Wendy S' K' Doyle, Ann C' Gentile

1
Extracting Information from DataEase Data
Analysis Development with FCLib

• Wendy S. K. Doyle, Ann C. Gentile
• W. Philip Kegelmeyer
• Sandia National Laboratories, USA
• NECDC06
• October 24, 2006
• Los Alamos, NM

Sandia is a multiprogram laboratory operated by
Sandia Corporation, a Lockheed Martin
Company,for the United States Department of
Energy under contract DE-AC04-94AL85000.
2
Data is NOT Information yet

• Information is actionable answers questions and
  informs decisions
• Data (just the facts, maam) needs to be
  interpreted to become information
• The distinction can be fuzzy
• Data analysis helps turn data into information

3
Data Analysis Example

• Overall Process create a model, stress it,
  evaluate (repeat)
• Model can with a hatch attached by screws
• Stress crush at an angle
• Evaluate Is the damage within acceptable limits?

4
Data vs. Information

• Damage Information
• How many screws break?
• How many screws are damaged?
• How many tears (dead element regions)?
• Do any tears penetrate?
• How much gapping?

5
Data vs. Information

• Use FCLib to convert data to information

• The analyst can now act on the information.

6
Data Analysis Development

• Issues
• Analysis development is iterative
• Data size is growing
• Humans are better at creating information than
  computers
• Problems with current methods
• Specialized and one-off tools cant be extended
• By hand methods are slow, hard to quantify
  have poor repeatability
• Our solution is to help the human create
  information
• Automate as much as possible
• But also allow flexibility

7
Feature Characterization Library (FCLib)

• FCLib Toolkit of data analysis building blocks
• C Library
• Open Source (BSD) http//fclib.ca.sandia.gov
• FCLib Features
• Provide variety of simple building blocks that
  can be composed into complex analyses
• Support feature-based analysis (feature region
  of interest)
• Minimize low-level processing (automation)
• Generality - applicable to a variety of science
  domains
• Current File Formats ExodusII, SAF LSDyna
• Provide an elegant interface - simple but not
  constrained

8
Presentation Outline

• Introduction
• FCLib Overview
• Example 1 Damage Assessment in Weapon Drop
  Simulation
• Example 2 Support for Machine Learning
• Example 3 Feature Tracking
• Summary Conclusions
• Future Work

9
Key Features of the Data Model

• Spatio-temporal data stored on unstructured
  meshes

• Minimal set of abstract data objects
• Types are deeper in the interface
• Subsets are full-fledged data objects
• Mesh ownership of data
• Time-varying data available per variable instead
  of per step

10
Analysis Building Blocks I

• Mesh Topology
• Get mesh entity children/parents/neighbors
• ,
  ,
• Segment into separate connected components
• Get skin
• Mesh Geometry (Spatial)
• Edge lengths/surface areas/region volumes
• Get mesh entities within bounding box or sphere
• ,

11
Analysis Building Blocks II

• Variable
• Variable Math
• Statistics Min/Max/Mean/StandardDeviation/Sum
• Decompose vectors into normal tangent
  components
• Threshold find subset
• Subset
• Set operations AND, OR or XOR to get new subset
• Time
• Feature tracking - track regions of interest
  (ROIs) over time
• Build your own!

12
How to Build an Analysis

• Explore stress in compressed regions
• 01 / get handles to objects of interest /
• 02 fc_loadDataset(some_data.ex2, dataset)
• 03 fc_getMesh(dataset, some_mesh, mesh)
• 04 fc_getVariable(mesh, pressure,
  pressureVar)
• 05 fc_getVariable(mesh, stress, stressVar)
• 06
• 07 / analysis string together building blocks
  /
• 08 fc_threshold(pressureVar, gt, 0., subset)
• 09 fc_getSubsetVariableMeanSdev(subset,
  stressVar,
• 10 mean, sdev)
• 11 printf(Stress g /- g\n, mean, sdev)

13
Example 1 Damage Assessmentin Weapon Drop
Simulation

• Three tools were used by W80 analysts on 150
  datasets
• screwBreaks
• tears
• gaplines
• Usage modes
• Generate a human readable report.
• Use single measure for sensitivity plots.
• Use single measure to drive optimization of drop
  angle to produce maximum damage.
• Tools were developed by FCLib developers in close
  communication with analysts.

14
About The Data

• Example Data
• 2.5 Gb
• 5 parts, 640,000 elements
• Largest 400,000 elems
• 18 screws
• 13 time steps

• Real Data
• 50 Gb
• 100 parts, 1.8M elements
• Largest 350,000 elems
• 80 screws
• 4-30 time steps

15
How Many Damaged or Broken Screws?

• Visual inspection
• fclib/bin/screwBreaks d3plot

16
screwBreaks Output

• Mesh Screws Screw 0 Step 12 BR
  0.00 ( 0.00/ 6.00)
• Mesh Screws Screw 1 Step 12 BR
  0.00 ( 0.00/ 6.00)
• Mesh Screws Screw 2 Step 12 BR
  0.00 ( 0.00/ 6.00)
• Mesh Screws Screw 3 Step 12 BR
  0.00 ( 0.00/ 6.00)
• Mesh Screws Screw 4 Step 12 BR
  0.00 ( 0.00/ 6.00)
• Mesh Screws Screw 5 Step 12 BR
  0.31 ( 1.89/ 6.00)
• Mesh Screws Screw 6 Step 12
  still broken
• Mesh Screws Screw 7 Step 12 BR
  0.27 ( 1.64/ 6.00)
• Mesh Screws Screw 8 Step 12 BR
  0.00 ( 0.00/ 6.00)
• Mesh Screws Screw 9 Step 12 BR
  0.00 ( 0.00/ 6.00)
• Mesh Screws Screw 10 Step 12 BR
  0.00 ( 0.00/ 6.00)
• Mesh Screws Screw 11 Step 12 BR
  0.00 ( 0.00/ 6.00)
• Mesh Screws Screw 12 Step 12 BR
  0.00 ( 0.00/ 6.00)
• Mesh Screws Screw 13 Step 12 BR
  0.00 ( 0.00/ 6.00)
• Mesh Screws Screw 14 Step 12 BR
  0.15 ( 0.93/ 6.00)
• Mesh Screws Screw 15 Step 12
  still broken
• Mesh Screws Screw 16 Step 12 BR
  0.27 ( 1.64/ 6.00)
• Mesh Screws Screw 17 Step 12 BR
  0.00 ( 0.00/ 6.00)

Status of each screw at each time step (only last
step shown)
Screw is broken.
Screw is damaged (BR breakage ratio)
Summary for mesh
Summary for dataset
17
How Bad are the Tears?

• Visual inspection
• fclib/bin/tears -d displacement d3plot

18
tears Output

• Tear characterizations for dataset 'd3plot'
• Tears criteria 'elem_death' lt 0
• Time step index 12
• 5 mesh(es)
• Mesh 0 'Shell' has 18 dead element region(s)
• Mesh 1 'Plate' has 26 dead element region(s)
• Mesh 2 'Cover Plate' has 0 dead element
  region(s)
• Mesh 3 'Horseshoe Plate' has 0 dead element
  region(s)
• Mesh 4 'Screws' has 6 dead element region(s)
• Combining of dead elem regions not requested
• Found 50 tears
• Sorting tears by region diameter (largest first)
  ...
• Tear 0
• numDeadElementRegions 1
• meshIDs 0
• meshNames 'Shell'
• numCell 280
• region volume 35.8475
• region diameter 19.8277

Input details
Per mesh summary
Tear details
19
More tears Output

• File tear-regions.bb gt bounding boxes of the
  tears

20
How Much Gapping?

• Visual inspection
• fclib/bin/gaplines -d displ d3plot Shell Cover
  Plate 0.1

21
gaplines Output

• Dataset 'd3plot'
• Meshes 'Shell' and 'Cover Plate'
• Displ 'displacement'
• Min Dist 0.1
• Number of gap lines found 12482
• Number of sets of sides involved 2
• Stats for set 1 ('Shell_shape0_side18-Cover
  Plate_shape0_side2')
• numGapline 10628
• Step Gap Length
  .
• ID Value num min max
  mean stdev
• 0 0.000000 10628 0.000000 0.028636
  0.003368 0.005055
• ...
• 12 0.003000 10628 0.002291 6.851685
  1.527860 1.449041
• Step Normal Component of Gap
  Length .
• ID Value num min max
  mean stdev
• ...
• 12 0.003000 10628 0.002291 6.851685
  1.527860 1.449041

Input details
Result Summary
Stats reported for each side and overall
Gap length stats
Resolved with respect to face normals
22
More gaplines Output

• File gaplines.ex2 gt the gaplines and length
  variables
• 2 adj. surfaces found gt Cover Plate sits in
  recess in Shell
• Location and amount of gapping easily visualized

23
Example 2 Support for Machine Learning (ML)

• FCLib used to manipulate data to and from ML
  algorithms
• At least 6 of the tools written by ML researchers
• varSmooth - smooth a variable by averaging values
  within given radius
• Reduces noise and other artifacts of ML
  algorithms
• connect_comp_regions gen_overlap_matrix -
  extract and match up regions from different
  datasets
• Generate accuracy measures for different ML
  algorithms
• Crossval - breaks the dataset into multiple
  partitions
• Generate random samples for cross validation
  training

24
Example 3 Feature TrackingWhats Going on in
the Ridges?
t0
t5
t10
t15
t25
t35
t20
t30
25
Find Track Features

• 01 / get handles to objects of interest /
• 02 fc_loadDataset(cancrush.saf, dataset)
• 03 fc_getMeshByName(datatset, can, mesh)
• 04 fc_getSeqVariable(mesh, eqps, numStep,
  stressVar)
• 06
• 07 / create a feature group and populate with
  features /
• 08 fc_createFeatureGroup(featureGroup)
• 09 for (i 0 i lt numStep i)
• 10 fc_threshold(stressVari, gt, 1.0,
  subset)
• 11 fc_segment(subset, 0, numROI, ROIs)
• 12 fc_trackStep(i, numROI, ROIs,
  featureGroup)
• 13
• 14 fc_writeFeatureGraph(featureGroup,
  graph.dot)

26
Reading the Feature Graph
t0
Feature 4
F8
F9
F6
t25
t25
t35
27
Analyze the Features

• 01 / loop over each feature /
• 02 fc_featureGroupGetNumFeature(featureGroup,
  numFeature)
• 03 for (i 0 i lt numFeature i)
• 04 fc_getFeatureROIs(featureGroup, i,
• 05 numROI, stepIDs,
  ROIs)
• 06 / loop over each time step the feature
  exists /
• 07 for (j 0 j lt numROI j)
• 08 / do analysis /
• 09 fc_getVariableSubsetMinMax(seqVarstepIDs
  j,
• 10 ROIsi, minsij,
  maxsij)
• 11 / print stats (not shown) /
• 12
• 13

28
Interpret Results Repeat
Feature Graph
29
Summary Conclusions

• FCLib is an open source toolkit of analysis
  building blocks.
• FCLib has been used by weapons analysts and
  machine learning researchers to create data
  analysis tools
• FCLib makes development of data analysis easier
• Automates low level processing
• Can assemble analyses from building blocks
• Developed building blocks can be reused
• Analyses are tweakable and extensible
• Analyses are quantitative and repeatable

30
Future Directions

• Make FCLib more usable, looking into
• Scriptable (Python?)
• GUI interface (like AVS Express or LabView)
• Output database (xml) to support varied query
  modes
• Grow FCLib community
• Create space for users to share code
• More building blocks! More data analyses!
• Dataset comparisons (very hard)

31
Acknowledgements

• Jay Dike, Tim Shelton Tim Kostka - SNL,
  Analysts
• Ken Buch - SNL, Machine Learning (ML)
• Robert Banfield, Larry Hall, et. al. - Univ. of
  S. Florida, ML
• This research was supported by ASCs Pre and Post
  Processing Environments (PPPE) Data Discovery
  (DD) Program.
• Contact Wendy Doyle ltwkoegle_at_sandia.govgt
• FCLib available at http//fclib.ca.sandia.gov