Multiscale Structural Reliability Analysis in high dimensions with Tensor Trains and Physics-Augmented Neural Networks

Aryan Tyagi 1 Alex de Beer 2 Tiangang Cui 2 Jan N. Fuhg 1,3,† 1 Department of Aerospace Engineering & Engineering Mechanics, The University of Texas at Austin, Austin, TX 2 School of Mathematics and Statistics, University of Sydney, New South Wales, Australia 3 The Oden Institute of Computational Science and Engineering, The University of Texas at Austin, Austin, TX † Correspondence: [email protected] · 2026

Read Multiscale Structural Reliability Analysis in high dimensions with Tensor Trains and Physics-Augmented Neural Networks by Aryan Tyagi 1 Alex de Beer 2 Tiangang Cui 2 Jan N. Fuhg 1,3,† 1 Department of Aerospace Engineering & Engineering Mechanics, The University of Texas at Austin, Austin, TX 2 School of Mathematics and Statistics, University of Sydney, New South Wales, Australia 3 The Oden Institute of Computational Science and Engineering, The University of Texas at Austin, Austin, TX † Correspondence: [email protected] (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-18776.txt.

Introduction

Tyagi and three colleagues tackle multiscale structural reliability for composites, where small scale material uncertainty must be carried to the large scale. Standard methods solve coupled partial differential equation systems again and again across scales, which is too costly in many dimensions. Their framework pairs a physics augmented neural network, which gives homogenized stiffness fast and within physical bounds, with a tensor train transport method that estimates the chance of failure. On a 3D benchmark, it gives steady failure estimates in dimensions up to 150.

Catalog record

CREATORAryan Tyagi 1 Alex de Beer 2 Tiangang Cui 2 Jan N. Fuhg 1,3,† 1 Department of Aerospace Engineering & Engineering Mechanics, The University of Texas at Austin, Austin, TX 2 School of Mathematics and Statistics, University of Sydney, New South Wales, Australia 3 The Oden Institute of Computational Science and Engineering, The University of Texas at Austin, Austin, TX † Correspondence: [email protected]
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE69103
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.18776

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Multiscale Structural Reliability Analysis in high dimensions with Tensor Trains and Physics-Augmented Neural Networks Aryan Tyagi 1 Alex de Beer 2 Tiangang Cui 2 Jan N. Fuhg 1,3,† 1 Department of Aerospace Engineering & Engineering Mechanics, The University of Texas at Austin, Austin, TX 2 School of Mathematics and Statistics, University of Sydney, New South Wales, Australia 3 The Oden Institute of Computational Science and Engineering, The University of Texas at Austin, Austin, TX † Correspondence: [email protected] Abstract Structural reliability evaluation for composites constitutes a fundamentally high-dimensional multiscale problem, as microscale material uncertainties (e.g., fiber volume fraction and constituent moduli) must propagate to the macroscale and can be quantified as high-dimensional random fields. Conventional approaches are computationally intractable, as they rely on repeatedly solving tightly coupled partial differential equation systems across scales while contending with the exponential complexity inherent in high-dimensional uncertainty quantification (UQ). This work introduces a scalable and physically consistent framework that addresses both bottlenecks simultaneously in the case of separation of scales and (anisotropic) linear elasticity. In particular, we couple a physics-augmented Voigt–Reuss Neural Network (VRNN) with the Deep Inverse Rosenblatt Transport (DIRT) method to estimate the posterior probability of multiscale structural failure. The VRNN is used to resolve the computationally expensive FE 2 scheme by providing a near-instantaneous evaluation of the homogenized stiffness tensor that is guaranteed to be symmetric, positive-definite, and strictly bounded within the Voigt–Reuss limits. This enables fast evaluation of the homogenized responses, which are required for the macroscale finite element model. The DIRT method constructs a sequence of functional tensor train approximations to efficiently store an approximation of the high-dimensional optimal importance sampling distribution for estimating the probability of failure under the VRNN surrogate. This mitigates the curse of dimensionality arising from the Karhunen–Loève (KL) expansion of the random fields. The framework is demonstrated on a three-dimensional heterogeneous benchmark problem, where the uncertainty in the microscale material properties is characterized by a Bayesian posterior distribution obtained from limited strain observations. Our results show that the DIRT-based estimator can provide low-variance estimates of failure probabilities in dimensions up to 150. This performance confirms that the combined VRNN-DIRT approach yields reliable failure probability estimates efficiently, providing a robust solution for scalable multiscale UQ in engineered composites. 1 Introduction Structural reliability analysis aims to quantify the safety of an engineering structure. This is typically achieved by modeling uncertain input parameters in a probabilistic framework and evaluating the failure domain through a Limit State Function (LSF) to estimate the probability of failure of the structure [ 57 , 5 , 27 ] . Over the past decades, a large number of numerical methods have been developed and used to efficiently obtain estimations of the probability of failure, such as Subset Simulation [ 3 ] , Importance Sampling [ 13 , 22 ] , and First and Second Order Reliability Methods (FORM and SORM) [ 2 , 9 ] . In mechanical systems, uncertainties in the parameters can be aleatoric (inherent) or epistemic (arising from measurement errors). Most classical reliability methods assume probabilistic models are known a priori, and therefore implicitly treat uncertainty as aleatoric [ 38 , 50 , 49 , 69 , 66 ] . However, recent efforts have also incorporated epistemic uncertainties through Bayesian updating of model parameters and uncertainties using noisy measurement data, leading to an updated (posterior) probability of failure [ 60 , 61 , 62 , 39 ] . This leads to the posterior probability of failure, which reflects both inherent randomness and updated knowledge inferred from measurement data. While probabilistic reliability methods quantify how uncertainties affect structural failure, a central question that remains is: where do these uncertainties actually originate and how should they be modeled in materials with complex internal architecture such as fiber-reinforced composites? Fiber-reinforced composite materials used in aerospace, mechanical, and civil engineering structures are often subjected to harsh operating conditions. In such composite materials, mesoscale fluctuations are present in the material properties of the structure [ 63 , 62 ] . However, the actual source of uncertainty is the microstructure of the composite. This can arise from manufacturing defects, variations in constituent moduli, or imperfections introduced during additive manufacturing. These considerations naturally motivate modeling effective microstructural properties as random fields. This makes structural reliability for composite materials fundamentally a multiscale problem. The microscale uncertainties must propagate to the macroscale before influencing the structural response. Under the assumption of statistical and mechanical scale separation [ 47 , 12 ] , such multiscale problems can be solved using a hierarchical approach [ 28 ] . In complex geometries, this naturally involves solving a system of coupled partial differential equations (PDEs). Here, the associated boundary value problems (BVPs) at the microscale (called homogenization problems) are typically solved using Finite Element Methods (FEM) [ 24 , 56 , 35 , 42 ] , or Fast Fourier Transform (FFT)-based solvers [ 44 , 43 ] . Using these approaches, several studies have incorporated uncertainty in the geometric descriptors of the microscale unit cell [ 70 , 4 ] . However, computationally quantifying the effects of these uncertainties is generally computationally prohibitive, especially for spatially varying macrostructures. In particular, FE 2 (FE at macroscale – FE at microscale) needs to solve a BVP at each integration point over the macroscale. While FE-FFT is faster than FE, increasingly fine voxel grids are required if the uncertainties arise due to small changes in the microstructure geometry (e.g., fiber radius). To address this, reduced-order and surrogate models for microstructural homogenization have been developed. These approaches replace the coupled PDE system with a macroscale PDE combined with a constitutive evaluation, similar to classical single-scale continuum mechanics. Historically, these reduced-order models were mostly based on intrusive projection methods based on proper orthogonal decomposition [ 36 ] . However, parameterized homogenization, like that required for uncertainty quantification, is not straightforward with these methods [ 52 , 53 ] . In the context of surrogate modeling, parameterization can be naturally included. Several machine learning architectures have therefore been applied, such as feedforward neural networks [ 37 , 8 ] , graph neural networks [ 64 ] , or Gaussian process regression [ 26 ] . However, the field has increasingly shifted towards the development of machine learning models that can intrinsically, i.e., by design, fulfill physical concepts and mechanistic assumptions instead of only being trained on data [ 30 , 31 , 32 , 40 , 23 ] . In such architectures, fewer training labels are needed since these assumptions act as regularizers in the parametric space. Additionally, they have the potential to remain more accurate when extrapolating data [ 25 ] . In the present work, we focus on fiber-reinforced composites with linearly elastic fiber and matrix materials. We rely on the recently introduced Voigt-Reuss Neural Networks (VRNNs) that structurally enforce that predicted stiffness matrices remain within the theoretical Voigt–Reuss bounds for a given microstructure [ 33 ] . In this work, we assume that the microstructure is spatially variable over the macroscale (under the assumption of scale separation) and follows an unknown probability distribution, which can be inferred from limited macroscale observations through Bayesian updating. Even with efficient homogenization, reliability analysis remains computationally prohibitive due to the

Read the complete text