Physics-informed Bayesian Optimization for Quantitative High-Resolution Transmission Electron Microscopy

This paper introduces a physics-informed Bayesian optimization framework that automates and significantly accelerates the quantification of atomic structures in high-resolution transmission electron microscopy, achieving a three-to-four order-of-magnitude improvement in time efficiency for determining 3D crystal structures from single images.

The Gist

Imagine you are trying to figure out the exact layout of a tiny, intricate city made of atoms. You have a single, slightly blurry photograph of this city taken from above. The problem is that the camera lens is imperfect (it has scratches and distortions), the photo is a bit grainy, and the city itself is 3D, but you only have a flat 2D picture.

In the past, figuring out the city's layout from this photo was like trying to solve a massive jigsaw puzzle by hand, one piece at a time, while blindfolded. You'd guess a piece's position, see if it fits, guess again, and repeat this thousands of times. It took experts days or even weeks to get it right, and they could only do it for a tiny corner of the city.

This paper introduces a super-smart, physics-savvy robot that can solve this puzzle in minutes instead of weeks. Here is how it works, broken down into simple concepts:

1. The Problem: The "Blindfolded Puzzle Solver"

High-Resolution Transmission Electron Microscopy (HRTEM) is a powerful microscope that lets us see individual atoms. But the images it produces are often distorted by the microscope's own imperfections (like a warped lens) and noise (like static on an old TV).

To understand what the atoms are actually doing, scientists have to run computer simulations to guess what the image should look like if the atoms were in a certain arrangement. They compare the simulation to the real photo. If they don't match, they tweak the guess and try again.

• The Old Way: A human expert would tweak a few settings, check the result, tweak a few more, and repeat. It was slow, tedious, and prone to human error.
• The Bottleneck: Because there are so many variables (lens settings, atom positions, thickness, etc.), this process is like searching for a needle in a haystack while the haystack is constantly moving.

2. The Solution: The "Smart Detective" (Bayesian Optimization)

The authors created a new method called Physics-Informed Bayesian Optimization (BO). Think of this not as a blind guesser, but as a super-smart detective.

• The Detective's Strategy: Instead of checking every single possibility (which would take forever), the detective uses a "probabilistic engine." It makes an educated guess about where the answer might be, checks that spot, and then uses the result to update its map of the "haystack."
• The Physics Twist: What makes this detective special is that it knows the rules of physics. It knows, for example, that atoms can't just float in empty space or that certain arrangements are energetically impossible. It uses these rules to instantly rule out bad guesses, saving massive amounts of time.
• The "Trust Region": Imagine the detective focuses on a small neighborhood (a "Trust Region") where the answer is most likely to be. If they find a good clue there, they expand the neighborhood. If they hit a dead end, they shrink the search area to focus tighter. This keeps the search efficient and prevents the detective from getting lost in a giant, confusing city.

3. The "Continuous vs. Discrete" Trick

One tricky part of the puzzle is that some variables are continuous (like the exact angle of a lens) and some are discrete (like the number of atoms in a column—you can't have 2.5 atoms).

• The Analogy: Imagine trying to tune a radio. You can turn the dial smoothly (continuous), but you can only pick whole stations (discrete).
• The Fix: The researchers invented a clever math trick (Continuous Relaxation) that treats the "whole stations" as if they were smooth dials for a moment. This allows the detective to use smooth, fast math to find the best spot, and then snaps the result back to a whole number. It's like using a smooth slider to find the perfect volume, then snapping it to the nearest preset.

4. The Result: From Weeks to Minutes

The team tested this on a crystal called Barium Titanate (BaTiO3), which contains heavy, medium, and light elements.

• The Achievement: They took a single 2D image and reconstructed the 3D atomic structure of the entire sample.
• The Speed: They did this 3 to 4 orders of magnitude faster than before. That means a task that used to take weeks or months now takes minutes.
• The Discovery: Because the method was so fast and accurate, they discovered something new: the atoms at the edge of the crystal were behaving differently than in the middle. The "electric polarization" (a key property of the material) was fading away at the edges, likely because the crystal was so thin. This is like noticing that the lights in a city dim out as you get closer to the edge of the map.

Why Does This Matter?

This isn't just about solving puzzles faster.

• Digital Twins: It allows scientists to create perfect "digital twins" of materials—exact 3D computer models that match reality.
• Watching Change: Because it's so fast, scientists can eventually use this to watch materials change in real-time (like watching a chemical reaction happen atom-by-atom) rather than just taking a snapshot and waiting days to analyze it.
• Automation: It paves the way for microscopes that can automatically analyze samples without needing a human expert to sit there and tweak knobs for weeks.

In short: The authors built a "smart detective" that uses the laws of physics to solve the complex puzzle of atomic structures in minutes, turning a slow, manual process into a fast, automated one. This opens the door to understanding materials in ways we never could before.

Technical Summary

1. Problem Statement

Quantitative High-Resolution Transmission Electron Microscopy (HRTEM) is essential for understanding atomic-scale structure-property relationships. However, extracting accurate atomic structural information from HRTEM images is hindered by:

• Imaging Artifacts: Residual aberrations, partial coherence, noise, and detector imperfections distort the image.
• Non-linearity: The relationship between observed peak intensity and atomic column configuration is non-linear due to electron phase-contrast imaging.
• Computational Bottleneck: Traditional model-based quantification relies on a multi-step, user-supervised iterative approach. This involves manually adjusting imaging parameters (e.g., defocus, astigmatism) and structural parameters (e.g., atomic positions, occupancies) to minimize the difference between experimental and simulated images.
• High Dimensionality: The parameter space is hyper-dimensional, especially when determining 3D structures of nanoscale crystals. Optimizing these parameters manually is tedious, time-consuming (taking days to weeks), and restricts analysis to small sample areas and experienced users.

2. Methodology

The authors propose a Physics-informed Bayesian Optimization (BO) framework to automate and accelerate HRTEM quantification. The core components include:

A. Bayesian Optimization Framework

• Surrogate Model: Instead of exhaustive sampling, the method uses a Gaussian Process (GP) regression model to approximate the objective function (Mean Squared Error between experimental and simulated images).
• Acquisition Function: A parallel Upper Confidence Bound (qUCB) function is used to balance exploration and exploitation, selecting the next batch of parameters to evaluate.
• Trust Region (TuR): A Trust Region Bayesian Optimization (TuRBO) strategy dynamically adapts the search domain. If the error decreases significantly, the region expands; otherwise, it contracts to focus on promising areas. This mitigates the "curse of dimensionality."

B. Physics-Informed Constraints & Handling Mixed Parameters

• Continuous Relaxation (CR): The optimization involves both continuous parameters (e.g., defocus) and discrete parameters (e.g., integer number of atoms in a column). The authors introduce a CR method using a Bernoulli distribution to map discrete variables to a continuous proxy. This avoids the high computational cost of Monte Carlo methods while maintaining differentiability for the acquisition function.
• Physical Constraints: Before generating new candidate parameters, the algorithm filters out physically impossible configurations (e.g., atoms violating geometric constraints or energy rules), ensuring the search remains within physically meaningful bounds.

C. Stepwise Optimization Workflow

To manage complexity, the workflow is divided into three stages:

1. Global Imaging Parameter Matching: An averaged experimental image (constructed from multiple Regions of Interest, ROIs) is used to optimize global imaging parameters (aberrations, image spread) and an average sample thickness.
2. Parallel 3D Structural Determination: The 3D structure of individual ROIs is determined in parallel. This step optimizes local thickness, local absorption factors, local defocus, and discrete atomic counts (occupancies) for each column.
3. Integrated 3D Reconstruction: The individual ROI models are stitched together. Local defocus differences are interpreted as relative $z$-position shifts to construct a unified supercell model for final validation.

3. Key Contributions

• Automation of HRTEM Quantification: Replaced the manual, user-supervised iteration with a fully automated BO loop.
• Efficiency in High-Dimensional Spaces: Successfully optimized a parameter space containing ~48 individual variables (17 unknown parameters) by breaking the problem into manageable sub-problems and using TuRBO.
• Handling Discrete Variables: Developed a computationally efficient Continuous Relaxation method specifically tailored for discrete atomic counts in BO, avoiding expensive Monte Carlo sampling.
• 3D Structure from 2D Images: Demonstrated the ability to reconstruct the full 3D atomic structure (including thickness variations and atomic occupancies) from a single 2D HRTEM image.

4. Results

The method was validated using Barium Titanate (BaTiO$_3$) single crystals containing heavy (Ba), medium (Ti), and light (O) elements.

• Speed Improvement: The BO framework achieved a 3 to 4 orders of magnitude improvement in time efficiency.
  • Traditional Method: Days to weeks.
  • BO Method: ~75 seconds for global imaging parameters; ~300 seconds for full 3D structural determination of 12 ROIs.
• Accuracy:
  • The Mean Squared Error (LMSE) converged to ~0.0007, slightly above the vacuum noise level (0.0004), indicating an excellent fit.
  • Normalized Cross-Correlation (NCC) peaks reached 98.1% for averaged regions and 95–98% for individual ROIs.
  • The reconstructed 3D model correctly identified a wedge-shaped thickness variation (from ~2.1 nm to ~3.5 nm) caused by sample preparation.
• Physical Insights:
  • The method revealed a reduction in Ti–O relative displacement at the edges of the BaTiO$_3$ specimen.
  • This indicates a local suppression of ferroelectric distortion and a tendency toward a cubic-like configuration in thin regions, attributed to depolarization fields and finite-size effects.
  • The approach was also successfully applied to YAlO$_3$ (YAO), demonstrating robustness across different material systems.

5. Significance

• Scalability: The method enables the analysis of larger sample volumes and complex heterogeneous structures, which were previously intractable due to computational time.
• In Situ/Operando Applications: The drastic reduction in post-acquisition analysis time (from weeks to minutes) makes it feasible to track structural evolutions (e.g., phase transitions, defect migration) occurring on the scale of seconds to minutes, paving the way for automated, real-time structural quantification.
• Digital Twins: The ability to rapidly reconstruct accurate 3D atomic models from single images provides a robust pathway for generating "digital twins" of materials for simulation and design.
• Generalizability: By incorporating physical priors and constraints, the approach overcomes the data-hunger and generalization issues common in pure deep-learning approaches, making it suitable for diverse experimental conditions.