Progress Report (Post)

Posted by Almambet Iskakov, Robert Pienta on October 7, 2015

Progress Report (Post)


Almambet Iskakov, Robert Pienta

12 00 AM ,Wed, Oct 07 2015

Quick Recap

  • Directional Solidification of Al, Ag, Cu Eutectic Alloy
    • Made from state-of-the-art simulations
    • Multiple solidification velocities and volume-fractions
    • Plenty of data, plenty of challenges to overcome

3D-DS

Data: (12x) 800x800x300

3D-DS

We can characterize these points using:

  • Two volume fractions (the third is dependent)
  • Solidification velocity

Curious Parameters

  • We have more than 12 simulations at this resolution
    • 2 have solidification velocities around 0.6
    • All other experimental velocities are 0.09 or less
    • Is this suspicious?

    • Possibly more data in future

Our Workflow

3D-DS

Steps with unanticipated challenges: dataflow2

2-Points Everywhere

  • 3 phases (0 - Al, 1 - Ag-Al, 2 - Al-Cu)
  • Assume periodic microstructure (based on simulation)
  • Finding a representative microstructure

3D-DS * Which microstructure is representative?

Example two-point statistics (autocorrelation) 3D-DS

Steady-State Solidification

Finding representative microstructure (preliminary method) 3D-DS

  • Compute the difference between spatial correlations to identify convergence/trends

Comparing autocorrelation 3D-DS

  • High initial variation; steady in full simulation result?
  • Explore other methods (RVE, etc.)

PCA

Simulated Volumes in PCA-Component-Space pca (grouped by solidification vel.)

Cumulative Variance by PCA Component tradeoff Encouraging singular-value fall-off characteristics

Linkage Overview

Multivariate regression problem: 3D-DS

Linkage and its application

linksteps

From a vol-fraction and solidification vel. to “a microstructure”.

Ongoing Work

  • We are currently working on cross-validation for our pipeline.
    • Originally wanted k-fold cross validation, but…

    • That’s leave-one-out for ~10 data points.

  • We have not completed the reconstruction code, but can produce everything up to it.

Challenges

  • Representing each volume with a microstructure
    • Choosing an RVE
    • Doing an expensive 3D 2-pt statistic
  • Choosing which correlations to use as PCA inputs
  • No control over simulation data