Machine-learning interatomic potentials with global charge embeddings and multifidelity training predict the ground-state structures and thermodynamics of charged defects in the solar absorber ${\text{Sb}}_2{\text{Se}}_3$ with hybrid-DFT accuracy.