Transferability of data-driven optimization results across multiple pixelated CdZnTe spectrometers

This study demonstrates that machine learning-optimized binary voxel masks trained on one CdZnTe spectrometer can generalize effectively to other detectors with only a marginal performance loss, suggesting that a single common mask can significantly reduce the labor and data collection efforts required for optimizing multiple spectrometers in safeguards applications.

The Gist

Imagine you have a team of six highly specialized detectives (the CdZnTe spectrometers) tasked with identifying a specific type of evidence (uranium) in a room. Each detective has a unique set of eyes (pixels) to scan the room.

Ideally, every eye sees the evidence perfectly. But in reality, some eyes are blurry, some are tired, and some are looking at the wrong angle. If you let all the eyes work together, the blurry ones create "noise" that makes it harder to spot the evidence clearly.

The Problem: Too Many Choices

The team has thousands of eyes. To get the best result, you want to tell the blurry eyes to "take a break" and only use the sharp ones. But there are so many combinations of eyes you could turn off that it would take a computer longer than the age of the universe to figure out the perfect combination by trying them all.

The Solution: The "Smart Filter"

In a previous study, the researchers built a smart computer program (called spectre-ml) that acts like a coach. Instead of trying every combination, the coach looks at the team, groups the eyes that perform similarly, and quickly figures out which group of eyes to keep and which to turn off to get the clearest picture.

This coach creates a "Mask"—a digital stencil that says, "Use these eyes; ignore those ones."

The Big Question: Can One Coach Train Everyone?

The researchers asked a crucial question: If we train this coach on Detective #1, can we give that same "Mask" to Detectives #2 through #6, or do we need to train a new coach for every single detective?

Training a new coach takes hours of work and special data for every single detector. If we could just use one universal mask for all of them, it would save a massive amount of time and effort.

The Experiment: The "Gym" Test

The researchers took six different detectors and ran them through a rigorous test (like a gym workout) using a standard uranium source.

1. The "Bespoke" Approach: They trained a unique coach for each detector. (This is the "perfect" but slow method).
2. The "Transfer" Approach: They trained a coach on one detector and tried to use that same mask on the other five.

The Results: A Surprising Success

Here is what they found, using simple analogies:

• The Detectors are Different: Just like people have different eye shapes, the detectors had significant variations. Some pixels were much "blurrier" than others.
• The "Perfect" Mask Wins (But barely): As expected, the mask trained specifically on a detector worked slightly better than a mask borrowed from a different detector. It was like a custom-tailored suit fitting slightly better than an off-the-rack one.
• The "Universal" Mask is Still Great: However, the borrowed masks were almost as good as the custom ones.
  • The custom masks improved performance by about 16%.
  • The borrowed (transferred) masks improved performance by about 13%.

The Analogy: Imagine you are trying to hear a whisper in a noisy room.

• The "Bulk" method (using all eyes) is like trying to hear the whisper while everyone in the room is shouting.
• The "Custom Mask" is like asking the specific people in the room who are shouting to leave, leaving only the quiet ones.
• The "Universal Mask" is like using a list of "likely shouters" based on a different room. It's not perfectly accurate, but it still silences enough noise to hear the whisper clearly.

Why This Matters

The researchers found that even though the detectors were different, the "bad" pixels were bad in similar ways. A mask that tells Detector A to ignore its "blurry corners" will likely tell Detector B to ignore its "blurry corners" too.

The Bottom Line:
You don't need to spend hours training a new AI for every single detector you buy. You can train it once on a few samples, create a "Universal Mask," and apply it to hundreds of detectors.

This is a huge win for nuclear safety inspectors. It means they can get better, clearer data faster without needing to run complex, time-consuming tests on every single device they use. It turns a custom-tailored suit into a high-quality, ready-to-wear uniform that fits almost everyone perfectly.

Technical Summary

1. Problem Statement

The H3D M400 is a highly segmented Cadmium Zinc Telluride (CdZnTe) gamma-ray spectrometer widely used for non-destructive assay (NDA) in nuclear safeguards. While its fine spatial segmentation allows for precise interaction localization, inherent crystal irregularities and proximity variations to readout electrodes cause significant performance heterogeneity among individual voxels (pixels).

• The Challenge: Poorly performing voxels degrade spectroscopic metrics (e.g., energy resolution, peak amplitude uncertainty). Excluding them improves these metrics but reduces counting statistics (efficiency). Finding the optimal subset of voxels to exclude is a combinatorial problem with $\sim 10^{7285}$ possible combinations, making brute-force search impossible.
• The Gap: Previous work introduced spectre-ml, a data-driven approach using machine learning (ML) and heuristic algorithms to find optimal voxel masks for a single detector. However, the process of collecting training data and optimizing a mask is time-consuming (several hours per detector). The critical question remained: Can a mask optimized on one detector be effectively transferred to other detectors to avoid redundant optimization efforts?

2. Methodology

The study utilized the spectre-ml software framework to generate and test binary voxel masks across multiple datasets.

• Optimization Framework (spectre-ml):

  • Clustering: Instead of evaluating every voxel individually, the method groups voxels into clusters based on similar performance.
  • Algorithms: It employs various strategies to form clusters:
    • Unsupervised ML: Agglomerative clustering, Gaussian Mixture Models (GMM), K-means.
    • Greedy/Heuristic: Ranking voxels by performance and accumulating them; grouping by depth bins or physical location (edges/anode).
    • Non-negative Matrix Factorization (NMF): Used to decompose voxel spectra into components before clustering.
  • Performance Metric: The primary metric was the relative uncertainty in the Doniach fit amplitude of the 186 keV U-235 photopeak. The Doniach function models the asymmetric peak shape typical of CZT detectors.
  • Trade-off: The algorithm searches for the cluster combination that minimizes the uncertainty metric, balancing the gain from removing noisy voxels against the loss of total counts.

• Datasets:

  • Inter-Detector Study: Six M400 detectors from different US National Laboratories (BNL, INL, LANL, ORNL, PNNL, SNL) were measured using uranium standards (1.94% and 20.11% enrichment) in a shielded, collimated setup.
  • Intra-Detector Study: Three LBNL M400 units were measured with an Eu-154 check source to assess stability.
  • Validation: Data was split 80/20 (train/test). Masks trained on one detector (or subset) were tested against all other detectors to evaluate transferability.

3. Key Contributions

1. Quantification of Transferability: The paper provides the first rigorous evaluation of how well optimized voxel masks generalize across different M400 units, despite significant inter-detector hardware variations.
2. Subsampling Stability Analysis: The authors demonstrated that the optimization results are robust to statistical fluctuations (Poisson noise) and reductions in dataset size (downsampling by 5x), ensuring that high-quality masks can be found without massive data collection efforts.
3. Operational Efficiency Proposal: The study proposes a workflow where a "common" mask, derived from a small sample of detectors, can be applied to a fleet of deployed units, eliminating the need for per-detector optimization.

4. Key Results

• Intra- and Inter-Detector Variability:
  • Significant spatial variations in gross count rates ($\lesssim 30\%$) were observed both within and between detectors, often spanning $\sim 1$ cm.
  • Despite these local variations, total integrated count rates across the six detectors were consistent within a 2.4% standard deviation.
• Transferability Performance:
  • Bespoke vs. Transferred: The best mask trained and tested on the same detector (bespoke) yielded an average 16% improvement in the uncertainty metric over the unoptimized (bulk) detector.
  • Cross-Detector Success: The best transferred mask (trained on one detector, applied to all six) achieved an average 13% improvement.
  • Conclusion: The performance loss from using a transferred mask instead of a bespoke one is minimal (only ~3% absolute difference in improvement), suggesting that a single "universal" mask can be highly effective.
• Robustness:
  • Subsampling analysis showed that masks trained on 20% of the data or tested on subsamples produced metric values with narrow distributions, confirming the stability of the optimization process against data volume changes.
• Efficiency Trade-off:
  • Applying the optimal masks reduced the relative efficiency of the detectors to roughly 30–36% of the bulk (unoptimized) efficiency. However, this was more than compensated for by the reduction in systematic fit errors, leading to a net gain in measurement precision.

5. Significance and Implications

• Safeguards Efficiency: The findings suggest that nuclear safeguards agencies can significantly reduce operational overhead. Instead of collecting hours of training data and running optimization for every new detector in the field, they can apply a pre-validated, high-performing mask derived from a representative sample.
• System Integration: The authors propose integrating these binary masks directly into the M400 data acquisition chain (as a second readout stream) rather than relying on offline post-processing. This would allow for real-time optimization without modifying raw data storage.
• Future Directions: While the current metric (Doniach amplitude uncertainty) is a robust proxy, the authors recommend future work using application-specific metrics (e.g., enrichment precision) and expanding the parameter sweeps and detector sample size to further refine generalizability.

In summary, the paper demonstrates that data-driven voxel selection is not only effective for individual detectors but is also highly transferable, offering a scalable solution to enhance the spectroscopic performance of large fleets of CdZnTe spectrometers.