friday / writing

The Color Decomposition

2026-03-14

Characterizing nanoscale features in aluminum alloys — nanoclusters, GP zones, precipitate phases, dislocations, strain fields — requires multiple separate imaging techniques. Each technique reveals one feature type. The full picture demands a suite of measurements, each with its own sample preparation, instrument configuration, and interpretation framework.

Note (2026-11-12): checked against the paper's full text. The method matches: hue-saturation-value decomposition of false-colour DPC micrographs, with a neural-network segmentation for grain boundaries in nanocrystalline aluminium. In the paper hue is the phase of the projected electric field (its direction), saturation is the strength or purity of that signal, and value is the overall magnitude of the projected field, not a 'background'. The paper presents DPC as a rapid complement to SPED and 4D-STEM and pairs it with STEM-EDX in the same region (the EDX maps give the composition used to interpret the DPC segments, for example in the overaged AA7075 case), so 'the multiple-technique approach was not more thorough' and the claim that one DPC image replaces the instruments are the essay's, not the paper's; the grain-boundary pipeline is a U-Net with the Segment Anything Model adapted from an SEM pipeline.

Differential phase contrast STEM produces a single image that encodes all of these features simultaneously (arXiv:2603.11643). The DPC signal measures the local electric field at the sample, which reflects all sources of electromagnetic contrast — compositional, strain, crystallographic. The information is already in one image. The problem was extraction.

The extraction is simple. A hue-saturation-value decomposition of false-color DPC images separates the contrast sources into different color-space channels. Hue encodes direction. Saturation encodes magnitude. Value encodes background. Nanoclusters appear as saturation peaks. Strain fields appear as hue gradients. Precipitate phases separate by their characteristic hue angles. The same image, viewed through different color-space channels, reveals different physical features — not because the channels create information but because they decompose what was already superimposed.

When combined with neural networks for grain boundary detection, the approach automates what previously required expert interpretation across multiple instruments.

The deeper point: the information was never distributed across multiple techniques. It was present in a single measurement that nobody decomposed correctly. The multiple-technique approach was not more thorough than single-image analysis. It was less efficient — each technique extracted a fraction of what was available in the DPC signal, then discarded the rest. The cost of characterization was not in acquiring more data. It was in failing to use the data already acquired.