Machine Learning prediction of maximum phonon frequency in inorganic fluoride and oxide compounds for rare-earth-doped nanoparticle host screening
Preprint
- 16 June 2026
- preprint
- Published by Elsevier in SSRN Electronic Journal
Abstract
Rare-earth-doped nanoparticles are versatile photonic platforms whose performance in upconversion luminescence imaging and photothermal therapy is controlled to a large extent by the maximum phonon frequency of the host lattice ωmax. Hosts with ωmax below approximately 400 cm-1 suppress non-radiative relaxation and favour bright upconversion emission, whilst those above 600 cm-1 promote efficient phonon-assisted heat generation. Despite this importance, systematic ωmax data are available for only a small number of benchmark compounds, and density functional perturbation theory calculations are too expensive and time-consuming to apply across the full inorganic chemical space. Here we train an XGBoost regressor on 1154 inorganic compounds with density functional perturbation theory phonon data from the Materials Project database, using 49 compositional, structural, and electronic descriptors. The model achieves a held-out test R2 of 0.975 and a mean absolute error of 30.6 cm-1. Inference on 55243 rare-earth-compatible fluoride and oxide compounds identifies 2702 upconversion luminescence candidates and 39338 photothermal therapy candidates. SHapley Additive exPlanations (SHAP) analysis reveals that the minimum bond length is the single most informative descriptor, consistent with the bond stiffness origin of optical phonon frequencies. A transfer learning experiment in which the model is pre-trained on Debye temperatures before fine-tuning on ωmax produces lower test accuracy (R2 = 0.937), demonstrating that acoustic-mode proxies and optical phonon cutoffs carry distinct structural information that should not be conflated in multi-target phonon models. Per-anion calibration against experimental Raman phonon energies for eleven benchmark host compounds yields a mean absolute error of 18.4 cm-1, with seven of eleven predictions falling within reported literature ranges, and correctly reproduces the fluoride-versus-oxide phonon energy ranking.Keywords
This publication has 40 references indexed in Scilit:
- Locations of the energy levels of lanthanide ions in Gd2O3 single crystalsOptical Materials, 2023
- Decorating rare-earth fluoride upconversion nanoparticles on AuNRs@Ag core–shell structure for NIR light-mediated photothermal therapy and bioimagingJournal of Rare Earths, 2022
- Synthesis of Yb3+/Ho3+ co-doped Y2O3 nanoparticles and its application to dye sensitized solar cellsJournal of Molecular Structure, 2021
- DScribe: Library of descriptors for machine learning in materials scienceComputer Physics Communications, 2020
- Machine learning models for the lattice thermal conductivity prediction of inorganic materialsComputational Materials Science, 2019
- Luminescence decay-based Y2O3:Er phosphor thermometry: Temperature sensitivity governed by multiphonon emission with an effective phonon energy transitionJournal of Luminescence, 2019
- Strong emission via up-conversion of Gd2O3:Yb3+, Ho3+ nanopowders co-doped with alkali metals ionsJournal of Luminescence, 2014
- Fluoride materials for optical applications: Single crystals, ceramics, glasses, and glass–ceramicsJournal of Fluorine Chemistry, 2011
- Thermal expansion, molar volume and specific heat of diamond from 0 to 3000kJournal of Electronic Materials, 1996
- Raman spectra of lanthanide sesquioxide single crystals with A‐type structurePhysica Status Solidi (b), 1979