 Good afternoon. Good evening to my family watching this at home. It's about midnight for them, so they promised to take a look. And good hello to future generations who will look at this. I'd like to acknowledge first, before starting, PhD student of mine, Gerard Dam, who did some work on this presentation. And myself, of course, Bruce Jeffey from the USDS Santa Cruz, and Donald Roofing, and this is then apart from the organizing committee of this very nice symposium. I will talk about S3 Mova Dynamics, because that's my specific area of interest. Let me start with this... Well, I'm getting a little bit nostalgic showing this movie to you. I'm supposed to click right here, because it's eight years ago that I made this movie, and maybe you'll remember. This show is basically a DELF 3D run over about 3,000 years. What you see here is basically a tidal basin. There's no inflow from here. At this side, a delta is forming. There's tidal movement only, one-half meters, tidal amplitude. And you can see that there's morphological updating, and you can see nice patterns evolving. There's a delta, but I don't consider deltas. It's kind of collateral damage to me in this movie. So you see a nice meandering pattern evolving. And the nice thing about this is that this highly nonlinear system with the shallow water equations of DELF 3D and all the sediment transports involved is an Englut-Hansel type of transport. Bad slope effects included, is able to reproduce a kind of equilibrium situation. It's not chaotic, it's something that's gentle, and well, for a very schematized case, gives you a kind of confidence in what DELF 3D is able to do. But of course it would be nice to do that also for real life, real case studies. I'm pressing a button here, which I'm not supposed to press. If you translate the molar results in a different way, you can see here the translation into energy dissipation levels. In the beginning you have a lot of energy dissipation, so energy dissipation means high shear stresses, high sediment transports, large amount of sediment being transported. Over app and over flood distributed, and this is then over the time scale of about 3,000 years. And you can see that the energy dissipation levels, they drop considerably, and also that it's more symmetric, distributed over app and flood after 3,000 years. That means there's less more of a dynamic activity, and there's kind of trend towards equilibrium. If you want to look at this in a real life case, you need a very, well, a bathymetric data set of 3,000 years, so we don't have that. First, I want to take you to the weather forecast, going to Boulder. I was very much interested in basically the temperature. You have all kinds of temperature, weather forecast, temperature models, that in the beginning they predict very nicely, quite nicely, and in the end, after two weeks, there's a lot of variation between the different models that you use. You can translate that towards more dynamic modeling with the L3D. You have the uncertainty trumpet, where you start with fluid dynamics, then you include sediment dynamics in your calculation, you update your bath, you include sediment transport, more for dynamics, and, well, you're increasing the error that you introduce into your system. That is what we would think of intuitively. And the question is, is this true or not? What is a model like the L3D? What is it doing in a realistic case? What kind of line is it following? To explore that, we focused on the Western Scheldt. I have two case studies. One is the Western Scheldt in the Netherlands, the other is San Pablo Bay in San Francisco Bay. First discussed the Western Scheldt. It's a very nice estuary because there's not so much sediment supply. Tidal forcing is relatively constant. Or is constant, there's not so much storms that mess up things. It's a sandy environment, so it's an ideal case study to test the model. And I'm going to reproduce the patterns of shoals and channels in this estuary. Well, this is then the phenome model, which is basically the L3D. I will skip that a little bit. This is explaining the resolution. So it's a finite element method. And you can see that the major morphodynamic features are covered by the resolution of the model. That's the important thing here. On a very regular basis, we have bathymetry starting from 1860. So that links with the previous... I was glad that I can make the link to the previous presentation because I saw the hind cast from 1960 onwards. I do the same. And after 1955, I have bathymetries every two years. So let's take a look. Here you can see the modeled patterns and here the measured patterns after 18 years. 30 years, 45 years, 71 years, 95 years, 100 years and 110 years. You can see if I go back, that in the beginning it doesn't look good. There are some similarities, but if you continue a little bit, then you can see striking similarities. We can translate that also in a Brice-Kill score. Brice-Kill score means that if it's one, it's perfect. You have the perfect model. If it's zero, you are as good as your initial conditions. If it's below zero, you should not have started your modeling effort at all. And if it's between zero and one, it's going from bad to good. You can also translate it in the kind of error because the error is then the difference between the measured profile and your modeled profile and the signal is then the difference between your initial condition and your measured development. And this is the Brice-Kill score over time. So you see indeed that in the beginning, the first 18 years or the first 30 years almost, you have negative Brice-Kill scores. And counter-intuitively, the Brice-Kill score increases over time a lot up to levels which you could refer to as reasonable to fare. I would say it's quite promising. It's quite good. If you translate that into errors and signals, you can see in this plot that the error is indeed increasing. So that is what we intuitively would expect. There's an increase of error. There's a build of errors. But the reason why the Brice-Kill score is increasing as well is that our signal, that's the red line, is increasing more than the error. One of the reasons for this is in the beginning, the initial conditions are not good. I only apply one grain size. In reality, there's mud. There's bed slope effects. There's non-erodable layers. All these things are happening. And the model first adjusts to its very poor, defined initial conditions. That's what you see in the beginning. But what takes it over on the longer term is probably the interaction of the main title constituents with the plan form of the basin. And starting from a flat bed, you see that process. I put in the same amount of sediments as was present in 1998. And the patterns that you see evolving over here really look like the patterns of the 1998 bathymetry. So the shape, the plan form of your basin determines to a large extent how the morphological development is going to take place. If I start at different points in time with the same data set, I see quite a large similarity in the development of the Brice-Kill score. And also if I translate it into errors and signals, you can see that after 30 years the signal becomes larger than the error. That means you need a kind of 30-year morphodynamic spin-up before you can trust your model results. If you remember the first picture I showed you, the first move was this decrease in energy dissipation levels. You can see that here also. So this is the model development in energy dissipation levels. And the straight line is the calculated energy dissipation levels based on the measured bathymetries. And you see in both systems that they have decreasing energy dissipation levels. And from this point onwards there's a lot of dredging involved. And that may a little bit explain the discrepancy between measured and model development. So we are basically modeling with our model the blue line. So the error increases, but the signal increases even more. The model is then particularly suitable for longer-term developments. If you want to study climate change's impact of sea level rise, then this model is perfectly suited for that. And that's a kind of a general research question of how to decrease the morphodynamic spin-up time. You could increase the skill or decrease the morphodynamic spin-up time by including, for example, mud, which is also present in this S3 or maybe tidal fluctuations, neapspring tidal cycles, and that will increase the skill score a little bit more, I guess, here. So going back to America, California, we are in this huge catchment of the Sacramento River and the San Joaquin River. They confluence in an area referred to as the Delta. And here you can see, in subsequent, all the river flows into Soussumbes and Pablo Bay before discharging through the Golden Gate into the ocean. About 150 years ago, they found gold there and huge amounts of blasted sediments, so they're washing out the gold from the sediments by blasting mountains and by hydraulic mining and then adding some mercury to get the gold out. So a lot of huge amounts of contaminated sediments were flowing downhill and towards the S3. This has done a little bit overview of the S3, Soussumbes and Pablo Bay Golden Gate. And here you can see measured developments. So you see this is the 1850 to 1890, so this is exactly the period of the hydraulic mining. You see a lot of deposition occurring and this is the period 1950 to 1980, where the hydraulic mining stopped in about 1890. They constructed lots of dams in the catchment and, well, that resulted in a very limited supply of sediments from the Delta towards the S3 and that large parts of the Bay became erosional as a result. The data is, again, very, very detailed. They did a good job from 1860 onwards. So basically there are three characteristic periods, the hydraulic mining period, lots of deposition, the erosional period because of the limited sediment supply and the impact of climate change in the future. So about 8 million cubic meters per year are deposited in San Pablo Bay in the depositional period and in typically the erosional period there was an erosion of about an order of magnitude smaller, 0.8 million cubic meters per year, still a considerable amount of sediments. A focus is then on reproducing these patterns with Delta 3D. So this is a nice view of what happened. You see a lot of deposition on the shoals here and narrowing of the channel. So the aim of the study was to to hind-cast the bathering changes in San Pablo Bay to assess the uncertainty levels and to predict future developments. With Delta 3D, I'm not going into detail here, we have to schematize the tide, we have to schematize the river flow and the river supply of sediments. There's wind that we needed to include. There is sediment transport of sand but also by mud. In total we have about 53 unknown parameters that we are dealing with. So this causes again a lot of uncertainty about the model results. Some of them we have to guess, we have to intelligently guess, we have to schematize. Yeah, can we trust them, the model outcomes? Well, this is the model outcomes. So this is the measured patterns, deposition period, erosional period and this is the model patterns and the difference then between the deposition period and the erosional period is indeed the supply, basically the sediment concentration at the river boundary over here. And you see that Delta 3D is doing a pretty good job in reproducing these patterns and you can even make predictions here. By the time I started this study this was still a forecast about five years ago, now it's a hind-cast. But you can run this continuously and make predictions for the coming century and also then include what could be the effect of sea level rise on these patterns. A nice thing about Delta 3D is you can use it as a virtual laboratory now and you can discriminate between wet season and dry season and we'll see what processes are when important and where's the sediment coming from and all kinds of nice exercises you can read all about that in some publications. But we are here to discuss uncertainty levels and you can see a range of model parameters that are varied. The black line here indicates the measured volume for the deposition period, erosional period in the future and here you can see the reproduction of the model and based on that you can already see that if I vary my model parameters within reasonable range I can still have a similar type of volumes and also the Bryce Killscore shows a little bit the same development of it. There's not so much variation given the variations of the input parameters. Maybe the wind here is most sensible or the river discharge. There's some difference between the deposition period because the wind is in an erosional period because during the erosion period the wind is how do you model the wind is very sensitive here and the river discharge, the supply of sediment has most uncertainty over here. How much percent of them the model results has skill, so how well is the model reproducing the observed patterns and what is the confidence level that we have given the input parameter uncertainty. We again have the Bryce Killscore but we also have a kind of confidence index which from an ensemble of runs that I showed you varying the model input parameters we have a mean value but also a standard deviation. So this is the measured pattern this is the model pattern of the ensemble averaged values this is a standard deviation this is mu over sigma over mu if you turn it around the Bryce kill and this is the pattern that you get if you take all the area that has a Bryce kill that's larger than 0.5 or a confidence index that's larger than 0.3 and if you then multiply this pattern with your ensemble averaged you can calculate basically the volume that answers both criteria and you can get a graph like this so for the depth of this period there's a lot of confidence in the model outcomes there's not so much variation that confidence is decreasing for the erosional period because the wind skimmitization which is a really unknown parameter in this setting is very important for the erosional patterns that you see that is one explanation the other explanation is that for erosion basically data of how is the sediment eroding not on the top layer but also on the lower layers and that's very difficult data to get aware of the skill score is about 57% of the modelled erosion and sedimentation volume fulfills the Bryce kill score of 0.5 that's also decreasing for the erosional period so if you combine confidence index and Bryce kill score you get about 53% so 53% of the volume is we are confident and it has skill you can also use this methodology then to investigate what parameters have caused the most uncertainty in your model outcome so we're doing then an ensemble not of all the model parameters but per model parameters for example the critical shear stress or the river discharge and by varying then the confidence index threshold and here you can see the percentage of the volume that is entering that threshold you can see here that during the depositional period the parameter that causes the highest uncertainty is the river discharge and during the erosional period the parameter causing the largest uncertainty is the wind with also the erosion factor so if you want to do measurement campaigns in order to perform to have a better performance of your model you should do measurements about these two model parameters it's a very helpful tool so some concluding remarks as we move on the development is better predictable on long time scales that exceed basically the time of the morphodynamic spin up which is typically decades after decades the model skills core becomes significant and even in very complex environments there's uncertainty levels by uncertain model input parameter settings but they remain limited I think both for the Western Scheld and St. Parlor Bay what's important is the plan form of your estuary that determines to a high degree where your sediment is depositing and that is then the last remark so thank you for your attention I have here a list of it's a very eco-centric list by the way of all my publications over the past six years but I can promise you that these list includes a lot of very useful references to work of others as well so thank you for your attention it's just been that I missed it but I wondered if you could hold forth a little more on the actual qualitative difference between confidence and skill in your model yeah, that's this one but can you give us in translated into how those two things could be different because intuitively you think they would correlate very closely with each other yeah well the skill is defined as how well is your model reproducing measured developments and the confidence is how confident are you in your model results given the large variations or the uncertain model input parameters I may have missed it but basically the results when you are showing us this comparison is this result of your calibration or its validation did you run your model on an independent data set? I would say yes I mean it's with very default settings of DELF3D you get these results and of course the results depend especially for San Pablo Bay a lot of gassing also the forcing conditions and the settlement supply but also if I vary then model parameters within reasonable bounds reasonable range the impact is not so high that's basically what you see here I mean there are some variations if I change the erosion factor by factor 2 or the fall velocity by factor 2 the roughness by 10% diffusion by 10% these are reasonable variations I mean these are typically variations that you would expect given a certain uncertainty of course I can make the range larger but what I then would see is that these patterns are similar the patterns remain similar maybe the deposition volumes will be twice as small or twice as big but the pattern and that well makes me to conclude that basically the interaction of the plan form and the main title constitutes the term to the high degree how these patterns develop you're saying that the model is not sensitive to parameters so therefore calibration is not even doing anything well I'm saying that does something but it is not I mean if you look here oops you see that there is variation and I did I had to do some calibration here that's true but on the other hand you can also argue from the other side that I don't show here the sensitivity parameter I don't include it in this presentation but the patterns remain the same so yeah this simulation how many grid points did you use the resolution yeah but it's the resolution and what is it a body fitted grid or about the resolution so the typical resolution is about 100 by 200 meters I see so the total number of grid would be less than a million right yeah so I didn't have your computing facilities so a typical run of reproducing these things would cost me one day on a decent PC and because of the large amount of runs that I had to do I did some higher resolution to get it published the patterns didn't look much different