🔬 optics

Reducing acquisition time and radiation damage: data-driven subsampling for spectro-microscopy

This paper introduces data-driven subsampling strategies that leverage spectral and spatial importance to reconstruct full spectro-microscopy datasets from as little as 4–6% of measurements, thereby significantly reducing acquisition time and radiation damage while preserving essential chemical information.

Original authors: Maike Meier, Lorenzo Lazzarino, Boris Shustin, Hussam Al Daas, Paul Quinn

Published 2026-02-05

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Maike Meier, Lorenzo Lazzarino, Boris Shustin, Hussam Al Daas, Paul Quinn

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to take a high-resolution photograph of a delicate, ancient painting. You want to see every tiny brushstroke and color variation, but the bright light of your camera flash is so intense that it starts to fade the paint the longer you look at it. In the world of science, this is exactly the problem researchers face with a technique called spectro-microscopy.

Scientists use this technique to "see" the chemical makeup of materials (like energy materials or biological samples) by shining X-rays on them. To build a complete picture, they usually have to scan the sample row-by-row, like a printer head moving across a page, at dozens of different energy levels. This takes a long time and bombards the sample with so much radiation that it can get damaged or destroyed before the scan is finished.

This paper proposes a clever solution: Don't scan everything. Just scan the most important parts, and use math to fill in the rest.

Here is how the paper breaks it down, using simple analogies:

1. The Problem: The "Full Scan" is Too Slow and Damaging

Think of a spectro-microscopy experiment like trying to map a city by walking every single street at every hour of the day.

The Issue: If you walk every street (scan every pixel) at every hour (every energy level), it takes forever. Plus, the "sun" (the X-ray beam) is so hot that the city (the sample) might burn up before you finish.
The Goal: The researchers wanted to walk only a few streets, but still be able to draw a perfect map of the whole city.

2. The Secret: The Map Has "Shortcuts"

The paper explains that these chemical maps aren't actually as random or complex as they look. They have a hidden structure.

The Analogy: Imagine the city map is actually made of just three main colors (Red, Blue, and Green) mixed in different amounts. Even though the map looks like millions of different shades, it's really just a combination of those three base colors.
The Math: Because the data is built from a few "core ingredients" (chemical states), it is low-dimensional. This means there is a lot of redundancy. If you know the pattern of the Red areas, you can guess where the Blue areas are. You don't need to measure every single pixel to understand the whole picture.

3. The Solution: "Smart Sampling" vs. "Random Guessing"

Previous methods tried to solve this by picking random rows to scan (like closing your eyes and pointing at the map). This worked okay, but you still had to scan about 10–20% of the map to get a good result.

The authors propose two new Data-Driven strategies. Instead of guessing, they use the data itself to decide what to measure.

Strategy A: The "Highlighter" Method (RISS)

How it works: The researchers first take a few quick, full scans to get a "rough draft" of the map. They analyze this draft to see where the most interesting changes are happening (where the colors are shifting the most).
The Action: They then go back and scan only the rows that are most "important" or "active."
The Result: They fill in the missing gaps using a mathematical algorithm (called LoopedASD) that acts like a smart auto-complete feature, predicting the missing pixels based on the patterns it found in the important rows.
The Win: They can get a high-quality map by scanning only 4–6% of the data, instead of 10–20%.

Strategy B: The "Cornerstone" Method (CURISS)

How it works: This method is based on a mathematical trick called CUR decomposition. Imagine you have a giant spreadsheet of data. This method says, "If I measure just a few specific rows and a few specific columns (the corners and edges), I can reconstruct the entire spreadsheet."
The Action: They use the "importance" data to pick the absolute best rows and columns to measure.
The Result: Because they are picking the most informative pieces, they can reconstruct the whole image even more reliably at very low sampling rates (as low as 3–4%).
The Win: This method is faster and more stable than the "highlighter" method when you are scanning very little data.

4. The "Adaptive" Upgrade

The paper also suggests a way to be even smarter. Imagine you start with a small sketch. If the sketch looks a bit blurry, you don't just guess; you add a few more specific details where the blur is worst.

The researchers created an Adaptive version (ACURISS) that checks its own work. If the reconstruction isn't good enough, it automatically decides to scan one more "important" row to fix the error, then stops when the picture is clear enough.

The Bottom Line

The paper claims that by using these data-driven methods, scientists can:

Drastically cut scan times: They can get the same information by measuring only 4% to 6% of the usual data points.
Save the sample: Because the X-ray beam is on for a fraction of the time, the sample suffers much less radiation damage.
Keep the quality: Despite measuring so little, the reconstructed images are just as accurate as the full scans for identifying chemical states.

In short, instead of reading every single word in a book to understand the story, these methods teach the computer to read the most important sentences and chapters, allowing it to perfectly summarize the rest of the story without ever reading the boring parts. This saves time and protects the precious book (the sample) from falling apart.

Technical Summary: Reducing Acquisition Time and Radiation Damage via Data-Driven Subsampling for Spectro-Microscopy

Problem Statement
Spectro-microscopy, which combines X-ray Absorption Spectroscopy (XAS) with microscopy to map chemical states and spatial variations, is often constrained by long acquisition times and radiation damage to samples. A standard experiment involves rastering a sample through a monochromatic X-ray beam across multiple energy levels, generating a dataset of $n_E$ spectra for $N$ spatial pixels. This results in a massive data matrix $D$ of size $n_E \times N$ . The authors note that while techniques like Principal Component Analysis (PCA) reveal that these datasets possess an underlying low-dimensional structure (due to a limited number of chemical states), the mechanical necessity of scanning full spatial rows (or columns) for every energy level leads to significant redundancy. Previous work by Townsend et al. (2022) demonstrated that random raster sampling could reduce the measurement ratio to 10–20% using matrix completion, but the authors argue that further reduction is possible through more intelligent, data-driven selection strategies.

Methodology
The paper proposes two novel, data-driven subsampling strategies that leverage the low-rank nature of spectro-microscopy data to select which measurements to acquire. Both methods rely on the concept of leverage scores, which quantify the "information richness" of specific rows (energies) and columns (spatial locations) in the data matrix.

Spectral Importance Distribution: To determine which energy levels to sample, the authors utilize a dictionary of known absorption spectra ( $M_{dictionary}$ ). By calculating the leverage scores of this dictionary, they construct a probability distribution that prioritizes energies explaining the most variance. If a dictionary is unavailable, uniform random sampling is used as a fallback.
Spatial Importance Distribution: To select which spatial rows to scan, the authors first acquire a few full raster scans at the selected energies. These scans are stacked horizontally to form a matrix $F$ , where rows correspond to spatial sample rows. An Adaptive Randomized Pivoting (ARP) algorithm is then applied to $F$ to iteratively select spatial rows with high leverage scores, ensuring that diverse and information-rich areas of the sample are prioritized.

Based on these distributions, the authors introduce two specific acquisition schemes:

RISS (Raster Importance Sampling for Spectro-microscopy): A data-driven variant of the random raster sampling approach. For each energy, a subset of spatial rows is measured according to the spatial importance distribution. The missing data is reconstructed using the LoopedASD (Alternating Gradient Descent) matrix completion algorithm.
CURISS (CUR Importance Sampling for Spectro-microscopy): A scheme designed specifically to utilize CUR matrix decomposition. It measures a set of full XRF scans (rows of the data matrix) and a set of spatial rows (blocks of columns) based on the importance distributions. The missing entries are reconstructed using the CUR factorization ( $D \approx CU^{-1}R$ ), which offers a non-iterative, optimization-free completion method.

The paper also introduces ACURISS, an adaptive version of CURISS. This approach starts with a small subsampling ratio and iteratively refines the approximation by adding new XRF scans until specific stopping criteria are met. These criteria include the "completion variation" (change in the reconstructed matrix) and "spectral variation" (change in the estimated cluster spectra) between iterations.

Key Contributions

Data-Driven Sampling: The authors move beyond uniform random sampling by proposing methods that use statistical leverage scores to identify and prioritize the most informative measurements.
Two Novel Schemes: The introduction of RISS (compatible with iterative matrix completion) and CURISS (optimized for CUR decomposition).
Adaptive Framework: The development of ACURISS with practical stopping criteria that allow the experiment to terminate once sufficient accuracy is achieved, rather than relying on a pre-defined, fixed budget.
Theoretical Justification: The paper establishes that the leverage scores of the exact data matrix are equivalent to those of the spectral dictionary, providing a theoretical basis for using prior spectral knowledge to guide sampling.

Results
Experiments were conducted on the I14 Hard X-ray Nanoprobe beamline at Diamond Light Source using mixtures of hematite and magnetite. The authors evaluated performance using "completion error" (Frobenius norm distance between the reconstructed and full dataset) and "spectral error" (distance between cluster spectra derived from the full vs. reconstructed data).

Subsampling Efficiency: The data-driven methods (RISS and CURISS) significantly outperformed uniform random sampling. While random sampling struggled to produce satisfactory reconstructions below 10% subsampling, the data-driven approaches maintained high accuracy at ratios as low as 4–6%.
Method Comparison: CURISS demonstrated superior performance at very low subsampling ratios (below 8%), particularly in preserving the spectral information required for cluster analysis. RISS performed better than random sampling but was generally outperformed by CURISS at the lowest ratios.
Robustness: The methods remained effective even when a spectral dictionary was unavailable, though CURISS showed greater robustness than RISS in the absence of prior spectral knowledge at very low sampling ratios (<6%).
Adaptive Performance: The ACURISS strategy successfully converged to accurate results using stopping criteria based on the variation between iterations, validating the feasibility of adaptive experimental design.

Significance and Claims
The authors claim that their data-driven approaches allow for the acquisition of spectro-microscopy data with 4–6% of the measurements required for a conventional scan while preserving the essential chemical and spatial information. This represents a substantial improvement over the 10–20% ratios previously achievable with random sampling.

The paper highlights that the CUR-based approach (CURISS) is particularly significant because it combines high reconstruction accuracy with computational simplicity and speed, avoiding the heavy optimization costs associated with gradient-descent-based matrix completion. The authors note that this is the first application of CUR decomposition specifically for experimental design in spectro-microscopy. They conclude that these methods have the potential to significantly reduce experiment times and radiation doses, thereby broadening the applicability of XAS spectro-microscopy to large regions and in-situ processes where sample degradation is a limiting factor.