The models assigned to this category are AKSE1, BVMO1, CENER1, DNV1/2/3, JMUC1, NLR1, and RAM1(a)/20/21/25(a) (Bredmose et al., 2024; De Lauzon et al., 2013; Gao et al., 2023; Karch et al., 2024; Knezevic et al., 2022; Touzon et al., 2025; Yim et al., 2026).

This category is used by most participants and appears to represent the current standard industry practice. It involves a sequential workflow where the ILA is conducted first, followed by the FEA.

In most models applied in this project, due to the quasi-static (QS) nature of the FEA, the structural dynamics of the hull are generally not captured in the obtained structural responses. The RAM20/21/25(a) models are an exception, where the global effect of hull flexibility is captured in the ILA and transferred to the QS FEA through inertia loads. In general, methods adopted for DNV1/2, JMUC1, and NLR1 would allow for consideration of structural dynamics via fully dynamic time-domain FEA; however, this feature is not utilized in this project.

By separating the ILA and FEA, the global-to-local link in these models remains relatively loose and relies heavily on the quality of the load transfer (mapping) approach. A critical aspect of this process is the inclusion of all ILA load components; this is essential for maintaining load balance within the FEA. Any imbalances resulting from transfer inaccuracies must be equilibrated through support reactions, inertia relief, or other techniques. However, if a load component is neglected and the resulting imbalance is non-negligible, the equilibration will likely fail to reflect the physical nature of the missing load, ultimately distorting the calculated structural responses. In this regard, it should be noted that models CENER1, JMUC1, and NLR1 do not transfer viscous drag to the FEA; hence, the applied load equilibration technique might become relevant (see Table 1). For other models, the FEA equilibration technique is less relevant as the imbalance is either proved or expected to be negligible.

A major differentiator of the models in Category 1 is the way the hydrodynamic loads from the ILA are converted into pressures (i.e. where the applied pressures originate from). Here, three options exist (see Table 1): a.

Pressures are obtained directly from the time-domain potential flow solution using the rankine source method (DNV1 and JMUC1).

The models assigned to this category are are CIMN1, 2, and 3 (García-Espinosa et al., 2023; Berdugo-Parada et al., 2024; Serván-Camas et al., 2025; García-Espinosa et al., 2026).

In the models of this category, a pure time-domain analysis approach is used, combining global response and structural FEA. Hence, the global-to-local link in these models is inherently strong. The structural model includes hull, tower, topside (i.e. rotor-nacelle assembly, or RNA), and mooring lines. The potential flow problem is solved directly in the time domain using the FEM on a fluid volumetric mesh. In the structural solver, a dynamic analysis is performed in the time domain, inherently capturing structural flexibility and structural dynamic effects. Because the fluid and structural solvers are mathematically coupled at every time step, load equilibrium is inherently satisfied without the need for artificial boundary conditions or inertia relief techniques.

The fluid–structure interaction is considered on two levels: a.

The hull is coupled as a rigid body to the hydrodynamic potential flow solution, while elastic deformations have no effect on the fluid pressure (one-way coupling) (CIMN1).

The model assigned to this category is PRI1 (Defoy et al., 2018; Quideau et al., 2025).

In this category, no external FEA model is used, and no load transfer from ILA to shell-model FEA is done. Consequently, the discussed concepts for pressure calculation/reconstruction, load equilibration, etc., are not applicable. The global-to-local link is inherent but simplified in the way that local structural responses (stresses) are obtained from the global responses/loads. The ILA model differs significantly from the OrcaFlex reference model (see Sect. 6.1); the hull is modelled using skeleton beams, with potential flow multi-bodies attached to them at appropriate locations. The beams are joined using rigid connections, and their stiffnesses are defined based on computer-aided drafting (CAD) geometry to approximate the global structural flexible behaviour. Hence, global structural dynamics are inherently captured within the ILA. The structural responses (stresses) are derived from sectional beam loads (axial and bending) in a simplified way based on classical beam theory. With that, only longitudinal stress components can be retrieved, and the accuracy is expected to be limited around beam joints; i.e. no stresses are available inside the joints.

Disregarding the above toolchain categorization, the different models are further distinguished in Table 1 by the vertical extent of fluid pressure application. Most models (BVMO1, DNV1/2/3, NLR1, RAM1/20/21/25, and CIMN1/2) follow a linearized approach, applying both hydrostatic and hydrodynamic pressure components up to the initial, undisplaced still-water level (SWL). For models based on unit loads/responses, such linearization is usually a common limitation.

In contrast, other models (especially those applying pressures step by step) can capture specific non-linear surface effects:

Hydrostatic non-linearity. AKSE1 applies hydrostatic pressures to the instantaneous SWL in time domain, considering roll, pitch, and heave motions and accounting for a physically more correct pressure distribution. The unit-load-based models RAM1a and RAM25a also capture this effect by employing a data-driven reduced-order model, utilizing orthogonal eigenpressure modes as unit loads to reconstruct the hydrostatic pressures to instantaneous SWL in the time domain.

To better align the FE models used in the different toolchains, a reference structural model of the VolturnUS-S semi-submersible was created at the start of the project (see Fig. 2). The structural layout, including plate thicknesses and internal stiffening, was based on minimum scantling requirements and created solely for educational purposes of this project and to maintain consistency across the participants, with no intention of fully complying with any structural rule requirements (i.e. not to be used for a real engineering project).

The original VolturnUS-S semi-submersible design has ballast in columns and pontoons to achieve the required operating draft. While modelling of the ballast is straightforward in global response analysis (ILA), its consideration in structural assessments can be challenging. Some tools distribute the ballast mass to nodes of the FE mesh, while others model it as mass points connected to the FE mesh via springs or as pressure distributions. The deviations in the evaluated structural responses introduced by the different modelling techniques can be difficult to isolate and assess and would require another dedicated study. Therefore, the global effect of the ballast was replicated in this study in a simplified way by increasing the material density in different areas of the FE mesh.

The reference FE model consists of 420 672 shell elements (i.e. no 1D beam elements or solid elements) and only includes the hull. The tower and mooring lines are not modelled in FE and are applied as interface loads during ILA-to-FEA load transfer. The model is supported at the base of its three columns, constraining 3 DOFs (x, y, z) at column 1, 2 DOFs (y, z) at column 2, and 1 DOF (z) at column 3. The support conditions primarily affect the static analyses in LC2.1 and LC2.2; in the free-floating cases, the external loads should be in equilibrium, making FE support reactions negligible.

The reference FE model was created by Ramboll in ANSYS and shared with participants in various export formats and as a CAD geometry file to accommodate the different toolchains involved. In all toolchains using a sequential workflow (Category 1), the reference mesh definition could be adopted directly, or similar meshes could be created based on the CAD geometry. The integrated workflow (Category 2) applied in CIMN models required a structural model that included the entire floating structure (hull, tower, RNA, and mooring lines) and therefore had to be extended accordingly. The simplified workflow (Category 3) applied in the PRI1 model does not require a separate structural FEA model.

The alignment of the applied FEA models was verified based on the following:

mass and inertia check (LC1.1 in Table 3), which was performed by the participants during FE model creation and is not further detailed in this paper;

To investigate the global-to-local (i.e. ILA-to-FEA) load transfer, it is important that the global basis (ILA) is well aligned between all participants. Therefore, at the start of the project, Ramboll provided the project participants with a reference global response analysis (ILA) model in OrcaFlex software combined with model descriptions and parameter overviews. The modelling details aimed to suffice the common floating wind industry requirements and practices (e.g., hull hydrodynamics based on BEM, consideration of viscous drag, non-linear mooring) while at the same time being relatively simple to better align with capabilities of available load transfer approaches, allowing more institutions to participate (e.g. rigid hull, exclusion of ballast, exclusion of second-order wave loads).

This reference model was used directly by most participants for their analyses. Where specific toolchains required a different ILA software (Opera, Sima, DeepLines Wind, SeaFEM), the respective models were created to adopt (if possible) the OrcaFlex reference model features. Furthermore, some model variations were created by participants on purpose to investigate specific modelling aspects (e.g. hull flexibility, pressure application extent).

For this project that largely focuses on the ILA-to-FEA load transfer, the (hydrodynamic) modelling of the hull in the ILA is of particular interest; mooring lines and tower/RNA modelling are less relevant, as these components are usually not included in the same FEA and considered only via respective interface loads.

In the OrcaFlex reference model, the hull is modelled as a single rigid body. The first-order potential flow loads (radiation/diffraction) are reconstructed from frequency-domain coefficients obtained with BEM: radiation forces are calculated via the Cummins convolution integral, while wave excitation loads are applied via superposition of linear transfer functions. Hydrostatic stiffness is linearized about the initial (undisplaced) floating position. Viscous drag is modelled non-linearly using Morison elements, with drag forces integrated up to the instantaneous water surface using Wheeler stretching. Second-order wave loads are not considered. Mooring lines are modelled with finite elements (OrcaFlex line objects). The tower was modelled as a flexible beam and the RNA as a mass point with inertia and pre-computed load time series (see Sect. 4) applied to it in selected load cases (see Table 3). Wind drag was neglected.

The alignment of the global response models was verified based on the following:

static analyses in freely floating conditions (LCs 2.3, 2.4, and 2.5);

For the comparison, two groups of models should be distinguished where deviations are expected:

The first is models aimed to replicate the same features as the OrcaFlex reference model (rigid hull): AKSE1, BVMO1, CENER1, DNV1/2/3, JMUC1, NLR1, RAM1(a). In models RAM1(a), the hull was divided into 14 segments and treated as a multi-body system in the BEM solution. This is necessary for the global influence superposition load transfer approach (Karch et al., 2024). However, since the segments are rigidly connected, the global performance behaviour is the same as that of the single-body reference model.

Figure 11 shows panel-averaged stress components Sxx (in longitudinal pontoon/column direction) for panels CentCol, P1_S1, P1_S3, and Col_1 (see Fig. 3). As additional information, fatigue damage indicators (FDIs) are included as red dots in the statistics plots. The FDIs are essentially damage ratios relative to the RAM1 model, which is used as the reference for comparison. Damage is calculated based on rainflow counting of stress component time series, considering the S–N curve exponent of m=3. The FDIs are intended solely as qualitative indicators of fatigue damage accumulated in the different models and do not represent a rule-compliant fatigue assessment.

The time series plots in Fig. 11 reveal discrepancies between the global analysis models discussed in Sect. 6.2. The models based on the OrcaFlex reference model form one cluster (see zoom-in region 300–315 s), with the other rigid-hull models close by, except for BVMO1. All flexible-hull models deviate from this cluster. The statistics plots show that most models are relatively well aligned. Those in the cluster in particular show good consistency in mean and standard deviation values and mostly in FDIs. The statistics of the flexible-hull models CIMN1/2/3 and RAM20/21/25 generally align quite well with these. The flexible-hull PRI1 model aligns well on standard deviation values, while deviating on the mean values.

By comparing RAM20/21/25 with RAM1 (all using linearized hydrostatics) and RAM25a with RAM1a (both using non-linear hydrostatics), it is evident that considering hull flexibility increases the standard deviation values, particularly the FDIs. The amount of structural damping included in the RAM20/21/25 hull models (see Table 1) clearly impacts the FDIs.

A comparison between the CIMN1 and CIMN2 models shows that considering two-way hydrodynamic coupling – where elastic deformations affect hydrodynamic pressures – accounts for the hydrodynamic damping caused by waves radiating by elastic displacements, leading to reduced FDIs and a potentially more realistic representation of energy dissipation mechanisms.

Furthermore, a significant reduction in FDI can be seen in panels P1_S1, P1_S3, and Col_1 of models CIMN3, AKSE1, CENER1, JMUC1, RAM1a, and RAM25a. These models differ from similar ones in how hydrostatic pressures are applied in the FEA. While models such as BVMO1, DNV1/2/3, NLR1, RAM1/20/21/25, and CIMN1/2 apply a linearized distribution, models CIMN3, AKSE1, CENER1, JMUC1, RAM1a, and RAM25a use a non-linear formulation. This leads to different load path distributions within the hull structure (see Fig. 12).

In models with linearized distribution, the hydrostatic restoring moment is generated through negative pressures on the fore column. In models with non-linear distribution, however, it is generated through positive pressures on the aft columns. For LC4.1, the resulting force vector on the fore column is greater for models with a linearized distribution. This increases the sectional loads in the column base and throughout the pontoon sections, consequently increasing the panel-averaged stresses in the respective panels.

The PRI1 model uses a non-linear pressure formulation, and the deviations in stress statistics are likely partially related to this approach. However, the impact of the pressure formulation on the CentCol panel is unexpected; therefore, the remaining deviations are likely due to the simplifications of the beam-based structural model.

The impact of pressure formulation becomes very evident when looking at the sectional bending moments at the base of fore column (Col_1); see Fig. 13 (for CENER1 and BVMO1 no results were available).

The panel-averaged stresses (Sxx) for the CentCol and P1_S3 panels are evaluated for load cases LC6.2 (without RNA load) and LC6.3 (with RNA load), with irregular waves and Hs = 1.0 m, on a case-by-case basis. This allows the impact of the RNA loads to be assessed directly. Evaluation of the other panels leads to similar conclusions regarding the overall behaviour of the different models and is therefore omitted.

Figure 14 shows the stresses in panel CentCol. Without an RNA load (LC6.2), the tower leans forward, resulting in compressive stresses and small, motion-induced fluctuations from the low-wave-amplitude seaway with Hs = 1.0 m. The stress spectrum is governed by wave excitation and, for flexible models CIMN2 and RAM20/21, exhibits distinct peaks around the first fore–aft bending mode. The RNA load (in LC6.3) leans the tower in the opposite direction and results in tensile stresses with much more pronounced fluctuations (higher standard deviation values). The stress spectrum shows much higher energy content and is dominated by RNA-induced pitch motions.

While some differences (especially in mean values) exist, the overall agreement between the models is good. In particular, rigid-body models show good agreement in standard deviation and FDI values. Here, BVMO1 deviates from the others. In LC6.2, the flexible models RAM20/21/25 and CIMN1/2 show significantly higher standard deviation values and FDIs compared to rigid-body models. The RAM20 model without structural hull damping experiences excessive excitations at the first fore–aft bending mode, which reduce with the increased structural damping in RAM21/25 models but are still pronounced. The RNA load in LC6.3 stabilizes the structural response of the flexible models, making the results and conclusions very similar to those from LC4.1 (RNA only).

Figure 15 shows the stresses in panel P1_S3. Similarly to panel CentCol, the stress spectrum is dominated by wave excitation in LC6.2 and by RNA-induced pitch motions with a higher energy content in LC6.3. However, the excessive impact of the structural dynamic effects observed in the flexible models in panel CentCol is not seen in panel P1_S3.

Comparing AKSE1, CENER1, JMUC1, and PRI1 results and directly comparing RAM1 vs. RAM1a, RAM25 vs. RAM25a, and CIMN2 vs. CIMN3, it is evident that the hydrostatic modelling strategy (linearized vs. non-linear) has no impact on the CentCol panel (see Fig. 14). In contrast, in panel P1_S3 (close to the fore column), a clear impact of the hydrostatic modelling strategy can be observed in LC6.3 (with RNA load), where the pitch angle is considerable, consistent with the observations from LC4.1 (RNA only).

Including fluid–structure interaction (i.e. two-way coupling; see CIMN2 vs. CIMN1) appears to amplify the FDIs in LC6.2 (wave only), suggesting an increase in wave loads and/or structural oscillations resulting from them. By contrast, in LC6.3, where platform motions are mainly enforced by external RNA excitation, the inclusion of fluid–structure interaction has a damping effect through waves radiated by elastic displacements, thereby reducing the FDIs.

Similar to Sect. 7.2, the evaluation of the panel-averaged stresses (Sxx) for load cases LC7.2 (without RNA load) and LC7.3 (with RNA load), with irregular waves with Hs = 11.0 m, is limited to panels CentCol and P1_S3. Evaluation of other panels leads to similar conclusions and is therefore omitted.

Figure 16 shows the stresses in panel CentCol. Without an RNA load (LC7.2), the stress spectrum is completely governed by the high-amplitude wave excitation and resulting floater pitch response, with small peaks around the first fore–aft bending mode appearing in flexible models. The low-frequency RNA load shifts the distribution in the PSD in LC7.3, and the highest peaks occur at the RNA load frequency and the pitch natural frequency. Still, the wave energy content is very pronounced.

The time series plots in Fig. 16 reveal good overall agreement between most of the models. The time series for all rigid-body models follow a very similar course, except for BVMO1. This is due to the weak alignment of its global analysis model, which is potentially caused by the deviating formulation of the drag coefficients. The flexible models are generally closer to the rigid-body ones and are less scattered than in other LCs. This good agreement is also reflected in the statistics plots, with most mean and standard deviation values being relatively close. This indicates a good alignment between the global analysis models in terms of wave response. Most models compare very well on the applied hydrodynamic loads (radiation, diffraction, and drag), which clearly govern the loads in LCs 7.2 and 7.3.

Rigid-body motions dominate structural responses and structural dynamics, which makes the effect of hull flexibility less visible in time series and standard deviation values. Nevertheless, the FDIs highlight its importance; see the results for CIMN1/2 and RAM20/21/25(a). The FDIs of the flexible PRI1 model do not align with these models, which is probably due to the generally lower accuracy of the beam-based structural model.

The AKSE1 and JMUC1 models have higher FDI values than other rigid-body models. This cannot be attributed to their non-linear treatment of hydrostatic pressures, as this should have no impact on the CentCol panel and is also not observed in other models, such as CENER1, RAM1a, and PRI1, which treat hydrostatic pressures in a similar way.

The results for panel P1_S3 in Fig. 17 lead to similar overall observations. While the time series and the mean and standard deviation value statistics are very well aligned, the FDIs show discrepancies that are, in some cases, significant, even among similar rigid-body models (see BVMO1, DNV2/3, NLR1, and RAM1). These discrepancies cannot be explained by differences in the hydrostatic pressure treatment: RAM1 and RAM25 are similar to RAM1a and RAM25a, respectively, while e.g. AKSE1 significantly deviates from these. They also cannot be attributed to structural dynamics: rigid-body RAM1 behaves similarly to the flexible-body RAM20/21/25, while CIMN1/2 is significantly different.

The effect of fluid–structure interaction (see CIMN2 vs. CIMN1) in LCs 7.2 and 7.3 is much less pronounced than in LCs 6.2 and 6.3 (Hs = 1.0 m) and is only observed in the increased FDIs of the CentCol panel in LC7.2. This suggests that rigid-body hydrodynamics are the primary contributor to structural loading.

The effect on non-linear hydrodynamics (i.e. the Froude–Krylov pressures applied to the instantaneous wave surface) was expected to be visible in LCs 7.2 and 7.3 (Hs = 11.0 m). However, the relevant models (CENER1, JMUC1, and PRI1) are too inconsistent and do not allow clear trends to be derived in this regard. For CIMN3, no results were available.

One remaining possible explanation for the inconsistencies observed in the calculated FDIs in LCs 7.2 and 7.3 is the difference in FEA load equilibration techniques. This becomes relevant if residual loads arise from the ILA-to-FEA load-mapping process, which is more likely to occur in LC 7.2 and 7.3 environments with high loads and motions.

The comparisons across LCs 4.1, 6.2/6.3, and 7.2/7.3 show that the influence of the modelling technique depends strongly on the governing excitation mechanism. For RNA-load-dominated cases, the treatment of hydrostatic pressures and structural flexibility has a clear effect on stress ranges and fatigue damage indicators (FDIs). For wave-dominated high-sea-state cases, global hydrodynamic loading and rigid-body motions dominate the stress response, while differences in FDI remain sensitive to structural dynamics and local load-mapping details. The main trends are summarized in Table 4.