I do not think i can give you an accurate answer. However, training on simulated data and deployment on a world-like environment is usual practice in the High Energy Physics environment. They use Monte-Carlo high statistics simulations to train the model and evaluate their prediction performances using real-data collected by experiments:
An example
One thing you can do to limit the experimental-simulation disagreement would be to implement a domain adaptation layer into your network:
This is a nice article explaining domain adaptation
All these considerations may depend on the type of architecture you are going to develop, for which specific task (binary classification or other?), and how reliable are your simulations.
与恶龙缠斗过久,自身亦成为恶龙;凝视深渊过久,深渊将回以凝视…