Deep Neural Network-based Inverse Analysis with Application to a Rockfill Dam

• Published in: KSCE journal of civil engineering Vol. 28; no. 1; pp. 155 - 168
• Main Authors: Shahzadi, Gullnaz, Soulaimani, Azzeddine
• Format: Journal Article
• Language: English
• Published: Seoul Korean Society of Civil Engineers 2024; Springer Nature B.V; 대한토목학회
• Subjects: Algorithms; Artificial neural networks; Availability; Civil Engineering; Comparative analysis; Comparative studies; Computer applications; Dam construction; Engineering; Evolutionary computation; Finite element method; Genetic algorithms; Geomechanics; Geotechnical Engineering; Geotechnical Engineering & Applied Earth Sciences; Heterogeneity; Inclinometers; Industrial Pollution Prevention; Laboratory tests; Mathematical models; Methods; Neural networks; Numerical models; Objective function; Optimization algorithms; Oscillations; Parameter identification; Parameters; Particle swarm optimization; Rockfill dams; Safety engineering; Stochasticity; 토목공학; Inverse analysis; Stochastic algorithms; Deep neural networks; Uncertainty analysis; Rockfill dam
• ISSN: 1226-7988; 1976-3808
• DOI: 10.1007/s12205-023-0355-y

The availability of significant computational resources has played an essential role in developing advanced numerical models for the design and safety evaluation of complex structures such as rockfill dams. Determining the geomechanical parameters is deemed a crucial but challenging task for effective modelling. The general approach involves using in situ or laboratory tests or experimental relationships from the literature to assess these parameters. These measures, however, do not accurately reflect the actual situation. This paper proposes a data driven approach using deep neural networks and non-deterministic optimization algorithms to identify the soil parameters that will lead to displacements that best approximate the measured data. The methodology is applied to a rockfill dam recently built in Quebec, for which some measurements of inclinometer displacements are available. A finite element model (FEM) of two dimensions generates numerical solutions. A comparative study is performed to account for the heterogeneity of the materials by decomposing the computational domain into subdomains. Subsequently, the inverse analysis uses the surrogate model instead of the full FEM model for rapid computations. A suitable objective function is defined to account for large oscillations in the measurement data. Non-intrusive stochastic optimization algorithms such as Genetic algorithm (GA), Particle Swarm Optimization (PSO), and Differential evolution (DE) are evaluated for the minimization problem. Finally, the case study confirms the capability of the proposed methodology to identify the relevant dam parameters, provide some insights into the performance, and compare three optimization algorithms.