Index-Based Neural Network Framework for Truss Structural Analysis via a Mechanics-Informed Augmented Lagrangian Approach DOI Creative Commons
Hyeonju Ha, Sudeok Shon, Seungjae Lee

и другие.

Buildings, Год журнала: 2025, Номер 15(10), С. 1753 - 1753

Опубликована: Май 21, 2025

This study proposes an Index-Based Neural Network (IBNN) framework for the static analysis of truss structures, employing a Lagrangian dual optimization technique grounded in force method. A is discrete structural system composed linear members connected to nodes. Despite their geometric simplicity, large-scale systems requires significant computational resources. The proposed model simplifies input structure and enhances scalability using member node indices as inputs instead spatial coordinates. IBNN approximates forces nodal displacements separate neural networks incorporates equations derived from method mechanics-informed constraints within loss function. Training was conducted Augmented Method (ALM), which improves convergence stability learning efficiency through combination penalty terms Lagrange multipliers. accuracy were numerically validated various examples, including trusses, square grid-type space frames, lattice domes, domes exhibiting radial flow characteristics. Multi-index mapping domain decomposition techniques contribute enhanced performance, yielding superior prediction numerical compared conventional methods. Furthermore, by reflecting structured nature problems, demonstrates high potential integration with next-generation network models such Quantum Networks (QNNs).

Язык: Английский

Discrete Physics-Informed Training for Projection-Based Reduced-Order Models with Neural Networks DOI Creative Commons
N. Ruiz, S. Ares de Parga, J. R. Bravo

и другие.

Axioms, Год журнала: 2025, Номер 14(5), С. 385 - 385

Опубликована: Май 20, 2025

This paper presents a physics-informed training framework for projection-based Reduced-Order Models (ROMs). We extend the original PROM-ANN architecture by complementing snapshot-based with FEM-based, discrete residual loss, bridging gap between traditional ROMs and neural networks (PINNs). Unlike conventional PINNs that rely on analytical PDEs, our approach leverages FEM residuals to guide learning of ROM approximation manifold. Our key contributions include following: (1) parameter-agnostic, loss applicable nonlinear problems, (2) an architectural modification improving accuracy fast-decaying singular values, (3) empirical study proposed process ROMs. The method is demonstrated hyperelasticity problem, simulating rubber cantilever under multi-axial loads. main accomplishment in regards residual-based its applicability problems interfacing software while maintaining reasonable times. modified outperforms POD orders magnitude snapshot reconstruction accuracy, formulation not able learn proper mapping this use case. Finally, application ANN-PROM modestly narrows data accuracy; however, it highlights untapped potential residual-driven optimization future development. work underscores critical role construction calls further exploration architectures beyond PROM-ANN.

Язык: Английский

Процитировано

0
Index-Based Neural Network Framework for Truss Structural Analysis via a Mechanics-Informed Augmented Lagrangian Approach DOI Creative Commons
Hyeonju Ha, Sudeok Shon, Seungjae Lee

и другие.

Buildings, Год журнала: 2025, Номер 15(10), С. 1753 - 1753

Опубликована: Май 21, 2025

This study proposes an Index-Based Neural Network (IBNN) framework for the static analysis of truss structures, employing a Lagrangian dual optimization technique grounded in force method. A is discrete structural system composed linear members connected to nodes. Despite their geometric simplicity, large-scale systems requires significant computational resources. The proposed model simplifies input structure and enhances scalability using member node indices as inputs instead spatial coordinates. IBNN approximates forces nodal displacements separate neural networks incorporates equations derived from method mechanics-informed constraints within loss function. Training was conducted Augmented Method (ALM), which improves convergence stability learning efficiency through combination penalty terms Lagrange multipliers. accuracy were numerically validated various examples, including trusses, square grid-type space frames, lattice domes, domes exhibiting radial flow characteristics. Multi-index mapping domain decomposition techniques contribute enhanced performance, yielding superior prediction numerical compared conventional methods. Furthermore, by reflecting structured nature problems, demonstrates high potential integration with next-generation network models such Quantum Networks (QNNs).

Язык: Английский

Процитировано

0