Bulletin of the American Physical Society
APS March Meeting 2018
Volume 63, Number 1
Monday–Friday, March 5–9, 2018; Los Angeles, California
Session F54: Machine Learning in Nonlinear Physics and MechanicsFocus Session
|
Hide Abstracts |
Sponsoring Units: GSOFT GSNP Chair: Yohai Bar-Sinai, Harvard Univ Room: LACC 514 |
Tuesday, March 6, 2018 11:15AM - 11:51AM |
F54.00001: A unified perspective on disorder in atomic systems: machine learning material properties and design Invited Speaker: Ekin Cubuk With its ability to leverage rapidly increasing amounts of data and computational resources, machine learning (ML) has the potential to be an indispensable tool in condensed matter physics and materials science. Our recent work is an application of ML to the study of the physics of metastable phases and their dynamics. We start by addressing the perennial question of whether disordered solids have localized defects. Using ML, we develop a structural model that is significantly more predictive than previous attempts, while still being interpretable and generalizable. Next, we embed the resulting ML representation in theoretical models of several phenomena in disordered solids, e.g. fragility, fragile-to-strong transition, out-of-equilibrium dynamics, aging, glassy thin film dynamics, mechanical response, and grain boundaries in polycrystals. This approach leads to a unified perspective on disordered particle arrangements, from atoms to macroscopic grains spanning seven orders of magnitude in particle size. In addition, I will present the application of similar approaches to the design of new materials and phases with desired properties. Finally, I will discuss concerns with the application of ML to atomic systems, particularly with regards to accuracy, safety and generalization. |
Tuesday, March 6, 2018 11:51AM - 12:03PM |
F54.00002: Predicting the dynamics of crumpling with machine learning Christopher Rycroft, Jordan Hoffmann, Jovana Andrejevic, Lisa Lee, Shmuel Rubinstein The simple process of crumpling a sheet of paper with our hands results in a complex network of interconnected permanent creases of many sizes and orientations. On subsequent crumples, the sheet preferentially bends along these creases, introducing history dependence to the process. Here, we study the dynamics of crumpling using machine learning methods to analyze the complex crease networks. We perform experiments to systematically apply successive crumples to a sheet, from which we extract the crease networks using high resolution scans of mean curvature. We use the scans as input to machine learning methods to ask questions such as predicting hot spots for further crease generation. We discuss broader questions about how best to design experiments to take advantage of machine learning approaches. |
Tuesday, March 6, 2018 12:03PM - 12:15PM |
F54.00003: Detection and Characterization Techniques for Signatures of Crumpling History Jovana Andrejevic, Jordan Hoffmann, Lisa Lee, Shmuel Rubinstein, Christopher Rycroft When a sheet of paper is crumpled, unfolded, and re-crumpled, the total distance etched by creases grows in a predictable manner. Experiments of repeated crumpling of a thin elastoplastic sheet remarkably demonstrate that this measure is not history dependent. However, crease networks traversing equal distances but obtained by distinct crumpling protocols exhibit structural differences. How can we identify signatures of the crumpling history from the global statistics of such networks? We begin with a creative repurposing of the radon transform, an integral transform used commonly for data reconstruction in tomography, as a noise reduction tool to detect creases from scans of crumpled sheets. In addition to recovering clean contours of the crease network, this technique equips us with a measure of crease directionality. We discuss how this insight next informs a feature-based machine learning approach to evaluate possible indicators of the crumpling history. |
Tuesday, March 6, 2018 12:15PM - 12:27PM |
F54.00004: Using Machine Learning to Understand the Evolution of Damage Networks in Thin Sheets Lisa Lee, Jovana Andrejevic, Jordan Hoffmann, Christopher Rycroft, Shmuel Rubinstein The evolution of creases in a repeatedly crumpled thin sheet is incredibly complex. Nevertheless, some characteristics of this system have recently been shown to be very simple, such as the ability of total crease length to describe the evolution of damage via the equation of state. Yet, the equation of state cannot predict important properties such as when and where new creases will appear. Machine learning may be a good candidate for uncovering such hidden spatial-temporal patterns. However, systematic analysis of complex systems requires the appropriate alliance of inherently data-limited lab experiments with the big-data nature of machine learning. Thus, I will discuss aspects of how the inclusion of machine learning into the data analysis modifies how the experiment must be set up and how to develop new strategies for acquisition of data. |
Tuesday, March 6, 2018 12:27PM - 12:39PM |
F54.00005: Identifying Structural Defects in Disordered Packings of Elongated Particles using Machine Learning Matt Harrington, Andrea Liu, Douglas Durian Structural defects within a solid determine where plastic material failure and local rearrangements occur under a driving force. While dislocations and vacancies can be readily identified in a crystalline solid, similar defects can be difficult to find in a disordered solid or packing based solely on structural information. We present an experimental study of a model driven disordered solid--a dry two-dimensional granular pillar under compression--comprised of either circular or elongated particles. A machine learning approach that incorporates a multitude of spherical structure functions can classify circular particles as rearranging or non-rearranging using a linear support vector machine. When constituent particles are asymmetric, spherical structure functions are no longer sufficient; however, our machine learning approach can be modified to yield similar predictive performance for both grain shapes. This method also generates a scalar field, 'softness,' that quantifies how likely each particle is about to undergo a rearrangement. We discuss spatial and temporal characteristics of the rearrangement and softness fields. |
Tuesday, March 6, 2018 12:39PM - 12:51PM |
F54.00006: Reservoir computer predictions for the Three Meter magnetic field time evolution Artur Perevalov, Ruben Rojas Garcia, Itamar Shani, Brian Hunt, Daniel Lathrop The source of the Earth's magnetic field is the turbulent flow of liquid metal in the outer core. Our experiment's goal is to create Earth-like dynamo, to explore the mechanisms and to understand the dynamics of the magnetic and velocity fields. Since it is a complicated system, predictions of the magnetic field is a challenging problem. We present results of mimicking the three Meter experiment by a reservoir computer deep learning algorithm. The experiment is a three-meter diameter outer sphere and a one-meter diameter inner sphere with the gap filled with liquid sodium. The spheres can rotate up to 4 and 14 Hz respectively, giving a Reynolds number up to 1.5*10^8. Two external electromagnets apply magnetic fields, while an array of 31 external and 2 internal Hall sensors measure the resulting induced fields. We use this magnetic probe data to train a reservoir computer to predict the 3M time evolution and mimic waves in the experiment. Surprisingly accurate predictions can be made for several magnetic dipole time scales. This shows that such a complicated MHD system’s behavior can be predicted. |
Tuesday, March 6, 2018 12:51PM - 1:03PM |
F54.00007: Using image Super-Resolution techniques as a coarse-graining method for physical systems Yohai Bar-Sinai, Michael Brenner, Pascal Getreuer, Jason Hickey, Stephan Hoyer, Peyman Milanfar One of the most generic probelms in theoretical physics is that of coarse graining - how to represent the behavior of a physical theory at long wave lengths and slow frequencies, by "integrating out" the irrelevant degrees of freedom which change rapidly in time and space. This is usually done by writing effective long-wavelength governing equations which take into account the small scale physics (a few examples: renormalization approaches, eddy viscosity, geometrical optics, the lubrication approximation and Taylor dispersion). The problem of single image Super-Resolution -- upscaling a given image to sub-pixel resolution -- is of similar nature, but in the opposite direction: the goal is to infer the short-wavelength behaviour from the long-wavelength data. Recently, machine learning approaches has been used to solve the upscaling problem in a data-driven manner, by learning the coarse-to-fine mapping from natural images (e.g. RAISR). Inspired by these techniques, we use similar data-driven approaches to write effective coarse-grained equations to dynamic PDEs, extracting effective equations and physical constants. Possible applications of these results will be discussed. |
Tuesday, March 6, 2018 1:03PM - 1:15PM |
F54.00008: Predicting Emergent Crystalline Structural Order from Building Block Geometry Yina Geng, Greg Van Anders, Sharon Glotzer Quantitatively determining how building block attributes drive materials systems to form ordered target crystals is a fundamental challenge. Addressing this challenge is particularly difficult for systems that exhibit emergent order. Here, we combine inverse design with machine learning to construct a model that correctly classifies the emergent, entropy-driven crystallization of more than ten thousand convex polyhedral shapes into a small number of structures with an accuracy of greater than 90% using only two parameters. Our results demonstrate that the emergent, self-assembly of entropic crystals is controlled by a remarkably small number of parameters, and provides a quantitative model for predicting the expected behavior of colloidal self-assembly experiments. |
Tuesday, March 6, 2018 1:15PM - 1:27PM |
F54.00009: Intelligent, autonomous parameter space exploration of self-assembly simulations Matthew Spellings, Julia Dshemuchadse, Sharon Glotzer Researchers studying self-assembly are often plagued with a problem: the systems they study exhibit emergent behavior that could result from any combination of the many independent parameters they can tune. The typical solution to this problem is to restrict a study to the few most interesting variables and perform a screening experiment on a grid in this space. But what can be done when it is not clear which variables are most important? The scale of these studies quickly gets out of hand as analysis, visualization, and even selecting new experiments are multiplied by the dimensionality of the parameter space. Here we discuss approaches to incorporate machine learning methods into the experimental design and analysis loop of exploratory self-assembly simulations in order to optimize computational time spent simulating interesting and novel behaviors. By harnessing these methods, we can begin to probe the behavior of even more complex design spaces. |
Tuesday, March 6, 2018 1:27PM - 1:39PM |
F54.00010: Distilling the logic of behavioral dynamics using automated inference Bryan Daniels, William Ryu, Ilya Nemenman Biological systems that produce stereotyped, reproducible dynamics are often still difficult to model because they are controlled by a large number of unknown heterogeneous interactions. Phenomenological coarse-grained fits can be useful as descriptors, but they are often unprincipled or not interpretable as dynamical systems. In contrast, recent innovations in statistical inference allow for the principled discovery of dynamical systems that reproduce given time series data, even when details about the underlying interaction structure are unknown. This allows for the prediction of responses to unseen dynamical stimuli, and more importantly provides a window into the phase space structure that defines the system's coarse-grained logic. We demonstrate this approach using data from the stereotyped movement of C. elegans in response to a heat stimulus. The resulting dynamical models predict the existence of distinct behavioral states that are not directly observed in the time series data. |
Tuesday, March 6, 2018 1:39PM - 1:51PM |
F54.00011: Computational tools for data-driven design of soft robots Mohammad Khalid Jawed, Xiaonan Huang, Amarbold Batzorig, Carmel Majidi We report a numerical simulation framework for the mechanics of soft robots comprised of slender structures, for application in data-driven structural design, material selection, and adaptive control. Owing to prohibitive computational cost, soft robots are often designed solely based on empirical laws through a tedious trial-and-error process with no quantitative guideline. Inspired by fast and efficient modeling of hair and clothes in the animation industry, we adapt algorithms for physically-based simulations from computer graphics. We extend the Discrete Elastic Rods method to develop a simulator for a wide class of robots comprised of multiple slender structures (rods or shells) with compliant joints. In parallel with computation, we perform experiments with biomimetic robots composed of soft thermal actuators and confirm the validity of the simulation tool. Given the large number of parameters and high degree of nonlinearity associated with the performance and functionality of soft robots, emergent machine learning techniques offer a promising avenue for their computational design and optimization. The robustness and speed of our simulation can enable data-driven analysis of a broad range of smart programmable structures beyond soft robots. |
Tuesday, March 6, 2018 1:51PM - 2:03PM |
F54.00012: Deep Learning Physical Phenomena Joseph Gomes, Amir Barati Farimani, Vijay Pande Transport phenomena studies the exchange of energy, mass, momentum, and charge between systems,\cite{bird2007transport} encompassing fields as diverse as continuum mechanics and thermodynamics, and is used heavily throughout all engineering disciplines. Here, we show that modern deep learning models, such as generative adversarial networks, can be used for rapid simulation of transport phenomena without knowledge of the underlying constitutive equations, developing generative inference based models for steady state heat conduction and incompressible fluid flow problems with arbitrary geometric domains and boundary conditions. In contrast to conventional procedure, the deep learning models learn to generate realistic solutions in a data-driven approach and achieve state-of-the-art computational performance, while retaining high accuracy. Deep learning models for physical inference can be applied to any phenomena, given observed or simulated data, and can be used to learn and predict directly from experiments where the underlying physical model is complicated or unknown. |
Tuesday, March 6, 2018 2:03PM - 2:15PM |
F54.00013: Visualizing theory space: Isometric embedding of probabilistic predictions, from the Ising model to the cosmic microwave background Katherine Quinn, Francesco De Bernardis, Michael Niemack, James Sethna We develop an intensive embedding for visualizing the space of all predictions for probabalistic models, using replica theory. Our embedding is isometric (preserves the distinguishability between models) and faithful (yields low-dimensional visualizations of models with simple emergent behavior). We apply our intensive embedding to the Ising model of statistical mechanics and the ΛCDM model applied to cosmic microwave background radiation. It provides an intuitive, quantitative visualization applicable to renormalization-group calculations and optimal experimental design. |
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. |
© 2025 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