Availability of the tangent linear model (TLM) to the deep learning system of Bocquet et al. (2024)

Determine whether the deep learning-based analysis operator a_theta used in Bocquet et al. (2024) had access to the tangent linear model of the Lorenz-96 system during training or execution, in order to clarify the methodological assumptions and potential advantages relative to approaches that explicitly use the tangent linear model.

Background

The paper introduces two algorithms that reconstruct state error covariance solely from a state estimate, both relying on the tangent linear model (TLM) of the Lorenz-96 system. The authors note uncertainty about whether the deep learning-based system (DLS) of Bocquet et al. (2024) had access to the TLM, which bears on comparisons of methodological capabilities and fairness.

Because the forward model was available in the referenced setting, the authors argue that obtaining the TLM should not be considered an unfair advantage, as it can be derived from the forward model via linear operations. Nonetheless, the actual availability of the TLM to the DLS remains explicitly unresolved.