I am working on reduced-order modeling of CFD simulation data.

• I performed POD on temperature fields (structured grid) and extracted the POD coefficients.
• I also tried an Autoencoder approach to obtain latent variables.
• Then I mapped these coefficients/latents to input parameters using several ML models.

Problem:

In both approaches, the variation in coefficients/latents with respect to inputs is extremely small. Most coefficients appear almost constant, making ML prediction very poor (low R²).

What I have tried:

• Checked reconstruction quality → it’s good.
• Increased number of modes / latent dimensions → same issue.
• Used several ML regressors (linear, MLP, XGBoost) → still poor predictive performance.

Are there best practices for ensuring the reduced-order basis captures input-sensitive variations?

Any advice from those experienced with ROMs, POD, or CFD-to-ML mapping would be appreciated ?