The moment tensor potentials (MTPs) are trained, tested and validated for various single- and mixed-type point defects in LaZrO pyrochlores using the training datasets produced from first-principles molecular dynamics simulations. The reliability of the trained MTPs for defective LaZrO pyrochlores is demonstrated by the good agreement between the DFT method and machine learning potentials in calculating structural parameters, elastic constants and radial distribution functions. The temperature renormalized phonon spectra are obtained for all studied structures, and the results are compared with phonon dispersions calculated using harmonic lattice dynamics, showing the importance of temperature effects on the dynamical stability of defective crystal structures. Phonon modes in defective LaZrO pyrochlores are further classified into propagons, diffusons and locons based on the calculated phonon participation ratio (PR) and the Ioffe-Regel limit for each vibrational mode, revealing a broad distribution of wavelike vibrational states in defective structures. The lattice thermal conductivity and phonon quasi-particle lifetimes are calculated using both the perturbative linear Boltzmann transport equation (LBTE) and fully non-perturbative equilibrium molecular dynamics coupled with the Green-Kubo formula (EMD-GK) for temperature ranging from 300 K to 1500 K. The results reveal that some single- and mixed-type point defects including V48fO, Sn16cZr, Yb16cZr + Hf16cZr + V48fO and Yb16cZr + Sn16cZr + V48fO are the most prominent point defects in reducing phonon lifetimes, thus suppressing the lattice thermal conductivity and improving the performance of thermal barrier coatings (TBCs).