Numerical investigation of response-conditioning wave techniques for short-term rare combined loading scenarios

Response-conditioning wave techniques are a rational way to link wave excitation environments with return- period extreme loading responses. By retaining the wave excitation which leads to a design response, these techniques can also define extreme combined loading scenarios. For complex or novel hull forms, combined loading may be relevant for evaluating structural reliability or adequacy. But using combined loading scenarios as inputs to high-fidelity structural or dynamic modeling tools implies that such load scenarios are realistic for the defined return-period. This paper investigates three response-conditioning wave techniques: a modified Equivalent Design Wave method, a modified Conditioned Random Response Wave method, and the Design Loads Generator, to evaluate how well they reproduce combined loading statistics for a 1000-hr return-period as compared to stochastic brute-force simulations. The investigation is carried out for extreme combined loading scenarios on a 110 m trimaran hull. The Design Loads Generator produces the most realistic extreme combined loading statistics as compared to the brute-force approach with a significant reduction in computation time based on combined load conditional probability density functions, cumulative density functions, and individual stochastic load vectors.


Introduction
To ensure adequate lifetime structural performance, many classification societies suggest using combined loading scenarios as possible loading conditions to test a vessel's structural compliance, see, e.g. Lloyd's Register (2017), Bureau Veritas (2018a), DNV-GL (2018), Bureau Veritas (2017, 2018b, American Bureau of Shipping (2016Shipping ( , 2005, IACS (2018), Horn et al. (2013), ClassNK (2013) and Lloyd's Register (2014). The intent of using such combined loading scenarios is to capture the effect of stochastic loads acting in different planes of the vessel. However, defining such loading scenarios is not simple, especially for rare events. Even defining a return period for multivariate extremes can be ambiguous, see, e.g. Feld et al. (2019).
To define simultaneous load effects due to combined loading may require the underlying wave profile which excites these stochastic loads. Rare combined loading scenarios may be associated with long exposure periods, making brute-force simulation techniques, such as Monte Carlo Simulations (MCS) potentially intractable. Multiple Response-Conditioning Wave Techniques (RCWTs), such as the Equivalent Design Wave (EDW), Conditioned Random Response Wave (CRRW), and the Design Loads Generator (DLG), have been developed to construct wave profiles expected to excite specific loading scenarios.
H.C. Seyffert et al. The paper is structured as follows: Section 2 gives some background to the RCWTs examined in this investigation. Section 3 establishes the trimaran case study and benchmarks to compare the combined loading scenarios from the RCWTs with MCS. Sections 4-7 evaluate the techniques via the established benchmarks. Finally, Section 8 discusses the trade-offs between the RCWTs and Section 9 offers some recommendations and conclusions.

Response-conditioning wave techniques
This section gives a short background on some response-conditioning techniques which define a wave excitation profile based on a specific load response with a given return period. It is worth acknowledging the many other methods that examine combined loading, including, but certainly not limited to: combination formulas like Turkstra's rule (Turkstra, 1970) or the Ferry Borges-Castanheta method (Ferry-Borges and Castanheta, 1971), load coincidence models (Wen, 1993), point-crossing methods (Huang and Moan, 2008), and cross-spectral methods (Alfred Mohammed et al., 2012Mohammed et al., , 2016. For a recent investigation of such load combination models and their sensitivity when performing a hull girder reliability analysis, see Piscopo and Scamardella (2019).
This paper, however, only examines a few response-conditioning wave techniques which are chosen based on either their present usage as defined by classification societies (EDW), potential for expanding on these classification society methods (CRRW), or expected accuracy in capturing lifetime conditional loading statistics (DLG). In general, all of these techniques have a similar aim: to construct wave profiles which lead to extreme responses of a defined load scenario. Consider a stochastic wave profile time series ( ): where = √ 2 ( ) ( ) = single-sided wave energy spectrum = phase between − and If the phases are uniformly distributed between − and , as the number of frequency components approaches infinity the random variable expressed by the random process ( ) approaches a Gaussian random variable. In general, RCWTs condition a wave profile expected to lead to a pre-determined response value at a pre-determined time, say = 0. In this case, the phases may not be uniformly distributed between − and but will likely be tuned based on a transfer function of the desired response. Consider now some response of interest ( ) which is excited by the wave elevation ( ), expressed as: The random variable similarly approaches a Gaussian random variable as approaches infinity and if are uniformly distributed between − and . The RCWTs focus on defining the wave phases such that the resulting ( = 0) response is a rare occurrence. Of course, Eqs. (1)-(2) are linear relationships, but in general, response conditioning wave techniques are based upon linear theory and the assumption that a nonlinear response is a correction of the associated linear response. It is expected, though, that the processes employed in this investigation and conclusions would be similar for nonlinear, non-Gaussian models.

Equivalent design waves
Equivalent Design Waves are regular waves constructed to lead to a specific load value, here referred to as a design value, in a given heading at a specified instant in time, again at = 0. To define a load combination case, a single load is chosen as the primary load, called , which is the load the EDW is constructed to maximize. This EDW is constructed using the response amplitude operator (transfer function) of ( ( )) in a particular heading. The frequency of the EDW, called * , is the encounter frequency corresponding to the peak value of ( ). By construction, the EDW and resulting load profiles are regular waves with this same frequency * . The value of the primary load and phase at this frequency * are used to define the resulting EDW. Interested readers are referred to Seyffert and Kana (2019) and Seyffert et al. (2018) for more discussion on the implementation of the EDW technique for this investigation.

Implications of EDW assumption
The major assumption of the EDW method is that the wave which excites a specified loading scenario is regular and that the resulting loads are regular waves with the same frequency. This makes for a very simple construction of the EDW and resulting load profiles but means that the resulting load profiles have a limited range. Consider a vector of time-varying load profiles excited by an EDW. Since all of the load profiles are phase-shifted cosines with the same frequency * , the implication is that the combined load vector in -dimensional space (each dimension relating to a specific global load) will be a 2− ellipse.
The shape of an EDW-driven load vector is pre-determined because any phase-shifted cosine with a given frequency can be represented as a weighted linear combination of a sine and cosine wave at that same frequency. For load vectors excited by an EDW, this means that the range of an -dimensional load vector excited by an EDW has the maximum dimension of 2. Graphically, an -dimensional load vector excited by an EDW plotted in -dimensional space will trace out an ellipse which lies on a 2 − plane. The range of that ellipse has dimension 2 and the null-space has dimension − 2, regardless of the number of loads excited by the EDW. 1 The major question for evaluating the EDW as a response-conditioning wave technique is then: do the limitations stemming from the regular wave assumption outweigh the simplicity of application? Can the EDW lead to realistic combined loading scenarios as compared to more complex RCWTs and the benchmark MCS?

New waves, most-likely waves, and conditioned random response waves
Many response-conditioning wave techniques have been developed to take the general idea of the EDW (to construct a wave which leads to 1 Consider the set of functions on R that are a sum of a sine and cosine wave with the same frequency: i.e. ( ) + ( ), where and are real constants. is a vector space because linear sums of elements of and scalar multiples of elements of are also members of . Since ( ) and ( ) are maximally linearly independent in , meaning that there are no other linearly independent functions in the vector space , they form a basis of V. The implication of this is that a vector of combined loads excited by an EDW has at maximum 2 linearly independent dimensions. Such results can readily be confirmed from linear algebra or signal processing textbooks, see, e.g. Smith (2002). Indeed, the fact that ( ) and ( ) form a basis of is an underpinning of the Fourier series.
an extreme system response of interest) and correct for its shortcomings (namely, that the EDW is a regular wave). These techniques do not determine such wave profiles via brute-force simulation techniques but rather linear and probabilistic theory based on expectations of wave profiles given certain conditional criteria. Lindgren introduced the New Wave profile, giving a form for the expected (i.e. most-likely) profile of a Gaussian process conditioned on an extreme peak value at a given time instant (Lindgren, 1970); this result was also demonstrated by Tromans et al. (1991). Friis-Hansen and Nielsen developed the Most-Likely Wave (MLW) by expanding the New Wave profile to condition on both the wave amplitude and instantaneous frequency (Friis-Hansen and Nielsen, 1995). If this instantaneous frequency is the mean frequency = 1 ∕ 0 , where is the th moment of the wave energy spectrum, the MLW simplifies to the New Wave profile.
But since this New Wave profile is simply the autocorrelation function of the wave spectrum, Taylor et al. added randomness to the expression by embedding this most-likely wave profile into a random seaway (Taylor et al., 1997). However, a single extreme wave crest may not excite extreme dynamic responses. Airy wave theory suggests that the wave leading to the worst-case system response is the one which excites a response where all frequency components have a peak at the same time. Torhaug developed the Critical Wave Episode model which identifies wave profiles via a linear model that are expected to lead to extreme system responses. This model uses simulations based on design sea state histories to screen for critical wave characteristics (Torhaug, 1996). For example, a critical wave characteristic might be extreme wave height, meaning that the constructed Critical Wave Episode is based on the New Wave model.
Other authors have constructed wave profiles which lead to specific vessel responses by using the linear transfer function that links the wave excitation and response. Dietz combined the Most-Likely Wave (MLW) profile from Friis-Hansen and Nielsen with the Most-Likely Extreme Response (MLER) method from Adegeest et al. (1998) to develop the Most-Likely Response Wave (MLRW) (Dietz, 2004). The deterministic MLRW is conditioned on a linear vessel response taking a specified extreme value at a given instant in time with a given instantaneous frequency. If this instantaneous frequency is the mean frequency, the MLRW is the same as the MLER wave. When this MLRW/ MLER wave includes a random background, it is referred to as a Conditioned Random Response Wave (CRRW).
The MLWR (and therefore the CRRW) is similar to the EDW approach in that the linear response transfer function defines the resulting phase of the conditioned wave component at each frequency component. The MLRW uses the range of frequencies from the response transfer function, instead of the peak frequency like the EDW, to define the amplitude of each wave component. The MLRW incorporates the envelope process of the response along with an instantaneous frequency requirement at the time of the response event. The CRRW further includes normally distributed coefficients in the wave amplitudes, leading to an irregular wave profile. The details of the derivation of the CRRW can be found in Dietz (2004); the formulation of the CRRW technique employed here follows Drummen et al. (2009). It is worth noting, though, that the CRRW does not include randomness in the extreme loading event itself as this value is fixed a priori.

Generating a distribution of extreme responses from EDW & CRRW
By construction the EDW and CRRW techniques are designed to lead to a single a priori defined load value. While the CRRW includes randomness in the constructed wave profile, there is no randomness included in the extreme load response of interest at = 0. In reality though, a vessel experiences a distribution of extreme loading based on the exposure time in the operational profile. This is reflected when running MCS and in waves constructed by the Design Loads Generator, as discussed in the next section (Section 2.4): a distribution of extreme responses is collected. To compare with the DLG and MCS, the EDW and CRRW techniques are adjusted to construct an ensemble of waves which lead to a distribution of extreme load values.
A closed-form distribution can be defined for extreme values of a Gaussian process given its zero-crossing period and the return-period. Then, an ensemble of extreme load values can be chosen such that their empirical histogram approaches the theoretical Gaussian extreme value distribution (EVD) for the defined exposure. As deep-water waves can be assumed Gaussian and the loading transfer functions are linear, this is a reasonable estimation of the extreme loading distribution. This logic constructs the DLG wave profiles: an ensemble of waves that lead to a distribution of extreme responses whose empirical histogram approaches the Gaussian EVD. Based on this ensemble of extreme load values which follow the Gaussian EVD, an ensemble of Equivalent Design Waves and Conditioned Random Response Waves can be constructed which excite these load values. This approach distinguishes the modified-EDW (m-EDW) and modified-CRRW (m-CRRW) approaches used in this investigation. This modification allows a direct comparison between the m-EDW, m-CRRW, DLG, and MCS approaches for the different combined loading scenarios considered, as all RCWTs construct ensembles of waves leading to distributions of extreme load responses.

Design loads generator
The DLG was developed to construct an ensemble of irregular wave profiles which are expected to lead to extreme return-period linear system responses based on a loading transfer function and an exposure time (Alford, 2008;Kim, 2012). Unlike the response-conditioning wave techniques discussed in the previous subsections, the DLG does not fix the extreme load value a priori and then back-calculate a wave profile which leads to this response based on summing phase-shifted regular wave components via Airy wave theory. Instead, the question that drove the development of the DLG was: what is the distribution of wave phases in Eq.
(1) that leads to extreme responses of ( ) at time = 0 which belong to the Gaussian EVD for a given exposure? The main differences between the RCWTs discussed earlier and the DLG are: • The DLG does not require an a priori definition of the extreme load value at = 0 to construct the underlying wave profile, unlike EDW and CRRW. The extreme load value excited by a DLG wave is by construction a member of the Gaussian EVD for the defined exposure period. The DLG constructs an ensemble of short irregular wave profiles which excite extreme return-period response values which belong to the theoretical Gaussian EVD based on a characteristics period of the response (e.g. zero-crossing period) and the return-period. • The phases in Eq. (1) of the DLG waves are not limited based on the phases of the associated linear transfer function which describes the response of interest.
-The EDW can be seen as a simplification of Eq. (1) where only the frequency * which corresponds to the peak of the load is included and where the height of this wave is scaled to lead to a specific design value. By construction then, the EDW wave phase is defined by the transfer function phase at the frequency * , − .
-The MLRW/ CRRW wave uses more frequency components than the EDW and conditions based on an instantaneous frequency, but also has the phase of each regular wave component defined completely by the load transfer function; i.e. = − . Therefore, both EDW and CRRW construct waves with phases at each frequency component which only take on values contained in the range of phases in the load transfer function. In contrast, the DLG wave components at each frequency can take phases between − and .

Case study: Trimaran
For the numerical investigation of how well different responseconditioning wave techniques describe lifetime combined loading scenarios, a trimaran hull is examined. Combined loading is relevant for trimaran structural design, especially the cross deck structure. Increasing advances in high-fidelity structural modeling allow an in-depth study of such complex structural responses. But the limited historical and anecdotal knowledge about trimarans makes efficiently applying such models difficult. Therefore, response conditioning wave techniques may be relevant to define a finite number of lifetime combined loading scenarios for use in such tools. Indeed, some classification societies have already suggested this procedure to evaluate the adequacy of trimaran structures, e.g. Lloyd's Register (2017).
To compare load combination scenarios defined by the different RCWTs, the trimaran hull is the same as used in Seyffert and Kana (2019), Seyffert et al. (2018), which examined combined loading scenarios defined by the EDW technique. The trimaran parameters and operational profile for the load scenarios are given in Table 1. For this analysis, the rareness of the load combinations is limited by the exposure period which can reasonably be simulated by MCS to give converged statistics. Therefore, the chosen exposure time is 1000-h. Although a designer may be interested in rarer combined loading events, it is relevant to evaluate whether the RCWTs accurately describe load scenarios at a shorter return-period like 1000-h. It is not expected that EDW or CRRW will become more accurate for describing rarer responses or for flexible hulls where non-linearities will be more dominant, as compared to brute force stochastic simulations or physical model tests, see, e.g. Drummen et al. (2009). The DLG waves will still be statistically valid for events associated with longer return-periods, as any return-period can be related to a Gaussian EVD, though the Gaussian process assumption may break down for such rare responses due to non-linearities.

Combined loading scenarios on a Trimaran
Typically combined loading scenarios are based on conditional maxima. For all of the techniques described in Section 2, the load scenarios are defined by the simultaneous responses of all loads when a specific global load attains its 1000-h extreme value. These global loads are defined in Fig. 1. The vertical wave bending moment, , is the wave pressure distribution integrated across the depth of the vessel and the horizontal bending moment, , is the pressure distribution integrated across the vessel breadth.   For this investigation, combined load scenarios are defined when each of the global loads in Fig. 1 attains its extreme value in a given heading. The linear version of the high-order potential flow code Aegir (Kring et al., 2004) was used to generate the trimaran transfer functions of the global loads. These transfer functions are adapted from Seyffert and Kana (2019) and Seyffert et al. (2018), where the minimum earth-coordinate frequency was 0 = 0.4859 rad/s. Here, the transfer functions are linearly extrapolated so that the minimum frequency is 0 = 0 rad/s. The amplitudes are linearly extrapolated so that the amplitude at 0 = 0 rad/s is zero. The phases are extrapolated so that all phases at 0 ≤ 0.4859 rad/s equal the transfer function phase at 0 = 0.4859 rad/s. This extrapolation is necessary to conduct MCS to examine lifetime loading scenarios. The MCS analysis uses the Fast Fourier Transform (fft) algorithm to simulate the stochastic ocean environment. The fft algorithm assumes the frequency range begins from 0 = 0 rad/s, requiring the extrapolation for this investigation. The extrapolated transfer functions are shown in Fig. 2.

Design values of global loads
As described in Seyffert and Kana (2019) and Seyffert et al. (2018), the heading of each load scenario is based on the heading where the specific global load attains its largest 1000-h extreme value. In each heading, the most-likely extreme value of each load is determined using extreme value theory, see, e.g. Ochi (1990), with a zero-crossing period based on spectral moments. These 1000-h most-likely extreme values based on heading are given in Table 2. The largest most-likely extreme value of a load over all the headings is defined as the design value , and given in bold. Not all transfer functions were constructed for all headings. In those cases, the most-likely extreme value for that load case is marked as ''−''. Load combination factors, or LCF s, can be defined as the resulting load combination where each load is normalized by its respective design value.

Evaluating RCWT-defined lifetime combined loading scenarios
Defining realistic combined loading scenarios is of vital importance for an adequate structural design of novel vessels like trimarans. If these load scenarios are used in high-fidelity modeling tools to assess a design's structural adequacy, it is important that these load scenarios are statistically relevant and physically realistic. The main question is then: does it matter which technique is used to define these load scenarios (m-EDW, m-CRRW, DLG, MCS)? Do all of these methods give comparable statistics of lifetime combined loading scenarios? Clearly, any load combinations derived from statistics of brute-force MCS associated with the desired exposure period will be the most realistic because these situations are not based on any load conditioning, only the system physics as described in the stochastic simulations.
While a brute-force MCS approach is not a feasible approach for daily use by engineers to define lifetime load combinations, it can certainly be used as a benchmark to compare the other techniques. For each heading, 500, 1000-h MCS are run and statistics on the conditional maxima for each global load in Fig. 1 are collected in the headings suggested by Table 2. As well, the wave profile and resulting load vector time series around the time of the 1000-h maximum of each global load is saved for later comparison with the m-EDW, m-CRRW, and DLG analyses. The points of comparison between these techniques will be: 1. How do conditional probability density functions (pdf 's) for each load case defined by the RCWTs compare with MCS (Section 4)? This comparison will give good insight on whether the RCWTs accurately describe the relative magnitude and variance of the extreme combined loading distributions for this trimaran hull to evaluate the potential variation in short-term extreme combined loading scenarios. 2. How do conditional cumulative density functions (cdf 's) for each load case defined by the RCWTs compare with MCS (Section 5)? This comparison can illustrate the risk factors associated with these load scenarios, and whether increasing the risk factor on the primary load in a load case affects all loads equally. 3. How do load vectors associated with the most-likely 1000-h extreme primary load response from the RCWTs compare with MCS (Section 6)? This comparison will give insight on the simultaneous load correlations excited by the RCWTs. 4. What is the computational effort associated with employing RCWTs versus MCS (Section 7)? This comparison will help balance the simplicity and efficiency of applying the RCWTs with the quality of the resulting combined load scenarios.

Conditional pdf's from RCWTs vs. MCS
The first way to compare the RCWTs in terms of defining realistic combined loading scenarios is by considering the conditional pdf 's assembled by each technique for each case. In each case, a single global load called the primary load is maximized at time = 0 in the heading directed by the design values in Table 2. The different RCWTs (m-EDW, m-CRRW, DLG) assemble wave profiles which lead to 500 ( = 0) values belonging to the Gaussian EVD where is the 1000-h most-likely extreme value for the operational profile in Table 1. These resulting wave profiles can be used to assemble distributions of all ( = 0) responses conditioned on the extreme ( = 0) events. To describe these conditional pdf 's constructed by the different RCWTs and by MCS, the nomenclature in Eq. (3) in Box I is used.
By construction then, all RCWTs will lead to a 1000-h extreme value ( = 0) distribution with a most-likely extreme value at the load design value (i.e. = 1). Based on the conditioning technique, the conditional ( = 0) distributions will differ. The load pdf 's from the RCWTs and MCS are compared when normalized both as a and in physical dimensions to illustrate how well the RCWTs capture the potential load variance for the different cases.

Head seas
In head seas, both and experience their design value. and is about ∕2 radians. When is maximized at = 0 by the EDW, is almost ∕2 radians out-of-phase with the response, resulting in a load distribution at = 0 centered around = 0 with low variance.

Beam seas
In beam seas, only experiences its design value. Conditional pdf 's based on 1000-h extreme ( = 0) responses in beam seas are shown in Fig. 4. All of the RCWTs give reasonable estimates of the MCS pdf 's conditioned on extreme ( = 0) responses in beam seas as compared to MCS. All techniques indicate that ( | ) is marginally positive and close to = 0. The m-EDW technique seems to slightly over-estimate the peak of the ( | ) response distribution.

Oblique seas
In oblique seas, both and experience their design value. Conditional pdf 's based on 1000-h extreme ( = 0) and ( = 0) responses in oblique seas are shown in Figs. 5(a)-5(b), respectively. There are some striking differences between the ( | ) (Fig. 5(a)) and ( | ) (Fig. 5(b)) distributions assembled by the RCWTs and by MCS.  distributions are significantly narrower than the respective MCS distributions, although they do generally track the relative magnitude of the MCS distributions. The m-CRRW technique tracks the relative variance of the MCS distributions but in some instances noticeably over-predicts the load magnitude (negative m-CRRW ( | ) and negative m-CRRW ( | ) ) or under-predicts that load magnitude ( m-CRRW ( | ) ). The most surprising difference when comparing the RCWT and MCS conditional pdf 's is that the m-CRRW technique predicts the incorrect sign for both ( | ) and ( | ) as compared to MCS. This sign difference from m-CRRW for these loads may be explained by the phase restriction of the m-CRRW based on the range of the load transfer function phases. Fig. 2 indicates that and are nearly radians out-of-phase for about half of the frequency components: 0 ≤ ≤ 0.7. Based on the m-CRRW construction then, where the regular wave components have phases defined by the load transfer function phases (i.e. = − ), a significant portion of these regular wave components phased to maximize ( = 0) in oblique seas will excite simultaneous negative ( = 0) responses, and viceversa. This can explain why m-CRRW suggests the conditional extreme ( = 0) and ( = 0) responses have opposite signs. The m-EDW approach also has this phase limitation, but only for a single frequency based on the peak of the primary load . It is worth reiterating, though, that all RCWTs as well as MCS use the same transfer functions from Fig. 2. The DLG and MCS do not limit the wave phases in Eq. (1) based on the range of the transfer function phases like m-CRRW and m-EDW.

Conditional cdf's from RCWTs vs. MCS
The conditional load pdf 's assembled by the RCWTs and by MCS in Figs . This is unsurprising considering that the ensemble of waves constructed by m-EDW only differ by the wave height, which is tuned to excite a specific primary load value. The wave phase remains the same across all waves constructed by the m-EDW technique.
In contrast, m-CRRW, DLG, and MCS result in non-monotonic conditional cdf 's with respect to increasing . Whereas the m-CRRW regular wave phases remain unchanged for the frequency components across the ensemble of constructed waves, random coefficients stochastically alter the wave amplitudes, leading to the non-monotonic conditional cdf 's. When the randomness is not included in the CRRW (i.e. reverting back to the MLRW), the resulting conditional cdf 's are monotonic with increasing . As a way to quantitatively compare the RCWT and MCS cdf 's, conditioned on maximized ( = 0) in head seas, consider Figs. 7(a)-7(b). Fig. 7(a) examines the percentage of conditional cdf values assembled by each RCWT that are within some error bound of the respective MCS conditional cdf, with this bound measured as a percentage of each load's design value, considering all 500 extreme load responses. If a RCWT cdf were exactly the same as the MCS conditional cdf, 100% of its cdf values would have a 0% difference from the design value of the MCS cdf. This comparison considers how well the RCWT conditional cdf matches the MCS conditional cdf on a global sense. The RCWT leading to the best fit with MCS will have the highest percentage of its conditional cdf values within the smallest percentage difference from the respective MCS cdf (i.e. fastest ascent to 100 on the y-axis). Fig. 7(b) compares the RCWT and MCS conditional cdf 's on a more point-by-point basis by examining the difference between the conditional cdf 's at the 50 values included in the conditional cdf 's. In this case the RCWT leading to the best fit with MCS will have the minimum difference at each sampled value. A perfect fit with the MCS cdf would result in zero difference at each value. Fig. 7(a) indicates that the RCWT cdf 's conditioned on maximized ( = 0) in head seas all have 100% of their samples within 3% of from the MCS cdf. Since by construction all RCWTs lead to a distribution of ( = 0) values which follow the 1000h Gaussian EVD for in head seas, this is not a surprising result. It is far more interesting to consider how well ( | ) and ( | ) from the RCWTs compare with the MCS cdf 's. The middle inset of Fig. 7(a) indicates that m-EDW has the best match with MCS ( | ) , the DLG has a slightly worse comparison, and m-CRRW has the worst comparison with MCS. From Fig. 6(a), it is not surprising that m-CRRW ( | ) is the worst match to MCS ( | ) , but it is not so obvious why the m-EDW does better than the DLG for this conditional cdf.
Consider though, that m-EDW ( | ) does capture the relative magnitude of MCS ( | ) , just not the fact that it is nonmonotonic. Since m-EDW ( | ) is monotonic and has a range similar to MCS ( | ) , it has more values that are within a small distance to the MCS cdf than does the DLG cdf, which is non-monotonic. But if a m-EDW conditional cdf does not capture the average magnitude of the respective MCS conditional cdf, this comparison suffers, as seen for ( | ) in the right inset of Fig. 7(a). In this case, both DLG ( | ) and m-CRRW ( | ) are a far better approximation to MCS ( | ) than m-EDW ( | ) . Fig. 7     ). But in these instances, the DLG comparison with MCS is favorable and still captures the non-monotonic nature of the MCS conditional cdf 's. Surprisingly, m-CRRW performs worse than m-EDW when comparing with MCS for all conditional cdf 's in these oblique seas cases except

Effect of increasing primary load on secondary loads
The non-monotonic nature of the MCS conditional cdf 's illustrates an important point about combined loading, especially relating to safety factors. That is: a primary load response associated with a higher is not necessarily associated with larger simultaneous secondary loading values. In this case, a safety factor applied to the primary load may not actually lead to a load scenario which is harsher in a combined loading sense. Such a result implies that combined loading scenarios may be better defined based on a relevant limit state, versus a conditional maximum approach. The probability of not exceeding a specific load value, like primary loads which define these load scenarios, may be misleading when considering the magnitudes of the simultaneous loading conditions. It may be more relevant to examine the probability of exceeding specific combined loading scenarios which are relevant for a design limit state, as examined in Seyffert  In that way, a safety factor may be applied to the combined loading scenario as a whole, rather than on a primary load which defines a conditional maximum. Based on the results of this investigation, increasing the safety factor associated with a primary load for conditional maximum may not have the same effect on the secondary loads.

Load vectors associated with most-likely extreme primary load response from RCWTs vs. MCS
Another interesting comparison between combined loading statistics from the RCWTs and MCS is of individual load vectors around the lifetime combined loading event at time = 0. The following subsections consider these load vectors associated with the most-likely 1000-h extreme primary load response, or the response with a = 0.368, where ( = 0) = . Note that this wave for CRRW is the irregular CRRW associated with the ( = 0) response with = 0.368, not the MLRW. All load vector time series (from the RCWTs and MCS) have the same duration: about 22 s, centered around time = 0 when the ( = 0) = response occurs. A few things can be observed from comparing the load vectors excited by the different RCWT and MCS waves. The m-EDW load vector is a periodic ellipse and lies on a 2 − plane, unlike the m-CRRW, DLG, and MCS load vectors. This restriction of the m-EDW technique was noted in Section 2.1.1: because the m-EDW and the resulting loads are regular waves with the same frequency, the load vector is a 2− ellipse. Visually, the DLG load vector seems to be the most representative of the MCS load vector.

Head seas
Comparing m-EDW , m-CRRW , DLG , and MCS in Eq. (A.10) explains some of the visual differences between the RCWT and MCS-excited load vectors. DLG tracks the best with MCS in terms of the sign and relative magnitude of each correlation value, corresponding to the best visual agreement in Fig. 14(a). Comparing the determinants of the difference matrices, the DLG gives the best match to MCS, followed by m-EDW and then m-CRRW. All correlation matrices are full-rank except for m-EDW . Section 2.1.1 indicated that an m-EDW excited load vector has a 2 − range, or a correlation matrix with at maximum 2 linearly independent columns, relating to a maximum m-EDW rank of 2. Due to this rank restriction of m-EDW , the vector of m-EDW excited loads has a nullspace dimension of −2. This rank restriction of m-EDW might also have consequences for constructing a joint pdf of these combined loads. A reasonable joint pdf choice could be a joint Gaussian, but since m-EDW is singular, it cannot be used to assemble such a pdf.
Similar trends can be noticed for the load vectors excited by the m-EDW, m-CRRW, DLG, and MCS wave constructed for the ( = 0) response in head seas with = 0.368, shown in Fig. 14(b). The load vector correlation matrices and difference between the RCWT and MCS correlation matrices are given in Eq. (A.11) in the Appendix while the rank and absolute value of the determinants of these difference matrices are given in Eq. (6). Visually the DLG-excited load vector seems to give the best agreement with the MCS load vector. In the same way, DLG gives best agreement to MCS in terms of relative magnitude and sign of the correlation terms. In both of these head seas cases, m-EDW better tracks the correlation matrix of the MCS load vector than m-CRRW.
Load correlations for ( = ) response with = . : rank( DLG ) = 3 |det( DLG − MCS )| = 7.67 -5 rank( MCS ) = 3 Fig. 15 shows the load vector normalized as a which is excited by the m-EDW, m-CRRW, DLG, and MCS wave that excites the ( = 0) response in beam seas with = 0.368. The load vector correlation matrices and the difference between the RCWT and MCS correlation matrices are given in Eq. (A.12) in the Appendix, where order of the elements is , and . The ranks of the correlation matrices and absolute value of the determinant of the difference matrices are given in Eq. (7). Comparing the determinants of the difference matrices in Eq. (7), the DLG best captures the MCS load correlations for this case, followed by m-CRRW and then m-EDW.

Beam seas
of the resulting data cloud centered around the MCS mean would be exactly zero.       In the oblique seas case, the range limitation of the m-EDW excited load vector is more apparent, as the load vector still lies on a 2 − plane in the 5 − space. This is reflected by m-EDW , which still has rank 2, even though this case includes all 5 global loads. Despite this limitation, m-EDW still tracks the MCS load correlations quite well. But overall, the DLG has the best load correlation match with MCS for this case, supported by the remarkable visual comparison between the MCS and DLG wave leading to the ( = 0) response with = 0.368. The DLG wave excites very similar load vector shape patterns as MCS, especially in the − , − , − , and − planes (2nd, 5th, 6th, and 8th columns). Considering the comparison is of load vectors excited by single irregular waves, rather than an ensemble average wave, it is impressive that the DLG recovers such a similar simultaneous loading response.
Although the m-EDW excited load vector corresponding to the ( = 0) response in oblique seas with = 0.368 does not visually track as well with the MCS load vector in terms of the details of the load vector shape, it still gives a favorable estimation of the load correlations. Similar to the comparisons of the conditional cdf 's, although the m-EDW wave does not capture the details (e.g., the nonmonotonic nature of the cdf 's or the randomness of the load vector), it does globally describe the combined loading scenario when considering the load correlations. m-CRRW does lead to an irregular load vector like MCS, but does not estimate these load correlations as well as m-EDW. Fig. 17 shows the load vectors excited by the RCWTs and MCS associated with the ( = 0) response in oblique seas with = 0.368. The load correlation matrices, difference matrices, ranks of the correlation matrices, and determinants of the difference matrices are given in Eq. (A.14) and Eq. (9), respectively. Again in this case, the DLG proves the best visual match to the MCS load vectors, especially in the − , − , and − planes (2nd, 3rd, and 8th columns). Based on the determinant of the difference matrices in Eq. (9), the DLG by far best matches the MCS load correlations.

Computational effort associated with RCWTs vs. MCS
The final comparison to distinguish the RCWTs is the computational effort associated with constructing the lifetime combined loading scenarios. Mainly, does a RCWT give up accuracy at the expense of efficiency, or vice-versa? Table 3 tabulates the computational effort associated with generating the 500 wave profiles via the RCWTs and MCS  and collecting the conditional maxima statistics for the 5 combined loading scenarios investigated in this paper. The m-EDW, m-CRRW, and DLG waves and load statistics were generated via a MacBook Pro 2.3 GHs Intel Core i5. The MCS were run on an Ubuntu desktop with 12x Intel(R) Xeon(R) CPU E5-2609 v3 @ 1.90 GHz. The difference in computational effort between m-EDW and m-CRRW is essentially negligible for engineering purposes. Although the DLG takes longer, it is still efficient compared to MCS.

Discussion of RCWT performance
Section 3.3 introduced four points of comparison to evaluate the RCWTs. This section summarizes the results from Section 4 -Section 7 to give an overall comparison of the different RCWT performances as compared to MCS.

RCWT leading to best pdf match
Examining the combined load conditional pdf 's from the RCWTs with MCS in Section 4, the DLG gives the best visual comparison with MCS, in terms of load magnitude and variance, for all cases except beam seas, where DLG and m-CRRW seem to give equally good comparisons. Comparing just m-CRRW and m-EDW, m-CRRW much better captures the load variance than m-EDW. But both the m-EDW and m-CRRW technique can far over or under-estimate a load magnitude as compared to MCS, which could be problematic if these load cases are used as inputs to high-fidelity structural models to test structural compliance with classification society criteria.
Examining the conditional pdf 's also indicates that there may be significant variability in short-term extreme combined loading responses. It may not be possible to accurately describe short-term extreme combined loading scenarios by a single combination, which would not describe the potential load variability within that load case. Rather, a distribution of extreme combined loading responses may need to be considered.

RCWT leading to best cdf match
Comparing the conditional load cdf 's from the RCWTs with MCS in Section 5, the DLG gives either the best, or a close second best, comparison with the MCS for all ( | ) where ≠ . The DLG captures the non-monotonic nature of the MCS conditional cdf 's with relatively low error. In fact, the DLG best matches the MCS conditional cdf 's in 7/14 cases, only considering ( | ) where ≠ . In 5/14 of these conditional cdf 's, the EDW proves the best match but the DLG comes in a close second. Surprisingly, m-CRRW only gives the best match with MCS for the two beam seas cases where ≠ .
Comparing just the m-CRRW and m-EDW conditional cdf 's ( | ) where ≠ , m-CRRW only out-performs m-EDW in 5 cases: . But a visual inspection of these cdf 's in Figs. 5 ( Fig. 6(a), 9, and 11(b)) indicates that for the cases when m-CRRW outperforms m-EDW, only in three cases ( ( | ) ) the comparison of m-CRRW to MCS is better than m-EDW to MCS, but m-CRRW still does not give an accurate representation of the respective MCS cdf 's in terms of load magnitude. The m-CRRW technique does capture the non-monotonic nature of the MCS conditional cdf 's, but does not reliably capture the MCS load magnitude for every load scenario.

RCWT leading to best load correlation match
Examining the load vectors associated with the most-likely 1000-h extreme ( = 0) responses in Section 6 again suggests DLG gives the best comparison with MCS. The DLG waves exciting the 1000h most-likely extreme ( = 0) responses led to the closest load correlation comparison to MCS for all of the examined cases. The DLG waves also led to very similar load vector shapes as the associated MCS waves, even though this comparison was of single stochastic simulations and not ensemble averages. Comparing just m-CRRW and m-EDW, m-CRRW only out-performs m-EDW for the load vectors based on extreme ( = 0) response in beam seas and extreme ( = 0) response in oblique seas. On the other hand, m-EDW out-performs m-CRRW for matching the equivalent MCS load correlations extreme ( = 0) and ( = 0) in head seas and extreme ( = 0) in oblique seas, even though the m-EDW correlation matrices are rank-deficient.

Efficient vs. realistic
Clearly the m-EDW and m-CRRW are the most efficient RCWTs, although the computation time for the DLG is still minimal when compared to brute-force MCS and for engineering purposes. Considering that the m-EDW and m-CRRW methods are based on a similar idea (i.e. Airy wave theory, or phase-shifting frequency components to lead to maximum load responses), the added computation time to include all wave frequencies (m-CRRW) versus using a single wave frequency (m-EDW) is negligible. But based on the combined loading statistics evaluated for this trimaran hull, it is not clear that using the m-CRRW technique results in more realistic extreme combined loading scenarios than does the m-EDW technique, as compared to MCS.
Waves from the m-CRRW technique are certainly more physically realistic because they are irregular. But the waves from the m-CRRW technique did not reliably excite more realistic combined loading scenarios for the examined trimaran hull than the waves from the m-EDW technique. Surprisingly, the combined loading scenarios from the m-EDW technique seem to better represent the MCS load scenarios than m-CRRW, despite the fact that the m-EDW is truly a simplification of the m-CRRW technique. At the very least, m-CRRW did not present a significant improvement over m-EDW when defining extreme combined loading scenarios for this trimaran hull.

Conclusions
This paper investigated 1000-h extreme combined loading statistics on a trimaran generated by three response-conditioning wave techniques: a modified Equivalent Design Wave (m-EDW) approach, modified Conditioned Random Response Wave (m-CRRW) approach, and the Design Loads Generator (DLG), and compared the results with load statistics from brute-force Monte Carlo Simulations (MCS).
This investigation illustrated m-EDW and m-CRRW approaches which construct ensembles of waves that lead to a distribution of extreme loading responses. It is at least recommended to use these m-EDW and m-CRRW techniques instead of the traditional EDW and CRRW techniques to account for the possible extreme load variability of short-term combined loading scenarios, as the variance in the simultaneous loading distributions can be substantial. Compared to the analyses in Seyffert and Kana (2019) and Seyffert et al. (2018), the m-EDW technique represents a major improvement over using a single EDW to describe combined loading scenarios on this trimaran hull. Comparing just m-EDW and m-CRRW, it is not clear if m-CRRW offers any significant advantage over m-EDW in realistically describing extreme combined loading scenarios for the examined trimaran hull.
Overall, the DLG provides the best comparison with MCS when describing extreme combined loading scenarios for this trimaran, with a significant reduction in computation time. The DLG captured the load magnitude and variance of the conditional extreme loading pdf 's, the non-monotonic nature of the conditional cdf 's, and the irregular loading profiles and resulting load correlations for a specific risk parameter (i.e. extreme primary load response with = 0.368). This investigation confirms the suitability of the DLG, based on efficiency and comparisons with brute-force stochastic simulations, to construct wave environments leading to rare return-period responses, even for rare combined loading scenarios on a complex hull like a trimaran.