Differentiable CoCiP: exact sensitivities of contrail energy forcing with DIFFCOCIP v1.0
Abstract. Reducing contrail warming requires knowing how a flight's climate effect responds to changes in its trajectory, the weather and the contrail model's assumptions. The Contrail Cirrus Prediction model (CoCiP) computes this effect as energy forcing, EF, but its reference implementation does not provide these sensitivities. Finite differences could estimate them by perturbing one input at a time, but this is impractical for weather fields holding millions of values. We present diffcocip v1.0, a reimplementation that applies reverse-mode automatic differentiation to obtain all these sensitivities in a single backward pass. It reproduces the reference model's EF to the accumulated floating-point round-off and returns derivatives with respect to waypoint altitudes, gridded humidity and model parameters. Three applications assess the use and limits of these derivatives. In a four-season fleet campaign, altitude gradients guide the placement of candidate manoeuvre windows, and every candidate is then evaluated with the unmodified model. For about one third of reprofiled flights, the selected manoeuvre reduces EF despite a gradient prediction of an increase, showing why finite manoeuvres must still be checked against the full model. Humidity gradients identify the small part of a four-dimensional weather grid that influences one flight's EF, including positions reached by its drifting contrail after the aircraft has passed. Among six model parameters, EF proves most sensitive to the humidity-correction coefficient, whose effect is about 7 times larger than that of the next most influential parameter. The derivatives are exact within each smooth region of the implemented model, but do not capture discontinuous changes across the formation, persistence and termination thresholds. diffcocip also includes an optional smoothed formulation of these criteria to support gradient-based optimisation; the results reported here use the unmodified model throughout.
The authors release an open source package of python code which numerically reproduces the pycontrails CoCiP implementation to high precision, and provide case-studies to explore its ability to provide gradient-driven sensitivity analysis for three usecases. Two of the usecases are weather fields (altitude which is effectively a proxy for interpolated weather fields, and separate analysis on specific humidity q), and the other usecase analyzes 6 CoCiP model parameters. The development of differentiable CoCiP is an exciting technical contribution for the contrail modeling community and it has the potential to have an excellent scientific impact.
While the manuscript does convincingly present faithful reproduction of the pycontrails CoCiP forward pass and fidelity of the DIFFCOCIP gradients to finite-differences thereof, the manuscript in its current form does not demonstrate a significant scientific result as I believe is required for acceptance by JECATS (though, I will defer to the editor in this regard). The altitude analysis has similarities with flight planning for contrail avoidance maneuvers, and so should likely include a comparison to baseline flight planning methods. Humidity sensitivity has already been identified by the cited Teoh et al as the most sensitive input for CoCiP so it is not unexpected to be confirmed by this alternate methodology.
I suggest the authors pursue one of two paths: improve scientific significance and resubmit to JECATS, or re-target the paper to a journal that has an article type more focused solely on model development with a less stringent scientific significance requirement. If resubmitting here, authors are advised to move the majority of the software details into appendix and focus on improving scientific significance of results which are obtainable in a novel way unlocked by the differentiability of this newly released software. If re-targeting to a different venue, JOSS or GMD may be appropriate. The re-targeting strategy can allow future papers, including from other groups, to build on top of the method released here. One such example of that path to impact is pycontrails CocipGrid paper (https://gmd.copernicus.org/articles/18/253/2025/) which was published in GMD. An example of the other approach where a software framework also had strong scientifically significant methodological results in its release manuscript is the JAX-MD paper (https://papers.neurips.cc/paper/2020/file/83d3d4b6c9579515e1679aca8cbc8033-Paper.pdf).
1. Code Availability and Reproducibility
Line 630 (https://github.com/ramondalmau/diffcocip): As of Oct 27 this URL is returning 404 not found; further review of correctness/performance of this manuscript is difficult without access to the code. Also, note that JECATS requires a DOI be generated for citable software packages, the authors may consider using Zenodo or another service for generating this.
2. Major comments
3. Math formulation/notation
4. Figures and Visual Presentation
5. Manuscript Structure, Terminology, and Editorial Clarity