[R] Where to look for resources on developing metrics for generative model for science?

An AI researcher is developing a new generative foundation model for a scientific field, handling multimodal data like videos and time series. They currently lack robust evaluation methods, relying only on Mean Squared Error (MSE) and subjective human feedback. The core challenge is finding or developing suitable metrics for each modality, moving beyond field-specific standards like FID for images or PESQ for audio to properly assess the novel model's performance.