The document discusses using machine learning to design point defects in 2D materials for applications in quantum information science and nonvolatile memory. It describes using deep transfer learning on over 100,000 defect structures to predict defect properties like formation energy and electronic structure. This high-throughput screening identified over 100 quantum emitters and 15 atomically thin memristors. The document concludes that layered metal chalcogenides, hexagonal nitrides, and metal halides are promising materials for defects that could enable quantum emitters and nonvolatile resistive memory.
1. Machine Learning-Enabled Design
of Point Defects in 2D Materials
Nathan C. Frey
Shenoy Group, Materials Science & Engineering
University of Pennsylvania
2020 MRS Spring/Fall Meeting
n.frey@seas.upenn.edu
2. How do we engineer defect states
that are electrically, magnetically,
and optically addressable in solid-
state devices?
3. Why layered materials?
3(EPFL/G.Pizzi)
Tunable electronic, magnetic,
topological properties
Low-power, ultra-compact, flexible
Heterostructure and integrated solid-
state devices
n.frey@seas.upenn.edu
4. Point defects in 2D materials
4
Atomically thin resistive
memory (atomristors)
Deji Akinwande (UT Austin); Pawel Latawiec (Harvard University)
Quantum emitters
n.frey@seas.upenn.edu
5. Deep transfer learning in 2D
5
Pretrained
Model
~104 – 105 Bulk Crystals
~103 2D Materials
𝐸𝑓
𝑚𝑎𝑡
, 𝐸 𝑏𝑔, 𝐸 𝐹𝑒𝑟𝑚𝑖
Deep
Transfer
Learning 2D Structure Graphs
B
N
Mo
S
𝑅2 = 0.98
𝑀𝐴𝐸 = 0.06
Predicted Ef
mat (eV/atom)
TrueEf
mat(eV/atom)
n.frey@seas.upenn.edu
6. High-throughput screening
6
150 materials x 70 defects
≈ 10,000 defect structures
Ni
Cu
Pd
Ag
Al
Au
Electronic Structure
Machine
Learning
Engineered
Point Defects
~1,000
Band
Structures
~100
A
BB
B B
MTarget 1
Formation/binding energy
𝐸
𝒌
𝐸 𝑏𝑔Target 2
Defect level position
n.frey@seas.upenn.edu
7. Machine learning defect properties
7
Conduction Band
Valence Band
Δ𝐶𝐵 > 𝑘 𝐵 𝑇
Δ𝑉𝐵 > 𝑘 𝐵 𝑇
Target 1
Target 2
⋅⋅⋅
Random Forest Model
Host Material
Properties
Defect
Properties
Average All
Predictions
Tree 1 Tree n
Prediction 1 Prediction n
Prediction i
𝑟𝑆𝑛 ≫ 𝑟𝑆𝑒
𝜒 𝑆𝑛 ≪ 𝜒 𝑆𝑒
Sn Se
n.frey@seas.upenn.edu
8. Minimal physical model of defects
8
M
𝜖
1
𝑟
e-
𝐸 𝐸𝜖 Strain terms
Electrostatic terms
𝜖
𝑞2/𝑟
0.67 eV MAE for defect formation energy
0.93 F1 score for shallow vs deep defect level classifier
n.frey@seas.upenn.edu
10. Nonvolatile resistive memory score
10
High
stability
Low switching
voltage
100+ quantum emitters
15 atomically thin memristors
n.frey@seas.upenn.edu
11. Takeaways
11
1
Point defects in 2D materials are platforms for quantum
information science and nonvolatile memory.
2
Layered metal chalcogenides, hexagonal nitrides, and metal
halides including GeS, h-AlN, and MgI2 are promising,
unexplored systems.
n.frey@seas.upenn.edu