MODELLING AND SIMULATION
add in extra parallelisation functionality via an MPI (message passing interface) to run it on our Cray supercomputers.’ ‘It was well worth doing this extra
parallelisation, however, as the model training time went down from a couple of days to around an hour or less. So, this significantly improved our productivity,’ he added. As a result, RSI’s petrophysicist experts no longer need to complete this mundane and time-consuming work. Instead, the machine learning approach can ‘very quickly’ highlight whether a well is likely to be of interest, according to Brown.
Corrosion protection system of an oil platform using sacrificial aluminium anodes modelled using Comsol Multiphysics
g
interpret each well to perform a full petrophysical interpretation and generate this high-level information. Brown added: ‘When you bear in mind there is raw data for many thousands of wells, then the fact that it is such a manual and time- consuming process to interpret each one becomes a major barrier to being able to exploit all this information appropriately.’ A machine-learning algorithm has been
developed to speed up the process. But, due to the nature of geology, different measurements can change significantly from one level to the next. As a result, data was missing from the system and this was a tricky challenge to overcome, as Brown explained: ‘Techniques such as interpolation were not particularly useful here, and instead a machine learning technique, which supported missing data values, was needed. ‘We adopted boosted trees using the
XGBoost library, otherwise known as gradient boosting. This approach relies on the idea of decision tree ensembles, where a model consists of a set of classification or regression trees and features of the problem are split up among tree leaves.’ Here, each leaf holds a score associated with that feature, and as one walks the tree, scores are combined which then form the basis of an overall prediction answer. Usually, a single tree is not sufficient for the level of accuracy required in practice and so an ensemble of trees, where the model sums the prediction of multiple trees together, is used.
26 Scientific Computing World August/September 2019
“There is a wealth of subsurface data collected by borehole drilling, but as this is real-world data for each specific well, the data itself is noisy and bits can be missing”
‘As one trains a boosted trees model,
the trees are built one at a time, with each new tree helping to correct the errors made by previously built trees. This is one of the factors that makes boosted trees so powerful, and they have been used to solve many different machine-learning challenges,’ Brown said. While boosted trees can handle the missing data values, there are still some challenges to overcome. For example, there is still uncertainty in the resulting predictions and this information needs to be available to the petrophysicist when they are making a decision around the risk of drilling in a specific location. The boosted tree models are also
extremely sensitive to hyperparameter selection. Brown said: ‘There are around 10 interconnected hyperparameters that significantly impact the accuracy of prediction of the model and it’s really hard to find the right combination. We used the hyperopt library that automatically searches the hyperparameter space to find optimal hyperparameters, but had to
Speed matters Speed of calculation at the scales used in the oil and gas industry is another key issue, as Yen explained: ‘As you can imagine, oil and gas works on really large scales. As a result, the ability of the simulation tools to handle large models with a very high fidelity is important.’ He added: ‘It is one thing for the
products to have the ability to solve these massive problems, and another to be able to solve them quickly. With this in mind, Altair has invested significantly in providing our customers access to our solvers on appliance, which can be hosted on-premise and in the cloud. These technologies enable our customers to run these huge models in a seamlessly stretchable environment.’ However, the oil and gas industry also
works on small scales too, and it can be difficult to realise the speed of calculation required because of the multi-scale nature of the systems requiring simulation. Cai said: ‘The system scale varies from the atomic level to kilometres of grids. The multi-scale modelling capabilities and the flexibilities of coupling arbitrary physics in Comsol Multiphysics are in high demand for this purpose.’
g
Analysing vortex-induced vibrations in straked risers
@scwmagazine |
www.scientific-computing.com
Comsol
Comsol
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32
