[Paper Review] Nonlinear Reduced Order Modelling of Soil Structure Interaction Effects via LSTM and Autoencoder Neural Networks

Thomas J. Simpson, Nikolaos Dervilis+3 more|arXiv (Cornell University)|Jan 1, 2022

Structural Health Monitoring Techniques1 citations

TL;DR

This paper proposes a data-driven nonlinear reduced order model (ROM) for soil-structure interaction (SSI) in wind turbine monopiles using autoencoder and LSTM neural networks. The method learns nonlinear dynamics from full-order finite element simulations (Abaqus), reducing computational cost by over 300× while maintaining 98.13% accuracy in steady-state response prediction across multiple depths.

ABSTRACT

In the field of structural health monitoring (SHM), inverse problems which require repeated analyses are common. With the increase in the use of nonlinear models, the development of nonlinear reduced order modelling techniques is of paramount interest. Of considerable research interest, is the use of flexible and scalable machine learning methods which can learn to approximate the behaviour of nonlinear dynamic systems using input and output data. One such nonlinear system of interest, in the context of wind turbine structures, is the soil structure interaction (SSI) problem. Soil demonstrates strongly nonlinear behaviour with regards to its restoring force and has been shown to considerably influence the dynamic response of wind turbine structures. In this work, we demonstrate the application of a recently developed nonlinear reduced order modelling method, which leverages Autoencoder and LSTM neural networks, to a nonlinear soil structure interaction problem of a wind turbine monopile subject to realistic loading at the seabed level. The accuracy and efficiency of the methodology is compared to full order simulations carried out using Abaqus. The ROM was shown to have good fidelity and a considerable reduction in computational time for the system considered.

Motivation & Objective

• To address the high computational cost of repeated nonlinear finite element analyses in structural health monitoring (SHM) for wind turbine systems.
• To develop a data-driven reduced order model (ROM) that accurately captures nonlinear soil-structure interaction (SSI) behavior without requiring explicit knowledge of system physics.
• To demonstrate the feasibility of using deep learning—specifically autoencoders and LSTMs—for constructing efficient, high-fidelity ROMs of complex SSI systems.
• To evaluate the ROM's accuracy and computational efficiency against full-order Abaqus simulations on a realistic monopile model with py-curve-based nonlinear soil springs.

Proposed method

• Trained an autoencoder neural network to perform nonlinear dimensionality reduction on displacement time series from full-order simulations.
• Used the bottleneck layer of the autoencoder (4 latent variables) to represent the reduced-order state space.
• Trained an LSTM network to learn the temporal dynamics of the system in the reduced latent space from input forcing and latent state sequences.
• Reconstructed physical-space responses using the decoder component of the autoencoder after LSTM prediction.
• The entire framework is purely data-driven, relying only on input forcing and output response data from FOM simulations.
• The ROM was validated on unseen forcing inputs and compared to Abaqus simulations in both time and frequency domains.

Experimental results

Research questions

• RQ1 Can a deep learning-based ROM accurately predict the nonlinear dynamic response of a wind turbine monopile with SSI under realistic loading?
• RQ2 How does the performance of the autoencoder-LSTM ROM compare to full-order Abaqus simulations in terms of accuracy and computational speed?
• RQ3 To what extent can a 4-dimensional latent space capture the essential dynamics of a complex SSI system with nonlinear soil behavior?
• RQ4 Does the ROM maintain fidelity during transient response regimes, despite limited training data on such dynamics?
• RQ5 Can the method be generalized to other nonlinear structural systems with minimal prior knowledge of system physics?

Key findings

• The ROM achieved a normalized mean squared error (NMSE) of 1.87% in steady-state response prediction across all degrees of freedom, indicating high fidelity.
• The computational time for a 500-second simulation was reduced from 2272 seconds (Abaqus) to 6.99 seconds (ROM), representing a 325× speedup.
• The ROM accurately captured the frequency content of the response in steady-state, as confirmed by spectral comparisons in both time and frequency domains.
• The model showed reduced accuracy during the initial transient phase due to abrupt loading, but this was deemed acceptable given the unrealistic nature of the loading onset.
• The autoencoder successfully compressed the system response into a 4-dimensional latent space while preserving essential dynamic characteristics.
• The framework demonstrated strong generalization to novel forcing inputs not seen during training, confirming its robustness for inverse SHM applications.