Machine Learning for Materials Discovery

0 downloads 0 Views 6MB Size Report
May 3, 2018 - 38k alloys: . > 1 mm. 945k alloys: 1 distance > 30at% .... REST API for models. 42. You can do anything I showed today with ...
Machine Learning for Materials Design Logan Ward Postdoctoral Scholar Computation Institute, University of Chicago 3 May 2018

Importance of Materials Design 2

Examples of “Aluminum” Application

Key Properties

Conductivity

Low cost Shapeable

Alloy

Al + Mg

Al + Mg,Mn

Castable Strong

???

Al + Si,Cu

???

How do we tailor materials for new technologies? How can we do this quickly?

Computational Materials Engineering 3

Goal: Accelerate design of materials Method: Replace experiments with computers

Many Established Tools: • • • •

Density Functional Theory Phase Field Finite Element Analysis Computational Thermodynamics

Emerging Field: Data-driven Models Materials Data

Machine Learning

Predictive 𝜎𝑌 = 𝑓(𝑥) Model

FEA Image: http://www.icams.de/content/research/index.html

Machine Learning: What? Why? 4

What? Why?

Algorithms that generate computer programs Create software too complex to write manually 𝑦 = 𝑓(𝑥)

General Task: Given inputs, predict output Common Algorithms: y

𝑓(𝑥) = 𝑚𝑥 + 𝑏 𝑥 0? Training Set Dataset: 3000 entries from the OQMD

Simple ML Model: Accuracy 84.1%

Game: palestrina.northwestern.edu/metal-detection/

Application to the OQMD 13

Dataset: 240000 DFT Calculations (OQMD.org) Eg

𝚫𝐇𝐟

R: 0.944 MAE: 80.5 meV/atom

R: 0.924 MAE: 0.21 eV

Ref: Ward et al. npj Comp. Mat., (2016) 28

V

R: 0.993 MAE: 0.452 Å3/atom

Predicting Glass Forming Ability 14

Application: Metallic Glasses Goal: Predict glass-forming ability Dataset: Landolt-Börnstein 6836 experimental measurements  295 ternary systems  Binary property: [Can Form Glass] | [Cannot Form] 

Model: Random Forest 

90% accurate in 10-fold cross-validation Ref: Ward et al. npj Comp. Mat., (2016) 28

Predicting Glass-Forming Ability 15

Test: Remove Al-Ni-Zr data from training data, try to predict Measured

Predicted X No glass ● Glass

Same representation, very different material. Ref: Ward et al. npj Comp. Mat., (2016) 28

Focus #2: Crystal Structure 16

Our Approach: Voronoi-tessellation-based attributes Atomic Characteristics: 1. Element identity 2. Coordination number 3. Bond length 4. Cell size … Atomic Characteristics + Descriptive Statistics = 275 Attributes Ref: Ward et al. PRB. (2017) 024104

Learning Rate Comparison 17

Dataset: 32k DFT Δ𝐻𝑓 from the OQMD Test: Remove 1000, train on N remaining

Cross-validation is great, but does not model real use

CM: Faber et al. Int J Quantum Chemisty. 2015; PRDF: Schutt et al. PRB. (2014)

Application: The Prototype Search 18

Common Method: Prototype Search 1.

Select a crystal structure

2.

Evaluate all possibilities with DFT Select only stable ones

3.

Challenge: Computational cost Possible Solution: Guide with ML

Ref: Ward et al. PRB. (2017) 024104

DFT

Ranking Candidate Materials 19

Training Set: 32k entries from OQMD Test: Select top 100 entries from test set

Ranks entries better than existing methods, good choice for accelerating combinatorial searches

Ref: Ward et al. PRB. (2017) 024104

Summary 20

Collect

Process

Composition-Based Attributes

Represent

Learn

Crystal Structure Attributes

How do I use this to find a new material?

21

Part 3: Using ML To Find Materials Acknowledgments: Liquidmetal: S. O’Keeffe, J. Stevick, G. Jelbert NIST: J. Hattrick-Simpers SLAC: F. Ren, A. Mehta TRI: M. Aykol

Application: Bulk Metallic Glasses 22

Metals with Amorphous Structures ?

V

Tg

No Dislocations -> High 𝜎𝑌 Glass Transition -> Net Shape Casting No Grain Boundaries -> Low Hysteresis

Applications: Surgical tools, flight control surfaces, …

What are the tradeoffs? Source: liquidmetal.com Source: Wikipedia.org

T

Problems with BMGs 23

Main Problem: Metastability of glassy phase

GFA 𝑫𝒎𝒂𝒙 𝚫𝐓𝐱

Few Alloy Systems Known to Form Glasses Properties Links Our Proposed Solution: Data analytics Goal: Optimize 𝐷𝑚𝑎𝑥 and ΔTx

Materials Informatics Workflow 24

Collect

Process

Represent

Learn

𝑋Ԧ = 𝑓(𝐶𝑜𝑚𝑝𝑜𝑠𝑖𝑡𝑖𝑜𝑛)

“General Purpose” Attributes + Cluster Packing Models + DFT Δ𝐻𝑓 from OQMD GFA

6315 entries

GFA

𝑫𝒎𝒂𝒙

5916 entries

𝑫𝒎𝒂𝒙

𝚫𝐓𝐱

621 entries

𝚫𝐓𝐱

Random Forest Random Forest + Additive Regression

Task #1: Tuning Known Alloys 25

Tuned Alloys: Pareto Analysis 26

Predicted alloys have superior properties

Results: Improved design space

Task #2: Discovering New Alloys 27

Search Space: Ternary alloys of 53 elements 26 million ternary alloys 945k alloys: 𝐿1 distance > 30at%

290k alloys: 𝑃(𝐺𝑙𝑎𝑠𝑠) > 95% 57k alloys: Δ𝑇𝑥 > 66 K 38k alloys: 𝐷𝑚𝑎𝑥 > 1 mm

Run Time: ~ 2 days

Chosen System: Cu-Hf-[Mg,Ti]

Cu-Hf-Mg

Cu-Hf-Ti

28

Cu-Hf-[Ti/Mg]: Current results 29

Cu-Hf-Mg

Cu-Hf-Ti

~4 mm

Sample crystallized and/or shattered during injection molding - Cu60Hf25Ti15 could be formed with copper mold casting*

Mg ignited during arc melting Lessons Learned: • Integrate data from target process • Screen based on processing technique *Inoue et al. Acta Mat. (2001), 2645

Task #3: Sputtered Glasses 30

Goal: Find new sputterable glasses Search Space: Ternary alloys of 24 elements (~2.4 Million Alloys) Method: Combinatorial co-deposition Ding, et al. Nat. Mat. (2014) 2.4 M alloys / 1000 alloys/day -> 6 years Need: Prioritize promising experiments Ref: Ren et al. Sci Adv. (2018), eaaq1566

Step 1: Account for Processing Method 31

Design Goal: Find new sputterable glasses Step 1: Include processing information

Melt Spinning Data

ML Wikipedia

𝐺𝐹𝐴𝑀𝑆 = 𝑓(𝑥𝐻 , 𝑥𝐿𝑖 , … ) “Stacking”

Sputtering Data

ML

𝐺𝐹𝐴𝑆𝑃 = 𝑔(𝑥𝐻 , 𝑥𝐿𝑖 , … , 𝑓(… ))

Ding, et al. Nat. Mat. (2014)

Ref: Ren for et al.processing Sci Adv. (2018),method eaaq1566 by training stacked models Account

Step 1: Account for Processing Method 32

Goal: Find new glasses via sputtering Step 1: Include processing information Melt Spinning

Sputtering

Tune for Sputtering

Ref: Ren et al. Sci Adv. (2018), eaaq1566

Step 2: Adding Theories 33

Yang and Zhang. Mat. Chem. Phys. (2012)

Laws et al. Nat. Comm. (2015)

Mixing Thermo.

K. Laws (UNSW)

Size Mismatch Agreement for Zr ~ 50at%

Disagreement for high/low-Zr

Ref: Ren et al. Sci Adv. (2018), eaaq1566

Next Step: Adding Theories 34

Add Theories

Step 3: Scan 2.4M alloys (~1 hr) -> Selected Co-V-Zr

Ref: Ren et al. Sci Adv. (2018), eaaq1566

Comparison to Experiment 35

J. Hattrick-Simpers F. Ren (SLAC) (NIST)

A. Mehta T. Williams (SLAC) (USC)

Decent agreement. ML model “Zr-lean”, but close enough for success

Repeat, with Improved Model 36 Original Model

Before Co-V-Zr

HiTp Data

Updated Model

After Co-V-Zr

Improvements for Many Systems 37

Test: Leave-Ternary-Out CV Test Set: 673 alloys, 35 ternary diagrams No Sputtering Data No Sputtering Data Landolt Bornstein + All HiTp Data

HiTp Data Improves ML Accuracy for all Systems

Improvements for Many Systems 38

39

Software Infrastructure

Goal: Tight integration with design 40

Future: Adaptive design with automated experiments Problem: Disconnect between method developers and model users

What do we need to share?: • Data • Software/Methods • Models

Need #1: Access to Data Databases Datasets APIs LIMS etc.

EP

Data Discovery

Data Publication

Distributed data storage

• Mint DOIs • Associate metadata • Persist datasets

EP

• Query • Browse • Aggregate

EP

Ian Foster (PI)

Ben Blaiszik

Jonathon Gaff

Need #2: Accessible Methods 42

Language: Java Key Features: 1. Fast implementations of representations 2. Interfaces to ML libraries 3. REST API for models

Language: Python Key Features: 1. Links to materials databases 2. Compatible with Scipy stack 3. Large library of representations

You can do anything I showed today with Magpie and (soon) matminer

Need #3: Sharable ML Models (DLHub) 43 Send Model

Discover Data

Discover Data

Train Model

User 1

Send Materials Retrieve Data User 2

Call Call DLHub DLHub Model Receive Properties New science!

Library of materials ML models in Fall 2018

DLHub

Receive DOI

Summary 44

General-Purpose Representations Using Composition: DFT Δ𝐻𝑓 , 𝐸𝑔 , 𝑉; Glass-forming ability Using Crystal Structure: 2x better than existing approaches

Example applications: ✓ ✓

Improved Δ𝑇𝑥 of 2 commercial BMGs alloys Discovered many new metallic glasses

Open Databases and Software: http://materialsdatafacility.org/ http://bitbucket.org/wolverton/magpie http://github.com/hackingmaterials/matminer

Thanks to our sponsors! ALCF DF

Parsl

Globus

DLHub U.S. DEPARTMENT OF

ENERGY

IMaD

Argonne LDRD