We use the Earth system model of intermediate complexity iLOVECLIM, which is a code fork of the LOVECLIM model (Goosse et al., 2010). It includes a representation of the atmosphere, ocean, sea ice, terrestrial biosphere vegetation and the carbon cycle. The ocean component (CLIO) consists of a free-surface primitive equation ocean model (3∘×3∘ horizontal grid, 20 depths layers) coupled to a dynamic-thermodynamic sea-ice model. iLOVECLIM includes a land vegetation module (VECODE) (Brovkin et al., 1997) and a marine carbon cycle model (Bouttes et al., 2015), both computing the evolution of 13C and 14C. Previous work has shown that the simulated oceanic δ13C and Δ14C distribution is in reasonable agreement with observations and with available general circulation model (GCM) simulations (Bouttes et al., 2015). We have implemented 231Pa and 230Th in the iLOVECLIM model following the approach of Rempfer et al. (2017), albeit without including an explicit parameterisation of boundary and bottom scavenging as detailed in the following. In the model, 231Pa and 230Th are homogeneously produced in the ocean by radioactive decay of their respective uranium parents. They are removed from the water column by adsorption on settling particles (reversible scavenging). Their radioactive decay is also taken into account. In its current version, iLOVECLIM does not explicitly simulate the biogeochemical cycle of biogenic silica (opal), which is thought to be an important scavenger for 231Pa (e.g. Chase et al., 2002; Kretschmer et al., 2010). Therefore, we used prescribed and constant particle concentration fields obtained running the PISCES biogeochemical model coupled to NEMO ocean model (van Hulten et al., 2018). Previous work has shown that these particle fields are reasonably consistent with present-day observations (van Hulten et al., 2018) and that a set-up with fixed particles is able to capture the major features (Gu and Liu, 2017). Like in Bern 3D (Rempfer et al., 2017; Siddall et al., 2005, 2007) and CESM (Gu and Liu, 2017), we considered one particle size with a unique sedimentation speed of 1000 m yr-1. We consider that Pa and Th interact with three particles types: CaCO3, POC and biogenic silica. The role of lithogenic particles in Pa and Th scavenging having been questioned (van Hulten et al., 2018; Siddall et al., 2005). The conservation equations for dissolved and particulate 231Pa and 230Th activities are the following (Rempfer et al., 2017): 1∂Adj∂t=TAdj+βj+kdesorpj⋅Apj-Kadsorpj+λj⋅Adj2∂Apj∂t=TApj-∂(ws⋅Apj)∂z-(kdesorpj+λj)⋅Apj+Kadsorpj⋅Adj, where Adj and Apj are respectively the dissolved and particle-bound activities (dpm m-3 yr-1) of isotope j (231Pa or 230Th), βj (dpm m-3 yr-1) is the water column production of isotope j by radioactive decay of its uranium parent, λj (yr-1) is the decay constant of isotope j, ws (m yr-1) is the particle settling speed, Kadsorpj and kdesorpj (yr-1) are the adsorption and desorption coefficient of isotope j onto particles, respectively, and T is the tracer balance evolution term (dpm m-3 yr-1) resulting from the water mass advection and diffusion terms computed by the CLIO ocean model. The values used in this study for each of the above-cited parameters are compiled in Table 1.

As in Bern 3D (Rempfer et al., 2017), we chose to apply a uniform desorption coefficient denoted Kdesorp hereafter. However, the adsorption coefficient depends in our model on the particle concentration and composition of each location and is calculated with the following equation (Rempfer et al., 2017): 3Kadsorpjθ,Φ,z=∑iσi,j⋅Fi(θ,Φ,z), where θ is the latitude, Φ the longitude, z the water depth, σi,j the scavenging efficiency for isotope j on particle type i (m2 mol-1) and Fi is the particle flux (mol m-2 yr-1).

The scavenging efficiency σi,j is related to the partition coefficient Kd, which defines the proportion of each isotope (231Pa or 230Th) lodged in the dissolved phase or bound to particles. Therefore one Kd can be defined for each isotope and each particle type considered in the model. 4Kd(i,j)=σi,j⋅ws⋅ρswM(i)⋅kdesorp, where Kd(i,j) is the partition coefficient for isotope j for particle type i, σi,j is the corresponding scavenging efficiency, ws is the settling speed, Kdesorp is the desorption coefficient, M(i) is the molar mass of particle type i (i.e. 12 g mol-1 for POC, 100.08 g mol-1 for CaCO3 and 67.3 g mol-1 for opal) and ρsw is the mean density of seawater (constant fixed to 1.03×106 g m-3). Additionally the way each particle type fractionates the Pa and the Th is defined by the fractionation factor F(Th/Pa)i. 5F(Th/Pa)i=Kd(Th,i)Kd(Pa,i)=σ(Th,i)σ(Pa,i) Kd and fractionation factors, F(Th/Pa), have been measured for both radionuclides in various areas of the modern ocean, and they show a rather large distribution (see Table S1 in the Supplement) (Chase et al., 2002; Hayes et al., 2015). Consequently, these values are currently considered as tunable parameters in modelling studies (Dutay et al., 2009; Gu and Liu, 2017; van Hulten et al., 2018; Marchal et al., 2000; Rempfer et al., 2017; Siddall et al., 2005). Considering three particle types for both radionuclides, there are thus six tunable σi,j parameters in our model.

To efficiently sample our parameter space, we used a Latin hypercube sampling (LHS) methodology (https://CRAN.R-project.org/package=lhs, last access: 17 April 2019). In order to only select parameters values consistent with observed F(Th/Pa), we chose to use the couples {σTh,j, F(Th/Pa)i} as input parameters for the LHS. The value of σPa,i is then deduced from those two following equation (Eq. 5). The parameter ranges used in the LHS are given in Table S2. The LHS allowed a relatively good exploration of the parameter space with a relatively small number of model evaluations. We performed 60 tuning simulations of 1000 years each under preindustrial boundary conditions (i.e. orbital parameters, ice-sheets configuration and atmospheric gas concentrations), as described in (Goosse et al., 2010).

Consistent with previous modelling studies (van Hulten et al., 2018), this simulation length was sufficient for Pa and Th to reach an approximate steady state at the surface and in the deeper ocean. The model performance was evaluated by comparing outputs with present-day particulate and dissolved water column Pa and Th measurements compiled in the GEOTRACES intermediate data product 2017 (Schlitzer et al., 2018) as well as sedimentary Pa/Th core tops data as compiled in (van Hulten et al., 2018) and references therein. We selected the ensemble member that best fits the observational constraints using five metrics corresponding to the root-mean-square error (RMSE) between observation data and the closest model grid cell average for particulate and dissolved Pa and Th as well as sedimentary Pa/Th (see Sect. S1 in the Supplement for additional information). Table 2 presents the best-fit σi,j values. The best fit simulation is then used to investigate the multiproxy response to abrupt circulation changes.

With the best fit σi,j values (see Table 2), we ran our model for 5000 years under preindustrial (PI) conditions (as described in Goosse et al., 2010) from a simulation with an equilibrated carbon cycle (Bouttes et al., 2015). The result of this equilibrium simulation is used as a starting point to perform hosing experiments of 1200 years duration. The freshwater was added in the Nordic Seas following the approach described in (Roche et al., 2010). Each simulation contains three phases: a control phase (300 years), a hosing phase (300 years) and a recovery phase (600 years). The control phase is used to assess the natural variability of the circulation and associated proxies under the PI climate state.

The main goal of our study is to assess the response of two carbon-based proxies (δ13C, Δ14C) and of the Pa/Th to an abrupt circulation change in a physically consistent framework. This work represents a first step toward a better understanding of marine multiproxy records across the last glacial abrupt events. Therefore, our model–data comparison focuses on the Atlantic Ocean, where the Pa/Th can be used as a kinematic circulation proxy (e.g. François, 2007; McManus et al., 2004). and on the simulated bottom particulate Pa and Th activities which can be directly compared to the Pa/Th ratio recorded in marine sediments. The water column concentration results are presented in Sect. S2 and Figs. S1 to S3 in the Supplement.

The particle-bound Pa/Th of the deepest oceanic model grid cells is shown in Fig. 1. The observations from core-top data are superimposed as circles. Even if the observational dataset is somewhat patchy, it generally shows lower sedimentary Pa/Th ratios (0.04 or lower) in the basin interiors compared to the coastal areas. Another interesting feature is the high sedimentary Pa/Th values in the Southern Ocean between 50 and 75∘ S (opal belt), where Pa is heavily scavenged to the sediments by opal. Our model generates higher sedimentary Pa/Th in this region compared to the deep basin interior but our modelled opal belt stands slightly northward compared to the observations (Fig. 1). In our best fit simulation, the adsorption/desorption coefficients fall in the upper range of observations (see Tables 2, S1), therefore, Pa and Th are very effectively scavenged to the sediments. Besides, while Pa is mostly scavenged by the opal particles, Th mainly reaches the sediments with CaCO3 particles (see the Supplement for detailed information). Overall, these results are comparable with GCM outputs (van Hulten et al., 2018). The modelled values display the first order characteristics observed in the modern ocean and the sediment core tops. In the following section, we test the sensitivity of the simulated sedimentary Pa/Th to abrupt circulation slowdown (hosing experiments).

We added a freshwater flux of 0.3 Sv (sverdrups) in the Nordic Seas over a period of 300 years, which was sufficient to cause a drastic circulation reduction (Roche et al., 2010). Under PI conditions, the maximum of the AMOC stream function is about 15 Sv in our model. During the hosing, North Atlantic water formation drops to nearly 0 Sv, and the upper overturning cell completely vanishes (see Fig. S4 in the Supplement). The AMOC recovers ∼200 years after the end of the water hosing and displays a small overshoot with the maximum Atlantic meridional stream function exceeding 20 Sv around 900 years (Fig. 5).

We evaluate the response of Pa/Th, δ13C and Δ14C to the hosing in the Nordic Seas as follows. We identify, for each model grid cell and each proxy, the simulation periods exceeding 80 years during which the proxy values are outside of their natural variability range. The latter is defined as the proxy variance (2σ) under PI conditions over the first 300 years of the simulation (control phase – see Fig. 2). In most cases, zero, one or two periods of significant response are detected. In some grid cells with high proxy variability, we detect up to four periods of significant proxy response, which are difficult to relate to either the hosing or overshoot timing of our simulation. Consequently, we excluded the grid cells depicting more than two periods of significant response from the subsequent analysis. For the time series containing two or less periods of interest, we call “time of maximum response”, the simulation year for which the absolute difference between the proxy value and mean proxy value during the control phase of the simulation is maximal. The proxy value at the time of maximum response is denoted “proxy response” in the following (Fig. 2). The single or dual proxy response is compared to the control proxy value (i.e. the mean proxy value over the first 300 years of the simulation). Figure 3 represents the zonal mean proxy response in the western Atlantic basin, the delimitation between the two basins being defined by the position of the Mid-Atlantic Ridge (see Figs. S5 and S6 for eastern basin).

The δ13C and Δ14C only display one single response in the deep western Atlantic, whereas the Pa/Th generally displays one or two responses in this part of the basin. Generally, the single or the first response of each proxy has the same geographical pattern (Fig. 3a–f), while in the case of two distinct responses, the late response has a radically different pattern (Fig. 3g–i). For δ13C and Δ14C, the late response corresponds to a general increase in the western Atlantic basin. For Pa/Th, the late response pattern consists of increased values in the southern Atlantic and decreased values in the North Atlantic and is the opposite of the early response pattern. For the three proxies, the late response pattern is generally consistent with the circulation overshoot observed around 900 simulated years.

Considering that the single and the first responses mostly represent the proxy response to the hosing, a consistent proxy response is observed in the following three zones of the western Atlantic basin: the surface and intermediate waters (0–1500 m), the deep North Atlantic waters (> 2000 m) and the Southern Ocean (south of 30∘ S). In the surface waters and intermediate waters the proxy response pattern is as follows. The δ13C decreases in the first 500 m and then increases around 1000 m. The Δ14C increases from the surface to 1500 m, and Pa/Th displays no clear trend. In the deep North Atlantic waters, the δ13C and Δ14C decrease, while the Pa/Th generally increases, the three proxies reflecting a reduction in the North Atlantic Deep Water (NADW) flow rate. In terms of magnitude, we observe the strongest proxy response between 60 and 40∘ N and between 1500 and 3000 m for the three proxies. In the Southern Ocean, the Pa/Th generally decreases, except in the deep southern basin (between 40 and 60∘ S, below 3000 m). The δ13C and Δ14C display a dipole pattern increasing at the surface and decreasing at depth. For δ13C, the depth boundary is around 1500 m, while for Δ14C the depth boundary is ∼1500 m north of 50∘ S and reaches 3000 m south of 40∘ S.

Looking at the proxy time of response (Fig. 4), we observe significantly different patterns for Pa/Th and the carbon isotopes. The δ13C and Δ14C time of response increases with depth as follows: in the surface and intermediate waters the response occurs roughly 300 years after the beginning of the fresh water addition, around 3000 m the response is delayed by 150 years, and towards the ocean bottom the delay increases up to 600 years. On the contrary, the Pa/Th displays much smaller time of response, the timing of the Pa/Th response being generally synchronous with the minimum of the stream function 300 years after the beginning of the freshwater addition. Consequently, in most of the western Atlantic basin, the response of the carbon isotopes lags behind the Pa/Th response by a few hundred years, especially in the deeper waters.

In the eastern Atlantic basin the general pattern of the proxy response is similar to that of the western basin (Figs. S4–S5). However, we note that the δ13C has frequently more than a single response, especially at depths > 2000 m. The δ13C displays overall the same pattern as in the western basin except in the south deep basin (between ∼20 and 30 ∘ S – below 2500 m) in the case of a dual response. The Δ14C shows the same pattern as in the western basin, with increased values during the hosing in surface and intermediate waters and lower values at depth. Besides, the Pa/Th has a more complex pattern with a large increase in the arctic basin, a moderate increase in the tropical basin, and a decrease in the northern and southern basins through the entire water column. Concerning the times of response, the same pattern as in the western ocean is observed. Interestingly, we note that the Δ14C has longer times of response in the eastern basin compared to the western basin in the case of one single response (Fig. S5). This is consistent with a lower ventilation of the eastern basin associated with a lower penetration of the NADW relative to the situation in the western boundary current.

Overall, the objective analysis of the multiproxy response allows the identification of three regions with different reaction patterns:

The clearest proxy response happens in the northwestern Atlantic between 40 and 60∘ N, at 1000 and 5000 m. In this area, the δ13C and Δ14C display marked decreases during the hosing, while the Pa/Th significantly increases. Another interesting feature is that the δ13C and Δ14C lag behind the Pa/Th by about 200 years (Fig. 5a). In our model, this zone corresponds to the region of NADW formation and first portion of southward flow at depth.

One of the most interesting features of our simulations is that the δ13C and Δ14C response lags behind the Pa/Th response by a few hundred years, in particular in the northwest Atlantic, bathed by the NADW, suggesting a shorter time of response for the Pa/Th compared to carbon isotopes.

First, let us point out that the physical mechanisms leading to the δ13C, Δ14C and Pa/Th signals are fundamentally different. On the one hand, Pa and Th have very short residence times in the water column. In our simulations, the particles fluxes remain prescribed and constant. Therefore, any change in the particle production due to the freshwater forcing is not accounted for and does not impact the sedimentary Pa/Th in our simulations. Instead, the sedimentary Pa/Th in the North Atlantic purely reflects the capacity of the NADW to transport Pa further south. Thus, the NADW cessation rapidly translates in a sedimentary Pa/Th increase, which ceases as soon as the NADW resumes and is able to transport Pa again, at the end of the hosing period (i.e. after year 600). On the other hand, the oceanic carbon isotopes depend on not only circulation changes but also the biological activity and air–sea exchanges. Carbon is further actively partitioned between different reservoirs (terrestrial biosphere, ocean and atmosphere), hence the longer time needed by carbon isotopes to adjust to any change.

In our simulations, the freshwater perturbation lasts 300 years, and the NADW formation is stopped for about 100 years (Fig. 5), which is likely too short for the isotopic systems to fully adjust to the change. However, our experimental set-up is relevant to the study of centennial- to millennial-scale events (DO or Heinrich events) across which the climate and isotopic systems likely did not have the time to fully adjust either. In line with previous hosing experiments (e.g. Kageyama et al., 2013; Mariotti et al., 2012; Menviel et al., 2008, 2015), the freshwater addition in the Nordic Seas produces a surface cooling of about 1 ∘C associated with an increase in the CO2 solubility in surface waters and a general marine productivity decrease in the North Atlantic. Subsequent to the NADW cessation, the nutrients accumulate in the Atlantic intermediate waters, shaping a negative δ13C anomaly and its counterpart positive phosphate anomaly in the North Atlantic around 50∘ N (Menviel et al., 2015). In our simulations, the biological productivity or air–sea exchanges response peak around year 600, when the AMOC is the weakest (Fig. S7). Therefore, the lag between the Pa/Th and carbon isotopes responses is not related to late biological or air–sea exchange responses themselves but rather to the time needed to transport the anomaly to the deep ocean. Indeed, in our simulations, the northern sourced δ13C anomaly progressively spreads to the deep Atlantic basin and finally reaches the bottom ocean (> 4000 m) around year 800, about 200 years after the AMOC has reached its lower value (Fig. S8).