Evaluation Metrics

RealPDEBench reports 9 metrics to assess prediction accuracy on real-world data, physical consistency, and (for sim-to-real transfer) training efficiency.

Data-oriented

  • RMSE (Root Mean Square Error)
  • MAE (Mean Absolute Error)
  • Rel \(L_2\) (Relative \(L_2\) Error)
  • \(R^2\) (Coefficient of Determination)
  • Update Ratio (training efficiency; finetuning only)

Go to Data-oriented metrics →

Physics-oriented

  • fRMSE (Fourier Space Error)
  • FE (Frequency Error)
  • KE (Kinetic Energy Error)
  • MVPE (Mean Velocity Profile Error)

Go to Physics-oriented metrics →

Metric Comparison

Metric Category Range Best Value
RMSE Data-oriented \([0, \infty)\) 0
MAE Data-oriented \([0, \infty)\) 0
Rel \(L_2\) Data-oriented \([0, \infty)\) 0
\(R^2\) Data-oriented \((-\infty, 1]\) 1
fRMSE Physics-oriented \([0, \infty)\) 0
FE Physics-oriented \([0, \infty)\) 0
KE Physics-oriented \([0, \infty)\) 0
MVPE Physics-oriented \([0, \infty)\) 0
Update Ratio Data-oriented \([0, \infty)\) 0

Notes: - \(R^2\) is higher-is-better; all other listed metrics are lower-is-better. - Update Ratio is only reported for Real-world finetuning (simulated pretraining → real-world finetuning). - Update Ratio values \(< 1\) indicate simulated pretraining reduces the number of updates needed to reach the best real-training RMSE.