Bias-aware data assimilation
Ensemble Kalman filtering for nonlinear and chaotic systems — including estimating the model error the model does not know it has.
Standard ensemble Kalman filters assume the model is unbiased: wrong only by random noise with zero mean. Real models are not. They are simplified, they miss physics, and they are wrong in structured, systematic ways.
If that bias is ignored, the filter fights it. Observations are repeatedly used to correct a discrepancy that keeps coming back, and the estimate degrades even as more data arrives.
We develop bias-aware filters that treat the model error as a quantity to be inferred alongside the state and the parameters. The regularized bias-aware ensemble Kalman filter (rBA-EnKF) estimates the bias with a reservoir computer trained online, and regularises it so that the split between “model” and “bias” stays identifiable.
Key publications
- Nóvoa, A., Racca, A., & Magri, L. (2023). Inferring unknown unknowns: Regularized bias-aware ensemble Kalman filter. Computer Methods in Applied Mechanics and Engineering, 418, 116502. doi10.1016/j.cma.2023.116502
- Nóvoa, A., Racca, A., & Magri, L. (2022). Bias-aware thermoacoustic data assimilation. In Proceedings of the 51st International Congress and Exposition on Noise Control Engineering (pp. 1924–1931). doi10.3397/IN_2022_0271