With the increasing energy consumption of data centers around the world, the recovery of low-grade waste heat in data centers has become increasingly important. To maximize the waste heat recovery of thermoelectric modules in data centers, physics-based machine learning optimization of the thermoelectric assembly was carried out. A theoretical model and machine learning model of a fan-heat sink-thermoelectric generator system for waste heat recovery were proposed for predicting the performance of a thermoelectric assembly. Genetic algorithms were used to find the best heat sink geometric parameters for the possible greatest power generation. Given a temperature difference constraint of 75 °C between the heat source and the environment, the greatest output power from the thermoelectric assembly was 3.836 (watts), with the optimal plate fin height of 60 (mm), 50 fins, and 1.7 (mm) of fin pitch over a 120 mm × 120 mm area under the determined parameters of the thermoelectric module with a 120 mm cooling fan. In addition, by comparing different machine learning algorithms, the results revealed the superiority of random forest regression in solving optimal problems. This work provides a convenient and practical means for the fast optimization of thermoelectric assembly.