Bulletin of the American Physical Society
APS March Meeting 2020
Volume 65, Number 1
Monday–Friday, March 2–6, 2020; Denver, Colorado
Session M34: Machine Learning in Nonlinear Physics and MechanicsFocus
|
Hide Abstracts |
Sponsoring Units: DSOFT GSNP DCOMP Chair: Shmuel Rubinstein, Harvard University Room: 506 |
Wednesday, March 4, 2020 11:15AM - 11:51AM |
M34.00001: Learning the onset of frictional motion Invited Speaker: Yohai Bar-Sinai The onset of frictional motion is a central process in modeling mechanical phenomena, from squeaking hinges to catastrophic earthquakes. The processes involved in it are complex and to a large extent not understood. Specifically, the simple notion of "static friction coefficient" is most probably not a well-defined material parameter, as experiments show that it is variable, even for a single interface under carefully controlled experimental conditions. I will show that, in a well-controlled laboratory setup with detailed interfacial measurements, relatively simple linear models can explain much of the observed variance, indicating that a significant portion of the uncertainty is non-stochastic and is encoded in the frictional interface. In addition, I will discuss recent data-driven efforts in predicting the onset of frictional motion in geophysical faults, i.e. earthquake forecasting, which is a long-standing and very challenging problem. However, the confluence of new kinds of measurements and advances in machine learning offers new and promising directions. |
Wednesday, March 4, 2020 11:51AM - 12:03PM |
M34.00002: Detecting Depinning and Nonequilibrium Transitions with Unsupervised Machine Learning Danielle McDermott, Cynthia Reichhardt, Charles Reichhardt Using numerical simulations of a model disk system, we demonstrate that a machine learning generated order parameter can detect depinning transitions and different dynamic flow phases in systems driven far from equilibrium. We specifically consider monodisperse passive disks with short range interactions undergoing a depinning phase transition when driven over quenched disorder. The machine learning derived order parameter identifies the depinning transition as well as different dynamical regimes, such as the transition from a flowing liquid to a phase separated liquid-solid state that is not readily distinguished with traditional measures such as velocity-force curves or Voronoi tessellation. The order parameter also shows markedly distinct behavior in the limit of high density where jamming effects occur. Our results should be general to the broad class of particle-based systems that exhibit depinning transitions and nonequilibrium phase transitions. |
Wednesday, March 4, 2020 12:03PM - 12:15PM |
M34.00003: Machine Learning and Benchtop Experiments Shmuel Rubinstein Machine learning has generated much recent excitement within the physics community and provides a powerful new tool to analyze and understand many physical systems. However, in the experimental study of complex physical systems, the usage of machine learning is still in its infancy. Specifically, it is not obvious which scientific questions are susceptible to machine learning disruption and, even more interesting, which questions are not? In this talk, I will address our approach to these questions, sharing our attempts to leverage lab models of complex systems for this study. I will discuss our tactics to holistically amalgamate experiments with simulations. |
Wednesday, March 4, 2020 12:15PM - 12:27PM |
M34.00004: Supervised Autoencoder for Inverse Kirigami Design Paul Hanakata, Ekin Dogus Cubuk, David K Campbell, Harold S. Park Recently, machine learning (ML) methods have shown successes in predicting mechanical properties of composite materials as a forward solver. While ML approach is much faster than the conventional physics-based solvers (e.g. molecular dynamics), most current ML techniques need to screen the entire library in order to perform inverse design. Thus, this approach might no longer be practical for a system with a very large design space. Here, we use a supervised-autoencoder (sAE) to perform inverse design in graphene kirigami where predicting ultimate stress or fracture point is known to be difficult due to nonlinear effects arise from the out-of-plane buckling. Unlike the standard autoencoder, our sAE is not only able to reconstruct cut configurations but also to predict mechanical properties of graphene kirigami and classify graphene kirigami with either parallel or orthogonal cuts. Furthermore, we find that by interpolating different configurations the sAE is able to generate new designs consisting of mixed parallel and orthogonal cuts while only being trained with kirigami structures with parallel and orthogonal cuts. This allows us to design and optimize materials in the latent space, which is more efficient than to perform optimization in the original representation. |
Wednesday, March 4, 2020 12:27PM - 12:39PM |
M34.00005: Softness Correlations in Low-Temperature Supercooled Liquids Rahul Chacko, François P Landes, Giulio Biroli, Olivier Dauchot, Andrea Jo-Wei Liu, David Reichman Local structure is known to play a dominant role in determining where structural relaxation occurs [1,2]. This can be quantified using a machine learning approach, yielding a linear model mapping local structure to "softness", a quantity that predicts the propensity of a particle to rearrange [3]. We find that this machine-learned weighted integral of the pair correlation function, when trained on an athermal system relaxing under gradient descent, performs surprisingly well when predicting the dynamics of a supercooled liquid. We use swap Monte Carlo [4] to study the evolution of the spatial correlation of the so defined softness, down to deeply supercooled temperatures. We then compare this length scale to other length scales that have been identified in the literature. |
Wednesday, March 4, 2020 12:39PM - 12:51PM |
M34.00006: Data-driven inference of thermodynamic properties from non-equilibrium stochastic fluctuations Yoon Jung, Junang Li, Nikta Fakhri Living systems are driven out of equilibrium via an input of cellular chemical energy, part of which is dissipated into the environment. For systems that exhibit non-equilibrium steady states, thermodynamic fluctuations encode signatures that can be used to infer properties of the system. For example, model-based approaches can be used to quantify how far the system is from equilibrium by measuring the dissipation rate. Here, we present a data-driven approach for inferring system properties based on scattering transforms. The method utilizes symmetries arising from the stationary nature of stochastic fluctuations which allows solving inverse problems with fewer measurements. The results from simulations and experimental data demonstrate that the proposed approach serves as an effective method to infer system properties from thermodynamic fluctuations in living systems. |
Wednesday, March 4, 2020 12:51PM - 1:03PM |
M34.00007: Inverse learning of material physics through in-situ image data and continuum modeling Hongbo Zhao, Brian D Storey, Richard Braatz, Martin Bazant With the availability of microscopic spatio-temporal image data of materials, there is a tremendous amount of hidden information about the material properties. Using a framework of PDE-constrained optimization, we demonstrate that multiple constitutive relations can be extracted simultaneously from a small set of images of pattern formation. Examples include state-dependent properties such as the diffusivity, kinetic prefactor, free energy, and direct correlation function in the Cahn-Hilliard equation, Allen-Cahn equation, or dynamical density functional theory (Phase-Field Crystal Model). Compared to the data-driven modeling approach and the recent work on PDE discovery, our approach provides clear physical interpretability by prescribing a general governing equation while achieving a high expressive power through nonlinear and/or nonlocal (integro-differential) constitutive relations. |
Wednesday, March 4, 2020 1:03PM - 1:15PM |
M34.00008: Using Machine Learning to analyze Defect Annihilation Dynamics in Smectic C films Matthew Glaser, Eric Minor, Stian Howard, Adam Green, Cheol Park, Noel Anthony Clark We demonstrate a method for training a convolutional neural network with simulated images for usage in the study of topological defect annihilation in freely-suspended SmC liquid crystal films. Modern machine learning methods require large, robust training data sets to generate accurate predictions. Generating these training sets requires a significant up-front time investment that is often impractical for small-scale applications. Here we demonstrate a ‘full-stack’ computational solution, where the training data set is generated on-the-fly using a noise injection process to produce simulated data characteristic of the experimental system. The experiment requires accurate observations of both the spatial distribution of the defects and the total number of defects at every time step, making it an ideal system for testing the robustness of the trained network. The fully trained network was found to be comparable in accuracy to identifying the defects by hand, with a four-orders of magnitude improvement in time efficiency. |
Wednesday, March 4, 2020 1:15PM - 1:27PM |
M34.00009: Experimental Realization of Reservoir Computing with Wave Chaotic Systems Shukai Ma, Thomas M Antonsen, Edward Ott, Sarthak Chandra, Steven Anlage The execution of machine learning (ML) software largely depends on the computing `substrate', which is often not optimized for running ML tasks. The invention of ML-tailored hardware may greatly improve the computing speed and power efficiency. Photonic devices are well-suited for ML due to the parallelism of light. Reservoir computing (RC) is essentially a one-layer neural network (NN) with nonlinear connections, but radically simpler than NN since only the coupling between the reservoir nodes and outputs is trained. Thus RC is well-suited for physical realizations. Here we utilize the complicated wave dynamics inside a chaotic-shaped overmoded electromagnetic cavity containing nonlinear elements to emulate the complex dynamics of an RC. We propose unique techniques to create virtual RC nodes by both frequency stirring and spatial perturbation. The computational power of the wave chaotic RC is experimentally demonstrated with the so-called observer task, where we predict the future evolution of chaotic Rossler y(t) and z(t) time series using the x(t) series as the input. Different tasks are executed with a single RC physical device by simply switching output couplers. |
Wednesday, March 4, 2020 1:27PM - 1:39PM |
M34.00010: Chaotic source separation solved by a tank of water through invertible generalized synchronization Zhixin Lu, Jason Kim, Danielle Bassett Statistical methods such as independent component analysis and principal component analysis were proposed to separate a set of source signals from a mixed signal by assuming the statistical independence of sources. Here, we focus on chaotic source separation where the sources are chaotic systems. We assume that one has no knowledge about the governing equations of the source signals, and that the mixed signal is simply the sum of the sources. From the perspective of dynamical systems, we propose a supervised learning framework that can solve this problem through an intermediate dynamical system. To demonstrate the power of this framework, we employ a simulated tank of water as the intermediate system, and train it to regenerate chaotic signals from a mixed signal that is the sum of any 2 chaotic trajectories from 6 chaotic systems. We elucidate the underlying mechanism as constructing a nonlinear state-observer utilizing the concept of invertible generalized synchronization. We predict that it is impossible to perfectly separate the chaotic sources if the two sources systems are governed by the same dynamical equations. |
Wednesday, March 4, 2020 1:39PM - 1:51PM |
M34.00011: The Dependence of Reservoir Computing on System Parameters Louis Pecora, Thomas L Carroll Reservoir computing (RC) has become an interesting topic for the nonlinear Dynamics field. More so because despite many demonstrations of RC abilities the actual dynamical reasons RCs work is still not well understood. We examined several aspects of RC using fitting and prediction of signals as a benchmark. Among the parameters we explored [1] were network heterogeneity of link weights, existence of symmetries in the networks, variations of node parameters and vector fields, and spectral radius of the network. One important result is that the number of independent signals that can be used from a driven RC is limited and independent of the number of nodes above a certain vertex number. |
Wednesday, March 4, 2020 1:51PM - 2:03PM |
M34.00012: Reduced network extremal ensemble learning (RenEEL) scheme for community detection in complex networks Kevin E. Bassler, Jiahao Guo, Pramesh Singh We introduce an ensemble learning scheme for community detection in complex networks. The scheme uses a Machine Learning algorithmic paradigm we call Extremal Ensemble Learning. It uses iterative extremal updating of an ensemble of network partitions, which can be found by a conventional base algorithm, to find a node partition that maximizes modularity. At each iteration, core groups of nodes that are in the same community in every ensemble partition are identified and used to form a reduced network. Partitions of the reduced network are then found and used to update the ensemble. The smaller size of the reduced network makes the scheme efficient. We use the scheme to analyze the community structure in a set of commonly studied benchmark networks and find that it outperforms all other known methods for finding the partition with maximum modularity. |
Wednesday, March 4, 2020 2:03PM - 2:15PM |
M34.00013: Visualizing statistical models in Minkowski space: an analytical coordinate embedding Han Kheng Teoh, Katherine Quinn, Colin B Clement, Jaron Kent-Dobias, Qingyang Xu, James Patarasp Sethna Dimensionality reduction techniques are often used to provide a lower dimensional description of high dimensional data. Quinn et. al [1] proposed an intensive isometric embedding, InPCA in visualizing probabilistic models manifold with the Bhattacharyya distance. It was observed that the InPCA manifolds form a hierarchy of cross-sectional spans that shrinks geometrically in Minkowski space, allowing the use of only a few principal components to capture most of the variation. Here, we show that for a large class of multiparameter models that takes the form of exponential families, a different intensive embedding- the isKL embedding, built on symmetrized Kullback Liebler divergence generates an explicitly and analytically tractable embedding in a Minkowski space of dimension equal to twice the number of parameters. In principle, this technique not only offers a great dimensionality reduction, it also allows one to uncover a hidden exponential family that describes an experiment or a simulation if the isKL embedding gives a cutoff after N+N dimensions. We will discuss the optimization of isKL embedding in producing a good visualization with several statistical models. |
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