| 08 Sep 2026
Sensitivity of the Lagrangian temperature anomaly decomposition to model choice: Lessons from the 2021 Pacific Northwest heatwave
Abstract. The Lagrangian temperature anomaly decomposition has emerged as a powerful tool for quantifying the relative importance of adiabatic, advective and diabatic processes during heatwaves. However, its sensitivity to the choice of Lagrangian model, whether pure-trajectory or dispersion-based, remains poorly understood. Using seven different trajectory sets, we apply this decomposition to the June 2021 Pacific Northwest heatwave to examine discrepancies in process attribution across model classes. Pure-trajectory models consistently identify diabatic heating as the dominant contributor, whereas dispersion models attribute the largest fraction to adiabatic warming. A trajectory cluster analysis and parcel position diagnostics reveal that this divergence does not stem from different geographic origins or the near-surface vertical distribution over the target region during the heatwave, which are broadly similar across trajectory sets, but from contrasting vertical histories. While dispersion-model parcels reside at higher altitudes before arrival, exposing them to subsidence warming, pure-trajectory parcels remain near the surface where land–atmosphere feedbacks and antecedent dry soils amplify diabatic heating. These findings demonstrate that this Lagrangian framework is fundamentally sensitive to model choice, particularly to the representation of turbulent motion, which induce parcel vertical dispersion absent in the resolved wind fields. While this study does not establish which Lagrangian model class is most suitable, it represents a first step toward understanding these discrepancies and cautions that heatwave attribution studies relying on a single Lagrangian model should interpret inferred process dominance carefully, as it may reflect methodological choices rather than the underlying physics alone.
This study compares different Lagrangian models in their capability to decompose different terms in the heat budget, at the example of a specific event. I have several concerns regarding this paper, which prevent me from recommending it for publication.
Most importantly, it is a “mechanical” intercomparison exercise that does not offer any physical interpretation. Differences are found but they are not explained. A scientific paper should go beyond simply comparing and should also offer explanations. In fact, I am wondering why this study is needed at all. It is clear that trajectory models do not account for turbulence. However, all trajectories end up in the boundary layer, where turbulence is extremely important and where trajectory models fail in capturing the real transport history of an air mass. So, in my opinion, they are simply not fit for the purpose, and I would avoid using them, unless you are driving them with the output of a large eddy simulation, where turbulent motions are captured in the input wind fields.
For the particular problem studied in this paper, the difference between trajectory models and dispersion models seems largely a result of the neglect of entrainment of free-tropospheric air into the boundary layer in the trajectory models (which is related to the lack of turbulence in the trajectory models). Warming of air in the boundary layer is the result of diabatic warming at the surface but, equally or even more importantly, entrainment of warm free tropospheric air at the boundary-layer top. Since the trajectory models do not capture the entrainment, it is not surprising that they attribute a larger fraction of the warming to diabatic processes, whereas FLEXPART attributes more to adiabatic warming (i.e., entrainment of descending free tropospheric air into the boundary layer). I am very surprised that entrainment is not even mentioned in the paper, although it seems like a logical explanation that should be explored specifically. It seems also plausible that trajectory models ignoring turbulence will result in a wrong process attribution. For dispersion models, their reliability will be controlled to a large extent by how realistically they capture boundary layer dynamics, in particular it will depend on the daily cycle of boundary-layer depth.
Another problem of the paper is the rather random choice of models. While different trajectory models are compared, all dispersion model calculations are based only on different versions of FLEXPART. Why did you add all these different versions of FLEXPART? What do you learn from this? The discussion of the results only focusses on the differences between trajectory models (as a group) and the FLEXPART ensemble. Differences between the trajectory models are not discussed at all, so why do you even need them all? To make the points the paper wants to make, it would have been completely sufficient to compare FLEXPART with and without parameterizations. I can’t see any discussions/conclusions that go beyond that, and what the different models (and model versions) used do add.
I have a few other specific points:
Lines 135-158: The domain-filling FLEXPART runs all use too few particles to obtain statistically robust results. This is mentioned for the LARA data set on line 200 but I think this is true for all data sets. The global domain-filling set-up (e.g. LARA) has been developed for climatologies but it is definitely not ideal (or even suitable) for case studies, unless you use hundreds of millions of particles. A better option here would be to run the model backward in time from the target domain, such that all particles contribute to the results.
Equation (1): The most interesting term is the last one (diabatic term). However, without splitting it further up into the different processes, it is not extremely useful, since you then don’t know which process is actually responsible (e.g., latent versus sensible heat flux, radiation).
Figure 1: The last column (Budget) is pretty useless, as it cannot be distinguished from the first column. Showing the differences would be much more meaningful.
Figure 4: I don’t think the cluster analysis provides any further insight. To me, it is not clear what it is supposed to show.
Lines 385-390: I do not understand the comments about closed boundaries with respect to turbulent exchanges. Of course, diabatic surface heating is only happening at the surface. That’s exactly how it should be!