M&M 2026: Closing the Feedback Loop: Live Polarization Analysis for Automated In-Situ 4D-STEM

Transcript

1. Closing the Feedback Loop: Live Polarization Analysis for Automated In-Situ 4D-STEM Fintan Hardy – f.hardy24@imperial.ac.uk
2. The evolution of information storage for computing Punched cards (1890s) Information stored with physical holes Semiconductor memory (1950-present) Information stored with voltage/current state Ferroelectric Memory (present-) Symmetry breaking of crystal Fast but amnesiac Quick and non-volatile Permanent but inert Ferroelectric memory for quick and energy efficient storage Imperial College London N. Hill, J Phys Chem B (2000) 2 12/08/2026
3. Ferroelectric materials overview Polar Domains – key ferroelectric property Polar vortices in [ 𝐏𝐛𝐓𝐢𝐎𝟑 𝟏𝟔 / 𝐒𝐫𝐓𝐢𝐎𝟑 𝟏𝟔 ]8 PFM polarisation domains of 𝑪𝒖𝟑 𝑩𝟕 𝑶𝟏𝟑 𝑪𝒍 Why 4D-STEM? Domain wall: The narrow interface separating regions of a ferroelectric material with differently orientated spontaneous polarisation Imperial College London Y. Shao et al. Nature Communications, (2023) R.G.P. McQuaid et al. Nature Communications (2017) 3 • Walls are only a few unit cells • Captures full diffraction pattern per pixel • Can map polarization as vector • In-situ biasing 12/08/2026
4. Automated microscopy Imperial College London S.V. Kalinin et, al. ACS Nano (2021) 4 12/08/2026
5. Automated 4D-STEM analysis Capture and analyse data Capture data Waiting for microscope time… Analyse data (off microscope) Decide next experiment Decide next experiment Current Workflow (Days) Imperial College London Now-possible Workflow (Minutes) 5 12/08/2026
6. Problems with analysing ferroelectric 4D-STEM data Common methods for ferroelectric analysis Polarisation mapped according to disk intensities Polarisation mapped according to Kikuchi band intensities 𝑷𝒃𝑻𝒊𝑶𝟑 /𝑺𝒓𝑻𝒊𝑶𝟑 [ 𝐏𝐛𝐓𝐢𝐎𝟑 𝟏𝟔 / 𝐒𝐫𝐓𝐢𝐎𝟑 𝟏𝟔 ]8 K. Nguyen et al. Microscopy & Microanalysis, (2018) Imperial College London Y. Shao et al. Nature Comms, (2023) 6 12/08/2026
7. Problems with analysing ferroelectric 4D-STEM data Sample tilt Natural sample tilt θ Inaccuracy from user Intrinsic bending of homogeneous material Limitations on stage tilt accuracy Imperial College London 7 12/08/2026
8. Problems with analysing ferroelectric 4D-STEM data Sample tilt θ P P Domain wall example 𝑪𝒖𝟑 𝑩𝟕 𝑶𝟏𝟑 𝑪𝒍 Boracite AFM Being on zone all the time is impossible for certain domain boundaries Imperial College London 8 12/08/2026
9. Problems with analysing ferroelectric 4D-STEM data How sensitive to tilt are these methods? Imperial College London F.G. Hardy et, al. In preparation 9 5 Å−1 0.5 Å−1 Polarisation signal using Kikuchi band method Polarisation signal using Friedel disk pair comparison method 12/08/2026
10. Problems with analysing ferroelectric 4D-STEM data How sensitive to tilt are these methods? Traditional symmetry breaking analysis in electron diffraction patterns become unreliable at extremely small tilting angles! We need quicker and more robust methods of analysing polarisation in 4D-STEM datasets! Imperial College London F. Hardy et, al. In preparation 10 12/08/2026
11. How can we reliably map polarisation of 4D-STEM datasets Existing routes around tilt, and what each one costs Multi-direction scanning in PFM Beam precession (4D-STEM) 0 mrad 20 mrad 47 mrad simulated Frame based precession (4D-STEM) None of these fixes are compatible with live, in-situ acquisition! Imperial College London K. Chu et, al. Rev Sci. instruments (2018) P. A. Midgley & A. S. Eggman, Topical Reviews (2014) S. M. Ribet, Microscopy &Microanalysis, (2025) 11 12/08/2026
12. Mapping polarisation with new reflections instead of intensity Superlattice peaks Cu Cu Cu Cu 0.25 Å-1 0.25 Å-1 Basically immune to tilting effects! Allows us to study boracite at different domain boundaries Can also be used to map ferroelasticity (e.g. 𝑆𝑟𝑅𝑢𝑂3 ) Imperial College London 12 12/08/2026
13. Analysing domain wall datasets Training a machine learning model Trained on ~100,000 simulated boracite diffraction images 0.25 Å-1 Predicted polarisation direction (Ground truth polarisation direction) Imperial College London 13 12/08/2026
14. Polarisation mapping on CuCl boracite – manual analysis vs Ghost hunter model 50 nm HAADF of 𝑪𝒖𝟑 𝑩𝟕 𝑶𝟏𝟑 𝑪𝒍 0.25 Å-1 0.25 Å-1 50 nm Polarisation mapping with ghost-hunter model (8,908 images/second on Nvidia RTX A6000 GPU) Imperial College London 14 12/08/2026
15. Imperial College London 15 12/08/2026
16. Ghost hunter examples In-Situ polarisation mapping / Strain mapping 𝐂𝐮𝟑 𝐁𝟕 𝐎𝟏𝟑 𝐂𝐥 Boracite 5D-STEM dataset Imperial College London 16 12/08/2026
17. Overview 1. Current methods for mapping ferroelectric materials are highly sensitive to tilt 2. New methods for polarization analysis in 4D-STEM datasets required 3. Live mapping on microscope with advanced vision models now possible Imperial College London 17 12/08/2026
18. Thank You Supervisors: Shelly Conroy, Aron Walsh WMD-Group - Informatics team Conroy group - Mariana Palos - Noah Schnitzer - Issa Saddiq - Geri Topore - Eda Colliver Fintan Hardy – f.hardy24@imperial.ac.uk Imperial College London 18 12/08/2026
19. Model Comparisons Model Params Mean Error Images/second Vision Transformer ResNetCNN 85.6 M 1.49˚ 1,315 11.6 M 2.29˚ 8,908 Equivariant CNN (GCNN) 6.0 M 4.69˚ 207 Imperial College London 19 12/08/2026
20. Tilt effects on ghost hunter model Imperial College London 20 12/08/2026