Skip to Main Content U.S. Department of Energy
Physics-Informed Learning Machines

Announcements

Kailai Xu presented a poster: Adversarial Numerical Analysis at Xpo.

We introduce a novel approach -- adversarial numerical analysis (ANA) -- for solving inverse problems involving probabilistic computations. The approach allows us to solve a variety of inference problem by leveraging the striking power of neural networks for classification and adversarial training algorithms for intractable probabilistic computation. We demonstrate its effectiveness and generality by solving various problems, including UQ, probabilistic inversion, parameter inference in stochastic processes and volatility calibration for option pricing.


New software released for MachineLearning & PDE's - DeepXDE First Release, the most powerful machine learning tool to solve PDEs.


George Em Karniadakis's paper "Quantifying the generalization error in deep learning in terms of data distribution and neural network smoothness" was recently highlighted in the weekly online magazine DeepAI


PhILMs workshop on May 31, 2019 at Sandia National Lab focused on Machine Learning.


CRUNCH group research highlighted by LANGMUIR as cover article

Langmuir cover image
This photo was reprinted with permission from the authors. Copyright 2019 American Chemical Society.

Research from the CRUNCH group has been featured on the cover of the prestigious ACS journal LANGMUIR. The research was a collaborative effort by Kaixuan Zhang, Zhen Li, Martin Maxey and George Karniadakis in partnership with Shuo Chen from Tongji University (Shanghai, China). Shown on the cover are different phases of droplet coalescence dynamics on a hydrophobic surface with hierarchical roughness, in which the colors on the rough wall indicate the distribution of surface structure height, and the wavelets on the liquid droplets represent the thermal capillary waves. View the article.

The study reports a novel self-cleaning mechanism of textured surfaces attributed to a spontaneous coalescence-induced wetting transition, which explains why droplets on rough surfaces are able to change from the highly adhesive Wenzel state to the low-adhesion Cassie-Baxter state and achieve self-cleaning. The study also reports that the spatial distribution of liquid components in the coalesced droplet can be controlled by properly designing the overall arrangement of multiple droplets. The findings offer new insights for designing effective biomimetic self-cleaning surfaces by enhancing spontaneous Wenzel-to-Cassie wetting transitions, and additionally, for developing new noncontact methods to manipulate liquids inside the small droplets via multiple-droplet coalescence.

CRUNCH group research highlighted by Soft Matter as cover article

softmatter cover image
Cover image reproduced with permission of the authors and the Royal Society of Chemistry".

Recent research from the CRUNCH group has been featured on the back cover of the prestigious RSC journal Soft Matter. This highlighted work was performed in the group of Prof. George Karniadakis from the Division of Applied Mathematics, Brown University, in collaboration with the group of Prof. Chao Yang from the Institute of Process Engineering, Chinese Academy of Sciences. Shown on the cover is a snapshot of multiscale simulation of a shear flow past an endothelial glycocalyx layer (EGL) in a microchannel, with the atomistic resolution locally (water molecules, heparan sulfate chains, lipid bilayer and transmembrane proteins) and the coarse-grained resolution in bulk domain. View the article.

The study reports the multiscale modeling of soft multi-functional surfaces. The endothelial glycocalyx layer is a soft multi-functional surface, coating the endothelial cells and lining the entire vascular system. Single brute-force atomistic simulation for this problem is prohibitively expensive and limited to small scales. This study shows that an efficient parallel multiscale method can bridge the atomistic and mesoscale regimes in modeling of soft mult-functional surfaces, from nanometer to micron and beyond.


Harnessing New Computing Tools to Solve Big Physics Problems

New $10 million collaborative grant to boost deep learning applications

philms

The Department of Energy's Pacific Northwest National Laboratory (PNNL) has been tapped to lead a $10 million, four-year effort to uncover hidden physics using new developments in deep learning, a computing technique that harnesses the power of machine learning and big data.

The Physics-Informed Learning Machines for Multiscale and Multiphysics Problems (PhILMs) Center, is a collaboration among PNNL and Sandia National Laboratories, with academic partners at Brown University, Massachusetts Institute of Technology, Stanford University, and the University of California, Santa Barbara.

The PhILMs collaborators are working together to adapt advanced computing techniques to help solve long-standing problems in science and engineering. 

"The most difficult problems in modern physics and engineering involve computing at vastly different scales and across large data sets," said George Karniadakis, the Center Director, who holds a joint appointment at PNNL and Brown University. "By combining the expertise of mathematicians, programmers, and specialists in machine learning techniques, and focusing that expertise on these physics problems, we are aiming to bring new tools to the global physics community."

The research team plans to create, test, and release new open-source computational tools developed through the collaboration and make them freely available through the team's PhILMs web site.

The initial research goals include developing computational frameworks based on deep learning algorithms for applications in combustion, subsurface, and earth systems, as well as machine learning tools for the design of functional materials with tunable properties. Ultimately, the researchers hope to establish scientific machine learning as a new meta-discipline at the intersection of computational mathematics, big data, and deep learning.

PhILMs leadership includes PNNL's Alexandre Tartakovsky, Center Co-Director; Michael L. Parks, Sandia National Laboratories; Mark Ainsworth, Brown University; Eric Darve, Stanford University; and Paul J. Atzberger, University of California, Santa Barbara.

The collaboration is supported by the Applied Mathematics Program within the U.S. Department of Energy, Office of Advanced Scientific Computing Research.


Paul Atzberger organized a mini-symposium, "Machine Learning Approaches for the Sciences and Engineering: Recent Developments", at the 2019 SIAM Computer Science and Engineering (CSE) Conference.


George Em Karniadakis kicked off the DOE ASC meeting on January 30th in Rockville.

PhILMs

Collaborations

Seminar Series

Computing Research at PNNL

Science at PNNL

Computing Research

View All Highlights

Contacts