Speaker
Description
Raman peak assignment in 2D and related materials is slow and error-prone, and language models can confidently invent symmetry-mode labels. We present Phaselenser, a physics-informed pipeline that converts Raman spectra into verified mode assignments and 3D atomic-displacement visualisations. Spectra are baseline-corrected and peak-picked (pybaselines, SciPy). Qwen then assigns each peak a symmetry label or “artifact” using literature-derived mode cards (MoS₂, quartz, calcite, anatase, silicon). A deterministic checker blocks forbidden modes and enforces one label per peak, and a Marimo interface renders the results. Unassignable peaks are reported as unassigned rather than guessed. Trained on 500 synthetic spectra, with calcite and 632.8 nm excitation held out, the LoRA fine-tuned model reached 94.4% mode accuracy (980/1038) and 100% artifact identification (196/196), including 99.8% on held-out calcite.
Other Fields of Research (Track) for Abstract
Material Science
| Nature of work | Computational |
|---|