Facial beauty prediction is a complex task that relies on subjective human perceptions, making it a challenging area of study within computer vision. In this paper, we propose a semi-supervised approach that using contrastive pre-training with SimCLR (simple framework for contrastive learning of visual representations) to predict facial beauty scores. By utilizing contrastive learning, our model learns robust representations through the self-supervised task of distinguishing between different views of the same image and between different images. We leverage a diverse dataset, SCUT-FBP5500, which comprises 5,500 annotated facial images, to develop a model capable of accurately predicting beauty scores. Our proposed method involves two primary phases: first, we pre-train the model using contrastive learning to acquire robust visual representations from a larger set of unlabeled images, and then we fine-tune it on the labeled SCUT-FBP5500 dataset. The results demonstrate that our model achieves a Pearson correlation coefficient of 0.9267, surpassing state-of-the-art methods in beauty prediction. These findings indicate the effectiveness of using contrastive pre-training for this application, as our model not only enhances prediction accuracy but also aligns more closely with human judgments of beauty. This study contributes to ongoing research in aesthetic assessment and highlights the potential of semi-supervised learning to improve performance in subjective evaluation tasks.