Post on 26-May-2015
description
Data Science Solutions by Materials Scientists
The Early Case Studies
Tony FastMaterials Data AnalystMaterials Informatics for Engineering Design
Woodruff School of Mechanical Engineering
Georgia Institute of Technology
*Any MINED shield is a link to a resource.
An Archival and Self Describing Data Format using HDF5
Data and Metadata stored in one file, Support in many languages, and Ideal support for high-dimensional data
*MXADataModel – Archival Data Format – ONR/DARPA Dynamic 3-D Digital Structures Program
HDF5 - The little zip file that could…
One Dataset – 1.6GB – 4 Experiments –with 160 Datasets each…..no long term value.
Volume
Variety Velocity
= Big DataPolymer - MD Titanium
Jacobs -GaTech
Bamboo
Martensitic Steel SiC/SiC Al-Cu Solidification
Frasier -OSU Wegst - Dartmouth
Gumbusch Ritchie- LLNL Voorhees - NW
Materials Science
The velocity that data is generated will rise and the speed that it will be analyzed in will decrease.
Rowenhorst, Lewis, Spanos, Acta Mat, 2010
β-Titanium
REDUCED OUTPUT:Grain sizeGrain FacesNumber of GrainsMean CurvatureNearest Grain Analysis
10 micron resolution with 4300 GrainsCompare with empirical models
Materials Science is a Big Data domain, but it is not treated that way.
Scalable, objective, parametric materials descriptorsManage data with care for the futureInteroperability, Sharing, and CollaborationEducate data scientists who can extract value from data using statistics, computation, and materials domain knowledge
Embrace complexity in big materials
data
Example Databases
AFLOW, Curtarolo Group Harvard Clean Energy Project Database
STRUCTURE INFORMATICSWORKFLOW
PHYSICS BASED MODELSSIMULATION EXPERIMENT
MICROSTRUCTURE (MATERIAL) SIGNAL
PROCESSING
ADVANCED & OBJECTIVE STATISTICAL ENCODING
DATA SCIENCE MODULES
INNOVATION ACCOUNTING
INTELL
IGEN
T
DESIG
N O
F EX
PER
IMEN
TS
Microstructure Informatics is a scalable, data-driven system to mine structure-property/processing connections from experimental and simulation materials science information; structure being the independent variable. The system is agnostic to material system and length scale, objectively quantifiable, and rapidly iterates in less cycles for both materials improvement and discovery.
DATA SCIENCEMODULES
MicrostructureMaterial Structure
ProcessingProperty
Data science modules are machine learning and statistical tools to extract rich bi-directional structure-property/processing linkages from encodings of materials & microstructure datasets. Mining modules create structure taxonomies, homogenization and localization relationships, ground truth comparison between simulation and experiment, materials discovery, and materials improvement.
ADVANCED & OBJECTIVE STATISTICAL ENCODING
THE MICROSTRUCTURE IS A SAMPLE IN AN IMMENSE STATISTICAL POPULATION.
α-β Titanium
SPATIALSTATISTICS
t t
t
Statistical correlations between random points in space/time which reveal systematic patterns in the microstructure. Contains the original μS within a translation & inversion. An objective encoding for most materials datasets.
CURRENT APPLICATIONSmetals, polymers, fuel cells, cmc, md, & a bunch of other
things
TYPES OF SIGNALS sparse, experimental, simulation, heterogeneous, surface,
bulk
The fidelity of the spatial statistics are impacted by how the material structure is parameterized as a signal.
Objective Microstructure Classification of α-β Titanium Images StatisticsMine with Principal Component Analysis
Mechanical Deformation of Polymer Chains
Molecular Dynamics of
Aluminum Atoms
MPL
GDL
X-CTFinite Element ModelingStatisticsRegression to connect the statistics with diffusivity values from FEM
Bottom-up Homogenization Relationships
exac
t fit
simulation
mod
el
FEMε=5e-4
Meta-modeling with Materials Knowledge SystemsTop-down localization relationships
The MKS design filters that capture the effect of the local arrangement of the microstructure on the response. The filters are learned from physics based models and can only be as accurate as
the model never better.
INPUT OUTPUTControl
Meta-modeling with Materials Knowledge SystemsTop-down localization relationships
The MKS design filters that capture the effect of the local arrangement of the microstructure on the response. The filters are learned from physics based models and can only be as accurate as
the model never better.
Any M
odel
OTHER APPLICATIONSSpinodal Decomposition, Grain Coarsening, Thermo-mechanical, Polycrystalline
Top-Down Localization Relationships for High Contrast Composites
The MKS is a scalable, parallel meta-model that learns from physics based models to enable rapid simulation at a cost in accuracy.
N2 vs. Nlog(N) complexity It learns top-down localization relationships to extra extreme value
events and enables multiscale integration.
Structure-Processing MKS
Processing History
Structure-Property
Homogenization
Structure-Property
Localization
Objective parametric descriptors and data science enable integrationof bi-direction structure-property/processing linkages.
Data enables bidirectional S-P/P, multiscale integration, and higher throughput
CORE TECHNOLOGIES TO FUEL THE DATA AGE OF MATERIALS SCIENCE
Open Access, Open Source Software, Scalable Databases, High-Statistical Throughput Simulation and Experiment, Image
Segmentation, Machine Learning, Scalable Databases, Metadata Integration, Mobile Technology, Visualization, High Performance Computing, Cyberinfrastructure/Collaboratories, Collaboration &
Sharing
Selected Links
Any shield in this presentation is a link
HDF5 http://www.hdfgroup.org/HDF5/whatishdf5.htmlHDFView http://www.hdfgroup.org/hdf-java-html/hdfview/MXADataModel http://mxa.web.cmu.edu/Background.htmlCurtarolo Group http://www.mems.duke.edu/faculty/stefano-curtaroloAFLOW http://materials.duke.edu/apool.htmlHarvard Clean Energy Project http://www.molecularspace.org/Serial Sectioned Titanium https://cosmicweb.mse.iastate.edu/wiki/pages/viewpage.action?pageId=753830MATIN http://www.materials.gatech.edu/matinMaterials Genome Initiative http://www.whitehouse.gov/mgi