Lagrangian Trajectory Modelling for a Person lost at Sea during Adriatic Scirocco Storm of 29 October 2018

On 29 October 2018 a windsurfer’s mast broke about 1 km offshore during a severe Scirocco storm in the Northern 1 Adriatic Sea. He was drifting in severe marine conditions until he eventually beached alive and well in Sistiana (Italy) 24 2 hours later. We conducted an interview with the survivor to reconstruct his trajectory and to gain insight into his swimming 3 and paddling strategy. We then attempted a Lagrangian simulation of his trajectory in two ways. Firstly by performing a lee4 way simulation using the OpenDrift tracking code using two object types: Person-in-Water-1 and Person-powered-vessel-2. 5 Secondly, we model the trajectory using our own Lagrangian tracking code FlowTrack. In both cases a high-resolution (1 km) 6 setup of NEMO v3.6 circulation model was employed for the surface current component and a 4.4 km operational setup of the 7 ALADIN atmospheric model was used for wind forcing. OpenDrift yields best results using Person-powered-vessel-2 object 8 type, indicating a relatively broad search and rescue area which covers 45 km after six hours and rises to 380 km after 24 9 hours. The simulated most probable SAR area envelops the reconstructed drift trajectory and is also temporaly consistent with 10 the reconstruction. FlowTrack yields a search and rescue area with a comparable lateral extent but with much less downwind 11 spread. While both Lagrangian models were able to envelop the reconstructed drift trajectory during this validation, we recom12 mend using OpenDrift for similar search-and-rescue missions in the future due to its flexibility and drifting object dependent 13 calibration on empirical data. 14

True Color image of the Gulf of Trieste from the day after the beaching, 31 Oct 2018 (obtained from Copernicus Open Access Hub: https://scihub.copernicus.eu). Turbid Soča/Isonzo river plume is clearly visible along the northern shore of the Gulf.
It is around 23 UTC that his drift turns north-east. After 23 UTC, he is located approximately on the Piran-Grado line. Sea 59 conditions get very severe, he is laying on the windsurf board, mostly facing southwest, away from the mean drift direction, 60 drifting backwards, clutching the footstraps on the surfboard. He estimates that every 50th wave breaks over him and pulls 61 the surfboard from under him. When this happens he needs to wait to reach the crest of the wave to re-locate the board and 62 catch it. In the morning, on 30 Oct 2019 07 UTC, he is located 2 -4 km south-southwest of the Soča/Isonzo river mouth. By 63 9-10 UTC he is located roughly 1-2 km south-southeast of the river mouth and the water gets significantly colder as he likely 64 enters the Soča/Isonzo river plume (visible in Figure 1 b) ). By the time of his entering the plume, the Soča/Isonzo runoff is at 65 a several-month maximum, as depicted in Figure 2. From 11 UTC on he is paddling actively toward northeast to overcome the 66 riverine westward coastal current until he reaches the beach near Sistiana at 16 UTC. 67 The drifting trajectory, reconstructed from above, is shown in the b) panel in Figure 1. In the present paper, we present 68 two attempts to simulate this trajectory using two different particle tracking models, OpenDrift and FlowTrack. Available 69 observations and general marine conditions during the drift are presented in Section 2; numerical models used for particle

High Frequncy Radar System
The HF systems deployed in the Gulf of Trieste consist of two WERA stations (Gurgel et al., 1999) Figure 3. Maximum vertical discretization stretch is located at 15th level to allow for appropriate vertical 99 resolution near the surface. In all regions shallower than 2 m, a minimum 2 m depth is enforced. Vertical level depths in meters 100 are 0. 50, 1.51, 2.55, 3.64, 4.83, 6.20, 7.94, 10.38, 14.18, 20.56, 31.68, 51.23, 84.58, 137.94, 215.83, 318.24, 440.67, 576.90, 101 721.55, 870.95, 1022.92, 1176.25, 1330.29, 1484.69,1639.28, 1793.97, 1948.71, 2103.47, 2258.25, 2413.03, 2567.81, 2722 interpolation is applied over 5 cells to reduce wind gradients between the two products. It is worth noting that the ALADIN 125 SI model is forced at the boundary by the same ECMWF product we use to provide winds to ECWAM outside (and south of)  particles N p (i.e. several thousand) are allowed. In this study N p = 408 particles were seeded in a 1 km ×1 km square around 168 the initial location. Once the particle is initialized in the wet cell of the model grid, it is subjected in each timestep to advection, 169 turbulent diffusion and, if applicable, fate. Lagrangian trajectory r p (t) of p−th particle (p = 1, . . . , N p ) is computed using a 170 second order Runge-Kutta method (Euler method is also available) to integrate the following initial value problem where r 0,p in Equation (2) denotes initial position of p-th particle.

172
Terms of the right hand side of equation (1) are as follows. Term u c (r p (t), t) denotes the Eulerian current at particle location 173 r p (t) at time t. This term is obtained from the NEMO circulation model. Term u w,p (r p (t), t) denotes the wind drift vector 174 of the p-th particle at particle location r p (t) at time t. Wind drift generally has lift and drag component, thus deviating from 175 the wind direction. This deviation is treated in FlowTrack by rotating the wind drift vector of p-th particle by an angle θ(p) ∈

176
[−L α , +L α ], where L α stands for the leeway divergence angle (Allen and Plourde, 1999). In this work L α was set to 20 • as 177 recommended for a person with surfboard in Table 8-1 of Allen and Plourde (1999). Since wind lift can generally act to the 178 left or to the right of wind, with both options having equal probabilities (Breivik and Allen, 2008), FlowTrack distributes θ(p), (3)

181
Each particle retains its respective angle of rotation throughout the simulation and does not jibe. The wind drift vector of the 182 p-th particle is then computed from ALADIN SI 10m winds u 10 (r p (t), t) at particle location as

184
Parameter q in equation (1)  does not alter its modulus, and hence disregards the wind lift force, we attempted to compensate for this by estimating q = 190 (q DW + q σ DW ) 2 + (q XW + q σ XW ) 2 ≈ 2.5 percent of the wind speed.

191
Term u in (1) represents random fluctuations in the velocity vector to simulate subgrid turbulent diffusion. In the present 192 paper, the modulus of fluctuations has been manually constrained to 2 · 10 −2 m s −1 . Term u s (r p (t), t) on the right hand side of 193 the equation (1) is the Stokes drift contribution, i.e. Eulerian mean of unclosed Lagrangian particle orbits in the surface gravity 194 wave field. It was however shown (see (Hackett et al., 2006;Breivik and Allen, 2008) for further references) that Stokes drift,195 while present in the motion of the water, has negligible impact on drift speed of objects whose typical dimension is more than 196 roughly six times smaller than surface gravity wave wavelength λ w . We can compute λ w from surface gravity wave dispersion 197 relation ω(k) = gk tanh(kH). Since wave vector is k = 2π/λ w , we can solve for λ w by iterating where T w is the mean wave period of the wave field obtained from the ECWAM model at a representative point along the  In this section we present a qualitative analysis of marine conditions from available observations, and also marine drift results 212 from both particle tracking models presented in Section 4. likely shows no notable inflow due to inertial westward coastal current from the Soča/Isonzo river, which manifests itself as an 247 outflow from the Gulf, confined to this part of the coast (see Figure 1 for the related river plume).

248
OpenDrift results for drifting object type Person in Water (PIW-1) are presented in Figure 8.   (1000) Active (735) Stranded (265) Initial (1000) Active (694) Stranded (306) Initial (1000) Active (586) Stranded ( (1000) Active (912) Stranded (88) Initial (1000) Active (831) Stranded (169) Initial (1000) Active (589) Stranded ( ADIN SI) and Lagrangian tracking models, used in an attempt to hindcast this trajectory. We present available measurements 293 from the regional coastal buoy Vida and HF surface current radar to qualitatively assess marine conditions in the Gulf of Trieste  wave data and Stokes drift computation prior to particle advection, but this option was not used for the simulations performed 302 in this paper. The reason for this is that wave-induced drift has been shown to be of importance only for objects whose typical 303 dimension is comparable to the wavelength of surface gravity waves scattering off the object. During the storm in question,

304
surface gravity waves wavelengths surpassed 50 m, a value ten to twenty times longer than the length of the board or a person.