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 L63: Machine Learning and Data in Polymer Physics IIFocus Live
|
Hide Abstracts |
Sponsoring Units: DPOLY DCOMP DBIO GSNP Chair: Jonathan Whitmer, University of Notre Dame |
Wednesday, March 17, 2021 8:00AM - 8:12AM Live |
L63.00001: Machine Learning of Phase Transitions and Dynamical Crossovers in Polymers Tarak Patra, Debjyoti Bhattacharya, Ashwin Bale Phase transitions and dynamical crossovers in polymers are governed by correlated microscopic interactions and relaxation of their large number of atoms and segments over multiple time and length scales. Given the wide variation in microscopic degrees of freedom and macroscopic properties exhibited by polymers, identifying new types of phase transitions and corresponding states can be challenging. Moreover, there is no universal order parameter or straight forward approach to characterize wide range of crossovers in polymers including metastable phase transitions, vitrification, jamming, gel formation and localization transition. Here, we report a generic machine learning framework for autonomous identification and characterization of phase transitions and dynamical crossovers in molecular dynamics trajectory of polymers. We demonstrate this framework for coil to globule transition, crystallization and glass formation during cooling of polymers, and provide new physical insights of these processes. This framework does not need any a-priory knowledge of the crossover and is extensible to predict other phase transitions and dynamical crossover during thermophysical processes such as cooling, drying, and compression of polymers. |
Wednesday, March 17, 2021 8:12AM - 8:24AM Live |
L63.00002: Rheology-Informed Neural Networks (RhINNs) for direct and inverse complex fluid modeling Mohammadamin Mahmoudabadbozchelou, Safa Jamali We present Rheology-Informed Neural Networks (RhINNs) architectures as alternative platforms to solve systems of Ordinary Differential Equations (ODEs) commonly used in rheological constitutive modeling of complex fluids. The proposed RhINN is employed to solve the constitutive models with multiple ODEs by benefiting from automatic differentiation in neural networks. We present direct and inverse solutions of a Thixotropic ElastoViscoPlastic (TEVP) constitutive equation for a series of different flow protocols by employing our RhINNs methodology. From a practical perspective, commonly an exhaustive list of experimentations is required to accurately parameterize a TEVP model based on a specific fluid of interest. Only then, the model can be utilized in order to predict the fluid response to a different flow protocol. Alternatively, in our inverse problem, we let the RhINN framework learn the model parameters by training on a series of limited experimental data. We show that the model can be extended to various models by including different systems of ODEs, solved for arbitrary geometries, and recover complex kymographs of kinematic heterogeneities and transient shear banding of thixotropic fluids. |
Wednesday, March 17, 2021 8:24AM - 8:36AM Live |
L63.00003: Design of Polymers for Energy Storage Capacitors Using Machine Learning and Evolutionary Algorithms Joseph Kern, Lihua Chen, Chiho Kim, Rampi Ramprasad Many applications, such as electric vehicles and switched-mode power supplies, require capacitors that have high energy density, operating temperature, dielectric breakdown strength, and failure tolerance. Modern polymer film capacitors are useful due to their high failure tolerance; however, they suffer from low energy density per volume and low thermal stability. By utilizing a genetic algorithm approach, we have designed hypothetical polymers with bandgaps above 5 eV, glass transition temperatures above 500 K, and dielectric constants above 4 at 100 Hz. These are useful properties, as a high bandgap can be used as a proxy for dielectric breakdown field strength, a high glass transition temperature indicates the polymer can function uniformly from low to high temperatures, and a high dielectric constant improves energy density per volume. Over 10,000 hypothetical polymers were designed, which have been further down selected (and recommended for synthesis) based on synthesis feasibility considerations. |
Wednesday, March 17, 2021 8:36AM - 8:48AM Live |
L63.00004: Phase diagrams of polymer-containing liquid mixtures with a theory-embedded
neural network Issei Nakamura We develop deep neural networks (DNNs) that consider the phase separation of polymeric liquids. In this talk, we discuss our new built-in function that is constructed through coarse-grained mean-field theory and the scaling laws in polymer physics. This characteristic theory-embedded layer enables us to perform the learning process efficiently with relatively small numbers of artificial neurons and hidden layers and provides the DNNs with reasonable predictive power. To demonstrate the efficacy of our DNNs, we will discuss the phase diagrams of polymer solutions, and the salt-free and salt-doped diblock copolymer melts. Moreover, we will show the predictive power of the DNNs by considering some experiments for the lithium salt-doped diblock copolymers such as PEO-b-PS. |
Wednesday, March 17, 2021 8:48AM - 9:00AM Live |
L63.00005: Neural Network Prediction of Polymer-Solvent Coexistence Curves Jeffrey Ethier, Rohan Casukhela, Josh Latimer, Matthew Jacobsen, Richard Arthur Vaia Solution phase behavior of polymers is fundamental to synthesis, processing, purification, and self-assembly. While there have been numerous attempts at predicting phase behavior from Flory-Huggins theory and equation of state methods, no such universal theory exists, and often additional fitting parameters are needed. Taking an orthogonal approach, we establish the feasibility of using experimental data and machine learning algorithms (neural networks, Gaussian Process Regression, and others) to predict polymer phase behavior. Focusing on available data (>2500 cloud points) on polystyrene in various solvents, we examine the precision of upper and lower critical solution co-existence predictions with various feature descriptors. We show that these models can predict cloud point temperatures for unknown concentrations, molecular weights, and solvents within experimental error. Furthermore, these models can be used to estimate unknown polymer-solvent properties, such as Chi, or Hansen solubility parameters from molecular “fingerprints”. This methodology demonstrates the potential to establish a community database, which can scale with user input (or automated data collection) and integrate prior knowledge to provide phase behavior estimates, as well as test theoretical concepts. |
Wednesday, March 17, 2021 9:00AM - 9:12AM Live |
L63.00006: Prediction of Block Copolymer Phase Behavior Using Machine Learning Nathan Rebello, Akash Arora, Tzyy-Shyang Lin, Sarah Av-Ron, Bradley Olsen Self-consistent field theory (SCFT) provides valuable insights into the driving forces behind microphase separation in block copolymers. While it is qualitatively highly accurate, it suffers quantitative limitations: increased error near the order-disorder transition, inability to capture strong asymmetry in experimental phase diagrams, and reliance on the Flory-Huggins interaction parameter to quantify block incompatibility with complex functional dependence on temperature, chemistry, and composition. Given the challenges in physics-based methods, we develop a purely data-driven model to predict phase behavior for neat diblock copolymers and compare with SCFT. First, we collect over 5000 experimental phase measurements from literature, recording the chemistry, molar mass, volume fraction, and temperature for each phase in a database. Then, we train a random forest classification model to predict the disordered, lamellar, cylindrical, gyroid, and spherical phases. This model, when adjusted for uncertainty in molecular parameters, is significantly more accurate than SCFT predictions with the Flory-Huggins interaction parameter calculated from group contribution theory. This work demonstrates the value that machine learning brings to soft matter materials design. |
Wednesday, March 17, 2021 9:12AM - 9:24AM Live |
L63.00007: Deep Learning and Self-Consistent Field Theory: A Path Towards Accelerating Polymer Phase Discovery Yao Xuan, Kris T Delaney, Hector D. Ceniceros, Glenn H Fredrickson A new framework that leverages data obtained from self-consistent field theory (SCFT) simulations with deep learning to accelerate the exploration of parameter space for block copolymers is presented. Deep neural networks are adapted and trained in Sobolev space to better capture the saddle point nature of the SCFT approximation. The proposed approach consists of two main problems: 1) the learning of the effective Hamiltonian as a function of the average monomer density fields and the relevant physical parameters and 2) the prediction of saddle density fields given the polymer parameters. There is an additional challenge: the effective Hamiltonian has to be invariant under shifts (and rotations in 2D and 3D). A data-enhancing approach and an appropriate regularization are introduced to effectively achieve said invariance. In this first study, the focus is on one-dimensional (in physical space) systems to allow for a thorough exploration and development of the proposed methodology. |
Wednesday, March 17, 2021 9:24AM - 9:36AM Live |
L63.00008: Thermal conductivity, heat capacity and speed of sound of epoxy resins guangxin lyu, Christopher Evans, David G. Cahill Epoxy resins with enhanced thermal conductivity are in great demand to improve the thermal management of electrical motors. However, the thermal conductivity of epoxy resin is typically low, comparable to 0.2 W/m-K, and a predictive understanding of the connection between molecular structure and thermal conductivity is not yet established. Moreover, epoxy resins are usually cured by epoxies and hardeners, so there are thousands of combinations. Machine learning is an effective way to predict thermal conductivity of different combinations if sufficient data is provided. We established one efficient way (less than 8 minutes for each sample) to measure thermal conductivity, heat capacity and speed of sound of epoxy resin by frequency domain probe beam deflection (FD-PBD) and time domain thermoreflectance (TDTR). 80 combinations of epoxy and hardener have been studied. Small molecular structure difference is found to have a large influence on thermal conductivity of epoxy resin. For example, 5-chloro-m-phenylenediamine has thermal conductivity of 0.27 W/m-K, which is ~100% higher than o-phenylenediamine after curing with resorcinol diglycidyl ether. The differences of the two molecules are link site of amine groups and additional chlorine on benzene ring. |
Wednesday, March 17, 2021 9:36AM - 9:48AM Live |
L63.00009: Using Machine Learning to Predict the Glass Transition Temperature of Polyimides Chengyuan Wen, Binghan Liu, Josh Wolfgang, Timothy Long, Roy Odle, Shengfeng Cheng To expedite the process of discovering new polyimides, an important high-temperature polymer, we apply a machine learning approach to predict their glass transition temperature (Tg), which controls their processability and possible temperature window of applications. We have collected the structure and Tg data for 225 polyimides. For each polyimide, 1342 features are generated using its composition-based SMILE notation. The 225 data points are separated into a training and a test set. A machine learning algorithm based on the LASSO regularization is applied to the training set to construct a predictive model of Tg. In this process, the training set is split further into a training and a test subset either randomly or with the training subset being statistically representative of the entire training set. The performance of the resulting predictive model of Tg is evaluated with the independent test data set that never enters the training process. The best predictive model is obtained if the training and test subsets are split statistically and the bagging approach is used to improve its stability. The model is further tested with results from molecular dynamics simulations of new polyimides yet to be synthesized and a good agreement is observed. |
Wednesday, March 17, 2021 9:48AM - 10:00AM Live |
L63.00010: BoltzmaNN: Predicting effective pair potentials and equations of state using neural networks Fabian Berressem, Arash Nikoubashman Neural networks (NNs) are employed to predict equations of state from a given isotropic pair potential using the virial expansion of the pressure. The NNs are trained with data from molecular dynamics simulations of monoatomic gases and liquids, sampled in the NVT ensemble at various densities. We find that the NNs provide much more accurate results compared to the analytic low-density limit estimate of the second virial coefficient. Further, we design and train NNs for computing (effective) pair potentials from radial pair distribution functions, g(r), a task which is often performed for inverse design and coarse-graining. Providing the NNs with additional information on the forces greatly improves the accuracy of the predictions, since more correlations are taken into account; the predicted potentials become smoother, are significantly closer to the target potentials, and are more transferable as a result. |
Wednesday, March 17, 2021 10:00AM - 10:12AM Live |
L63.00011: Data-driven tools to “fingerprint” soft material structuring in complex processing flows Patrick Corona, Barbara Berke, L. Gary Leal, Marianne Liebi, Matthew Helgeson Flows involving complex time-varying deformations are ubiquitous in polymer processing, and are important to engineering non-equilibrium structure in soft materials, yet design of such processes is challenged by the availability of accurate structure-based rheological models. The development of these models is biased toward the simple viscometric flows in which they are tested, limiting their applicability in complex processes. We introduce a new experimental methodology to “fingerprint” the microstructural response of complex fluids to nearly arbitrary flows and enable an alternative approach of data-driven modeling and design. The method involves scanning small angle x-ray scattering (sSAXS) in a fluidic four roll mill (FFoRM) device that can produce arbitrarily variable two-dimensional stagnation flows. Using measurements on rod-like polymers, we demonstrate how FFoRM-sSAXS can be used to generate thousands of Lagrangian trajectories that map structural response to the time history of deformation type and rate in a flow. We show how these large data sets can be used to understand the effects of flow history on material order, rigorously test physics-based constitutive models, emulate common processing flows, and directly synthesize process-structure-property relationships. |
Wednesday, March 17, 2021 10:12AM - 10:48AM Live |
L63.00012: Gaussian Processes and Deep Learning for Experimental Data Invited Speaker: Daniela Ushizima Several experimental disciplines depend upon exploring large and high-dimensional parameter spaces to search and find new scientific discoveries. For example, in materials science, the wide variety of pathways for synthesis, processing, and environmental conditions that influence material properties give rise to particularly vast parameter spaces. In order to improve efficiency during materials discovery, one of the main strategies is to increase automation of the exploration processes. Methods for autonomous experimentation have become more sophisticated recently, allowing for multi-dimensional parameter spaces to be explored and with minimal human intervention, thereby liberating the scientists to focus on interpretations and big-picture decisions. This presentation will showcase some of the advantages of using ML methods to search for materials configurations in large databases, e.g. using pattern recognition for polymeric films. We will also discuss algorithms for handling data from high-throughput experiments, such as Convolutional Neural Networks (CNN) and Gaussian process regression (GPR), with a focus on a DOE-funded software called “gpCAM” for autonomous data acquisition, which is based on GPR. This talk will give an introduction to the inner workings of the algorithms, how to use it, and will present a handful of examples. |
Wednesday, March 17, 2021 10:48AM - 11:00AM Live |
L63.00013: Meta-Reinforcement Learning as the Driver of Data Acquisition in Autonomous Polymer Discovery Sarath Swaminathan, Victoria Piunova, Krystelle Lionti, Chinyere Agunwa, Daniel Sanders, Dmitry Zubarev Discovery and development of polymer materials is driven by experimental data acquisition. Experiments unfold under conditions of delayed rewards on incredibly rich landscapes shaped by multiple experimental degrees of freedom, including continuous (concentration, temperature, radiation, time) and categorical (monomers, catalysts, initiators, solvents) [1,2]. Deep reinforcement learning (RL) emerges as an appealing approach with a capability to interact with lab equipment, handle delayed rewards, and find non-trivial research strategies under realistic constraints of discovery/development projects. We report development of an end-to-end RL approach applied to preparation of spin-on-glasses (SOGs). The primary focus of the talk is meta-learning strategies [3] that ensure generalizability of the RL agent performance, and associated task of data augmentation at the training stage. |
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