Here, we introduce the concept of using synthetic data to train large models for automated neural signal processing and curation of detected neural spikes. We show that by leveraging the diversity in physically accurate synthetic data, machine learning models can be trained very effectively. The trained models can be easily transferred to real-world applications without the need of expensive real-world ground truth data. We use the automatic spike sorting curation as an example to explain the general pipeline of integrating synthetic data into training. Considerations and strategies for generating synthetic data with diversity are also discussed.