Professor Lehigh University
Real-Time Hybrid Simulation of Soil-Foundation-Structure Interaction Systems Using AI-Based Surrogate Models
Co-Authors: Faisal Nissar (Lehigh University), Davide Gorini (University of Trento), and Maryam Rahnemoonfar (Lehigh University)
Abstract: Real-time hybrid simulation (RTHS) is an experimental–computational methodology in which a structural system is partitioned into an analytical substructure, modeled numerically, and an experimental substructure, tested physically in the laboratory. The two substructures are dynamically coupled at their interface degrees of freedom, with compatibility and equilibrium enforced in real time to evaluate system response under dynamic loading. RTHS has traditionally been used to assess the dynamic performance of structural systems, often neglecting soil–foundation–structure interaction (SFSI) effects due to experimental and computational limitations. SFSI can significantly influence seismic response through inertial interaction, coupling of translational and rotational deformation modes, modification of energy dissipation, and irreversible soil and foundation behavior. However, direct experimental modeling of soil–foundation systems is impractical, and high-fidelity continuum soil models are computationally prohibitive under strict real-time constraints. This study introduces an AI-based macroelement surrogate modeling framework that enables RTHS of structural systems together with their interacting soil–foundation systems. The surrogate models are trained using high-fidelity numerical simulation data generated with high-performance computing resources at the Texas Advanced Computing Center (TACC), providing a multi-fidelity representation of SFSI behavior suitable for real-time execution. When employed within an RTHS framework, the resulting surrogate models enable simulations that both employ and generate data supporting natural hazard simulation, uncertainty quantification, and performance-based system evaluation. The results demonstrate an efficient pathway for extending RTHS to complex multi-physics SFSI problems, improving the ability to quantify earthquake-induced impacts on the built environment and assess infrastructure performance and community resilience.
Assistant Professor Texas Tech University
Augmented-Input Variational LSTM for Nonlinear Structural Response Histories Metamodeling Accounting for Aleatory and Epistemic Uncertainty
Co-Authors: Manisha Sapkota (Texas Tech University) and Min Li (Rensselaer Polytechnic Institute)
Abstract: Accurate uncertainty propagation is essential for nonlinear dynamic structural systems subjected to stochastic excitation, particularly in performance-based analysis and risk-informed decision-making. Conventional simulation-based approaches, such as Monte Carlo sampling, become computationally prohibitive when repeated nonlinear time-history analyses are required. Recent advances in deep learning offer new opportunities for efficient surrogate modeling; however, most existing approaches primarily focus on deterministic prediction or capture only record-to-record variability, while neglecting parametric system uncertainty and epistemic uncertainty associated with model inadequacy and limited data. This study proposes a deep learning-powered uncertainty propagation framework for nonlinear dynamic structural systems under stochastic excitation. The framework employs sequence-to-sequence neural networks to learn the mapping from stochastic excitation to structural response histories, while explicitly accounting for both aleatory and epistemic uncertainties. Parametric system uncertainties are incorporated through augmented network inputs, and epistemic uncertainty is quantified using Monte Carlo dropout, which provides a scalable approximation to Bayesian inference without incurring significant computational overhead. The proposed approach enables efficient estimation of response distributions, and predictive confidence intervals. Numerical examples involving nonlinear structural dynamics demonstrate that the proposed framework achieves high accuracy in response prediction while providing reliable uncertainty quantification. The results highlight the potential of deep learning-based uncertainty propagation as a practical and scalable alternative to conventional simulation-based methods for nonlinear dynamic analysis of structures subjected to stochastic excitation.
PhD Student UC Berkeley
Physics-Informed Adaptive Fourier Neural Operator Framework for Nonlinear Seismic Response Prediction from SDOF to 5-Story MDOF Systems
Co-Author: Luis Ceferino (University of California, Berkeley)
Abstract: This work develops a fragility-oriented Physics-Informed Adaptive Fourier Neural Operator (AFNO-PINN) framework for efficient prediction of nonlinear seismic structural response and downstream risk quantification. We began with bilinear elasto-plastic SDOF systems and established an end-to-end simulation pipeline, including ground-motion selection/augmentation, parameter sampling, and nonlinear time-history analysis (Newmark-based integration) to generate displacement, velocity, acceleration, and restoring-force responses. Building on these datasets, we designed multimodal AFNO architectures that combine time–frequency ground-motion features (FFT/STFT, envelopes, and statistical descriptors) with structural parameters and learned time encodings. To improve robustness and physical plausibility, we incorporated physics-informed residuals (equation-of-motion constraints), curriculum-style multi-stage training, normalization, and adaptive loss balancing, and benchmarked against carefully controlled data-driven baselines. We implemented engineering-focused diagnostics beyond pointwise errors, including hysteresis-loop fidelity, spectral agreement, phase/energy consistency, and elastic–inelastic masking to better assess performance under yielding. We then extended the framework to a 5-story MDOF shear-building setting by integrating spatial coupling (GNN modules), conditional modulation (FiLM), and vectorized physics residuals across floors. Finally, we connect fast surrogate predictions to fragility analysis by enabling scalable intensity-measure studies (e.g., stripe analysis) to extract demand metrics such as peak inter-story drift ratio and collapse indicators, supporting the construction of fragility curves for performance-based earthquake engineering.
Assistant Professor University of Miami
An LLM-powered Chatbot to Automate Structural Analysis
Co-Authors: Ziheng Geng (University of Miami), Yinglao Liu (Independent researcher), and Jiachen Liu (University of Miami)
Abstract: Structural analysis is fundamental to the design and assessment of civil structures and infrastructure, yet its practice often relies on complex workflows that demand significant expertise and manual effort. Recent advances in large language models (LLMs) offer new opportunities to transform how engineers interact with structural analysis tools. This research presents an LLM-powered chatbot designed to support structural analysis by translating natural language into executable codes. Rather than replacing engineers or existing engineering software, the chatbot functions as an intelligent interface and assistant to engineers. Users specify analysis tasks (e.g., analysis of plane frames), and the chatbot performs the structural analysis by generating code that can be executed in established software such as OpenSees and SAP2000. The chatbot builds on an LLM-based multi-agent system that decomposes the modeling process into a sequence of subtasks and then executes them. The system is evaluated on 20 benchmark problems and consistently outperforming state-of-the-art general-purpose LLMs, including Gemini-2.5 Pro and ChatGPT-4o. The chatbot has been tested in structural analysis class, in senior design, and in companies. The chatbot has been tested in structural analysis courses, senior design projects, and industry settings. The presentation will share results and insights from these initial tests.
Assistant Professor Utah State University
A physics-informed machine learning approach based on knowledge distillation to support performance-based design
Abstract: The complexities of performance-based seismic design (PBSD) have limited their widespread use for advancing community resilience. PBSD requires nonlinear finite element models, dynamic simulations, and integration with fragility and consequence functions, which makes them computationally demanding and difficult to scale. Recently, Machine learning (ML) provided a promising solution by serving as surrogate models to estimate nonlinear responses, damage states, or losses. However, ML-based approaches often face challenges of generalizability and physical inconsistency.
This study develops a novel physics-informed ML (PIML) approach for PBSD using knowledge distillation. In this method, physics-based relationships and domain knowledge are treated as auxiliary information to guide the training of teacher models, which subsequently transfer knowledge to computationally efficient student models. Unlike traditional PIML implementations that rely on partial differential equations in the loss function, this approach is better aligned with structural design problems where knowledge cannot be expressed solely in equation form. Algorithmically, the framework introduces two key innovations: (1) it extends generalized distillation into the PBSD domain, and (2) it incorporates a transformer-based architecture capable of capturing complex input–output mappings with high accuracy using limited information. From an application perspective, the study advances PIML in PBSD by leveraging generative AI for knowledge distillation, enabling scalable, physically consistent, and design-oriented surrogate models.
PhD Student University of California, San Diego
Long Short-Term Memory Networks-Recursive Averaged Multistep Sequence-to-Sequence (LSTM-RAMSS) Model as Emulators for Finite Element Models of Nonlinear Structural Dynamic Systems
Co-Authors: Zhen Hu (University of Michigan-Dearborn) and Joel P. Conte (University of California, San Diego)
Abstract: Performance-based seismic design (PBSD) and reliability and risk assessment of structural systems require thousands of nonlinear response history analyses of high-fidelity FE models, making repeated simulation a major computational bottleneck. Data-driven surrogate models, such as LSTM-based, offer a practical path to accelerate and/or enable these probabilistic workflows, but the standard LSTM surrogates, such as Sequence-to-Sequence (Seq2Seq)-based and Nonlinear AutoRegressive eXogenous (NARX), that often degrade when applied to realistic high-dimensional FE model of nonlinear structural dynamic systems. This work presents LSTM-RAMSS, a unified recurrent framework that integrates (i) recursive prediction, leveraging prior outputs as inputs to propagate system dynamics (LSTM-NARX architecture); and (ii) multistep sequence-to-sequence prediction (LSTM-Seq2Seq architecture) to stably predict response histories. The proposed workflow introduces three key elements: (1) an optimized convolutional autoencoder (CAE) strategy combined with the use of an engineering intensity measure for selecting the broadest possible range of real ground motions characteristics and the full range from quasi-linear to strongly nonlinear response regimes, (2) a dilation strategy to capture nonlinear inelastic behavior across multiple time scales, and (3) an intermediate averaging mechanism to improve robustness during recursive rollout. Performance is demonstrated on two tiers of application: (a) academic benchmarks (linear 8-story shear building, nonlinear inelastic SDOF system, nonlinear inelastic 3-story shear building) and (b) realistic design-scale models (RC bridge column, 2D 9-story steel frame, 3D RC frame). Results show that LSTM-RAMSS accurately emulates both univariate and multivariate global/local structural time histories across increasing complexity and nonlinearity, enabling scalable surrogate-based PBSD and reliability and risk analyses.
PhD Student University of Michigan
Metamodeling of multi-hazard nonlinear stochastic dynamic systems with adaptive hybrid neural operator schemes
Co-Authors: Somdatta Goswami (Johns Hopkins University) and Seymour M.J. Spence (University of Michigan)
Abstract: Simulating high-dimensional, nonlinear dynamic structural systems subjected to extreme natural hazards is computationally demanding, particularly for uncertainty quantification and performance-based design. Recent advances in neural-network-based metamodeling have shown promise in reducing this cost; however, many existing approaches, including physics-informed machine learning frameworks, do not enforce physical constraints during inference, limiting their reliability for practical structural engineering applications.
This study proposes a novel hybrid framework that integrates neural operators—a class of models designed to learn mappings between infinite-dimensional function spaces—with the Newmark–beta implicit time-integration scheme. While implicit integration methods inherently satisfy governing physical constraints, their computational expense is dominated by repeated evaluations of restoring forces. To mitigate this bottleneck, the proposed framework uses a neural operator to efficiently approximate restoring forces based on prior structural states, thereby accelerating the time-integration process.
To maintain physical consistency during inference, the structural response is periodically corrected by solving the governing equations using the full Newmark–beta scheme after an adaptively determined number of hybrid time steps. The proposed approach is validated on a steel moment-resisting frame subjected to stochastic seismic and wind excitations. Generalization capability is evaluated by training the metamodel on one hazard type and testing it on the other, a scenario in which conventional machine-learning-based metamodels often fail.
Results demonstrate that the proposed framework achieves substantially higher accuracy than conventional approaches, yielding more than an 85 times reduction in mean absolute error compared to LSTM-based models, while providing a computational speed-up of approximately 2 to 4 times relative to traditional implicit integration.
PhD Student Stanford University
Enabling Higher-Resolution Nonlinear Structural Analysis Through GPU-Accelerated Solvers
Co-Author: Barbara G. Simpson (Stanford University)
Abstract: Solving systems of linear equations is the foundation of many algorithms in nonlinear structural analysis. Although GPU-accelerated iterative solvers have demonstrated transformative speedups across various scientific computing applications, most structural engineering finite element frameworks continue to rely on legacy CPU-based direct solvers. This work extends the OpenSees framework with programming interfaces for two state-of-the-art GPU-accelerated linear solver libraries: AmgX (iterative methods) and cuDSS (direct methods). Both libraries include configurable solver settings that can be tailored to different structural models and GPU hardware, including preconditioners, convergence tolerances, single precision options, storage formats for iterative solvers, and factorization/reordering strategies for direct solvers. Performance is evaluated using a simply supported beam discretized with brick elements and a J2 plasticity constitutive law, subjected to static and dynamic analyses. Depending on model resolution, results show that speedups of up to an order of magnitude can be achieved without notable loss of accuracy when solver parameters are carefully selected. Trade-offs between iterative and direct solvers are assessed in terms of convergence behavior, memory usage, and sensitivity to ill-conditioning. Findings support the use of GPU-accelerated solvers and expand on the algorithms available to reduce computational costs and enable higher-fidelity simulations of nonlinear structural dynamic response.