Scribd
Upload
52 views

Machine Learning Dislocation

This document discusses using machine learning to classify dislocation structures from molecular dynamics simulations. It presents extracting dislocation segments and connected structures from simulation data using the Dislocation Extraction Algorithm. The goal is to classify structures as stacking fault tetrahedra or not. It considers representing the 3D geometry as point clouds and using deep learning methods like PointNet for classification. The work plan is to decide on a descriptor, test machine learning methods, and preprocess the simulation data.

Uploaded by

Terence Deng
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
52 views

Machine Learning Dislocation

This document discusses using machine learning to classify dislocation structures from molecular dynamics simulations. It presents extracting dislocation segments and connected structures from simulation data using the Dislocation Extraction Algorithm. The goal is to classify structures as stacking fault tetrahedra or not. It considers representing the 3D geometry as point clouds and using deep learning methods like PointNet for classification. The work plan is to decide on a descriptor, test machine learning methods, and preprocess the simulation data.

Uploaded by

Terence Deng
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 9

Classification of Dislocation

Structures
Machine Learning in Materials Science
Project introduction

09.10.2018

Jari Järvi, Jarno Laakso, Marvin Schmidt, Xuan Yang


Dataset

● Equiatomic multicomponent alloy


(EAMC) data from MD irradiation
simulations

● 15 x 3500 xyz-files

● 500 000 atoms per file


Dislocation structures

● Dislocation Extraction Algorithm


(DXA) from OVITO

➔ Dislocation segments

➔ Connected dislocation structures


Stacking-Fault Tetrahedra (SFT)

SFT Maybe SFT Definitely not SFT


ML Motivation
● One measure of radiation damage is the concentration of SFTs

● Difficult to count SFTs with conventional methods


○ Especially true for imperfect structures

● Too much data for manual counting


Dislocation structure descriptor
Segment 12: type = 3, length = 8.612626, connected: (261,280)
● How to describe a 3D [[-74.2065506 -30.5234127 87.20332336]
[-72.25637054 -32.31211472 87.74546051]
dislocation geometry in ML? [-67.94432068 -36.35557175 87.70750427]]
Segment 261: type = 3, length = 9.038182, connected: (260,279)
[[-74.12023163 -36.71182632 -95.46377563]
[-71.38674164 -36.96691132 -92.64537048]
[-67.94432068 -36.35557175 -88.92739105]]
● Point clouds? Segment 280: type = 3, length = 8.441553, connected: (260,281)
[[-74.2065506 -30.5234127 -89.43157959]
[-71.96852875 -30.25783157 -91.05818939]
[-67.78326416 -30.6616745 -94.85032654]]
Segment 260: type = 3, length = 8.330929, connected: (12,279)
[[-67.78326416 -30.6616745 -94.85032654]
[-67.36207581 -32.25524521 -93.60269928]
[-67.4023056 -34.1496315 -91.57362366]
[-67.94432068 -36.35557175 -88.92740631]]
Segment 279: type = 3, length = 8.836716, connected: (280,281)
[[-74.12023163 -36.71182632 81.17112732]
[-71.03252411 -33.85820007 80.98238373]
[-67.78326416 -30.6616745 81.78457642]]
Segment 281: type = 3, length = 8.669982, connected: (12,261)
[[-74.12023163 -36.71182632 81.17112732]
[-74.40336609 -34.8782196 83.18630219]
[-74.39068604 -32.92400742 85.05884552]
[-74.2065506 -30.5234127 87.20332336]]
Point cloud descriptor

Pi = (x, y, z, type) Descriptor: D = (P1 , … , Pn )

Qi et al., PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation, arXiv:1612.00593 (2017)
Shabat, 3D Point Cloud Classification using Deep Learning – Recent Works (2017)
ML method

● Objectives
○ Classify a dislocation structure (1 = SFT, 0 = not SFT)

● Supervised?
○ How to label training data? Manually?

● Unsupervised?

● Previous ML studies for example:


○ Dislocation dynamics Salmenjoki, Predicting the behaviour of dislocation systems with
machine learning methods (2017)
○ 3D geometry recognition Bronstein et al., Geometric deep learning: going beyond
Euclidean data, arXiv:1611.08097 (2017)
Work Plan

● Decide on the descriptor and ML method --> Test

● Preprocessing the data (DXA analysis etc.)

You might also like