Online Beam Current Estimation in Particle Beam Microscopy

This paper proposes and validates a novel online method for estimating particle beam current directly from secondary electron count data using time-resolved measurements and joint estimation techniques, thereby eliminating the need for offline calibration and significantly improving micrograph accuracy and sample milling outcomes.

The Gist

Here is an explanation of the paper "Online Beam Current Estimation in Particle Beam Microscopy," translated into simple language with creative analogies.

The Big Picture: The "Flickering Flashlight" Problem

Imagine you are trying to take a high-resolution photo of a tiny, delicate object (like a virus or a microchip) using a super-powerful microscope. But instead of light, this microscope uses a beam of tiny particles (like ions) to scan the surface.

The Problem:
Think of the beam like a flashlight. To get a perfect photo, the flashlight needs to shine with a steady, constant brightness. However, in these microscopes, the "flashlight" often flickers or changes brightness unexpectedly.

• Why? The source gets old, or dirt builds up inside the machine.
• The Result: When the beam gets brighter, the photo looks too bright in that spot. When it dims, it looks dark. Because the microscope scans line-by-line (like a printer), these brightness changes create ugly horizontal stripes across your image, ruining the picture.

Usually, scientists have to stop the machine, take it apart, and calibrate it (like tuning a radio) to fix the brightness. This is slow, expensive, and stops work.

The Solution:
This paper proposes a clever trick: Don't fix the flashlight; just figure out how bright it is while you are taking the picture.

The authors developed a mathematical "detective" that looks at the data the microscope is already collecting and says, "Wait, I can tell exactly how much the beam flickered at every single moment, and I can correct the image instantly."

---

How It Works: The "Rain and Umbrella" Analogy

To understand how they do this, let's use an analogy.

Imagine you are standing outside in the rain, holding an umbrella.

• The Raindrops are the particles in the beam (the beam current).
• The Puddles you see forming are the "Secondary Electrons" (the signal the microscope detects to make the image).
• The Ground is the sample you are looking at.

The Old Way:
You assume it's raining at a steady rate. You count the puddles and say, "Ah, this ground is wet because it's a rainy day."

• The Flaw: If the rain suddenly gets heavier (beam current spikes), you think the ground is just really wet (high yield). If the rain stops, you think the ground is dry. You can't tell the difference between a "wet ground" and "heavy rain." This leads to the stripe artifacts.

The New Way (Time-Resolved Measurement):
Instead of waiting for the whole minute to pass, you look at the rain in tiny, rapid bursts (milliseconds).

• You count how many drops hit your umbrella in the first millisecond, then the next, then the next.
• Even if the rain is heavy, you might get a burst where no drops hit. If the rain is light, you might get a burst where many drops hit.
• By looking at the pattern of these tiny bursts, you can mathematically reverse-engineer two things simultaneously:
  1. How wet the ground actually is (The Sample's properties).
  2. How hard it was actually raining (The Beam Current).

The paper proves that by breaking the measurement time into these tiny "sub-bursts," you have enough information to solve for both variables at the same time, even though they are mixed together.

---

The "Detective's Toolkit": Two Different Cases

The authors realized that different microscopes have different "personality types" regarding how their beams flicker. They created two different detective strategies for each type:

1. The "Smooth Jazz" Beam (Helium & Electron Beams)

Some beams flicker smoothly, like a jazz musician slowly changing volume. The brightness drifts up and down gradually.

• The Strategy: The algorithm assumes the beam current is "smooth." If the current was 10 units at the last pixel, it's probably around 10.1 or 9.9 at the next one, not 50.
• The Trick: They use a mathematical "smoothness filter" (Total Variation Regularization). It's like telling the detective: "Don't guess that the rain suddenly turned into a hurricane unless you have overwhelming proof. Assume it's a gentle drizzle that slowly gets heavier."
• Result: This removes the stripes and creates a crystal-clear image, even if the beam was fluctuating wildly.

2. The "Light Switch" Beam (Neon Beams)

Some beams (like Neon ion beams) are notorious for being unstable. They don't drift; they just snap back and forth between two specific settings (like a flickering light switch: ON, OFF, ON, OFF).

• The Strategy: The algorithm treats this like a Hidden Markov Model. It's a game of "Guess the State."
• The Trick: The algorithm knows the beam can only be at "Level A" or "Level B." It looks at the pattern of the rain (the data) and asks: "Is it more likely the beam is at Level A right now, or Level B?" It uses the history of previous pixels to predict the next one.
• Result: It can pinpoint exactly when the beam switched states and correct the image instantly.

---

Why This Matters: The "Superpower"

Why should anyone care about fixing stripes in a microscope image?

1. No More "Stop and Fix": You don't need to stop the machine to calibrate it. The microscope fixes itself in real-time.
2. Better "Milling": These microscopes are often used to cut or carve tiny structures (like making microchips). If the beam gets too strong, it burns the sample. If it's too weak, the cut is shallow. Knowing the exact beam current allows the machine to adjust its speed automatically to prevent damage.
3. New Instruments: There are powerful new microscopes (like Neon beam microscopes) that are amazing but hard to use because their beams are so unstable. This technology makes them usable for everyone, not just experts.

The Bottom Line

The authors took a problem where two unknowns (Beam Brightness and Sample Texture) were mixed together like a smoothie, making it impossible to taste the ingredients separately.

By tasting the smoothie in tiny, rapid sips (Time-Resolved Measurement) and using smart guessing rules (Markov models and smoothness filters), they proved you can separate the ingredients perfectly. They built a software "detective" that sees the invisible flickering of the beam and cleans up the image instantly, turning a blurry, striped mess into a perfect, high-definition picture.

Technical Summary

Here is a detailed technical summary of the paper "Online Beam Current Estimation in Particle Beam Microscopy" by Seidel et al.

1. Problem Statement

In particle beam microscopy (including Scanning Electron Microscopes (SEM) and Focused Ion Beam (FIB) microscopes using Helium, Neon, or Gallium ions), image quality and milling accuracy depend critically on a stable beam current ($\lambda$).

• The Issue: Beam intensity often fluctuates due to source aging or contamination. When the beam is raster-scanned across a sample, these fluctuations create horizontal "stripe artifacts" in the resulting micrograph.
• Current Limitations: Conventional microscopes do not measure beam current in real-time. They rely on offline calibration or assume a constant current. Post-processing algorithms exist to remove stripes, but they do not provide the actual beam current value, which is essential for controlling sample damage and milling rates.
• The Challenge: Estimating both the Secondary Electron (SE) yield ($\eta$, which determines image contrast) and the beam current ($\lambda$) simultaneously from a single noisy measurement is theoretically impossible without additional assumptions, as the expected signal is the product $E[Y] = \lambda\eta$.

2. Methodology

The authors propose estimating $\lambda$ and $\eta$ online using the same data used to form the image, leveraging Time-Resolved (TR) measurements.

A. Measurement Models

Instead of a single cumulative count per pixel, the dwell time is split into $n$ shorter sub-acquisitions.

• Conventional Model: Measures total SEs $Y = \sum X_i$. This follows a Neyman Type A distribution where $\lambda$ and $\eta$ are inseparable.
• Time-Resolved (TR) Model:
  • Continuous-Time (CT): Idealized model with perfect temporal resolution of SE bursts.
  • Discrete-Time (DT): Practical model where the pixel dwell time is divided into $n$ sub-intervals. The vector of counts $\mathbf{y}_k = [y^{(1)}_k, \dots, y^{(n)}_k]$ contains rich information about the variance of the process, allowing the separation of $\lambda$ and $\eta$.

B. Theoretical Analysis (Cramér–Rao Bound)

The authors derived the Cramér–Rao Bound (CRB) for the joint estimation of $\eta$ and $\lambda$.

• Key Finding: At high SE yields ($\eta$), the joint estimation problem is only marginally more difficult than estimating $\eta$ when $\lambda$ is known.
• Low Yield Challenge: At low $\eta$, the problem becomes significantly harder, requiring higher doses or inter-pixel correlations to achieve accurate estimates.

C. Estimation Algorithms

The paper proposes Maximum Likelihood (ML) and Maximum A Posteriori (MAP) estimators tailored to different beam behaviors:

1. Single-Pixel Estimators:

   • Derive ML estimates for $\eta$ and $\lambda$ using CT and DT models.
   • Limitation: Requires high dose for accuracy, especially at low $\eta$.

2. Smoothly Varying Beam Current (Electron & Helium Ion Beams):

   • Model: Beam current is modeled as a first-order Gaussian Autoregressive (AR) process (smoothly varying).
   • Algorithms:
     • Causal: Estimates current at pixel $k$ using measurements up to $k$ and the previous estimate $\hat{\lambda}_{k-1}$.
     • Non-Causal: Uses the entire image sequence (Forward-Backward approach) for global optimization.
     • Regularization: Both causal and non-causal versions incorporate Total Variation (TV) regularization on $\eta$ to enforce piecewise smoothness in the micrograph, improving reconstruction quality.

3. Discrete Jumping Beam Current (Neon Ion Beams):

   • Model: Beam current toggles between two known discrete values (modeled as a Hidden Markov Model).
   • Algorithms:
     • Causal: Uses the Forward algorithm to compute belief states.
     • Non-Causal: Uses the Forward-Backward algorithm to compute the most probable state sequence given all data.

D. Mathematical Tools

• The paper provides novel numerical methods for handling the Neyman Type A distribution.
• It derives expressions for the derivatives of the log-likelihood using Touchard polynomials, which are essential for implementing gradient-based optimization (e.g., gradient descent, proximal gradient methods).

3. Key Contributions

1. Feasibility Proof: Demonstrated via CRB analysis that joint estimation of beam current and SE yield is theoretically feasible using TR data, particularly at moderate-to-high SE yields.
2. Algorithm Development: Proposed causal and non-causal joint estimators that exploit spatial correlations (AR models) and temporal correlations (Markov models) of the beam current.
3. Superior Performance: Showed that incorporating explicit beam current estimation outperforms state-of-the-art post-processing methods (like frequency-domain filtering) and standard baseline methods.
4. Mathematical Derivations: Provided closed-form derivatives and approximations for Neyman Type A likelihoods using Touchard polynomials, enabling efficient numerical optimization.

4. Results

The methods were tested on synthetic microscopy data simulating Helium Ion Microscopes (HIM) and Scanning Electron Microscopes (SEM).

• Image Quality (SE Yield $\eta$):

  • The proposed joint estimators (especially with TV regularization) produced micrographs with Root Mean Squared Error (RMSE) nearly indistinguishable from the "oracle" case (where true beam current is known).
  • HIM Example: The non-causal estimator with TV regularization achieved an RMSE of 0.2298, significantly outperforming the baseline (1.1129) and the frequency-domain filter (0.9817).
  • SEM Example: Similar improvements were observed, reducing RMSE from ~0.097 (baseline) to ~0.032 (non-causal with TV).
  • Stripe artifacts were effectively eliminated.

• Beam Current Estimation ($\lambda$):

  • The estimated beam currents closely matched the ground truth.
  • Neon Beam (Discrete): The non-causal HMM estimator achieved a 0.21% error rate in identifying the correct beam state, compared to 0.89% for the causal estimator.
  • Helium Beam (Continuous): The causal and non-causal estimators tracked the smooth variations of the beam current with low RMSE.

• Robustness: The algorithms were shown to be robust to variations in tuning parameters (regularization strength) and correlation coefficients.

5. Significance and Impact

• Artifact Removal: Eliminates horizontal stripe artifacts without needing offline calibration or post-processing that assumes constant current.
• Instrument Diagnostics: Provides real-time feedback on beam health, allowing operators to detect source degradation or contamination immediately.
• Milling Control: Accurate online knowledge of beam current allows for dynamic adjustment of dwell times, preventing sample damage and improving milling precision.
• Enabling New Technologies: Specifically addresses the stability challenges of Neon Ion Beam microscopes, potentially accelerating their adoption by solving the critical issue of unstable beam currents.
• Generalizability: The framework applies to any particle beam system where TR measurements can be acquired, offering a path toward fully autonomous, self-calibrating microscopy instruments.