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 X05: Machine Learning in Nonlinear Physics and Mechanics IFocus Live
|
Hide Abstracts |
Sponsoring Units: DSOFT GSNP DCOMP Chair: Christopher Rycroft, Harvard University; Shmuel Rubinstein, Harvard University Room: 05 |
Friday, March 19, 2021 8:00AM - 8:36AM Live |
X05.00001: Calculating the entropy of physical systems with Machine Learning Invited Speaker: Yohai Bar-Sinai Characterizing the entropy of a system is a crucial, and often computationally costly, step in understanding its thermodynamics. It plays a key role in the study of phase transitions, pattern formation, protein folding and more. We present a novel method, termed MICE (Machine-learning Iterative Calculation of Entropy) for calculating the entropy by iteratively dividing the system into smaller subsystems and estimating the mutual information between each pair of halves. The estimation is performed with a recently proposed machine learning algorithm which works with arbitrary network architectures that can fit the structure and symmetries of the system at hand. We show that our method can calculate the entropy of various systems, both thermal and athermal, with state-of-the-art accuracy. Specifically, we study various classical spin systems, and identify the jamming point of a bidisperse mixture of soft spheres. Lastly, we suggest that besides its role in estimating the entropy, the mutual information itself can provide an insightful diagnostic tool in the study of physical systems. |
Friday, March 19, 2021 8:36AM - 8:48AM Live |
X05.00002: Predicting Erosion Channel First Passage with Machine Learning Isaac Khor, Li Han, Arshad Kudrolli We investigate statistical and machine learning approaches to predicting erosive headward growth in rivers with bimodal source distributions. Porosity maps generated by multiscale simulation methods calibrated with experiments [1] are used to generate large data sets amenable to statistical analysis. We focus on complementary methods to analyze the statistical features of the spatio-temporal porosity maps and perform image analysis with convolution neural networks (CNN). The accuracy of each method in predicting channel passage from a single sink point to two source points located across a rectangular domain as a function time and distance to source will be discussed. CNN methods are found to consistently provide the earliest prediction with greatest confidence. We will further discuss the robustness of the predictions depending on the degree of domain disorder, data size, and network architecture. |
Friday, March 19, 2021 8:48AM - 9:00AM Live |
X05.00003: Machine Learning Prediction of Avalanche-like Events in Knitted Fabric Adèle Douin, Frederic Lechenault, Jean-Philippe Bruneton Knitted fabric is a thread-based metamaterial that exhibits crackling noise when deformed. Discrete events of broadly-distributed sizes are displayed both in the force signal and in the deformation field. The occurrence of those intermittent, scale-invariant events cannot be accurately predicted at the time being, as in most systems with similar avalanche-like behavior. However, Machine Learning methods have proven to be a useful tool when dealing with time-series predictions and image processing, giving hope to forecast fault failures in our system. We study the feasibility of predicting quantities such as next event amplitude or fault failure times by training a Neural Network algorithm to infer information from a series of past force signal and deformation field. Furthermore, since knits display spatially extended avalanche-like yielding events, we studied their statistical properties and compared them to other known systems to assess whether they can be considered an analogous seismic model. |
Friday, March 19, 2021 9:00AM - 9:12AM Live |
X05.00004: What makes a clog: characterizing 2D granular hopper flows using machine learning methods Jesse Hanlan, Douglas J Durian In hopper flow grains discharge at a constant rate, independent of fill height, and can form stable arches that clog the system and arrest flow. Thomas and Durian (PRL 2015) supported a Poissonian formulation of hopper flow by measuring the fraction of flow microstates, F, which cause a clog. New states are brought to the outlet with a constant sampling time, until the flow randomly finds a stable state which forms an arch. Koivisto and Durian (PRE 2017) then showed the random states being sampled depend on configuration, rather than momenta, by measuring similar scaling of F with outlet diameter between submerged and dry hoppers. We expand on this work by characterizing the individual microstates. We use a vertical, 2D hopper to image and track individual states, a new protocol to isolate unique configurations, and a novel machine learning analysis to leverage all the data, whether it is flowing or clogging. We explore the ability to predict whether a state will cause a clog to form based on only it’s image. |
Friday, March 19, 2021 9:12AM - 9:24AM Live |
X05.00005: Predicting Plasticity in 3D Model Glasses Using the Local Yield Stress Method Dihui Ruan, Sylvain Patinet, Michael Falk The Local Yield Stress (LYS) method was developed to probe local regions and quantify their susceptibilities to plasticity by measuring the incremental stress required to trigger a local rearrangement (). This method identifies shear transformation zones in amorphous materials, which play a role analogous to dislocations in crystalline materials. We apply the LYS method to 3D Kob-Anderson glasses each with one million atoms in a ~ simulation box. We limit ourselves to a single probing perfectly aligned with the loading on the boundary. Correlations between the local yield stresses and the plastic events are observed to persist up to ~1/3 of the yielding strain at the optimal probing region size, a radius of (600-700 atoms). From a preliminary sampling on the local yield surface at the location of the 1st plastic event, we find that the local yield surface is highly anisotropic, and the most susceptible local rearrangement doesn’t necessary align perfectly with the applied loading on boundary. This implies that the correlations may be improved if a complete local yield surface were to be sampled at each probing site. |
Friday, March 19, 2021 9:24AM - 9:36AM Live |
X05.00006: Predicting nonlinear stochastic and quantum dynamics without PDEs Alasdair Hastewell, Jorn Dunkel Physics aims to make quantitative predictions given information about initial and boundary conditions. Predicting the behavior of quantum and nonlinear stochastic systems often involves formulating and solving linear operator equations that have a characteristic eigenbasis. Once this basis has been identified, prediction reduces to a straightforward series evaluation. In many practically relevant situations, however, the natural bases cannot be calculated analytically. Here, we introduce a method that skips the model formulation step of the traditional prediction pathway by directly discovering the fundamental bases from data through an interpretable matrix factorization. Given the basis, predictions can be made by computing a simple combination of matrix products, providing a fast forecasting procedure when no explicit dynamical equation is known. |
Friday, March 19, 2021 9:36AM - 9:48AM Live |
X05.00007: Soft Matter Physics for Machine Learning: Dynamical loss functions Miguel Ruiz Garcia, Ge Zhang, Sam Schoenholz, Andrea Liu The neural network architectures, loss functions or optimizing protocols that are used in deep learning often stem from laborious trial-and-error design. This has triggered great interest in improving theoretical understanding of the connection between the structure of the loss function landscape and the performance of the optimizing protocol, or algorithm. So far, most effort has focused on improving the algorithm (e.g. stochastic gradient descent). We take a different approach by exploring new loss functions. In particular, we explore the effect of dynamical loss functions that change during training. Preliminary results show that this new approach can outperform results obtained with static loss functions for particular cases. We use the Hessian and Neural Tangent Kernel spectrums to understand how topographical changes of the loss function landscape can improve learning. |
Friday, March 19, 2021 9:48AM - 10:00AM Live |
X05.00008: Large-scale visualization with machine learning of dislocation networks in colloidal single crystals Ilya Svetlizky, Seongsoo Kim, Seong Ho Pahng, Agnese Curatolo, Michael Brenner, David Weitz, Frans A Spaepen Understanding the formation and evolution of dislocation networks and their effect on the mechanical properties of solids is challenging, as it spans a vast hierarchy of length and time scales. Hard-sphere colloidal suspensions provide a unique model system to address these difficulties. |
Friday, March 19, 2021 10:00AM - 10:12AM Live |
X05.00009: Statistical properties of ridge networks in crumpled sheets Catalin Veghes CVeghes@clarku.edu, Li Han, Arshad Kudrolli We study a folding model of crumpled paper using a flat folding and d-fold algorithm introduced by Hofmann, et al. (2019). The progression of ridge length and angle distributions obtained from the model and complementary experiments are analyzed in terms of crumples generated with random flat-fold or d-cone seeded algorithms. The ridge length distribution and scaling of number of vertices and edges of the domains enclosed is further compared with complementary sequential folding-unfolding algorithms which also induces creases. The number of edges of the domains is found to saturate with increasing number of folds in all cases. We then demonstrate that patterns generated by the two algorithms can be distinguished using convolution neural network (CNN) classification scheme. |
Friday, March 19, 2021 10:12AM - 10:24AM Live |
X05.00010: Machine Learning of Mechanisms in Combinatorial Metamaterials Ryan van Mastrigt, Corentin Coulais, Martin Van Hecke, Marjolein Dijkstra Combinatorial metamaterials are metamaterials designed by combining fundamental building blocks, unit cells, picked from a discrete set. This discretized design space allows us to explore the limitless structural complexity of metamaterials in a controlled manner. However, analytical and conventional numerical approaches have difficulty in efficiently navigating this large design space, which grows exponentially with system size. Here we employ machine learning techniques to explore combinatorial metamaterial design. We show that a trained convolutional neural network is able to classify never before seen configurations into those supporting system-wide periodic deformations in one dimension, line modes, and those who do not with an accuracy of 99.96%. This suggests that the network has correctly learned to identify the set of design rules for the presence of a line mode. To study the scalability of this long-correlation classification problem we establish the relation between network complexity and a set of simplified, imperfect order parameters. Our work provides insight into fundamental learning behaviour and application of neural networks with regard to complex discrete structure-property maps. |
Friday, March 19, 2021 10:24AM - 10:36AM On Demand |
X05.00011: Simplifying Physics Informed Neural Networks in case of periodicity to address low quality and sparse data while solving differential equations : an application in fluid dynamics. Gaétan Raynaud, Frederick P. Gosselin, Sébastien Houde To address most scientific problems, two approaches usually stand : mathematical modelling or experimentation. Physics Informed Neural Network [1] (PINN) is a recent numerical method that closes this gap using multi-layer perceptrons that approximate physical quantities. This allows resolution of ill-posed problems with a light formalism by optimizing residuals of differential equations and fitting provided data. However classical PINN can have difficulties to converge. To tackle that issue in the presence of periodic dynamics, we introduce ModalPINN where a truncated Fourier decomposition is enforced directly into the neural network’s structure. Application of this technique is given for a periodic flow using simulated experimental data. Gains up to 2 orders of magnitude in precision and robustness regarding added noise or delay are observed. |
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