Dust, soot and sea spray can arrive in the same air sample. A team led by Jasmita Khadgi trained an AI system to read each speck through three measurements: its image, size and elemental composition.[1]
In the accepted Scientific Reports study, published on 19 October, attention-based fusion achieved 89.6% accuracy under internal stratified fivefold cross-validation, compared with 61% for the best chemistry-only model and 58% for images alone. The 1,622 labelled particles covered seven classes; testing did not hold out an independent city or campaign.[1]
Keeping the evidence attached to the object
An ingredient list for an entire sample loses the pairing. Suppose two samples contain the same amounts of mineral and salt. In one, they occupy separate grains; in the other, salt coats the mineral. The bulk totals match, but the surfaces exposed to air differ. This is an illustrative distinction, not a result from the new classifier.[2]
Researchers used electron micrographs and X-ray spectra. Their explanatory analysis points to cubic shapes plus sodium and chlorine for sea spray. Carbon-rich signatures sometimes pulled soot towards the organic-rich category, despite soot’s aggregated shape.[1]
The new dataset excluded 49 complex internally mixed particles, along with samples lacking elemental data. Its result covers selected particles with paired measurements; collection and microscopy remain part of the work.[1]
The atmosphere makes mixtures
Earlier fieldwork shows why that pairing matters. In a 2022 study of 239 samples collected at Zeppelin Observatory in Svalbard between 2017 and 2021, Kouji Adachi and colleagues found mineral dust carrying sea salt and soot coated with sulfates. Their microscopy combined images with elemental maps, preserving the connection between shape and ingredients.[2]
Sea salt was present in 35 of 82 examined dust particles recovered from cloud droplets or ice crystals, compared with 23 of 125 dust particles sampled from ambient air. The researchers left open how the attached salt influences ice formation.[2]
A mixture complicates a request for one label. Calling a coated grain “dust” may retain its core material while concealing its surface. Calling it “sea salt” can lose the mineral underneath. The category chosen should preserve the property relevant to the scientific question.[2]
Sources & further reading
- Advancing automatic classification of single fine particles via multimodal deep learning framework Khadgi and colleagues / Scientific Reports · 19 Oct 2026 · Accessed 19 Oct 2026
- Composition and mixing state of Arctic aerosol and cloud residual particles from long-term single-particle observations at Zeppelin Observatory, Svalbard Adachi and colleagues / Atmospheric Chemistry and Physics · 10 Nov 2022 · Accessed 19 Oct 2026