JOBTALK Flashcards
title
Good morning everyone. I’m Stephen and thank you so much for inviting me here today. It’s been a while since I’ve returned to Melbourne and it’s definitely my pleasure to be back here.
For today’s presentation, I’m going to talk about the development process of OxCaisson, which is a family of new design methods for suction caisson foundations that I have developed for the offshore wind industry.
But before I do that, let me give you some background information so that you can have a better idea of the context in which I develop this work.
intro offshore wind
offshore wind energy is a popular source of clean energy in Europe, China and increasingly, worldwide.
At the moment, most offshore wind farms are located in shallow water sites and for these sites, monopile foundations are the dominant choice as they are typically the most economical.
However, as we look into the future, offshore wind farms are expected to move into deeper waters, and for such sites, the suction caisson or suction bucket foundation may be more economical.
For e.g. the recent hywind floating wind farm in Scotland uses suction caissons to anchor the floating system into the ground
suction caisson
For those who are not familiar, a suction caisson foundation is like an inverted bucket
Its unique selling point is the installation process, which is faster, quieter and cheaper than that for the monopile foundation
To install a monopile foundation, you typically require
expensive hammers to drive them into the ground. This is going to be a problem when you go into deeper waters as monopiles will get larger and you would need bigger hammers.
In constrast, to install a caisson foundation, all you need is a pump to suck water out from within the caisson, which creates a pressure difference that drives the caisson into the ground.
Furthermore, this process can be reversed for easy uninstallation.
However, despite the potential advantages, there is still a major obstacle to the adoption of suction caissons in offshore wind.
And that is the lack of suitable design methods that meet the needs of the offshore wind industry.
let me explain by giving you a quick overview of the typical design process for an offshore wind farm.
wind farm design
First you have some ground data which you use to size your foundations for each of your hundred or so turbine locations.
Then, this process is repeated many times as you try different layouts in order to find the most optimal wind farm design.
As you can see, this entire process requires a lot of computations and hence the design method used to size the foundations must be very efficient.
So this brings me to the main research problem that I am addressing,
research problem
Which is the need for a fast and accurate design method for suction caisson foundations, that can handle the large number of computations required to optimise an offshore wind farm design.
Ideally, it should also be general enough to handle any arbitrary loading conditions and ground profiles.
Existing design methods typically do not meet these 3 criteria. For e.g. the popular 3D finite element model is very accurate and general purpose, but it is computationally expensive. In contrast, the existing macro-element models for caissons are fast and reasonably accurate but they are not general enough.
Hence the research goal is to develop a design method that satisfies these 3 criteria
why important?
It is important that we develop such a design method as it would enable us to search the foundation parameter space more efficiently and more thoroughly and this reduces opportunity costs.
Since the cost of foundation constitutes about 18% of the total cost for a wind farm, any cost reduction here goes a long way in helping reduce the cost of offshore wind energy, which would help accelerate our transition towards a low carbon future
oxcaisson
When I first thought about this problem, I asked myself ‘is it possible to develop a design method that not only meets the 3 criteria but is easily adoptable by industry?’
Through my conversations with industry practitioners, I noticed that they are very fond of 2 types of design tools. They like the 3D FE model for its accuracy and the Winkler model for its efficiency and flexibility.
So I thought, why not combine these two methods to get the best of both worlds?
And that is essentailly the philosophy behind OxCaisson.
OxCaisson is a family of Winkler-based models that can efficiently approximate the 3DFE predictions of the caisson behaviour under any 6DoF loading conditions.
The model is efficient as the physics has been simplified. The caisson is modelled using 1D frame elements and the soil response is modelled using Winkler-type soil reactions, which represents the net force or moment applied by the soil at each depth.
The source of accuracy for the model lies with the soil reactions, which have been specially calibrated using the results from the 3dfe simulations
let me describe how this calibration process works.
oxcaisson calibration
First, a series of 3DFE simulations is run to study the caisson behaviour under 6DoF loading conditions.
Then, the soil reactions are extracted from the stress components of the surrounding soil elements.
Thereafter, suitable formulations are identified to approximate the mapping between the inputs of the simulations and the corresponding soil reaction outputs.
depending on what type of soil constitutive model you use in the simulations, you will end up with different formulations for the soil reactions.
for this research, I’ve explored 3 different soil models and each of them focus on different aspects of the foundation response.
oxcaisson design methods
The first soil model is a linear elastic model, which focuses on capturing the initial foundation response.
The second soil model is a custom, non-linear elastic model, which focuses on capturing the small strain, non-linear behaviour of real soil.
The third soil model is a linear elastic, perfectly plastic model, which focuses on capturing the strong interaction effects between combined loading at the ultimate limit state.
From the calibration process, I’ve developed 3 design methods that approximate the 3dFE predictions for each soil model and each design method is used for different types of analysis.
The first design method is called OxCaisson-LE, and it is used for FLS assessment
The second design method is called OxCaisson-NLE, and it is used for SLS assessments.
The third design method is called OxCaisson-LEPP, and it is mainly used for ULS assessments.
For all of these design methods, I’ll just be able to give a brief overview but I want to give you some idea of my processes and describe how I went about to solve the problems that arise during the development process
oxcaisson-le
For the first design method, the main research question that I’m interested in is:
How is each soil reaction related to the local displacements?
And surprisingly, the result is not as expected.
virtually all of the existing Winkler models today assume that the soil reactions are dependent only on its primary displacement.
However, the results from the 3dfe simulations show that there is actually coupling between the lateral and moment soil reaction and so, the existing Winkler models are not capturing the full physics of the soil foundation interaction.
From the calibration process, i have determined closed form formulations for each of the soil reaction stiffness components and this is the basis of the oxcaisson le model.
oxcaisson-le results
I’ve validated the design method by using it to predict the global foundation stiffness predictions for rigid and flexible caissons and they agree well with the 3DFE predictions, with a max. error of about 6% for the coupling stiffness
However, this design method is not suitable for SLS assessmenta as it is widely known that real soils behave non-linearly at increasing strains.
So the next research question I have is: How can I build a design method that can better capture the non-linear behaviour of real soil?
stiffness degradation curve
In industry, the small strain non-linear behaviour of soil is typically described using the soil stiffness degradation curve shown here, where the shear modulus decreases with increasing strain.
This curve is commonly parameterised using a 2-parameter formulation shown here which applies to a wide range of soil including clay, silt and sand.
When I was trying to solve this problem, I got thinking: since this particular formulation is commonly used to describe the non linear behaviour of soil in lab testing, why dont i just develop a design method that predicts the caisson behaviour in a material that behaves in accordance to this formulation?
So that’s exactly what I did.
oxcaisson-nle in
I ran 3DFE simulations of the caisson problem with a custom non-linear elastic soil constitutive model that behaves according to the formulation shown here.
This formulation is the same as the previous formulation, except that the shear strain has been replaced by the deviatoric strain in order to generalise to the full three-dimensional strain space.
After calibrating the soil reactions, the resultant design method is called OxCaisson-NLE, and it has several advantages.
First, it is a unified design method for clay, sand or silt.
Second, it does not require any additional work, as the required soil parameters are readily available from the existing lab test results.
Third, it provides more realistic site specific predictions of the caisson behaviour that better corresponds to the soil sample behaviour that you observe in the lab.
Here are some examples of the predictions of OxCaisson-NLE and you can see that they agree well with the 3DFE predictions.
Nevertheless, this design method cannot be used for ULS assessments as it cannot capture the interaction effects between combined loading at the ultimate limit state.
So the next research question is: how can I develop a design method that can capture these interaction effects?
couple linear elastic with plastic yield surface
To address this problem, I’ve developed an elastplastic winkler model called oxcaisson lepp which is calibrated using 3DFE simulations based on a linear elastic, perfectly plastic soil model.
to approximate the ultimate response of undrained clay and drained sand, I’ve adopted the von mises and drucker prager yield criterion
This design method couples the linear elastic soil reactions of the OxCaisson-LE model with local plastic yield surfaces, which are calibrated using the ultimate response of the soil reactions that were extracted from 3DFE simulations.
But before i can actually start calibrating the soil reaction yield surfaces, one question I have to deal with is: how can i obtain the yield surface data efficiently
seq swipe
To address this problem, I’ve used the sequential swipe technique, which is a robust and efficient method to obtain high-res yield surface data.
Below are some examples of the soil reaction yield surfaces that were obtained under different loading conditions using this technique.
Don’t worry too much about the actual loading conditions at which these data were obtained.
I just want you to see that these yield surfaces vary greatly in shape and size with different loading conditions
So that brings me my next research question: how can i efficiently identify yield function formulations that can capture this complicated evolution of shapes and sizes with different loading conditions?
The traditional approach for this is to try to find a suitable formulation by trial and error and hence, it is very time consuming and labour intensive.
convex yield
to address this problem I’ve developed a new framework that can automatically derive a robust convex yield function that best fits the yield surface data.
It is based on a novel sum-of-squares optimisation technique, which is a new development from the field of convex optimisation.
The key benefits of this framework is that it supports any multi-dimensional load space, it is fast. It is automated, and the formulations it produces have nice mathematical properties such as global convexity, thermodynamics consistency and C2 continuity, which makes the resultant elastoplastic Winkler model numerically robust
convex yield examples
here are some examples of the predictions of the yield functions produced by the framework.
They agree very well with the actual yield data and they manage to capture the complicated evolution of shapes and sizes with different loading conditions.
So this completes the calibration of the soil reaction yield surface
Now, let’s see how well the design method performs
failure envelopes ULS
here are some examples of the global failure envelopes predicted by OxCaisson-LEPP.
As you can see, they agree very well with the 3DFE predictions, but the key advantage is that you can determine these failure envelopes a lot faster, which enables rapid ULS assessments.
summary
So to summarise, I’ve set out to develop a fast design method for suction caissons that can provide accurate prediction under ahy general 6dof loading conditions
The end result is OxCaisson: a family of fast Winkler-based models for suction caisson foundations that can approximate the 3DFE predictions of the caisson behaviour in 3 soil models, which can used for FLS, SLS and ULS assessments.
Compared to the 3DFE simulations, using these design methods can save you at least 97% of the computational time, which makes them highly suited for offshore wind applications.
So that’s a quick overview of the development process of OxCaisson.
Looking forward, where do I see opportunities to further improve the way we design our offshore foundations?
future
One opportunity I see is using machine learning to replace the most time-consuming aspect of the modelling process, which is the formulation of the soil reactions.
Here is a workflow of a data-driven approach that I can developing for monopile design at the moment, where i use machine learning to predict the soil reactions for any arbitrary soil constitutive model
First, we have a probabilistic ML model that has been trained on the inputs and soil reaction outputs of the 3DFE simulations.
This ML model will then be able to predict the soil reactions corresponding to any input. Not only that, it will also tell you how confident it is with its prediction via its uncertainty measure
If the uncertainty is above some predefined threshold, a new 3dfe simulation will be run to reduce design risk and the new results will be added to the database and the ML model will be updated.
The idea is that with time, your database will grow and the need to run the time consuming 3DFE simulation decreases drastically, as there is less chance of encountering a new soil profile that is significantly different from the existing soil profiles in the database
So what you end up with is a fast, scalable data-driven design method that gives you 3DFE precision but with much higher efficiency.
thank you
with that, I would like to end my talk
But first I would like to thank my industry partner Orsted for kindly supporting and funding this work.
And I would like to thank you all for listening.
So thank you.
I’ll be happy to take any questions now
