Bulletin of the American Physical Society
APS March Meeting 2023
Volume 68, Number 3
Las Vegas, Nevada (March 5-10)
Virtual (March 20-22); Time Zone: Pacific Time
Session S53: Machine Learning |
Hide Abstracts |
Sponsoring Units: GDS DFD DMP Chair: Jennifer Hobbs, Zurich North America Room: Room 307 |
Thursday, March 9, 2023 8:00AM - 8:12AM |
S53.00001: Towards learning a Lattice Boltzmann collisional operator Alessandro Gabbana, Alessandro Corbetta, Vitaliy Gyrya, Daniel Livescu, Joost Prins, Federico Toschi In this work we explore the possibility of learning a custom collision operator, represented as a deep neural network, for the Lattice Boltzmann method by matching observable data. We present preliminary results in which a neural network is successfully trained as a surrogate of the single relaxation time BGK operator. |
Thursday, March 9, 2023 8:12AM - 8:24AM |
S53.00002: Learning closure models with neural operator-embedded differentiable CFD Varun Shankar, Venkat Viswanathan We discuss an approach using the differentiable physics paradigm that combines known physics with machine learning to develop closure models for the Navier-Stokes equations. Current ML turbulence modeling approaches often distinguish between offline a-priori training and online a-posteriori testing, resulting in additional generalization error that can be challenging to characterize. Advances in algorithmic differentiation and adjoint solvers are enabling a new class of models that embed neural networks into simulations, even during training, allowing the network to learn directly from the desired a-posteriori loss function. In parallel, neural operators that map between function spaces are gaining interest due to their discretization-invariant nature that allows for broad applicability without retraining. Thus, differentiable physics and neural operators form the ideal pairing to learn unknown closures in fluid modeling. We test our approach on a variety of fluid datasets and quantify error across a range of generalization parameters. We find that constraining models with inductive biases in the form of PDEs that contain known physics or existing closure approaches produces highly data-efficient, accurate, and generalizable models, outperforming state-of-the-art baselines. Addition of structure in the form of physics information also brings a level of interpretability to the models, potentially offering a stepping stone to the future of closure modeling. |
Thursday, March 9, 2023 8:24AM - 8:36AM |
S53.00003: A data-free partial differential equation (PDE) solver in the framework of physics-informed neural networks (PINN) Xiaoyu Tang, Boqian Yan Physics-informed neural networks (PINN) have been proposed to solve partial differential equation (PDE) given laws of physics and sparse training data. With PINN, the training data, laws of physics, boundary condition (BC) and initial condition (IC) are treated as parts of the total loss function, which is optimized. The derivative terms in PDEs are modeled by the auto-differential (AD) technique for easy implementation, and the BCs and ICs are imposed in a soft manner by optimization. Provided sparse solutions, the neural network is trained. However, the derivative term modeled by AD technique has lower accuracy than modeled by finite difference (FD) scheme, which will be demonstrated in this talk. The BCs and ICs could not be satisfied exactly due to the nature of the optimization of the total loss function, which affects the accuracy of the solution. In many cases, the training data are hard to obtain. In this talk, combining the advantages of finite difference method (FDM) and PINN, a new data-free PDE solver called PINN-FDM is introduced. In FDM-PINN, the derivative terms are modeled by FD instead of AD, and BCs and ICs are imposed exactly. FDM-PINN could solve the complicated PDE without training data, and the accuracy of the solution could be improved significantly. |
Thursday, March 9, 2023 8:36AM - 8:48AM |
S53.00004: Defending smart electrical power grids against cyberattacks with deep Q-learning Mohammadamin Moradi, Ying-Cheng Lai, Yang Weng A key to ensuring the security of smart electrical power grids is to devise and deploy effective defense strategies against cyberattacks. To achieve this goal, an essential task is to simulate and understand the dynamical interplay between the attacker and defender, for which stochastic game theory and reinforcement learning stand out as a powerful mathematical/computational framework. Existing works were based on conventional Q-learning to find the critical sections of a power grid to choose an effective defense strategy, but the methodology was applicable to small systems only. Additional issues with Q-learning are the difficulty to consider the timings of cascading failures in the reward function and deterministic modeling of the game while the attack success depends on various parameters and typically has a stochastic nature. Our solution to overcoming these difficulties is to develop a deep Q-learning based stochastic zero-sum Nash strategy solution. We demonstrate the workings of our deep-Q learning solution using the benchmark W&W 6-bus and the IEEE 30-bus systems, the latter being a relatively large-scale power-grid system that defies the conventional Q-learning approach. Comparison with alternative reinforcement learning methods provides further support for the general applicability of our deep-Q learning framework in ensuring secure operation of modern power grid systems |
Thursday, March 9, 2023 8:48AM - 9:00AM |
S53.00005: Using transfer learning to generate samples for large systems of the spin-fermion Hamiltonian Georgios Stratis, Pau Closas, Adrian E Feiguin We are expanding on our previous work where we used neural networks to generate samples for the spin-fermion Hamiltonian. The main bottleneck of our previous work is generating the training data set requires time and memory resources that scale unfavorably as the sytem's size. We present a transfer learning approach in which we train neural networks on smaller systems and by appropriately modifying them we build models that can generate samples for far larger systems. |
Thursday, March 9, 2023 9:00AM - 9:12AM |
S53.00006: Learning Coronal Nonlinear Force-Free Magnetic Fields through Differentiable Rendering Phillip Lo, Eric Jonas We consider the ill-posed problem of computing the 3D magnetic field above the surface of the sun (the corona) from the vector magnetic field on the surface (the photosphere) and 2D optical projections of plasma flowing through magnetic field lines in the corona. We approximate the coronal magnetic field using the commonly-used nonlinear force free field (NLFFF) model. Traditional iterative numerical PDE methods have struggled with this problem, leading us to take a deep learning approach. Using a parameterized NLFFF approximation by Aschwanden (2012), we build a differentiable renderer that is able to synthesize vector magnetograms and images of magnetic field lines. Since we wish to be able to compute the coronal magnetic field from real solar data (for which there is no ground truth), we take an unsupervised learning approach by using our differentiable renderer in an autoencoder designed to learn the underlying NLFFF parameters. In particular, our loss function operates only on the rendered magnetograms and field line images; our training routine does not access the ground truth NLFFF parameters. We believe that the coupling of simple physical models with differentiable rendering provides a valuable and novel way of solving ill-posed inverse problems. We will show prediction results of our autoencoder on simulated data as well as progress on real solar data measured by the Solar Dynamics Observatory (SDO). |
Thursday, March 9, 2023 9:12AM - 9:24AM |
S53.00007: Multiscale Perturbed Gradient Descent: Chaotic Regularization and Heavy-Tailed Limits Soon Hoe Lim Recent studies have shown that gradient descent (GD) can achieve improved generalization when its dynamics exhibits a chaotic behavior. However, to obtain the desired effect, the step-size should be chosen sufficiently large, a task which is problem dependent and can be difficult in practice. In this talk, we introduce multiscale perturbed GD (MPGD), a novel optimization framework where the GD recursion is augmented with chaotic perturbations that evolve via an independent dynamical system. We analyze MPGD from three different angles: (i) By building up on recent advances in rough paths theory, we show that, under appropriate assumptions, as the step-size decreases, the MPGD recursion converges weakly to a stochastic differential equation (SDE) driven by a heavy-tailed Lévy-stable process. (ii) By making connections to recently developed generalization bounds for heavy-tailed processes, we derive a generalization bound for the limiting SDE and relate the worst-case generalization error over the trajectories of the process to the parameters of MPGD. (iii) We analyze the implicit regularization effect brought by the dynamical regularization and show that, in the weak perturbation regime, MPGD introduces terms that penalize the Hessian of the loss function. Empirical results are provided to demonstrate the advantages of MPGD. |
Thursday, March 9, 2023 9:24AM - 9:36AM |
S53.00008: Data-driven discovery and interpolation of Green's functions Harshwardhan Praveen, Nicolas Boulle, Christopher J Earls To gain a deeper understanding of nature, we present a data-driven approach to mathematically model unknown physical systems, by learning a Green's function for its hidden, governing partial differential equations. The systems considered are observed as input-output pairs, by collecting physical responses under excitations drawn from a Gaussian process. Two methods are offered to learn the Green's function: 1) using the proper orthogonal decomposition modes of the system as a surrogate for the empirical eigenvectors of the Green's function and fit the eigenvalues using the data; and 2) using a generalization of randomized singular value decomposition to construct a low-rank approximation to the Green's function. These are demonstrated 1D examples: Poisson, Helmholtz, Airy, and multi-physics contexts. We also present a 2D demonstration, for the Poisson problem. Additionally, we propose a way to interpolate between Green's functions learned for different modeling contexts, by performing principled interpolation on a manifold. The interpolation is demonstrated on Airy's problem in 1D and Helmholtz problem in 2D. |
Thursday, March 9, 2023 9:36AM - 9:48AM |
S53.00009: Toward automated design of optimized high energy density material science experiments on the Z Machine Andrew J Porwitzky, Justin L Brown, William E Lewis High energy density dynamic compression materials science experiments in the Mbar regime can only be performed on a limited number of experimental facilities and thus often consist of one-off experimental designs. Due to the bespoke nature of such experiments, design is done via an expert process that is guided by high fidelity multiphysics computation. In a typical dynamic materials properties experiment, a tailored current pulse shape is generated by the Z Machine that quasi-isentropically compresses a sample material to high energy density states. Minor fluctuations in the firing time of laser-triggered gas switches (LTGS) – used to create the custom pulse shape – can result in undesirable shocked compression of samples. A computational framework has been developed that analyses the statistical spread in LTGS timings to calculate the probability of experiment failure. Recent additions to this framework have opened the possibility of automated experiment design that meets designated target metrics that has never before been achieved at the Z Facility. Exciting possibilities lay before us to optimize not only to intuitive metrics (sample input pressure, shockless compression probability, sample dimensions, etc.) but to non-intuitive metrics such as minimizing possible damage to the vacuum/insulator stack. Progress towards fully automated design of a candidate experiment will be presented to the data science community for broader engagement. |
Thursday, March 9, 2023 9:48AM - 10:00AM |
S53.00010: Using deep-learning to uncover physics of magnetic (charged particle) confinement in Magnetized Liner Inertial Fusion William E Lewis, Owen M Mannion, Christopher A Jennings, Daniel E Ruiz, Patrick F Knapp, Matthew R Gomez, Adam J Harvey-Thompson, Stephen A Slutz, Kristian Beckwith, Kristian Beckwith Magnetized Liner Inertial Fusion (MagLIF) is a magneto-inertial fusion (MIF) concept studied on the Z-machine at Sandia National Laboratories. In MagLIF an axially premagnetized and laser preheated gaseous deuterium (DD or DT) fuel contained in a cylindrical beryllium tube or liner undergoes quasi-adiabatic heating and flux compression to achieve fusion relevant conditions. The magnetic field-radius product (BR) near bang time determines the extent of confinement of charged fusion products and is of fundamental interest in understanding MagLIF performance. We built an artificial neural network surrogate trained on expensive physics calculations of magnetized fast charged-particle transport and associated secondary neutron emission in MIF plasmas used to diagnose BR. This enables Bayesian inference of BR for a series of MagLIF experiments that systematically vary inputs including laser preheat energy deposited, gas fill density, and target dimensions. We demonstrate flux loss consistent with Nernst advection of magnetic field out of the hot fuel and diffusion into the cold target wall under these changes to experimental conditions. |
Thursday, March 9, 2023 10:00AM - 10:12AM |
S53.00011: MADEM: Energy-efficient training of Deep neural networks using Memristor arrays Suin Yi, Suhas Kumar Energy-efficient and biologically plausible machine learning to train artificial deep neural networks is presented using hardware-algorithm co-optimization. Despite the success of backpropagation that computes the error gradient very efficiently through just two propagations (forward and backward) thanks to chain-rule, backpropagation has several shortcomings such as biological implausibility (e.g., nonlocality and symmetric synaptic weights). Also, the backpropagation-based gradient computation necessitates high precision calculations of neuron activities at least with 16-bit precision synaptic weights that necessitates high precision digital computing and energy-consuming. By using a novel learning algorithm compatible with analog in-memory-computing provided by memristor array, we demonstrate efficient training of deep neural networks, which is not only biologically plausible (local update rule), but also energy-efficient (5 orders of magnitude smaller), and faster (36 smaller latency). |
Thursday, March 9, 2023 10:12AM - 10:24AM |
S53.00012: Automatic detection of fake tweets about Covid-19 Vaccine in Portuguese Rafael Geurgas Zavarizz, Leandro R Tessler The Covid-19 pandemic induced an unprecedented wave of disinformation in social networks. This had dire consequences for society. The situation was particularly serious in Brazil due to official support for unproven treatments and denial of vaccine effectiveness. Political polarization helped to create an explosion of false tweets in Portuguese, which constitutes a threat to public health. An algorithm that could discriminate between true and false messages would be a very important tool to reduce or even stop the wave of disinformation. We developed BERTVacPort, an approach to automatically and reliably label tweets about vaccines in Portuguese as reliable or fake. The architecture relies upon a pre-trained Portuguese BERT-like transformer base Neural Network with two extra fully-connected layers. To train the implementation we collected almost 3 million tweets containing the word vacina, vaccine in Portuguese, over a 7 month period. We classified a fraction of the corpus (16,731 tweets) and used it to fine-tune the algorithm. The best results were achieved when retrained the last seven layers of the BERT-like network and the two additional layers. We obtained 74% f1-score and 74% accuracy. Considering the heterogeneity of the database and the ambiguity of many messages limited to 280 characters, the results are probably as good as possible and outperform a human reader. |
Thursday, March 9, 2023 10:24AM - 10:36AM |
S53.00013: Transferable Coarse Grained Free Energy Models Enabled by Active Learning Blake R Duschatko, Jonathan P Vandermause, Nicola Molinari, Boris Kozinsky Coarse graining techniques play an essential role in accelerating molecular simulations of systems with large length and time scales. Theoretically grounded bottom-up models are appealing due to their thermodynamic consistency with the underlying all-atom models. In this direction, machine learning approaches hold great promise to fitting complex many-body data. However, training models may require collection of large amounts of expensive data. Moreover, quantifying trained model accuracy is challenging, especially in cases of non-trivial free energy configurations, where training data may be sparse. We demonstrate a path towards uncertainty-aware models of coarse grained free energy surfaces. Specifically, we show that principled Bayesian model uncertainty allows for efficient data collection through an on-the-fly active learning framework and open the possibility of adaptive transfer of models across different chemical systems. Uncertainties also characterize models' accuracy of free energy predictions, even when training is performed only on forces. This work helps pave the way towards efficient autonomous training of reliable and uncertainty aware many-body machine learned coarse grain 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