Machine-Learning Phonon Spectra for Defect Luminescence

Defects in solids show promise for quantum networking and computation, but predicting their luminescence from first principles requires expensive evaluation of all phonon modes in supercells containing hundreds of atoms. We demonstrate a procedure for fine-tuning machine learning interatomic potentials to accelerate the calculation with negligible loss in accuracy. Crucially we show that the inherently required structural relaxation of the defect to obtain the equilibrium geometry provides a small dataset that is sufficient for foundation model fine-tuning. We also assess structure-generation procedures to supplement the relaxation dataset when higher accuracy is needed. We benchmark the approach against explicit computation and experimental data for several defects. In particular, we are able to accurately describe the fine details of coupling to local vibrational modes in the luminescence spectrum of the T center in Si, a prominent quantum defect.