Symposium – Frontiers in Machine Learning for the Physical Sciences

THANK YOU FOR
REGISTERING

Please update your Zoom client to the latest version (5.3.2) to participate in the poster session breakout rooms

Please note: the browser-only version of Zoom does not support breakout rooms -- please install Zoom on your computer

Please Review these breakout room instructions since the feature is new to Zoom 5.3

Please direct any questions meant for plenary session speakers to the moderator named "Q&A" in chat

ZOOM LINK

[ This event occurred in the past ]

PLENARY SESSION 1
Please note all times below are PDT (U.S. Pacific Daylight Time)

8:30 AM

Pramod Khargonekar
Vice Chancellor for Research, UCI
Introductions

8:45 AM

Keynote Address
Stanley Osher
Innovations in Mean-Field Game Theory for Scalable Computation and Diverse Applications

Abstract

Mean field games play essential roles in AI, 5G communications, unmanned aerial vehicle path planning, social norms, and controlling natural disasters, such as COVID 19. In this talk, we present several results by our MURI team in the year 2019-2020. We designed fast and reliable numerical algorithms with connections to AI and machine learning, and formulated models in mean-field inverse problems, velocity control for massive rotary-wing UAV’s, controlling COVID 2019 pandemic spreading, etc. Several numerical examples and engineering experiments will be presented. Future directions will be discussed. This is based on a joint work with many people at UCLA, University of South Carolina (Wuchen Li who just moved to UofSC), University of Houston, and Princeton University.

9:30 AM

Kipton Barros
Automated discovery of a robust interatomic potential for aluminum

Abstract

Machine learning is emerging as a powerful tool for emulating electronic structure calculations. I will discuss recent work in building interatomic potentials relevant to chemistry, materials science, and biophysics applications. A key idea is active learning, in which the training data is iteratively collected to address weaknesses of the ML model. This approach can achieve a surprising level of transferability, as will be demonstrated with a case study for elemental aluminum.

10:00 AM

Kieron Burke
Machine learning for electronic structure calculations and a new approach to warm dense matter simulations

Abstract

Electronic structure calculations using density functional theory have become common in many areas of science and technology. I will give a brief overview of the impact of machine learning in electronic structure theory over the past decade [1]. I will highlight the work at UC Irvine to employ machine learning to find new and better density functional approximations [2,3]. Toward the end, I will also explain a new methodology for improving simulations of warm dense matter that are relevant to the National Labs [4].

References

[1] von Lilienfeld, O.A., Burke, K. Retrospective on a decade of machine learning for chemical discovery. Nat Commun 11, 4895 (2020).
[2] Bogojeski, M., Vogt-Maranto, L., Tuckerman, M.E. et al. Quantum chemical accuracy from density functional approximations via machine learning. Nat Commun 11, 5223 (2020).
[3] Kohn-Sham equations as regularizer: building prior knowledge into machine-learned physics L Li, S Hoyer, R Pederson, R Sun, E Cubuk, P Riley, K Burke, arXiv:2009.08551
[4] Bypassing the energy functional in density functional theory: Direct calculation of electronic energies from conditional probability densities R J. McCarty, D Perchak, R Pederson, R Evans, Y Qiu, S R. White and K Burke, 2007.01890

10:30 AM

Lighting Talks (Video Presentation)

[ 15 Minute Break ]

11:15 AM

Poster Session

Click Here to view a list of posters that will be presented

Poster Session by Topic and Breakout Room

Breakout Room 1

Extreme Physics, Fundamental Physics/Astronomy and Astrophysics

Irina Espejo - Excursion - Active Learning and Gaussian Processes for black box inference
Jessica Howard - Foundations of a Fast, Data-Driven, Machine-Learned Simulator
Evan Jones - Tests of Catastrophic Outlier Prediction in Empirical Photometric Redshift Estimation with a Support Vector Machine
Bernadette Bosco - Machine Learning in Astronomy: Galaxies ML
Anne-Katherine Burns - A Machine Learning Solution to Computationally Intensive Problems in Big Bang Nucleosynthesis (BBN)

Breakout Room 2

Methods, Hardware/Software/Algorithms/ Mathematics

Eric Montoya - Spin torque oscillators for spintronic neuromorphic computing
Tess Smidt - e3nn: A modular PyTorch framework for 3D Euclidean neural networks
Truong Nguyen - An Expectation-Maximization accelerator for unsupervised learning of adaptive Gaussian Mixture models
Jiayi Li - Tropical Geometry for Understanding Expressivity of Neural Networks
Yonatan Dukler - Optimization Theory for ReLU Neural Networks Trained with Normalization Layers

Breakout Room 3

Biophysics

Cameron Movassaghi - Machine Learning Applications for Multiplexed Neurotransmitter Detection
Zixiao Zong - Topological Models of Amyloid Fibril Formation, and Identification of Fibril Topologies from Fibrillization Kinetics
Elizabeth - Diessner: Mapping the Mutational Landscape of the SARS-CoV-2 Main Protease: Molecular Modeling and Comparative Analysis
Eric Medwedeff - Towards Scalable Simulation of Dynamical Graph Grammar Biological Models, a Natural Arena for ML Model Reduction
Cory Scott - Efficient Learning of Cytoskeletal Dynamics with Multiscale Machine Learning and Optimized Projection Operators

Breakout Room 4

Many-body Physics

Mathieu Bauchy - End-to-End Differentiability and TPU Computing to Accelerate Materials’ Inverse Design
Yu Song - Deciphering the viscosity of nuclear waste immobilization glasses by deep learning
Azmain Abrawr Hossain - Extracting Exciton Binding Energy using Regression Techniques
David Rosenberger - Evaluating diffusion and the thermodynamic factor for binary ionic mixtures

Breakout Room 5

Weather and Climate; Science Image Analysis and Bio-image Analysis

Matthew Laffin - Physics-Constrained Neural Networks for Large-Scale Inference of Subglacial Topography under Greenland and Antarctica
Griffin Mooers - Generative Modeling of Atmospheric Convection
Nadia Ahmed - Remote Sensing for Severe Weather Detection
Silvia Miramontes - Accelerating Cell Counting with Quantitative Microscopy Based on U-Net

[ 15 Minute Break ]

PLENARY SESSION 2

12:30 PM

Brian Spears
Cognitive Simulation: Combining simulation and experiment with artificial intelligence

Abstract

Large-scale scientific endeavors often focus on improving predictive capabilities by challenging theory-driven simulations with experimental data. Yet, both simulation and experiment have become overwhelmingly rich, with a complex of observables including scalars, vector-valued data, and various images. At Lawrence Livermore National Laboratory (LLNL), we are using modern artificial intelligence (AI) technologies to combine predictive simulation models with rich experimental data. We call this set of methods cognitive simulation (CogSim). We will describe a strategic LLNL research effort aimed at using recent advances in deep learning, computational workflows, and computer architectures to develop improved predictive models.

We will present our progress toward our CogSim vision using work from a wide range of applications, including inertial confinement fusion research at the world’s largest laser, the National Ignition Facility (NIF). We will describe advances in multi-modal machine learning architectures, state-of-the-art tools for uncertainty quantification and model training at enormous scale, and efforts to explore next-generation computational platforms.

1:00 PM

Stephan Mandt
Machine Learning and Physics: Bridging the Gap

Abstract

Physicists and machine learners have much in common but speak different languages. Both communities deal with high dimensional state spaces and build sophisticated approximation schemes to develop tractable models. In the first part of my talk, I will show how (1) Bayesian matrix factorization methods can be used to improve over physics-based modeling of thermodynamic properties fluid mixtures in chemical engineering and how (2) variational autoencoders can help reveal hidden patterns in climate science data sets. As an example of how physics can influence machine learning, I will talk about spontaneous symmetry breaking in time series models and an improved learning algorithm, Goldstone gradient descent.

[ 15 Minute Break ]

1:45 PM

Gowri Srinivasan
Combining Graph Theory and Machine Learning to Characterize Fractured Systems

Abstract

Fractured systems are ubiquitous in natural and engineered applications as diverse as hydraulic fracturing, underground nuclear test detection, corrosive damage in materials and brittle failure of metals and ceramics. Microstructural information (fracture size, orientation, etc.) plays a key role in governing the dominant physics for these systems but can only be known statistically. Current models either ignore or idealize microscale information at these larger scales because we lack a framework that efficiently utilizes it in its entirety to predict macroscale behavior in brittle materials. We propose a method that integrates computational physics, machine learning and graph theory to make a paradigm shift from computationally intensive high-fidelity models to coarse-scale graphs without loss of critical structural information. We exploit the underlying discrete structure of fracture networks in systems considering flow through fractures and fracture propagation. We demonstrate that compact graph representations require significantly fewer degrees of freedom (dof) to capture micro-fracture information and further accelerate these models with Machine Learning. Our method has been shown to improve accuracy of predictions with up to four orders of magnitude speedup.

2:15 PM

Eric Mjolsness
AI approaches to graph dynamics for multiscale computational science

Abstract

A key operation in multiscale science is the mathematical change of scale under which many fine-scale variables get replaces with fewer coarse scale variables in an approximate, emergent dynamics. This inter-scale relationship has two important points of interaction with artificial intelligence: (a) it may be learned from fine-scale simulation data, and (b) it may require a substantial change of problem representation up to and including the introduction of dynamic graphs: collections of labelled nodes (vertices with spatial positions) and their connecting links (edges). I will briefly introduce three methods that point in these directions: [1] A combination of algebraic multigrid methods with graph neural networks applied to microtubule biomechanics; [2] A “Dynamic Boltzmann Distribution” method for learning model reduction of stochastic spatial biochemical networks; and [3] a meta-language for stochastic spatial graph dynamics, Dynamical Graph Grammars, which has an underlying theory related to operator algebras which may be understood combinatorially.

References
[1] C.B. Scott and Eric Mjolsness. In IOP Machine Learning: Science and Technology, accepted MS https://iopscience.iop.org/article/10.1088/2632-2153/abb6d2, August 2020.
[2] Ernst, Bartol, Sejnowski, Mjolsness, Phys Rev E 99 063315, June 2019.
[3] EM, “Structural Commutation Relations for Stochastic Labelled Graph Grammar Rule Operators”, arXiv: http://arxiv.org/abs/1909.04118, August 2019.

UC SOUTHERN HUB:

Frontiers in Machine Learning for the Physical Sciences

THANK YOU FOR REGISTERING

Please update your Zoom client to the latest version (5.3.2) to participate in the poster session breakout rooms

Please note: the browser-only version of Zoom does not support breakout rooms -- please install Zoom on your computer

Please Review these breakout room instructions since the feature is new to Zoom 5.3

Please direct any questions meant for plenary session speakers to the moderator named "Q&A" in chat

ZOOM LINK

Click Here to view a list of posters that will be presented

Breakout Room 1

Breakout Room 2

Breakout Room 3

Breakout Room 4

Breakout Room 5

THANK YOU FOR
REGISTERING