting Material Properties from Electronic Structure using Convolutional Neural Networks Alex Casey, Brian Barnes, Ilias Bilionis, Steven F. Son Developing numerical descriptions of complex objects, like molecular structure, is a difficult task. The accuracy of a machine learned model depends on the input representation. Ideally, input descriptors encode the essential physics and chemistry that influence the target property. Thousands of molecular descriptors have been proposed and proper selection of features requires considerable domain expertise. In contrast, deep learning networks are capable of learning rich data representations. This provides a compelling motivation to use deep learning networks to learn molecular structure-property relations from 'raw' data. We develop a convolution neural network capable of directly parsing the 3D electronic structure of a molecule described by spatial point data for charge density and electrostatic potential concatenated into a 4D tensor. The model is jointly trained on over 20,000 molecules that are potentially energetic materials (explosives) to predict dipole moment, total electronic energy, Chapman-Jouguet (C-J) detonation velocity, C-J pressure, C-J temperature, crystal density, HOMO-LUMO gap, and solid phase heat of formation. This work demonstrates the first use of complete 3D electronic structure for machine learning of molecular properties. |
Thursday, March 5, 2020 9:12AM - 9:24AM |
R20.00005: Optimization of Molecular Characteristic using Continuous Representation of Molecules by Variational Autoencoder with Discriminator Kyosuke Sato, Kenji Tsuruta Efficient molecular search contributes to an essential speedup of the development of organic devices and, in turn to the improvement of their characteristics. In the present study, we focus on the deep learning variational auto-encoder (VAE) model[1], where molecules represented by SMILES strings can be efficiently converted to multivariable continuous space. The VAE consists of two neural networks: an encoder and a decoder. The one-hot representation of SMILES is input to the encoder and mapped to the latent variable space. Here we further improve the output rate of valid SMILES of decoder by introducing a discriminator attached to the VAE stream. Adopting a molecular-mechanics method to calculate 3D structure from SMILES, we can optimize physical properties of the molecule by other simulation methods such as density-functional-theory calculations even when there is not enough data set. The range of physical property space covered by the SMILES representation is thereby expanded and the data-driven optimization using Kernel Ridge Regression method can be performed within the search space. In the presentation, we show the effectiveness of this method for optimizing a molecular HOMO-LUMO gap as an example. |
Thursday, March 5, 2020 9:24AM - 9:36AM |
R20.00006: An Initial Design-based Deep Learning Procedure for the Optimization of High Dimensional ReaxFF Parameters MERT SENGUL, Yao Song, Linglin He, Ying Hung, Tirthankar Dasgupta, Adri C.T. van Duin Atomistic level investigations are a significant part of today’s materials discovery research; however, most of the modeling methods that can describe chemical reactions are restricted to small molecular systems due to computational costs. The ReaxFF, an empirical interatomic potential, is capable of simulating reactions in larger molecular systems; however, the application of ReaxFF requires a significant preprocessing. One of the preprocessing steps is the optimization of functional parameters that are used to calculate interatomic interactions. This optimization process is complex due to high dimensionality. Here, we propose a deep learning (DL)-based procedure to be used in ReaxFF parameter optimization. The procedure is composed of three stages, which are: 1) data set creation; 2) DL model fitting and 3) local-minima detection. This DL procedure eliminates unfeasible regions in parameter space, which originate from the unphysical atomistic interactions, and constructs a more comprehensive understanding of a physically meaningful parameter landscape. The performance of the procedure will be evaluated by its application to molecular systems. |
Thursday, March 5, 2020 9:36AM - 9:48AM |
R20.00007: Feature Extraction Using Semi-Supervised Deep Learning. Muammar El Khatib, Wibe A De Jong Features are defined as measurable properties that characterize observed phenomena and represent a key part of machine learning (ML) algorithms. In materials sciences, ML has successfully accelerated atomistic simulations using man-engineered features for tasks such as energy or atomic forces predictions. These features fulfill physics constraints such as rotational and translational invariance, uniqueness and, locality (the sum of local contributions reconstructs a global quantity). However, these ML models are known to perform poorly when operating out of the training set regime because features are not representative of the underlying structure of the data. This could be improved if features are extracted with advanced hybrid architectures e.g. a variational autoencoder that is trained with physics constraints introduced with an external task and a loss function. We will explore how the use of semi-supervised learning techniques can be a powerful tool for the extraction of features for atomistic simulations. All results shown herein can be reproduced with ML4Chem: a free software package for machine learning in chemistry and materials sciences. |
Thursday, March 5, 2020 9:48AM - 10:00AM |
R20.00008: Unsupervised feature extraction in simple physical models through mutual information maximization Leopoldo Sarra, Florian Marquardt When studying systems with many degrees of freedom, a typical problem is to find the correct low-dimensional variables to describe them on a higher level of abstraction. However, sometimes it is not clear how to choose meaningful quantities. By defining relevant features as low dimensional variables that preserve the largest mutual information with the original coordinates of the system, we set up an unsupervised learning technique to automatically extract those features. A variational bound allows to estimate mutual information through deep neural networks. We show example applications to statistical mechanics and classical dynamics. |
Thursday, March 5, 2020 10:00AM - 10:12AM |
R20.00009: Integration of Neural Network-Based Symbolic Regression in Deep Learning for Scientific Discovery Samuel Kim, Peter Lu, Michael Gilbert, Srijon Mukherjee, Li Jing, Vladimir Čeperić, Marin Soljacic Symbolic regression is a powerful technique that can discover the underlying analytical equations describing data, which can lead to explainable models and generalizability outside of the training data set. Here we use a neural network for symbolic regression based on the EQL network and integrate it into other deep learning architectures such that the whole system can be trained end-to-end through backpropagation. We demonstrate this system on an arithmetic task involving MNIST digits and on prediction of dynamical systems. The architecture is able to simultaneously extract meaningful latent variables and find the underlying equations that generalize extremely well outside of the training data set compared to a standard neural network approaches, paving the way for scientific discovery. |
Thursday, March 5, 2020 10:12AM - 10:24AM |
R20.00010: Rapid machine learning-based solutions of partial differential equations on complex domains. Vikas Dwivedi, Balaji Srinivasan Conventional methods like FEM (finite element method) and FVM (finite volume method) are mesh-based. However, if the computational domain consists of a complicated polygon with very short line segments, then it puts serious restrictions on mesh generation methods like triangulation. Recently, Berg et. al [1] have developed a deep unified ANN algorithm to |
Thursday, March 5, 2020 10:24AM - 10:36AM |
R20.00011: Probabilistically-autoencoded horseshoe-disentangled multidomain item-response theory models Joshua Chang, Shashaank Vattikuti, Carson C Chow Item response theory (IRT) is a non-linear generative probabilistic paradigm for using exams in order to quantify latent traits. In multidimensional IRT, one requires a factorization of the test items. For this task, linear factor analysis methods are often used, making IRT a posthoc model. We propose skipping the initial factor analysis by using a sparsity-promoting horseshoe prior to perform factorization directly within the IRT model so that all training occurs in a single self-consistent step. By binding the generative IRT model to a Bayesian neural network (forming a probabilistic autoencoder), one obtains a scoring algorithm consistent with the interpretable Bayesian model. In some IRT applications the black-box nature of a neural network scoring machine is desirable. Within this problem, we investigate the translation of some regularization principles common in Bayesian modeling to neural networks. |
Thursday, March 5, 2020 10:36AM - 10:48AM |
R20.00012: Turbulence-generating networks Armando Garcia, Rao Gudimetla, Jorge Munoz The propagation of light through the atmosphere is often simulated with screens that impose a phase difference on an incoming wave. The screens are applied in series and have phase power spectral densities that are consistent with the strength of the turbulence to be simulated. We propose a new method to simulate the effect of turbulence on light propagation that uses mathematical graphs (networks) that are agnostic of spatial dimension or angle information. The intensity variation in the final screen is a measure of turbulence strength and we used several machine learning methods to adjust the probability distribution functions (PDF) of the weights of the edges of the networks to achieve the desired variations. We compare our results to the phase screen method for the case of the Kolmogorov spectrum using statistical measures such as the structure function. Finally, we explore avenues for improving the computational cost of producing turbulence-degraded images with this method as well as extending it to anisotropic turbulence. |
Thursday, March 5, 2020 10:48AM - 11:00AM |
R20.00013: SignalTrain: Modeling Time-dependent Nonlinear Signal Processing Effects Using Deep Neural Networks William Mitchell, Scott Hawley The advent of increased consumer computing power and graphics processing unit (GPU) usage over the last decade has made possible machine learning approaches to modeling problems once thought impractical.This work expands on prior research published on modeling nonlinear time-dependent signal processing effects associated with music production by means of a deep neural network1. The presented results show the progress in accurately modeling these effects through architecture and optimization changes, increasing computational efficiency, lowering noise, and extending to a larger variety of nonlinear audio effects. Unique contributions of this effort include the ability to emulate the individual settings or “knobs” you would see on an analog piece of equipment, and with the production of commercially viable audio, i.e. 44.1kHz sampling rate at 16-bit resolution. |
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