1. Radi AJJAJI : "Incrementality deficiency in ARPEGE 4d-var assimilation system"

In order to be able to work at home on 4d-var problems (those linked with incrementality aspects), I tried to adapt the configurations involved into 4d-var on IBM at Casablanca. As we could run 3d-var before, the work was not very hard. The huge problem that I encountered was the large memory required by 4d-var for minimisation at truncation T95 (incorporating simplified physics). This is due to the activation of physics especially radiation (the coefficients are stored in memory). As I was running just on one node (9 GB memory and 16 processors with a blocking factor NPROMA of 31), it was impossible to achieve T95 minimisation. The solution that allowed me to solve this problem was to distribute work on two nodes by specifying 8 processors on each node.

Now, it is possible to run the old 4d-var with 3 outer loops (T42, T63 and T95) on IBM for a reasonable cost (90 minutes execution on 16 processors with CY22T2). I succeeded in reproducing the wrong increments due "eventually" to the high incrementality in ARPEGE 4d-var.

I performed some experiments on internal DFI into 4d-var in order to catch the impact of strong or weak constraint on the increments. But no major sensitivity to notice.

As many changes have been introduced on 4d-var since CY22T2, I am working on adapting CY25T1 to be able to run 4d-var in a newer context. ODB and all the applications related to it (Batodb, Shuffle, screening, minimisation) are already adapted.