Thermal-expansion mismatch remains a major challenge for the long-term stability of solid oxide fuel-cell and electrolysis-cell materials, particularly for chemically complex perovskite electrodes operating at elevated temperatures. Reliable prediction of thermal expansion in these materials is difficult because anharmonic lattice dynamics, octahedral rotations, chemical disorder, and temperature-driven structural evolution can limit the applicability of conventional phonon-based approaches. Here, we combine quasi-harmonic phonon calculations (QHA) and on-the-fly machine-learning force-field molecular dynamics (MLFF-MD) to evaluate finite-temperature structural behavior in LaFeO${}_3$ and La${}_{0.6}$Sr${}_{0.4}$Co${}_{0.2}$Fe${}_{0.8}$O${}_{3-\delta }$. LaFeO${}_3$ was used as a benchmark material, and the QHA captures the qualitative expansion trend but overestimates the volumetric TEC, giving ${\alpha }_V=4.08\times {10}^{-5}$~K${}^{-1}$ compared with the high-temperature X-ray diffraction value of $3.52\times {10}^{-5}$~K${}^{-1}$. In contrast, MLFF-MD gives ${\alpha }_V=3.56\times {10}^{-5}$~K${}^{-1}$, in close agreement with the experiment. For La${}_{0.6}$Sr${}_{0.4}$Co${}_{0.2}$Fe${}_{0.8}$O${}_{3-\delta }$, QHA becomes unreliable due to phonon instabilities, while MLFF-MD provides stable finite-temperature trajectories and captures the experimentally observed increase in linear TEC across the rhombohedral-to-cubic transition. Specifically, MLFF-MD predicts an increase in ${\alpha }_L$ from $1.31\times {10}^{-5}$~K${}^{-1}$ in the low-temperature rhombohedral regime to $1.90\times {10}^{-5}$~K${}^{-1}$ in the high-temperature cubic regime, consistent with the structural evolution observed experimentally. The remaining quantitative discrepancy with HT-XRD at high temperature is consistent with the stoichiometric model used here, which isolates the intrinsic thermal expansion while excluding chemical-expansion contributions associated with oxygen nonstoichiometry under experimental conditions. Once trained, the MLFF reduces the cost of an individual molecular-dynamics step by roughly two orders of magnitude relative to direct molecular dynamics, enabling substantially longer trajectories and broader sampling of thermally accessible configurations while retaining close agreement with DFT reference energies and forces. These results show that on-the-fly MLFF-MD provides a practical first-principles framework for predicting thermal expansion and phase evolution in disordered solid-oxide perovskites beyond the regime where phonon-based QHA remains reliable.