Exploring the role of geometric and learning biases in Model Order Reduction and Data-Driven simulation