Novelty-Driven Target-Space Discovery in Automated Electron and Scanning Probe Microscopy

This paper introduces BEACON, a deep-kernel-learning framework that enables automated electron and scanning probe microscopy to actively discover novel scientific behaviors in target spaces by learning structure-property relationships during experiments, a method validated through rigorous offline benchmarking and successfully deployed on a scanning transmission electron microscope (STEM).

The Gist

The Big Problem: The "Needle in a Haystack" Dilemma

Imagine you are a detective trying to solve a mystery in a massive, dark warehouse filled with millions of boxes. You have a flashlight (the microscope) that can show you what the boxes look like on the outside (the image). But the real clues—the secret ingredients inside—are hidden and can only be found by opening a box and running a complex chemical test (the spectrum).

The Catch:

1. Opening a box and testing it takes a long time.
2. If you open too many boxes, you might damage the warehouse (beam damage).
3. You don't know which boxes contain the clues. They aren't labeled.

The Old Way (Classical Microscopy):
Traditionally, scientists would try to be thorough. They would open boxes in a strict grid pattern (Box 1, Box 2, Box 3...) across the whole warehouse. This is slow, expensive, and often misses the rare, weird boxes that hold the most exciting discoveries because they were buried in the middle of a boring area.

The New Way (The BEACON Framework):
The authors built a "Smart Detective" robot (an AI) that doesn't just open boxes randomly or in a grid. Instead, it learns as it goes. It looks at the outside of the boxes, guesses what might be inside, and then decides: "I've seen enough boring boxes; let's go find something totally weird and new."

---

How the "Smart Detective" Works

The paper introduces a system called BEACON (which stands for a fancy acronym, but think of it as a Bayesian Evolutionary Analysis for Cosmic Observation Networks). Here is how it thinks, broken down into three simple steps:

1. The "Feature Extractor" (The Eyes)

The robot uses a special brain (Deep Kernel Learning) that looks at the image of the box. It doesn't just see "a box"; it sees patterns, textures, and shapes. It learns to say, "This texture usually means something interesting is inside."

2. The "Elite Set" (The Hall of Fame)

As the robot tests boxes, it keeps a "Hall of Fame" of the most interesting results it has found so far.

• Old AI (The Optimizer): If the robot finds a box with a "good" result, it gets greedy. It thinks, "This is the best spot! I will stay here and test this one spot 50 times to be sure." This is bad because it misses other cool things elsewhere.
• BEACON (The Explorer): BEACON asks, "Is this result new?" It compares the new result to its "Hall of Fame." If the result is just "okay" or "similar to what I already have," it ignores it. But if the result is weird, unique, or different from everything in the Hall of Fame, it gets excited.

3. The "Gambler's Instinct" (Thompson Sampling)

Sometimes the robot isn't 100% sure what's inside a box. Instead of guessing the average, it uses a technique called Thompson Sampling. Imagine rolling a dice to guess the contents.

• If the robot is very unsure about a region, the dice roll might suggest a "crazy" outcome.
• BEACON loves "crazy" outcomes because they represent novelty. It sends the microscope to that spot just to see if the crazy guess was right. This prevents the robot from getting stuck in one boring area.

---

The Analogy: The Music Festival

Think of the microscope as a music festival with 10,000 stages.

• The Images are the posters on the doors.
• The Spectra are the actual music playing inside.

The Old Strategy (Grid Search):
You walk down every single aisle, peeking into every room for 10 seconds. You miss the hidden underground bands because you were too busy checking the main stage.

The "Optimizer" AI:
You hear a great song in one room. You decide, "This is the best song ever!" You stand in that one room for the rest of the day, listening to the same song over and over, ignoring the rest of the festival.

The BEACON Strategy:
You listen to a song. If it sounds like the last 10 songs you heard, you move on. But if you hear a genre you've never heard before (Jazz in a Rock festival?), you stop and investigate immediately. You are trying to map out every different type of music in the festival, not just find the "best" song. You want to know: "What weird sounds exist here that no one has ever heard?"

---

What Did They Prove?

The researchers tested this "Smart Detective" in two ways:

1. The Simulation (Offline): They used old data where they already knew the answers (like a map of the warehouse). They showed that BEACON found more "weird" and diverse spots much faster than the greedy "Optimizer" robots. The Optimizer got stuck in one corner; BEACON explored the whole map.
2. The Real Deal (Online): They actually hooked this AI up to a real, expensive electron microscope in a lab. They told it to look for specific types of nanoparticles.
   • Result: The AI successfully found the rare particles without human help. It didn't get stuck; it roamed the sample, finding diverse structures just like it did in the simulation.

Why Does This Matter?

In the past, "Discovery" in science often meant hoping to get lucky while looking at a grid of data.

• Before: "Let's scan this whole area and hope we see something cool later."
• Now: "Let's send a robot that actively hunts for the unknown."

This paper shows that we can teach microscopes to be curious. Instead of just being a camera that takes pictures, the microscope becomes a scientist that asks questions, learns the answers, and decides where to look next to find the most surprising discoveries. It turns the microscope from a passive tool into an active partner in discovery.

Technical Summary

1. Problem Statement

Modern automated microscopy (STEM and SPM) faces a fundamental challenge in materials discovery: the most critical scientific information often resides not in the immediately visible structural features (images), but in the target space of sequentially acquired spectra or functional responses (e.g., EELS, PFM hysteresis loops).

• Limitations of Current Methods: Traditional automated strategies rely on Bayesian Optimization (BO) for optimization, which seeks to maximize a known objective function or minimize prediction uncertainty. These methods often suffer from "trajectory collapse," where the algorithm rapidly converges on a single local optimum or a narrow region of the sample, failing to explore diverse physical behaviors.
• The "Scalarizer Bottleneck": Defining a reward function (scalarizer) requires deep domain expertise. Furthermore, standard BO focuses on optimizing the value of the scalarizer, rather than discovering novel behaviors or rare phases that may not be the global maximum but are scientifically significant.
• Constraints: Brute-force dense grid sampling is often impossible due to time constraints and beam damage risks.

2. Methodology: BEACON-DKL Framework

The authors propose BEACON (Bayesian Evolutionary Analysis for Cosmological Observation Networks), adapted here for microscopy using Deep Kernel Learning (DKL). The framework is designed to actively search for novelty in the target space rather than optimizing a specific value.

Core Architecture

1. Deep Kernel Learning (DKL) Surrogate Model:

   • Feature Extraction: A Convolutional Neural Network (CNN) processes raw image patches (e.g., HAADF-STEM or PFM topography) to extract high-dimensional structural features.
   • Latent Space Mapping: The CNN compresses these features into a low-dimensional latent space.
   • Gaussian Process (GP): A GP is trained on the latent features to predict the spectral/functional response (scalarizer). The kernel is a "deep kernel" where the input is the CNN output, allowing the model to learn structure-property relationships end-to-end.

2. BEACON Acquisition Function:

   • Unlike standard acquisition functions (e.g., Expected Improvement) that seek high values, BEACON seeks diversity in the target space.
   • Elite Set: The algorithm maintains an "elite set" ($\mathcal{E}$) of the top-performing measurements based on a user-defined physical criteria.
   • Novelty Score: For a candidate point $x$, the acquisition value $\alpha_{BEACON}(x)$ is calculated as the average distance to its $k$-nearest neighbors within the elite set in the target (response) space.
   • Thompson Sampling: To handle uncertainty, the algorithm draws stochastic samples from the posterior distribution of the GP rather than using the mean. This allows the system to explore regions with high uncertainty that might yield novel responses.

3. Active Learning Loop:

   • Initialize with random seed points.
   • Train DKL model on current data.
   • Identify the elite set and compute novelty scores for unmeasured locations.
   • Select the location with the highest novelty score.
   • Acquire new data, update the model, and repeat.

3. Key Contributions

• Shift from Optimization to Discovery: The paper introduces a paradigm shift from "optimizing known objectives" to "discovering unknown behaviors" by prioritizing novelty in the target space.
• Benchmarking Framework: The authors established a rigorous set of monitoring functions to evaluate discovery algorithms, including:
  • Target Space Coverage: The fraction of the ground-truth response distribution visited.
  • Latent/Feature Space Coverage: The diversity of structural motifs explored (using VAEs).
  • Surrogate Model Behavior: Analysis of Mean Absolute Error (MAE) and uncertainty stability.
• Operational Deployment: The workflow was successfully transitioned from offline validation on pre-acquired datasets to real-time implementation on a physical Spectra 300 STEM microscope.
• Open Source: All code, notebooks, and benchmarking tools are made publicly available to facilitate community adoption.

4. Results

The method was validated on two distinct datasets and one live experiment:

A. Scanning Probe Microscopy (PFM)

• Dataset: Ferroelectric domains in Lead Titanate (PTO).
• Comparison: BEACON vs. Expected Improvement (EI) and Maximum Uncertainty (MU).
• Findings:
  • Trajectory Collapse: EI and MU quickly collapsed into localized regions of the image, repeatedly sampling the same microstructural features.
  • BEACON Performance: BEACON maintained a spatially distributed trajectory, exploring diverse domain patterns.
  • Metrics: BEACON achieved significantly higher Target Space Coverage and Latent Space Coverage (diversity of structural motifs) throughout the 300-step budget. It also produced a more stable surrogate model with lower variance in MAE compared to the fluctuating EI model.

B. Electron Microscopy (STEM-EELS)

• Dataset: Plasmonic nanoparticles.
• Findings: While all methods eventually covered similar ranges of scalar values (Target Space), BEACON explored the structural manifold (feature space) much more broadly. EI and MU remained trapped in specific regions of the sample, whereas BEACON uniformly traversed the sample, identifying diverse structural contexts for the same physical response.

C. Live STEM-EDX Experiment

• Setup: Autonomous acquisition on a mixed nanoparticle library (CdS nanorods, CdSe nanoplatelets, Quantum Dots).
• Scalarizer: Integrated Selenium (Se) signal.
• Outcome: BEACON demonstrated superior patch space coverage compared to EI and MU over 100 steps. It successfully distributed measurements across structurally distinct regions of the sample, confirming the algorithm's ability to operate in real-time hardware environments.
• Efficiency: Computational overhead was negligible (~0.05s for BEACON acquisition vs. ~3s for hardware acquisition), making it viable for live experiments.

5. Significance

• Enabling "Unknown Unknowns" Discovery: The framework allows microscopes to autonomously identify rare defects, emergent phases, or unexpected physical regimes without requiring a pre-defined "best" target.
• Overcoming Human Bias: By decoupling the exploration strategy from the immediate visual appearance of the image and focusing on the response space, the algorithm avoids human bias toward visually obvious features.
• Scalability: The integration of DKL with active learning provides a scalable path for high-throughput materials discovery, moving beyond exhaustive grid scanning.
• Community Impact: By providing a transparent benchmarking suite and open-source code, the authors set a new standard for evaluating and comparing active learning algorithms in automated microscopy, fostering an "open era" of algorithmic-driven discovery.

In summary, this work demonstrates that novelty-driven acquisition is superior to traditional optimization strategies for exploring complex, heterogeneous materials, enabling the discovery of diverse physical behaviors that would otherwise be missed by algorithms prone to local convergence.