Poster presentation from Skunkworks members at the Regional Materials and Manufacturing Network (RM2N) meeting in Eau Claire, WI. The focus is on machine learning applications for materials impurity diffusion data.
1. Machine Learning For Predicting Missing Diffusion Data
Liam Witteman*1,Ben Anderson**2, Haotian Wu2, Aren Lorenson2, Henry Wu2, Dane Morgan2
1Department of Chemical and Biological Engineering,1415 Engineering Drive, Madison, WI 53706, University of Wisconsin - Madison
2Department of Material Science & Engineering, 1509 University Avenue, Madison, WI 53706, University of Wisconsin - Madison
*email: lwitteman@wisc.edu **email: bdanderson2@wisc.edu
Introduction:
Diffusion coefficients tell us how atoms move under a driving force
MAterials Simulation Toolkit (MAST) is an
automated high-throughput workflow
manager for first-principles diffusion
calculations
Diffusion is important in manufacturing items such as:
Data:
To cover just FCC hosts: M(FCC)-X
● Needs ~15m core-hours
●Only covered ~10% so far
Question: How to quickly
and cheaply calculate the
rest of the M-X dataspace?
Software Infrastructure
for sustained Innovation
(Si2) award No. 1148011
Neural Network: Gaussian Kernel Ridge Regression
Conclusions:
●Idea is to mimic the brain
●Great at finding trends and
patterns
●Idea is to move to a higher
dimensionality where a linear trend
can be found
Results:
Assessment of “smart
consensus method” using
RMS in 20% left out of
dataset (5 test data “cases”).
With a root-mean-square of less than 300 meV, both machine
learning tools have predictive capabilities that can provide useful
predictions of missing data and speed completion of the Dataspace.
Further validation of prediction error is underway.
Best and Worst fits of Leave Out
20% Test, Average RMS: 179 meV
Best: 118 meV Worst:
265 meV
References:
● www.mydailynew.com
● www.indusoft.com
● en.wikipedia.org/wiki/Fuel_cell
With 7 host-
impurity pairs,
rms drops
below 400
meV (300 on
average)
-First principle diffusion data of impurity X in host M
● www.texample.net
● http://www.eric-kim.net