Calculating vibrational entropy for large atomic systems (e.g., 100,000 atoms) is usually a daunting task that requires massive computer memory and time. To solve this bottleneck, this presentation introduces a fast surrogate model with linear scaling cost. First, we show that the total vibrational entropy of a system can be broken down into smaller, local parts called "site entropies". We also prove a "locality property," which means these site entropies mainly depend on nearby atoms and only weakly on far-away ones. Because of this local behavior, we don't need to look at the whole system at once. Instead, we can use the Atomic Cluster Expansion (ACE) method to directly learn and fit the entropy from these local atomic sites. Finally, we tested our model on Silicon configurations, and the results show that it converges quickly and makes highly accurate predictions.