Bulletin of the American Physical Society
APS March Meeting 2021
Volume 66, Number 1
Monday–Friday, March 15–19, 2021; Virtual; Time Zone: Central Daylight Time, USA
Session Y05: Machine Learning in Nonlinear Physics and Mechanics IILive
|
Hide Abstracts |
Sponsoring Units: DSOFT GSNP DCOMP Chair: Christopher Rycroft, Harvard University; Shmuel Rubinstein, Harvard University Room: 05 |
Friday, March 19, 2021 11:30AM - 11:42AM Live |
Y05.00001: Neuromorphics for network inference:
new techniques and validation in opto-electronic experiments Amitava Banerjee, Joseph Hart, Rajarshi Roy, Edward Ott We devise a machine learning technique to solve the general problem of inferring network links with time-delays purely from time-series data of the network nodal states. This task has applications in fields ranging from applied physics and engineering to neuroscience and biology. To achieve this, we train a type of machine learning system known as reservoir computing to mimic the dynamics of the unknown network and relate the parameters of the reservoir system to the unknown network structure. Our technique, by its nature, is non-invasive, but is motivated by the widely-used invasive network inference method whereby the responses to active perturbations applied to the network are observed and employed to infer network links (e.g., knocking down genes to infer gene regulatory networks). We test this technique on experimental and simulated data from delay-coupled opto-electronic oscillator networks. We show that the technique often yields very good results particularly if the system does not exhibit synchrony of the nodal dynamics. We also find that the presence of dynamical noise can strikingly enhance the accuracy and ability of our technique. |
Friday, March 19, 2021 11:42AM - 11:54AM Live |
Y05.00002: Reconstruction of Protein Structures from Single-Molecule Time Series Maximilian Topel, Andrew Ferguson Single-molecule experimental techniques can record a number of observables describing dynamics of molecules, but not an all atom conformational dynamics. Takens’ Delay Embedding Theorem asserts that, under certain conditions, time-delay vectors of scalar observations contain sufficient information to reconstruct full molecular configurations up to an a priori unknown transformation. Applying Takens’ Theorem and tools from statistical mechanics, manifold learning, neural networks, and graph theory, we establish an approach Single-molecule TAkens Reconstruction (STAR) to learn this Jacobian and reconstruct all atom trajectories from experimentally-measurable scalar observables. We apply STAR to molecular dynamics simulations of a C24H50 polymer chain and the mini-protein Chignolin. Trained models reconstruct molecular configurations from synthetic time series data of head-to-tail molecular distances with atomistic root mean squared deviation accuracies better than 0.2 nm. This work exhibits potential to reconstruct protein structures from time series of experimentally-measurable observables and establishes theoretical and algorithmic foundations to do so with real experimental data. |
Friday, March 19, 2021 11:54AM - 12:06PM Live |
Y05.00003: Deep learning enabled wavefront shaping in complex cavities with a binary tunable metasurface Benjamin Frazier, Thomas M Antonsen, Steven M Anlage Modern electronics have become more densely populated due to miniaturization and are expected to perform in increasingly complex environments. These environments give rise to extreme electromagnetic interference through noise and unwanted coupling between components. The ability to isolate or reject interference and to do so intelligently is critical for practical applications. We previously demonstrated the ability to create nulls in the transmission coefficient or induce coherent perfect absorption states at arbitrary frequencies with a binary programmable metasurface[Frazier, Antonsen, Anlage, and Ott, "Wavefront Shaping with a Tunable Metasurface: Creating Coldspots and Coherent Perfect Absorption at Arbitrary Frequencies”, arXiv:2009.05538, https://arxiv.org/abs/2009.05538]. In this work, we show how deep learning can be leveraged to optimize the metasurface commands without relying on a blind iterative optimization approach. |
Friday, March 19, 2021 12:06PM - 12:18PM Live |
Y05.00004: Self-learning machines based on time reversal Victor Lopez Pastor, Florian Marquardt
|
Friday, March 19, 2021 12:18PM - 12:30PM Live |
Y05.00005: Learning active hydrodynamics from particle simulations Rohit Supekar, Boya Song, Alasdair Hastewell, Alexander Mietke, Jorn Dunkel Recent advances in particle-based simulation methods and high-resolution imaging techniques have enabled the precise characterization of collective dynamics in various biological and engineered active fluids. In parallel, data-driven algorithms for learning interpretable continuum models have shown promising potential for the recovery of underlying PDEs from continuum simulations. By contrast, learning macroscopic hydrodynamic equations and closure relations from microscopic particle simulations remains a major challenge. Here, we present a framework that leverages sparse regression learning algorithms to discover PDE models from coarse-grained microscopic data, while incorporating the relevant physical symmetries. We illustrate the practical potential through an application to a polar active particle model with alignment interactions mimicking those of swimming sperm cells. For this experimentally relevant model system, our scheme succeeds in learning hydrodynamic equations that reproduce the characteristic vortex dynamics observed in the particle simulations. More generally, these results demonstrate how one can learn continuum theories directly from large-scale microscopic simulations and observations of complex systems that have thus far eluded analytical coarse-graining. |
Friday, March 19, 2021 12:30PM - 12:42PM Live |
Y05.00006: Machine learning active-nematic hydrodynamics Jonathan Colen, Ming Han, Rui Zhang, Steven Redford, Linnea M Lemma, Link Morgan, Paul Ruijgrok, Raymond Adkins, Zev Bryant, Zvonimir Dogic, Margaret Gardel, Juan De Pablo, Vincenzo Vitelli Hydrodynamic theories effectively describe many-body systems out of equilibrium in terms of a few macroscopic parameters. However, such hydrodynamic parameters are difficult to determine from microscopics. Seldom is this challenge more apparent than in active matter, where the hydrodynamic parameters are in fact fields encoding the distribution of the energy-injecting microscopic components. Here, we use active nematics to demonstrate that neural networks can map out the spatio-temporal variation of multiple hydrodynamic parameters and forecast the chaotic dynamics of these systems. By analyzing microtubule-kinesin and actin-myosin experiments as computer vision problems, our algorithms can determine how activity and elastic moduli change as a function of space and time, as well as ATP or motor concentration. They can also forecast the evolution of these chaotic many-body systems solely from image-sequences of their past using a combination of autoencoders and recurrent networks with residual architecture. Our work paves the way for artificial-intelligence characterization and control of coupled chaotic fields in diverse physical and biological systems even when no knowledge of the underlying dynamics exists. |
Friday, March 19, 2021 12:42PM - 12:54PM Live |
Y05.00007: Tracking Islands on Smectic Bubbles using Machine Learning Ravin Chowhury, Eric Hedlund, Adam AS Green, Cheol Park, Joseph MacLennan, Noel Anthony Clark The Observation and Analysis of Smectic Islands in Space (OASIS) mission was a series of experiments on the International Space Station probing the dynamics of smectic islands on curved fluid bubbles in microgravity. The analysis of the motion of these islands through traditional particle tracking technologies is challenging due to uneven lighting and geometric distortion. We apply modern advances in deep learning to identify and track these islands in a novel way. Our Neural Network uses a modified U-Net [Ronneberger, et al., Medical Image Computing and Computer-Assisted Intervention, pp 234-241 (2015)] image segmentation net to generate masks that classify our objects. Segmented masks provide a simplified representation of the objects on screen and a detection layer easily predicts locations of each island. The model is trained using synthetic data generated using the 3D rendering software Blender. This technique shows promise for a variety of problems including improving the detection of topological defects in liquid crystals [Green et al., Soft Matter (2020)]. |
Friday, March 19, 2021 12:54PM - 1:06PM Live |
Y05.00008: Data-Driven Classical Density Functional Theory: A Case for Physics Informed Learning Petr Yatsyshin, Serafim Kalliadasis, Andrew B Duncan In traditional sense, physical modeling is often associated with analytic derivations, followed by computation and validation against data. On the other hand, modern statistical inference offers principled means to accomplish similar goals numerically, whilst staying in touch with the data at all stages of modelling. In the present talk we explore the synthesis of both these paradigms, applied to modelling classical many-body systems. We propose a data-driven physics-informed inference framework for Helmholtz free energy functionals of such systems. Our approach is fully Bayesian and yields uncertainty quantification of the inferred model about its own predictions. The proposed algorithm trains humanly interpretable analytic free energy functionals using particle data, obtained from small-scale simulations. We focus on classical statistical-mechanical systems with excluded volume repulsive interactions and use a prototypical case of a one-dimensional fluid for algorithm validation. We are able to train canonical and grand-canonical representations of the underlying system. Extensions to higher-dimensional systems are conceptually straightforward. Using standard coarse-graining techniques, our results can also be made applicable to fluids with attractive-repulsive interactions. |
Friday, March 19, 2021 1:06PM - 1:18PM Live |
Y05.00009: Extracting Dynamical laws in Dusty Plasmas using Machine Learning Wentao Yu, Guram Gogia, Justin Burton Machine learning is concomitant with highly complex systems where billions of data readily available. For individual experimental physics labs, such data is often expensive or time-consuming. Instead, simulated data with known underlying governing equations are used to train models. Here we provide a tractable experimental system with complex dynamics and copious amounts of data. By tracking the 3D motion of hundreds of levitated microparticles in a “dusty” plasma, we can tease apart the known and unknown forces and “learn” their underlying physics. Dusty plasmas are ubiquitous in the space and industry and exhibit a rich spectrum of forces (electrostatic, hydrodynamic, ion wake, drag, and stochastic noise). We can reveal many of these forces by analyzing the “Brownian” motion of the particles using machine learning trained on simulated data. In simulations, the forces are linearized and features extracted from trajectories using conventional methods (Fourier transformation, Bayesian Inference, and mutual information, etc.) are then fed to machine learning models. We can predict linear coefficients with a two-fold precision over conventional methods, and uncover new surprises such as the heavy-tailness of stochastic noise, and vorticity in the background ion flow. |
Friday, March 19, 2021 1:18PM - 1:30PM Live |
Y05.00010: Learning the Constitutive Relation of Polymeric Flows with Memory Naoki Seryo, Takeshi Sato, John Molina, Takashi Taniguchi We develop a learning strategy to infer the constitutive relation for the stress of polymeric flows with memory. We assume that the relations are in differential form, functions of the velocity gradient and stress, but no assumptions are made on their functional form. The required training data is obtained from stress trajectories generated during microscopic polymer simulations. This data is then used within a Gaussian Process (GP) regression scheme, in order to infer the most likely constitutive equation. The GP prediction for the constitutive relation can then be used within macro-scale flow simulations, allowing us to update the stresses in the fluid in manner which satisfies the dynamics of the underlying microscopic model. We tested the method on a simple microscopic model (non-interacting Hookean dumbbells) and successfully recovered the exact constitutive relation. The resulting macroscopic flow simulations give the same level of accuracy as Multi-Scale descriptions at a small fraction of the cost [1]. |
Friday, March 19, 2021 1:30PM - 1:42PM Live |
Y05.00011: Defect Annihilation in Liquid Crystal Physics: Using Deep Learning to Probe the Dynamics of Defects Adam AS Green, Ravin Chowdhury, Eric Minor, Stian Howard, Cheol Park, Noel Anthony Clark Smectic liquid crystals are a fluid state of matter that supports long-lived topological defects. Classical defect annihilation dynamics in two dimensions are well described by the XY-model, which predicts dynamical scaling—a single length scale controlling the dynamics. Despite years of diligent experimentation by the community, the applicability of the XY model to the defect dynamics in liquid crystals remains an open question. Recent theoretical and experiment investigations on the dynamics of isolated defect pairs have already shown annihilation to be heavily dependent on hydrodynamic interactions, which are beyond the XY model. Machine learning methods have recently been applied to analyze dense textures of defects in freely suspended smectic C films [1]. In this work, we apply these machine learning methods to analyze high-speed microscopy images of defect annihilation in quenched films. Due to the power-law scaling of defect annihilation, deviations from the XY model are predicted to be largest at early times, making this a highly sensitive test of the applicability of the XY model to quasi-2D smectic LC. |
Friday, March 19, 2021 1:42PM - 1:54PM Live |
Y05.00012: Machine Learning approach to the discrimination of phospholipid gel and fluid states in lipid bilayers. Vivien Walter, Céline Ruscher, Carlos Marques, Olivier Benzerara, Fabrice Thalmann The two-states model of the phospholipid gel-fluid transition was introduced in the late seventies, and is still routinely used to interpret calorimetric data in the field of lipid membranes [1]. Rather than relying on traditional order parameters (segment orientation, membrane thickness) we propose to use Machine Learning classifiers to assign either a gel, or a fluid state, to any lipid molecular conformation arising from molecular dynamics simulations [2]. Using respectively the high and low temperature simulated phases as training sets, we investigate the behavior of the lipid assembly at intermediate temperatures, in the binary mixture cases and in the presence of various external compounds [4]. |
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