Finite elements and machine learning combined in a case study example

Properties of the problem geometry by Konstantinidis D., Kamas I., Zevgolis I., (2023), “Long-term Deformations of a Highway Trial Embankment: Case Study Example with Combination of Finite Elements and Machine Learning “

We recently attended the presentation of a conference paper titled “Long-term Deformations of a Highway Trial Embankment: Case Study Example with Combination of Finite Elements and Machine Learning” by Konstantinidis D., et al., which investigates the potential of machine learning in estimating long term settlements of soft soils beneath embankment loading.

Furthermore, the case study of a road test embankment in Switzerland was considered. This embankment was running parallel to a canal and was founded on a layered soil profile, which mainly consisted of silt and organic silt, along with other admixtures, and a relatively high water table.

The finite element formulation of the problem was done in PLAXIS 2D, using the Mohr-Coulomb linear elastic – perfectly plastic constitutive model, and compared with the measurements taken in-situ from the test embankment.

More specifically, comparison between the two methods was made for the right and left toes of the embankment, as well as the horizontal contact axis between the embankment and the subgrade.

These showed a great match between the vertical settlement trends of the two methods’ results on the horizontal contact axis, with the finite element method yielding a peak displacement value approximately 10% larger than the experimental one.

The values for the vertical displacement at the toes yet again showed comparable trends between the experiment and the finite element model, however, this time the results of the finite element method looked considerably exaggerated.

Finally, a machine learning approach was also implemented on the settlement prediction problem. This used three independent Gaussian process regression (GPR) models, which were trained with PLAXIS 2D displacement results, and were given varying input values regarding the top layer’s Young’s modulus (E₁) and poison’s ratio (v₁), as well as the problem geometry and the second layer’s Poisson’s ratio (v₂).

Results from Konstantinidis D., Kamas I., Zevgolis I., (2023), “Long-term Deformations of a Highway Trial Embankment: Case Study Example with Combination of Finite Elements and Machine Learning “

The results showed that PLAXIS 2D was capable of yielding good results regarding displacements of soft soils carrying surcharge loads, even when using simple models such as the Mohr-Coulomb one. They also showed that such results can be used for the training of a machine learning algorithm.

Explore PLAXIS

Watch free PLAXIS Webinar Recordings On-Demand

References:

Konstantinidis D., Kamas I., Zevgolis I., (2023), “Long-term Deformations of a Highway Trial Embankment: Case Study Example with Combination of Finite Elements and Machine Learning “