Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
The MGI and AIJames A Warren
Director, Materials Genome Program
Material Measurement Laboratory
National Institute of Standards and Technology
Executive Secretary, NSTC Subcommittee on MGI
www.mgi.gov
Acknowledgements
Alex
NIST Colleagues including C. Campbell, A. Dima,
Z. Trautt, R. Hanisch, M. Brady, E. Lin, G. Kusne,
M. Green, J. Hattrick-Simpers, E. Lass, B. DeCost
Bryce Meredig (Citrine)
So many others that I’ve overlooked.
Power CorruptsVint Cerf
Outline
Define MGI
Racing towards AI, what does that mean?
Background and Models
Approaches and Successes
Humans + Machines
The Materials Genome InitiativeA Multi-Agency Effort
It’s an InitiativeDOE, NSF, NIST, DOD, NASA, FDA, NIH…
Through the Transition
To decrease time-to-market by 50% while <$$
Develop a Materials
Innovation Infrastructure
Achieve National goals in
energy, security, and human
welfare with advanced
materials
Equip the next generation
materials workforce
Materials Genome Initiative for Global Competitiveness
Span the Continuum
The G in MGI
Metaphors
Data Models
Simulation
Experiment
Quantum MacroMicroNano
Materials w/ Targeted Properties
The MGI Approach
Creating and Capturing Knowledge of Materials
M in MGI
MATERIALS PROPERTIES ARE HISTORY DEPENDENT
SOUFFLE INGREDIENTS
• 2 tbsp (30 mL) butter
• 2 tbsp (30 mL) all-purpose flour
• 1/2 tsp (2.5 mL) salt
• Pinch pepper
• 3/4 cup (175 mL) milk (1%)
• 4 eggs
• 2 egg whites
• 1/4 tsp (1.25 mL) cream of tartar
MGI - MATERIALS RESEARCH
THE CENTRAL PARADIGM OF MATERIALS RESEARCH▸ Materials Research is about making useful materials (transistors, LED,
titanium turbine blades, sporting goods…)
▸ Composition (Ingredients)
▸ Processing (Recipe)
▸ Structure (Souffle or Mess?)
▸ Properties (Fluffy, Delicious vs. Rocklike Charcoal)
▸ Challenge is to design new materials from existing knowledge without resorting to raw “Edisonian” repetition.
A Material’s Properties Depend on its History
Materials Are Complicated SystemsModeling is a Challenge
Alloy cooled from 300 °C
Alloy cooled from 800 °C
The Decade of MGI?These ideas are not new
Milanese Loop Alloy
High Strength 18K Gold
Anodizable 7000 Aluminum
-Custom Magnetic Stainless Steel
-2X harder
-60% stronger Al
-30% lighter than 316L
Apple watch-Announced September 2014
Baseline: 316L Stainless Steel
-Cold-forged to 40% harder
-Special purity mirror finish
Prospective Economic Impact
The potential economic impact of improved
technology infrastructure supporting materials
innovation in the United States is conservatively
estimated to be between $116 billion and $240
billion per year:
mgi.nist.gov/keyReport
MGI IN SUM
• The MGI Is about improving our ability to design and deploy new materials (faster)
• Need better (or just any) data and models
• The MGI is essentially a direct consequence of our improvements in computational power and associated models, coupled to the disruptive consequences of the Internet.
• There are a limited number of ways to get the “knowledge” that is the fuel for the MGI
• High Throughput Computations, w/ published data and software
• High Throughput Experiments
• Get it from everywhere (Mine the literature, mine published data, if only it were published!)—- Change Publication and feed the models
WHAT IS AI (TODAY)
NOT THE COMPUTER SCIENCE
DEFINITION BUT INSTEAD THE
COLLQUIAL ONE
(MACHINE LEARNING, DEEP NEURAL
NETS, ETC)
BUT NOT ROBOT DOOM
Mastering the curse of
dimensionallity
This talk is aging quickly
Growth of Materials Informatics
0
750
1500
2250
2003 2005 2007 2009 2011 2013 2015
Men
tions in P
ape
rs
materials informatics materials genome "machine learning" & "materials science"
White House launches
Materials Genome
Initiative
Citrine InformaticsBryce Meredig
2017 Gartner Hype CycleEmerging Technologies
The Relationship between Models, Measurement, Data, and
PublicationAn old polemic, illustrating some of the key questions, gaps, obstacles to
realizing materials by design
Traditional Approach Example
Measure the modulus of a
steel coupon by a tensile test
(measurement of strain)
record values is some sort of
table (perhaps Excel)
perhaps publish
publication contains
metadata
What the #$&% ̂is a modulus
It’s only a (constitutive) model (not reality)
It’s an equation
What you think you’re measuring is NOT the modulus
You are measuring a strain and assuming the model is
true.
A model is a lie that helps you see the truth. - Howard Skipper
All Models are Wrong, but some are Useful -- George Box
You can’t measure without a model
Measurement without Models is nearly
MEANINGLESS
APS Physics
Oct 13, 2017
BETTER MODELS = LESS DATA
TAKE THE LHC AS AN EXAMPLE
DATA PRODUCED AT 1PB/SEC !
REDUCED DATA SAVED: ~ GB/S
THAT’S A DARN GOOD MODEL
OTHER END : BIOLOGY?
MATERIALS: IN THE MIDDLE
Where does AI fit in this picture?
Some Reflections
Importance of Models in general
AI is a method for creating a model (descriptors and predictions)
AI might be viewed as a method for “coarse graining”
Physics is a model creation machine
Correlation versus causation & Notions of induction and ground truth
THIS IS WHAT MAKES AI+SCIENCE SPECIAL (versus recommendations for
what to buy with your socks)
In other words, we can be DATA POOR (and we often are) because we
(often) have good models
MGI is both necessary to and will be influenced by the coming revolutions in Manufacturing
Digital Thread
Fab Lab
Some MGI ExemplarsBuckle up
Design of Cu-Ni-Zn-(Mn) alloys with tailored electrical conductivity
Eric A. Lass, Mark R. Stoudt, Maureen E. Williams, Carelyn E. CampbellMaterials Science & Engineering Division
National Institute of Standards and TechnologyGaithersburg, MD, USA
Tony YingUnited States Mint
Washington, DC, USA
ICME 2017May 21-25, 2017Ypsilanti, MI, USA
Olson systems design diagram
Secondary Processing
Secondary Processing
Secondary Processing
A new improved alloy
Constructed a Calphad-type model for color, identified the entire composition space in quaternary Cu-Ni-Zn-Mn possessing the correct conductivity, but with varying colors from white to yellow
Identified the “best” (cost effective) alloy meeting the customer’s requirements, and where to move in composition space if color-requirements are relaxed
Exemplars #2 Databasing
DFT base stories
Materials Project
AFLOW
OQMD
NOMAD
MARVEL/MAX
MAST-ML
Extremely Well Curated Data
Chuck Ward
GALILEO EXAMPLE
MICHAEL NIELSEN
ACADEMIC INCENTIVES
THE MOONS OF SATURNGalileo no doubt planned to publish this new
discovery in his next book, but in the meantime, how
could he preserve his priority and prevent others from
claiming the discovery as their own? His solution was
to circulate an anagram, s m a i s m r m i l m e p o e t
a l e u m i b u n e n u g t t a u i r a s. Others would
know that he had discovered something and when he
had discovered it, but they would not known what the
discovery was. The number of letters in the anagram,
37, was too small to allow him later to fudge and
change the solution to describe a discovery made by
someone else in the meantime. Before the days of
scientific papers (invented in the 1660s) this was an
effective (if not always foolproof) method of claiming
priority.
http://galileo.rice.edu/sci/observations/saturn.html
We need to both change the rules and give people tools to make that easier
Why Curate Your Data?
Increase your productivity/efficiency
Can you find it easily
Do you remember what the data means (the experimental/simulations
conditions)?
Improve reproducibility
Improve research continuity (next student/staff member knows where to start)
Increase your research impact; others can find it and re-use it.
Minimize data loss
Adhere to government mandates
ENABLE DATA DRIVEN RESEARCH
Material Measurement Laboratory, Jan 23, 2017, Z. Trautt
Motivated by the Astronomy Community
Value Multiplier
Astronomy vs. Material Measurement
Measurement Type
Po
siti
on
in t
he
sky
Singular data models Modular data models
e.g.
e.g.
Measurement Type
Mat
eria
l syn
thes
is a
nd
p
roce
ssin
g h
isto
ry
48WHAT’S OUR TELESCOPE?
AI Efforts
MML AI Projects
AI Self-QA Using Learning Curves in Feedback Loops
(Congo)
Autonomous Materials Laboratory: Phase Mapping
(Kusne)
Combining High-Throughput Experimental Science
and Machine Learning to Accelerate Materials
Innovation (Hattrick-Simpers)
Deducing Prior Deformation from Simple Mechanical
Analysis (Reid)
High-Performance Crystal Plasticity by ML-driven
Interpolation (Reid)
JARVIS-ML: Machine Learning Prediction of Material
Properties (Choudhary)
Develop and Test Novel Root and Rule-Based
Natural Language Processing (NLP) Tools for
Information Indexing and Searching (Sriram, Bhat)
Large-Scale Atomic Force Microscopy Data Mining
and Analysis (Persson)
Learning Networks for IoT and Smart City Data Sharing
(Kusne)
ML to Integrate Multiple Data Types to Characterize
Billions of Reference Values in NIST Human Genome
Reference Materials (Zook)
NIST Genetic Sensor Foundry (Ross)
Material Design Toolkit (Li)
Materials Informatics/Semi-Automatic Curation (Dima,
Becker)
OAR (Open Access to Research) – NIST Science Data
Repository System (Greene)
Physically Inspired, Highly Transferable Neural Network
Interatomic Potentials (Tavazza)
Polymer Property Predictor and Database (Audus)
AI/ML Bootcamps
AI/ML Bi-Weekly Interest Group
ML and ClassifiersBrian de Cost (PhD work with Holm, Jain, and Rollett)
Construct VLAD representations
1: select
image features
2: characterize
image features
3: encode
image features
B. DeCost, H. Jain, A. Rollett, and E. Holm
JOM (2016). doi:10.1007/s11837-016-2226-1Vector of Locally Aggregated Descriptors
Powder lot classification
60
Platform: ARCAM
B. DeCost, H. Jain, A. Rollett, and E. Holm
JOM (2016). doi:10.1007/s11837-016-2226-1
SEM powder classification
Cross-validation results
B. DeCost, H. Jain, A. Rollett, and E. Holm
JOM (2016). doi:10.1007/s11837-016-2226-1
independent samples
ML and Autonomous Materials Science
A. Gilad Kusne and Collaborators
Phase Mapping: High-Throughput Approach (APL Materials 2016)
Fabricate hundreds-thousands of samples -> HiTp
Synthesis
Measure all samples -> HiTp Characterization
Rapid phase mapping -> Machine Learning
Co
Fe
Ni
Combi Library for Ternary Spread
Diffraction Patterns
XRD
Estimated Phase Map
ClusterAnalysis
Co Ni
Fe
Phase Mapping: High-Throughput Approach
Measurement is a time / resource sink
For wafer of 500+ samples:
In Lab: Takes weeks-months
Synchrotron: Takes 5+ Hours (Every second
counts)
Mn-Ni-Ge library535 samples
Stanford Synchrotron Radiation Lightsource30 seconds per sample4.5 hours
Bruker D830 Minutes per sample2 weeks!
Autonomous Phase Mapping
Autonomous F-scoreSequential F-score
Estimated phase boundary
Theory-based sample
Query
Measured samples
Mis
clas
sifi
cati
on
P
rob
abili
ty
AI is controlling X-ray diffraction systems at SLAC & in the lab!
Solution
Why use AI to just analyze data? Put it on control of the equipment!
Automated Phase Mapping
Synthesize Characterize Analyze
Test case:Combinatorial Library
Physics Knowledge& Databases
Now: Place AI in control of Synthesis.
Questions?