Experiences with FEniCS/DOLFIN, Part 1

It’s hard to apply informatics to simulation data if you don’t have any simulation data.

I’ve recently started playing with FEniCS — a set of tools for solving PDEs with finite element methods. They include a python-like language with syntax very similar to mathematical expressions (Unified Form Language UFL) for posing and solving PDEs in variational form. Scripts written in UFL can be compiled and run on a variety of platforms including distributed memory clusters and multicore architectures. I think they have some support for GPUs, too. This is all done under the hood. The compiled code can link to and take advantage of powerful solver libraries like PETSc and Trilinos.

FEniCS is developed and maintained by a team of top-notch computational science researchers and software developers. Development is very active, and their support via Launchpad is astounding — prompt, helpful, and informative. I really can’t say enough good things about them.

Following the download-and-install instructions on the website was straightforward for both my 13-in MacBook Pro (dual core, 8G RAM) from late 2009 running OSX 10.6 and a Dell workstation with two quad core processors and 12G of shared memory running Ubuntu 12.04. The demos are remarkably helpful, and so is the accompanying book.

I’m working on getting it running on a distributed memory cluster. Stay tuned.

Some background material for UQ

I finished a set of background notes for the uncertainty quantification class (ME470) I’m teaching this quarter with Gianluca Iaccarino. A link to the notes is below. Let me know if you think I left anything out (or if you find any typos)!

ME470 background notes

‘Computer Experiments’ by Koehler and Owen (1996)

I’ve heard a few statisticians complain about the rise in popularity (and funding) of research in uncertainty quantification, claiming that it’s all just a rehashing of statistics. And I think they have a point, for the most part. At the very least, those of us working in UQ should be aware of the excellent resources from the statistics literature that address UQ-like problems.

‘Computer Experiments’ by Koehler and Owen (Chapter 9 from ‘Handbook of statistics 13,’ 1996) is one of my personal favorites. It discusses Kriging and least squares surrogate models for expensive computer simulations and a variety of experimental designs. I consider it a must-read for people interested in UQ.

Here’s a PDF.

INFORMS ICS, Santa Fe, January 2013

I gave a talk at the INFORMS ICS conference last week at the beautiful Eldorado hotel in Santa Fe. I gave a broad overview and introduction to uncertainty quantification — with a bit about model reduction. This was part of a session organized by Brian Adams and Laura Swiler from Sandia Labs. Here’s a link to my slides.


I’m going to start posting here more when interesting things happen. Stay tuned.

Emily Shuckburgh

At SIAM Annual this year, Emily Shuckburgh gave a really great talk about simulation informations in the study of ocean currents. They began with simulations of ocean currents and hypothesized the existence of unstable manifolds in the ocean around Antarctica. Based on this idea, they then looked for evidence of these unstable manifolds in satellite data, and found then. Finally, to VERIFY their existence, then went to Antarctica to drop floats in the ocean to track where they go. These floats followed the expected manifolds! 

That’s taking simulation informatics — the initial simulation studies — all the way to real world science!

See the outline here: http://meetings.siam.org/sess/dsp_programsess.cfm?SESSIONCODE=14996 

I’m looking for relevant papers to post, but don’t expect to get around to it soon, so don’t expect them to make their way onto this page.


Simulation Informatics @ SIAMUQ

We’ve had great audience participation for our minisymposium at SIAM UQ 2012 so far!

There were 20-30 people in the audience throughout the entire mini-symposium.

Here’s a log of what’s going on!

5:30pm We had to cutoff questions for Barnabas’s talk! Please discuss here if you are interested.

5:35pm Qiqi Wang is now talking about unconventional information in simulations

5:37pm Qiqi on data and simulation: “This is the way we envision the future is going to be”

5:40pm Still Qiqi … “When you go to a chaotic dynamical system, [the adjoint method] breaks down.”  “Let me tell you the solution we just worked out.”

5:42pm When you use time-averaging, you get a smaller magnitude of noise, but that makes the derivative of the noise larger (10^100)!.  The difference between the finite time average and the infinite time average suffers from the butterfly effect, and this is a huge problem.

5:45pm Qiqi’s talk is too interesting, I need to pay attention now.  The video will be online soon 🙂

5:55pm Qiqi is solving challenges on how to get information out of unsteady simulations.

6:00pm Chandrika Kamath is starting to talk about scientific data mining.

6:03pm Kamath – “80-90% of the time is spent in data processing” [in scientific data mining]

6:07pm Kamath – classifying Poincare orbits is a tough challenge due to multi-scale and fractal patterns in the data, not to mention noise

6:09pm Kamath – feature building for the Poincare orbits include using polar coordinates to exaggerate some features and using residuals in polynomial fits

6:13pm Kamath – using parallel plots help to identify useful regions in a object by feature matrix

6:15pm Kamath – using 5-6 features incredibly good performance

6:15pm Kamath – Rayleigh-Taylor instabilities, 30TB and 80TB of data.  These are big problems!

6:17pm The problem – finding bubbles and spikes in tubulence data – isn’t even well defined.  They used image analysis techniques.

6:22pm Darn, we lost the video (out of battery) for the end of Chandrika’s talk.