The future of the DIA
Gerbrant van Vledder Delft University of Technology
Contents
• Purpose
• Non-linear four-wave interactions
• History of the DIA
• Shortcomings
• Improvements
• Inter-comparison of methods
Purpose
• Discuss the future of the Discrete Interaction Approximation
(DIA) for computing non-linear four-wave interactions Snl4 in
discrete spectral models
• Provide an overview of developments and requirements in
the development of efficient algorithms for Snl4 in
operational wave models
• Discuss the concept ‘efficient algorithm’ in relation to
computational costs, wave model performance and types of application
Spectral wave modelling
Action balance equation,
N
=
N(
,
,x,y,t)
, ,
4 3
g x g y
wind wcap nl brk fric nl
N
c N
c N
c N
c N
S
t
x
y
S
S
S
S
S
S
S
Importance of S
nl4
• Phillips (1960) showed basic principle of non-linear four-wave
interaction
• Theory extended to random surface gravity waves by
Hasselmann (1962) and Zakharov (1968)
• JONSWAP experiment (1973) concluded that Snl4 is mainly
responsible for forward shift of peak frequency
• Shape stabilization and influence on spectral shape, both in
frequency and direction space (Young and Van Vledder, 1993)
Basic equation of S
nl4
• Rate of change of action density in wavenumber k1 function of four
wave numbers involved in a resonant interaction
• Six-fold integral over the wave numbers k2, k3 and k4
• Delta functions reflect resonance conditions and ensure
conservation of wave energy, action and momentum
• G complicated coupling coefficient
1 1 2 3 4 1 2 3 4 1 2 3 4 1 3 4 2 1 4 3 1 2 3 4,
,
,
n
G
t
n n n
n
n n n
n
d d d
k k k k
k
k
k
k
k k k
Discrete Interaction
Approximation (DIA)
• Discrete Interaction Approximation (DIA) evaluates one subset of
one subset of possible interactions configurations
•
• DIA quite successful in development of 3G wave models WAM,
Wavewatch, SWAN, …
1 2 1 2 3 4 1 1
k kBasic equation of DIA
4 11 4 2 4 4 2 42
1
1
2
1
1
1
nl nl nl nlS
S
C
g
f
S
E
E
EE E
E
Evaluation in a discrete spectrum
0 0.1 0.2 0.3 0.4 0.5 0 30 60 90 120 150 180 210 240 270 300 330 f (Hz) ( o )E
+E
E
-E
+E
-Central density E loops
over all spectral bins
E
+and E
-obtained by
bi-linear interpolation
S
nl+and S
nl-similarly
distributed over
surrounding bins
DIA is just bi-linear
interpolation !!
Variations in the DIA
0.25
0.05
0.35
Medium range
Short range
Long range
Comparison of DIA with exact
solution (Xnl)
Deficiencies of DIA
• too much transfer towards higher frequencies
• too broad spectra, both in direction and in frequencies
Figure by Erick Rogers using 1-point model
after 12 hours simulation
with U10=12 m/s
• deficiencies of DIA usually compensated by tuning of
DIA and spectral resolution
• Increasing spectral resolution does NOT improve accuracy of
DIA, as DIA is tuned to a certain spectral resolution
• Optimal resolution f/f=10%, =10
• Experiment of fetch-limited wave growth with f=4%
actually led to double peaked spectra
• Only solution is to add, additional configuration covering
Degenerate results of DIA
Van Vledder, et al. ICCE 2000
A simple case going wrong
• Luigi playing with SWAN; reproducing wave flume experiment
• Dimensions of grid: xlen=70 m, ylen 50 m, depth 3 m
• Uni-model wave propagation in a wave flume, D()=cos840()
Wave direction upwards, in y-direction
• Wave boundary conditions: Hs=0.068 m, Tp=1 s,
JONSWAP spectrum with =6
Diagnosis
• DIA spreads energy because its single wave number configuration is calibrated against required transfer
• The equivalent Xnl wave number configuration is much weaker
as it is one of thousands others
• Growth of energy due to non-conservative behavior of DIA for extremely narrow directional spectra
• Also occurs for mDIA, will it also occur in GMD?
• Will it affect narrow directional spectra, implications for swell propagation?
How to improve on the DIA
• Fundamental difference between Xnl and DIA
is its number of wave number configurations
• The way to go is to increase the number of
configurations to include more scales of
interaction
• Either by developing a multiple DIA or by
reducing workload Xnl
Accurate Incorrect
Time consuming Fast
Exact methods
Discrete Interactions
Full Reduced Extended Classic
X
nlGMD, mDIA DIA
Considerations how to proceed
• Challenge is find balance between accuracy
and computational requirements
• What is an optimal method for computing
S
nl4?
• Good representation of S
nl4for given spectra,
or, good model performance in which S
nl4cooperates with other source terms?
• Exact methods are used as benchmark, but
which ?
Exact methods for Snl4
•
• Rewrite transfer integral to eliminate delta-functions and to make
transfer integral computationally feasible
• At least three basic analytical transformations exist in literature:
Webb (1978) - Masuda (1980) - Lavrenov (2001)
• Methods differ in various ways:
• choice of integration variables, i.e. Webb uses k1 and k3,
Masuda uses k1 and k2
• treatment of singularities
• internal transformations and approximations
1 1 2 3 4 1 2 3 4 1 2 3 4 1 3 4 2 1 4 3 1 2 3 4 , , , n G t n n n n n n n n d d d
k k k k k k k k k k k• WRT – RIAM – GQM: Like their analytical masters these
computational methods differ in various (hidden, unknown) ways
• Do they provide the same answers? Which one to trust?
• They are used as reference in development of approximate
methods
• To resolve these issues an inter-comparison study for Snl4 is
now being carried out (Van Vledder, Benoit, Hashimoto, Resio, Tolman, …) in the spirit of the SWAMP study (1983)
Inter-comparison study for Snl4
• Comparison against individual spectra
• Sensitivity to spectral resolution, spectral shape, directional
properties, symmetries, depth, ….
• Reveal internal hidden features like quadrature methods,
integration ranges, assumptions (e.g. smooth spectra)
• Perform dynamic wave model runs in combination with other
source terms to find out about stability and overall model performance (Tolman developed such a set for his GMD)
• Methodology can also be applied to approximate methods like
Extension of the DIA
• Adding additional -configurations
• Van Vledder et al. (2000), 2 configurations
• Hashimoto & Kawaguchi (2001), up to 5 configurations
• The original Discrete Interaction Approximation (DIA) of Hasselmann
had two configurations: 1=0.25 and 2=0.15 with weights of 3000
and 375
• The second configuration was dropped because it’s added value in
terms of wave model performance was insufficient with respect to model efficiency !
Generalized Multiple DIA
• Generalized DIA with arbitrary configuration proposed by Van Vledder (2001); cast in symmetric form by
Tolman (2003)
• MDIA in principle able to represent
full transfer using multiple configurations
• Final GMD (Tolman, 2011): 1 2 2 1 1 2 3 4 1 1 1 k k 1 2 2 1 1 2 3 4 1 1 1 1 k k nl,1 nl,2
deep deep shal shal nl,3 d s nl,4 g1 1 g2 2 g3 3 g4 4 1 1 2 2 3 3 4 4 g1 1 g2 2 δS -1 δS -1 1 1 = C B + C B × δS 1 n n δS 1 c E c E c E c E + -σ k σ k σ k σ k × c E c E + σ k σ k c Eσ kg3 3 c Eσ kg4 4
Results of optimized GMD
• Results obtained by Tolman (2012) are good. It is a major
improvement over the classic DIA
• Tolman (2012) used Xnl based on WRT method as ground truth
• Optimal GMD configuration(s) depend on choice of other source terms,
characteristics of host model, spectral resolution, spatial discretization and set of model runs
Conclusions
• More and more shortcomings of DIA are being discovered
• Efficient and accurate algorithms are being developed, notably the GMD
• The concept ‘efficient algorithm’ for Snl4 must be viewed in relation to
model performance and computational requirements, … not only against
its ability to approximate Xnl
• Efficiency should also be considered in relation to types of model
application, host model including settings, and choice of other source terms
• Inter-comparison framework needed for objective judgment of exact as
The future of the DIA (epilog)
• The DIA has served us well
• Accuracy of Snl4 (DIA) is lagging behind other source terms
• Increased computational power makes other methods more attractive
• More attention to shape of spectrum