betaFeedback

WorksPhysics and Astronomy

Generative Learning on Manifolds for Probabilistic Seismic Performance Assessment and Structural Identifiability

Abstract

Seismic performance assessment relies on computationally expensive nonlinear simulations, limiting uncertainty quantification and inference of structural properties from observed responses. We develop a manifold-based generative framework that learns the joint distribution of uncertain structural parameters and seismic-response trajectories from limited high-fidelity simulations. Combining Diffusion Maps, score-based Diffusion Modeling, and Geometric Harmonics, the framework enables both forward probabilistic prediction and inverse inference. Applied to a nine-story steel moment-resisting frame, the results reveal which structural properties are identifiable from seismic performance and how uncertainty in these properties influences response. In particular, the learned representation captures post-capping degradation characteristics more effectively than individual strength parameters. Inverse inference identifies distinct structural configurations that produce similar seismic performance, with representative solutions independently validated using nonlinear simulations. These findings demonstrate how generative learning can support uncertainty-aware seismic assessment while revealing structural non-identifiability and alternative mechanisms underlying observed performance

Keywords: Earthquake engineering, Uncertainty-aware AI, Generative modeling, Manifold learning, Incremental dynamic analysis

Introduction

Natural hazard risk reduction increasingly depends on computational models that connect hazard intensity to the performance of exposed infrastructure. In earthquake engineering, performance-based assessment combines seismic hazard characterization, nonlinear structural analysis, and consequence evaluation to quantify how buildings and infrastructure respond across a range of intensity levels and uncertainty sources [1, 2]. A central computational challenge is that nonlinear response-history analysis (NLRHA) must be repeated over many ground-motion records, intensity levels, and uncertain structural models. This becomes particularly expensive in Incremental Dynamic Analysis (IDA), where each ground motion is repeatedly scaled until the response evolves from the elastic regime toward severe nonlinearity and collapse.

Artificial intelligence and machine learning methods offer an opportunity to reduce this computational burden, but natural hazard applications require more than point prediction. Reliable models must preserve uncertainty, remain meaningful when high-fidelity data are scarce, and respect the structure imposed by the underlying physics. In performance-based earthquake engineering, both epistemic uncertainty in structural modeling and record-to-record variability in ground motion can strongly influence response, especially in the nonlinear and near-collapse regimes [3]. Existing acceleration strategies include advanced Monte Carlo methods [4, 5], reduced-order modeling [6], fragility-estimation procedures [7], semi-analytical stochastic methods [8, 9], and surrogate models [10, 11, 12, 13, 14, 15, 16, 17, 18, 19]. Most surrogate approaches, however, are formulated primarily as deterministic or conditionally averaged input–output maps. Such models can be highly effective for prediction but do not directly provide a generative representation of the coupled distribution of uncertain structural properties and fixed-dimensional seismic-response representations.

Generative models provide a complementary perspective by learning probability distributions rather than only deterministic mappings. Score-based diffusion models are particularly attractive because they represent complex target distributions through a diffusion process and corresponding reverse dynamics [20, 21, 22, 23]. Their direct application can nevertheless be difficult when high-dimensional engineering data are concentrated near a lower-dimensional manifold and only a limited number of high-fidelity samples are available. Probabilistic Learning on Manifolds and related nonlinear manifold-learning methods address this setting by identifying intrinsic coordinates and performing probabilistic learning in the reduced representation [24, 25, 26, 27]. Recent work has combined score-based generative learning with Diffusion Maps and Double Diffusion Maps to sample distributions supported near learned manifolds and reconstruct ambient variables through Geometric Harmonics [28, 29, 30, 31].

Here we adapt this manifold-based generative framework to probabilistic seismic performance assessment. IDA is especially well suited to this setting because each uncertain structural realization produces a response trajectory rather than a single scalar output. Instead of approximating selected response quantities independently, we learn the joint probability distribution of uncertain structural parameters and the corresponding representative IDA response. The resulting representation supports both unconditional generation of structural parameter–response pairs and forward conditional inference of response distributions for prescribed structural properties. The principal contributions of this work are:

  1. We formulate probabilistic seismic performance assessment through the joint distribution of uncertain structural parameters and representative IDA responses, enabling uncertainty-aware inference beyond conventional deterministic input–output surrogate models.

  1. We develop a hybrid manifold-based generative framework that retains structural parameters explicitly while representing seismic-response trajectories through non-harmonic (essentially independent) Diffusion Maps coordinates. Combined with score-based diffusion modeling, this enables probabilistic learning from limited high-fidelity simulations without additional nonlinear analyses for generation.

  1. We investigate the capability of accurately inferring the structural parameter values of the model from observations of its behavior (“data”).

  1. We demonstrate forward and inverse probabilistic inference using the learned input-output joint distribution. Forward conditioning characterizes the influence of uncertain structural properties on seismic performance, while inverse conditioning reveals distinct structural configurations compatible with the same observed response.

The proposed framework is demonstrated using a nine-story steel momentresisting frame [32] with uncertain nonlinear structural properties, previously investigated in the context of incremental dynamic analysis (IDA) and uncertainty quantification [14, 19]. Unlike our previous studies, which primarily focused on forward surrogate prediction and dimensionality reduction [14, 19], the present work aims to learn the joint parameter–response distribution, enabling generative sampling, conditional probabilistic inference, and the investigation of parameter identifiability. In this context, identifiability concerns whether distinct combinations of structural parameters produce distinguishable structural responses. Specifically, if the forward mapping from model parameters to structural responses is injective, the parameters can, in principle, be uniquely recovered from the observed responses, and the model is said to be structurally identifiable. Conversely, if multiple parameter configurations yield identical responses, the inverse problem is nonunique, indicating structural unidentifiability. From an analytical perspective, local identifiability can be investigated by examining the local injectivity of the forward mapping, typically through the rank of its Jacobian with respect to the uncertain parameters. The proposed probabilistic framework complements this perspective by characterizing the conditional distribution of structural parameters given observed responses, thereby capturing parameter ambiguity, uncertainty, and potentially multiple plausible parameter configurations.

The remainder of the paper introduces the score-estimation and manifold-learning components, formulates the joint generative problem, presents the earthquake-engineering application and validation results, and concludes with implications, limitations, and directions for uncertainty-aware AI in seismic performance assessment.

Methods

Score-Based Generative Modeling

Score-based generative models define a forward stochastic differential equation (SDE) that progressively transforms a target data distribution into a tractable noise distribution and then use the corresponding reverse dynamics to generate samples from the target law [20]. A general forward SDE can be written as

dx=f(x,t) dt+g(t) dBt(1)\mathrm{d}x = f(x,t)\,\mathrm{d}t + g(t)\,\mathrm{d}B_t \tag*{(1)}

where f(x,t)f(x,t) and g(t)g(t) are the drift and diffusion terms and BtB_t is Brownian motion. The associated reverse-time SDE depends on the score ∇xlog⁡pt(x)\nabla_x \log p_t(x),

dx=[f(x,t)−g(t)2∇xlog⁡pt(x)]dt+g(t) dBˉt.(2)\mathrm{d}x = \left[f(x,t) - g(t)^2 \nabla_x \log p_t(x)\right]\mathrm{d}t + g(t)\,\mathrm{d}\bar{B}_t. \tag*{(2)}

Conventional diffusion models approximate the score with a neural network trained by score matching or denoising score matching [33, 34].

In this work, we instead use the Monte Carlo score-estimation strategy of Liu et al. [35]. For a forward process with Gaussian transition density

p(xt∣x0)=N(αtx0,βt2I)(3)p(\mathbf{x}_t \mid\mathbf{x}_0) = \mathcal{N}\left(\alpha_t\mathbf{x}_0, \beta_t^2\mathbf{I}\right) \tag*{(3)}

the score can be estimated directly from the available reference samples without training a separate score network. Denoting the fixed empirical reference set by {xn}n=1N\{\mathbf{x}^n\}_{n=1}^{N}, the estimator is

s^(xt,t)=∑n=1Nαtxn−xtβt2w‾t(xt,xn),w‾t=p(xt∣xn)∑m=1Np(xt∣xm).(4)\widehat{\mathbf{s}}(\mathbf{x}_t,t) = \sum_{n=1}^{N} \frac{\alpha_t\mathbf{x}^n-\mathbf{x}_t}{\beta_t^2}\overline{w}_t(\mathbf{x}_t,\mathbf{x}^n), \qquad\overline{w}_t = \frac{p(\mathbf{x}_t\mid\mathbf{x}^n)}{\sum_{m=1}^{N}p(\mathbf{x}_t\mid\mathbf{x}^m)}. \tag*{(4)}

For the generative state z\mathbf{z} used in the present implementation, the Gaussian forward transition is parameterized as

p(zt∣z0)=N(αtz0,βt2I)(5)p(\mathbf{z}_t \mid\mathbf{z}_0) = \mathcal{N}\left(\alpha_t\mathbf{z}_0, \beta_t^2\mathbf{I}\right) \tag*{(5)}

with

αt=1−γt,βt2=(1−γ)+γt,t∈[0,1],(6)\alpha_t = 1-\gamma t, \qquad\beta_t^2 = (1-\gamma)+\gamma t, \qquad t\in[0,1], \tag*{(6)}

where γ=0.999\gamma=0.999. Thus, α0=1\alpha_0=1, β02=10−3\beta_0^2=10^{-3}, α1=10−3\alpha_1=10^{-3}, and β12=1\beta_1^2=1. The nonzero value of β02\beta_0^2 provides a small variance floor that regularizes the empirical score near the data endpoint t=0t=0. The corresponding linear forward drift coefficient is

b(t)=α˙tαt=−γ1−γt,(7)b(t)=\frac{\dot{\alpha}_t}{\alpha_t}=-\frac{\gamma}{1-\gamma t}, \tag*{(7)}

and the diffusion variance satisfies

σ2(t)=dβt2dt−2b(t)βt2=γ−2b(t)βt2.(8)\sigma^2(t)=\frac{\mathrm{d}\beta_t^2}{\mathrm{d}t}-2b(t)\beta_t^2=\gamma-2b(t)\beta_t^2. \tag*{(8)}

The associated probability-flow ODE is

dztdt=b(t)zt−12σ2(t)∇zlog⁡pt(zt).(9)\frac{\mathrm{d}\mathbf{z}_t}{\mathrm{d}t}=b(t)\mathbf{z}_t-\frac{1}{2}\sigma^2(t)\nabla_{\mathbf{z}}\log p_t(\mathbf{z}_t). \tag*{(9)}

Generation starts at T=1T=1 from z1∼N(0,I)\mathbf{z}_1\sim\mathcal{N}(0,\mathbf{I}), which approximates the terminal marginal since α1=10−3\alpha_1=10^{-3}, and Eq. (9) is integrated backward toward t=0t=0. A uniform time grid

tk=1−kΔt,Δt=1Nt,k=0,…,Nt,(10)t_k=1-k\Delta t,\qquad\Delta t=\frac{1}{N_t},\qquad k=0,\ldots,N_t, \tag*{(10)}

is used with Nt=1000N_t = 1000. With explicit Euler integration,

zk+1=zk−Δt[b(tk)zk−12σ2(tk)s^(zk,tk)].(11)\mathbf{z}_{k+1} = \mathbf{z}_k - \Delta t \left[ b(t_k)\mathbf{z}_k - \frac{1}{2}\sigma^2(t_k)\widehat{\mathbf{s}}(\mathbf{z}_k,t_k) \right]. \tag*{(11)}

The drift and score are evaluated for k=0,…,Nt−1k = 0,\ldots,N_t - 1; therefore, the final update reaches t=0t = 0 without evaluating the score at the endpoint itself.

Manifold-Based Generative Learning of Seismic Response

We couple physics-based IDA with manifold-based generative learning to estimate a low-dimensional representation of the dominant coupled variability between uncertain structural properties and aggregated seismic-response trajectories. Let

xs∈Rns,xs∼fxs(⋅),(12)\mathbf{x}_{\mathrm{s}} \in\mathbb{R}^{n_{\mathrm{s}}}, \qquad\mathbf{x}_{\mathrm{s}} \sim f_{\mathbf{x}_{\mathrm{s}}}(\cdot), \tag*{(12)}

denote the uncertain structural-parameter vector, and let

xg(r)∈Rng,r=1,…,Ng,(13)\mathbf{x}_{\mathrm{g}}^{(r)} \in\mathbb{R}^{n_{\mathrm{g}}}, \qquad r = 1,\ldots,N_{\mathrm{g}}, \tag*{(13)}

denote the rrth ground-motion realization. The structural model evaluated within IDA is written as

G:(xs,xg(r),λ)⟼(D(r)(λ),Sa(r)(λ)),(14)\mathcal{G}:\left(\mathbf{x}_{\mathrm{s}},\mathbf{x}_{\mathrm{g}}^{(r)},\lambda\right) \longmapsto\left(D^{(r)}(\lambda),S_{\mathrm{a}}^{(r)}(\lambda)\right), \tag*{(14)}

where λ\lambda denotes the record scaling level, D(r)(λ)D^{(r)}(\lambda) is the corresponding maximum interstory drift, and Sa(r)(λ)S_{\mathrm{a}}^{(r)}(\lambda) is the associated 5%-damped first-mode spectral acceleration. For each structural realization, Ng=30N_{\mathrm{g}} = 30 single-record IDA curves are available. To obtain a fixed-dimensional representation, each record-level IDA is first converted to a single-valued first-passage relation Sa(r)(D)S_{\mathrm{a}}^{(r)}(D) and interpolated onto a common drift grid. The representative response for each structural realization is then defined as the pointwise median of the 30 interpolated record-level responses. For records terminating before the end of the common grid, the terminal spectral-acceleration value is retained over the remaining drift coordinates. The 30 record-level responses are then aggregated pointwise on the common drift grid. Let Dcom=[D1com,…,D18com]\mathbf{D}^{\mathrm{com}} = [D_1^{\mathrm{com}},\ldots,D_{18}^{\mathrm{com}}] denote the common drift grid, consisting of 18 prescribed drift levels between 0 and 0.20. At each drift coordinate DjcomD_j^{\mathrm{com}}, we compute

S~a,j=median⁡r=1,…,NgSa(r)(Djcom),j=1,…,18,(15)\widetilde{S}_{\mathrm{a},j} = \operatorname{median}_{r=1,\ldots,N_{\mathrm{g}}} S_{\mathrm{a}}^{(r)}\left(D_j^{\mathrm{com}}\right), \qquad j = 1,\ldots,18, \tag*{(15)}

with Ng=30N_g = 30. The representative response is then defined pointwise as

S~a=[S~a,1,…,S~a,18]∈R18.(16)\widetilde{\mathbf{S}}_a = \left[\widetilde{S}_{a,1}, \ldots, \widetilde{S}_{a,18}\right] \in\mathbb{R}^{18}. \tag*{(16)}

The representative response is therefore a monotone median IDA curve evaluated on the common drift grid, written as xr=S~a∈R18\mathbf{x}_r = \widetilde{\mathbf{S}}_a \in\mathbb{R}^{18}. The learning vector in physical coordinates is then defined as

x=(xs,xr)=(xs,S~a)∈Rns+18.(17)\mathbf{x} = (\mathbf{x}_s,\mathbf{x}_r) = \left(\mathbf{x}_s,\widetilde{\mathbf{S}}_a\right) \in\mathbb{R}^{n_s+18}. \tag*{(17)}

For the present application, ns=6n_s = 6, so that x∈R24\mathbf{x} \in\mathbb{R}^{24}. Ground-motion record-to-record variability enters the analysis through the 30 single-record IDA trajectories used to construct xr\mathbf{x}_r; the ground-motion variables and individual record-level responses are not explicit random variables in the learned generative representation.

Building on our previous manifold-constrained generative framework [28], the present implementation combines Diffusion Maps [31] with training-free Monte Carlo score estimation [35]. In contrast to applying dimensionality reduction to the complete structural parameter–response vector, Diffusion Maps are applied only to the 18-dimensional representative response. The six structural parameters remain explicit coordinates throughout the generative procedure. This avoids reconstructing independently sampled structural variables through the response-manifold lifting map while retaining a low-dimensional representation of the nonlinear IDA response. Starting from the training response vectors

Xr={xr,i}i=1Ntr⊂R18,(18)\mathbf{X}_r = \left\{\mathbf{x}_{r,i}\right\}_{i=1}^{N_{\mathrm{tr}}} \subset\mathbb{R}^{18}, \tag*{(18)}

we apply Diffusion Maps [26] to obtain a response latent representation

Φr={ϕr,i}i=1Ntr⊂Rdr.(19)\mathbf{\Phi}_r = \left\{\boldsymbol{\phi}_{r,i}\right\}_{i=1}^{N_{\mathrm{tr}}} \subset\mathbb{R}^{d_r}. \tag*{(19)}

Fifteen diffusion eigenpairs are evaluated, and local regression residuals are used to identify non-harmonic coordinates. Six response coordinates are retained. The objective is not strictly dimensionality reduction, but identification of a compact set of non-redundant coordinates that preserves the intrinsic response geometry. The generative state is then defined as

zi=(xs,i(std),ϕr,i)∈R6+6=R12.(20)\mathbf{z}_i = \left(\mathbf{x}_{s,i}^{(\mathrm{std})},\boldsymbol{\phi}_{r,i}\right) \in\mathbb{R}^{6+6} = \mathbb{R}^{12}. \tag*{(20)}

Although the response representation is reduced only from 18 to 6 coordinates, the effect on the joint generative problem is more substantial: the physical parameter–response space is reduced from 24 to 12 dimensions. This reduction is particularly important in the present small-sample regime, where only 151 training realizations are available. Reducing the effective dimension decreases the difficulty of estimating the empirical score and the underlying joint distribution from a limited number of samples, while the held-out reconstruction results indicate that the dominant response information is retained.

In the generative phase, samples y∼N(0,I12)\mathbf{y} \sim\mathcal{N}(0, I_{12}) are drawn and transported directly through the probability-flow ODE using the Monte Carlo score estimate. The generated structural coordinates are transformed directly back to physical units. Only the generated response latent coordinates are lifted to the 18-dimensional response space using Geometric Harmonics (GH) (see Appendix § 6). The resulting samples

{x~i}i=1N∗={(x~s,i,x~r,i)}i=1N∗∼fxs,xr(⋅)(21)\{\tilde{\mathbf{x}}_i\}_{i=1}^{N^*} = \{(\tilde{\mathbf{x}}_{s,i}, \tilde{\mathbf{x}}_{r,i})\}_{i=1}^{N^*} \sim f_{\mathbf{x}_s,\mathbf{x}_r}(\cdot) \tag*{(21)}

provide an expanded ensemble of structural parameter–response realizations. A schematic illustration of the approach is provided in Figure 1.

Proposed manifold-based generative framework for learning the joint distribution of structural properties and seismic response and enabling conditional sampling

Figure 1. Proposed manifold-based generative framework for learning the joint distribution of structural properties and seismic response and enabling conditional sampling.

Earthquake Application: Nine-Story Steel Moment-Resisting Frame

We consider a nine-story steel moment-resisting frame with a single-story basement, hereafter referred to as LA9, designed for Los Angeles in accordance with the 1997 NEHRP provisions [32]. The structure is represented by a centerline model developed in OpenSees, incorporating nonlinear beam–column connections that allow plastic hinges to form at the beam ends, while the columns are assumed to remain elastic. While uncertainties in material yield strengths may cause columns to yield before their connecting beams, previous analyses have shown that this effect is relatively minor for the LA9 structure, particularly when beam and column steel strengths are strongly correlated [3]. Moreover, the authors in [36] demonstrated that plastic-hinge yield strength has a greater influence on displacement-related structural responses than variations in story mass and stiffness, highlighting its importance in structural uncertainty quantification. PP–Δ\Delta effects and the internal gravity frames are included in the model. The fundamental period of the reference structure is T1=2.35T_1 = 2.35 s, with the first mode accounting for approximately 84% of the total mass. The structure is therefore predominantly first-mode dominated, while retaining sensitivity to higher-mode effects. The present study focuses on uncertainty in the beam-hinge properties governing drift-sensitive structural response.

Structural Properties

The beam hinges1 were modeled as rotational springs with a quadrilinear moment-rotation backbone that exhibits symmetry for both positive and negative rotations. The monotonic envelope of the hinge moment-rotation relationship exhibits hardening beyond a normalized yield moment aMya_{M_y}, with a non-negative slope aha_h extending up to a normalized rotation (or rotation ductility) μc\mu_c, where a negative stiffness segment begins. This drop, occurring at a slope aca_c, is halted by a residual plateau at a normalized height rr, which terminates abruptly at the ultimate rotational ductility μu\mu_u. In this study, we consider six normalized parameters defining the moment–rotation backbone,

xs=[aMy,ah,μc,ac,r,μu].(22)\mathbf{x}_{s} = [a_{My}, a_{h}, \mu_{c}, a_{c}, r, \mu_{u}]. \tag*{(22)}

The six parameter types are modeled as independent random variables following truncated Gaussian distributions to prevent non-physical values. Spatial variability among beam hinges is neglected; for each structural realization, the sampled parameter vector is assigned consistently to all corresponding beam hinges in the frame. Therefore, xs∈R6\mathbf{x}_{s} \in\mathbb{R}^{6}. To generate the initial data, we employed Latin Hypercube Sampling (LHS) [37] to draw samples from fxsf_{\mathbf{x}_{s}}. Figure 3 shows the 160 LHS realizations drawn from the truncated-normal parameter distributions summarized in Table 1. The different histogram shapes reflect the combined effects of the prescribed coefficients of variation, truncation bounds, and finite LHS sampling.

ParameterMeanCOV (%)Lower boundUpper boundDistribution
aMya_{M_y}1.0200.701.30Truncated normal
aha_h0.10400.040.16Truncated normal
μc\mu_c3.0401.204.80Truncated normal
aca_c-0.5040-0.80-0.20Truncated normal
rr0.50400.200.80Truncated normal
μu\mu_u6.0402.409.60Truncated normal

Table 1. Distribution properties of the uncertain structural parameters.

Histograms of the 160 realizations of the structural-parameter vector $x_s$

Figure 3. Histograms of the 160 realizations of the structural-parameter vector xsx_s generated using Latin hypercube sampling [37].

Structural Response

For the IDA calculations, a suite of 30 ordinary ground-motion records was used (Table 2). The records represent a relatively narrow earthquake scenario, with moment magnitudes between 6.5 and 6.9 and moderate source-to-site distances. The motions were recorded primarily on firm soil and were selected without strong directivity effects. For each record, IDA consists of repeated nonlinear response-history analyses in which the ground motion is progressively scaled over increasing intensity levels. The intensity measure (IM) is the

No.EventStationComponent (°)SoilMMRR (km)PGA (g)
1Loma Prieta, 1989Agnews State Hospital090C,D6.928.20.159
2Northridge, 1994LA, Baldwin Hills090B,B6.731.30.239
3Imperial Valley, 1979Compuertas285C,D6.532.60.147
4Imperial Valley, 1979Plaster City135C,D6.531.70.057
5Loma Prieta, 1989Hollister Diff. Array255–,D6.925.80.279
6San Fernando, 1971LA, Hollywood Stor. Lot180C,D6.621.20.174
7Loma Prieta, 1989Anderson Dam Downstrm270B,D6.921.40.244
8Loma Prieta, 1989Coyote Lake Dam Downstrm285B,D6.922.30.179
9Imperial Valley, 1979El Centro Array #12140C,D6.518.20.143
10Imperial Valley, 1979Cucapah085C,D6.523.60.309
11Northridge, 1994LA, Hollywood Storage FF360C,D6.725.50.358
12Loma Prieta, 1989Sunnyvale Colton Ave270C,D6.928.80.207
13Loma Prieta, 1989Anderson Dam Downstrm360B,D6.921.40.240
14Imperial Valley, 1979Chihuahua012C,D6.528.70.270
15Imperial Valley, 1979El Centro Array #13140C,D6.521.90.117
16Imperial Valley, 1979Westmoreland Fire Station090C,D6.515.10.074
17Loma Prieta, 1989Hollister South & Pine000–,D6.928.80.371
18Loma Prieta, 1989Sunnyvale Colton Ave360C,D6.928.80.209
19Superstition Hills, 1987Wildlife Liquefaction Array090C,D6.724.40.180
20Imperial Valley, 1979Chihuahua282C,D6.528.70.254
21Imperial Valley, 1979El Centro Array #13230C,D6.521.90.139
22Imperial Valley, 1979Westmoreland Fire Station180C,D6.515.10.110
23Loma Prieta, 1989Halls Valley090C,C6.931.60.103
24Loma Prieta, 1989WAHO000–,D6.916.90.370
25Superstition Hills, 1987Wildlife Liquefaction Array360C,D6.724.40.200
26Imperial Valley, 1979Compuertas015C,D6.532.60.186
27Imperial Valley, 1979Plaster City045C,D6.531.70.042
28Loma Prieta, 1989Hollister Diff. Array165–,D6.925.80.269
29San Fernando, 1971LA, Hollywood Stor. Lot090C,D6.621.20.210
30Loma Prieta, 1989WAHO090–,D6.916.90.638

Table 2. Suite of 30 ground-motion records used in the incremental dynamic analyses.

5%-damped first-mode spectral acceleration,

IM=Sa(T1,5%),(23)IM = S_a(T_1, 5\%), \tag*{(23)}

where T1=2.35T_1 = 2.35 s is the fundamental period of the reference LA9 structure. The corresponding engineering demand parameter (EDP) is the maximum interstory drift, EDP=Dmax⁡EDP = D_{\max}, consistent with the emphasis of the present study on deformation-sensitive structural and nonstructural response. Each record therefore defines an IDA trajectory relating Sa(T1,5%)S_a(T_1, 5\%) to Dmax⁡D_{\max}.

Component denotes the horizontal ground-motion component; Soil denotes the USGS/Geomatrix soil classification; MM is moment magnitude; RR is the closest distance to fault rupture; PGA is the unscaled peak ground acceleration. Adopted from [3].

The same suite of 30 ground-motion records was used for every structural realization, with Sa(T1,5%)S_a(T_1, 5\%) as the intensity measure and maximum interstory drift Dmax⁡D_{\max} as the engineering demand parameter. For each realization of xi\mathbf{x}^{i}, i=1,…,160i = 1,\ldots,160, the resulting 30 single-record IDAs were available from previous nonlinear analyses performed using the hunt-and-fill algorithm [38]. Each single-record IDA required approximately 12–14 nonlinear analyses, corresponding to approximately 360–420 nonlinear dynamic analyses per structural realization and approximately 57,600–67,200 analyses over the full 160-realization dataset. On a conventional laptop equipped with an 8-core Apple M1 Pro processor and 16 GB of unified memory, one complete 30-record adaptive IDA required on average approximately 42.3 min per structural realization, with individual runtimes ranging from 40.5 to 43.1 min. The 160 realization dataset corresponds to approximately 113 h of aggregate computation, or approximately 38 h of wall-clock time when three realizations are processed concurrently.

The present study constructs a fixed-dimensional response representation for each structural realization. Each record-level IDA curve is first converted to a first-passage monotone relation Sa(D)S_a(D) and interpolated onto the common 18-point drift grid. The 30 processed responses associated with a given structural realization are then aggregated pointwise by taking the median spectral acceleration at each drift coordinate. The resulting representative response is xr=S~a\mathbf{x}_r = \widetilde{\mathbf{S}}_a. Figure 4 illustrates the two sources of variability represented in the IDA dataset and the resulting median structural responses. Figure 4a fixes one structural realization and shows its responses to the 30 ground-motion records, thereby illustrating the record-to-record variability in the IDA response. The red curve denotes the pointwise median across the 30 records. Figure 4b instead fixes one ground-motion record and shows the responses of all 160 structural realizations, highlighting the variability induced by changes in the structural parameters; the dashed red curve denotes the pointwise median across these realizations. Finally, Figure 4c shows the representative median IDA curve associated with each of the 160 structural realizations, where each curve is obtained by taking the pointwise median across the 30 ground motion records.

Three plots illustrating record-to-record variability, structural variability, and median IDA responses

Figure 4. Illustration of the variability represented in the IDA dataset. (a) Record-to-record variability for a fixed structural realization; (b) structural-parameter variability for a fixed ground-motion record; and (c) representative median IDA responses across the full set of structural realizations.

Results

We evaluate the proposed framework using a nine-story steel moment-resisting frame subjected to earthquake ground motions (see § 2.3). Structural uncertainty is represented through six parameters describing the nonlinear strength and deformation characteristics of the structural components: the normalized yield moment (αMy\alpha M_y), strain-hardening ratio (αh\alpha_h), capping ductility (μc\mu_c), post-capping slope (αc\alpha_c), residual-strength ratio (rr), and ultimate ductility (μu\mu_u). For each structural realization, Incremental Dynamic Analysis (IDA) is performed using 30 earthquake ground-motion records (see Table 2), progressively scaled to characterize structural performance from the elastic to the highly nonlinear regime. The resulting Incremental Dynamic Analysis (IDA) curves describe the relationship between ground-motion intensity and structural deformation. Ground-motion intensity is quantified by the 5%-damped spectral acceleration Sa(T1,5%)S_a(T_1, 5\%), defined as the peak acceleration response of an equivalent linear single-degree-of-freedom oscillator with a natural period equal to the fundamental period T1T_1 of the structure and a damping ratio of 5%. Structural deformation is measured by the maximum interstory drift ratio Dmax⁡D_{\max}, defined as the maximum relative horizontal displacement between two consecutive floors, normalized by the corresponding story height, over the duration of the earthquake excitation. Each representative IDA curve is evaluated at 18 prescribed drift levels, providing a fixed-dimensional description of how the ground-motion intensity required to reach a given deformation level evolves from the elastic to the highly nonlinear structural response regime.

The dataset comprises 160 structural realizations, of which 151 are used for training and 9 are reserved for validation. The validation cases are selected using a maximin distance criterion in the standardized six-dimensional structural-parameter space to obtain a diverse set of parameter configurations. The Diffusion Maps algorithm (see Appendix § 5) is applied to the standardized response vectors to identify a compact nonlinear representation of the seismic-response manifold. This reduction exploits the intrinsic geometric structure of the IDA trajectories, allowing their dominant variability to be represented by a small number of non-redundant coordinates. In contrast, the six structural parameters are retained explicitly in the joint generative representation to preserve their physical meaning and enable direct conditioning on individual parameters. This hybrid representation reduces the dimensionality of the generative learning problem while supporting both forward probabilistic prediction and inverse inference of structural properties from observed seismic responses.

Identification and Validation of the Response Manifold

To identify a compact representation of the seismic-response manifold, Diffusion Maps are applied to the standardized 18-dimensional representative IDA responses. The method constructs a similarity kernel that measures the proximity of response trajectories, with a bandwidth parameter ϵr\epsilon_r controlling the scale over which neighboring responses are considered similar. Here, the bandwidth is selected as ϵr=10×median⁡i<j∥xr,i(std)−xr,j(std)∥2\epsilon_r = 10 \times\operatorname{median}_{i<j} \left\| \mathbf{x}_{r,i}^{(\mathrm{std})} - \mathbf{x}_{r,j}^{(\mathrm{std})} \right\|^2, where xr,i(std)\mathbf{x}_{r,i}^{(\mathrm{std})} denotes the standardized response vector for the iith training realization. This gives ϵr=172.72\epsilon_r = 172.72 for the present training dataset. The first 15 nontrivial diffusion eigenpairs are computed to examine the intrinsic structure of the response manifold. Because some diffusion eigenvectors may describe higher-order harmonics of coordinates already identified, retaining all of them would introduce redundant information. We therefore apply the local-regression criterion of [39], which evaluates whether each candidate diffusion coordinate can be locally approximated using the preceding coordinates. The resulting normalized regression residual rkr_k measures the extent to which the kkth coordinate contains information not explained by the previously considered coordinates. Larger residuals indicate potentially independent geometric directions, whereas smaller residuals suggest harmonic redundancy.

Figure 5a shows the normalized residuals for the evaluated diffusion coordinates. Six non-harmonic coordinates are retained (red bars), providing a compact representation of the dominant geometric variability in the IDA responses. The remaining coordinates are treated as predominantly redundant and are excluded from the subsequent generative learning procedure. The original parameter–response vector is x=(xs,xr)=(xs,S~a)∈R6+18=R24\mathbf{x} = (\mathbf{x}_s,\mathbf{x}_r) = (\mathbf{x}_s,\widetilde{\mathbf{S}}_a) \in\mathbb{R}^{6+18} = \mathbb{R}^{24}, where xs\mathbf{x}_s contains the six uncertain structural parameters and xr=S~a\mathbf{x}_r = \widetilde{\mathbf{S}}_a denotes the representative IDA response. Following dimensionality reduction, the joint generative state is defined as

Selection and validation of the response-manifold representation

Figure 5. (a) Normalized local-regression residuals rkr_k from the response-only Diffusion Maps analysis. Red bars indicate the six non-harmonic eigenvectors retained in the response representation. (b) Geometric-Harmonics reconstruction of the nine held-out representative IDA responses. Solid black curves denote the held-out median IDA responses and dashed red curves denote their reconstructions. Figure 5: Selection and validation of the response-manifold representation. (a) Non-harmonic Diffusion Maps coordinate selection using local-regression residuals. (b) Held-out reconstruction of the 18-dimensional representative IDA responses using the 6 retained coordinates and Geometric Harmonics.

z=(xs(std),ϕr)∈R6+6=R12,\mathbf{z} = \left(\mathbf{x}_s^{(\mathrm{std})}, \boldsymbol{\phi}_r\right) \in\mathbb{R}^{6+6} = \mathbb{R}^{12},

where ϕr\phi_r contains the six retained Diffusion Maps coordinates. This representation reduces the dimension of the generative learning problem while retaining the structural parameters explicitly for physical interpretation and conditional inference. To reconstruct the original seismic responses from the reduced coordinates, we employ Geometric Harmonics (GH) (see Section Appendix B), a kernel-based method that approximates the inverse mapping from the learned manifold coordinates to the original response space. GH is applied only to the response variables, lifting the six retained diffusion coordinates back to the 18-dimensional drift representation. The GH kernel bandwidth is set to 2.5 times the median squared pairwise distance in the response latent space.

The accuracy of this reduced representation is evaluated using the nine held-out structural realizations, which were not used to construct the response manifold or its reconstruction map. Figure 5b compares their original representative IDA curves with those reconstructed from the six retained coordinates. The overall root-mean-square error (RMSE) is 0.0106 g, with individual errors ranging from 0.0046 to 0.0147 g. Note that the RMSE is expressed in units of gravitational acceleration (gg), since the reconstruction error is evaluated in terms of spectral acceleration at the 18 prescribed interstory drift levels. These results indicate that the six-dimensional manifold representation preserves the principal characteristics of the seismic-response trajectories with limited reconstruction error, supporting its use in the subsequent joint generative learning procedure.

Physical Interpretation of the Manifold and Structural Parameter Recoverability

Having established that the six-dimensional Diffusion Maps representation accurately reconstructs the representative IDA responses, we next investigate the physical information encoded in the learned coordinates and their relationship to the uncertain structural parameters. Figure 6 shows pairwise projections of the retained Diffusion Maps coordinates, colored by the spectral acceleration SaS_a at selected maximum interstory drift levels. The response values vary smoothly across the embedding, indicating that nearby points tend to correspond to IDA trajectories with similar response characteristics. Moreover, the organization of spectral acceleration changes across drift levels, suggesting that the retained coordinates capture variations in the shape and evolution of the nonlinear IDA response rather than merely representing a single response amplitude. These observations provide qualitative evidence that the learned manifold preserves physically meaningful variations in seismic performance.

Pairwise projections of the retained response Diffusion Maps coordinates

Figure 6. Pairwise projections of the retained response Diffusion Maps coordinates, colored by the representative spectral acceleration at selected maximum interstory drift levels.

An important question is whether the learned response manifold contains sufficient information to infer the underlying structural properties from observed seismic performance. Parameter identifiability refers to the ability to uniquely infer the parameters of a model from observations of its behavior. More formally, identifiability concerns whether the mapping from model parameters to observable responses is injective, such that distinct parameter configurations produce distinct responses and an inverse mapping can be defined. When this condition is not satisfied, different parameter combinations may produce the same observable behavior, leading to structural non-identifiability. In the present setting, we consider the mapping from the six uncertain structural parameters, θ=(aMy,ah,μc,ac,r,μu)\theta= (a_{M_y}, a_h, \mu_c, a_c, r, \mu_u), to the six retained Diffusion Maps coordinates, ϕr=(ϕr,1,…,ϕr,6)\phi_r = (\phi_{r,1}, \ldots, \phi_{r,6}), which provide a compact representation of the seismic-response trajectories ϕr=F(θ)\phi_r = \mathcal{F}(\theta). From an analytical perspective, local identifiability can be investigated through the linearization of this mapping, characterized by its Jacobian,

J(θ)=∂F(θ)∂θ.\mathbf{J}(\theta) = \frac{\partial\mathcal{F}(\theta)}{\partial\theta}.

For a continuously differentiable mapping, a full-rank Jacobian provides a sufficient condition for local invertibility. Conversely, small singular values indicate parameter directions along which the observable response changes only weakly. Such directions may result in large parametric uncertainties and practical difficulties in distinguishing different structural configurations. Importantly, even when a mapping is locally invertible, this does not guarantee global uniqueness of the inverse solution. Beyond this local analytical perspective, manifold learning methods offer a data-driven approach to examining the relationship between model parameters and observable behavior. In particular, the intrinsic geometry of the response manifold may reveal that certain parameters, or combinations of parameters, are more strongly expressed in the observable responses than others. These identifiable combinations need not coincide with the individual physical parameters used to formulate the structural model.

Rather than explicitly computing the Jacobian or the Fisher Information Matrix [40, 41], we investigate parameter recoverability directly from the learned response manifold. We first examine how individual structural parameters vary across the Diffusion Maps embedding. We then attempt to reconstruct each parameter from the six retained diffusion coordinates using several nonlinear regression methods, including kernel ridge regression, support vector regression, random forests, extremely randomized trees, and gradient boosting. Predictive performance is evaluated using nested five-fold cross-validation, with hyperparameters selected exclusively within the inner training folds. The resulting out-of-sample coefficient of determination, R2R^2, quantifies how effectively each structural parameter can be recovered from the learned response representation. Higher R2R^2 values indicate stronger empirical recoverability, whereas lower values may reflect weak observable dependence, confounding between parameters, or limitations associated with finite data and regression accuracy. This analysis provides a data-driven assessment of parameter recoverability rather than a formal proof of structural identifiability. Importantly, it also motivates the investigation of whether physically meaningful combinations of backbone parameters may be more recoverable than the individual parameters themselves, as examined in the subsequent analysis of post-capping degradation.

Figure 7(a) illustrates the clearest relationship identified in this analysis. The capping rotation μc\mu_c varies systematically with the leading diffusion coordinate ϕr,1\phi_{r,1}: larger values of μc\mu_c are predominantly located at smaller values of ϕr,1\phi_{r,1}, whereas smaller values of μc\mu_c occur toward larger ϕr,1\phi_{r,1}. The points are additionally colored by SaS_a evaluated at Dmax⁡=0.20D_{\max}=0.20, revealing that the same region associated with larger μc\mu_c also corresponds to larger spectral acceleration. Thus, ϕr,1\phi_{r,1} is not simply an abstract latent coordinate, but organizes the IDA responses along a direction that is simultaneously related to a physically meaningful backbone parameter and to the intensity required to attain large drift. This interpretation is consistent with the role of μc\mu_c in controlling the transition from the hardening branch to post-capping degradation, a feature that can have a pronounced influence on the nonlinear IDA response.

Physical interpretation of the learned IDA-response manifold

Figure 7. Physical interpretation of the learned IDA-response manifold. Panel (a) shows that the leading diffusion coordinate is strongly associated with the capping rotation μc\mu_c and with the spectral intensity required to reach large drift. Panel (b) quantifies the recoverability of the six backbone parameters and shows that μc\mu_c is the parameter most strongly encoded in the learned manifold.

The quantitative reconstruction results in Figure 7(b) support this interpretation. Across the different nonlinear regressors, μc\mu_c was consistently the most recoverable structural parameter, with cross-validated R2R^2 values of approximately 0.72–0.75. In contrast, aMyaM_y and aca_c showed only modest recoverability, with best R2R^2 values of approximately 0.26 and 0.25, respectively, while aha_h, rr, and μu\mu_u produced R2R^2 values close to zero. The consistency of this ranking across different regression models indicates that their effects appear to be weakly expressed or non-unique within the median IDA responses used to construct the manifold.

An interesting result is that the post-capping slope parameter aca_c is more strongly recoverable from the learned response coordinates than either the hardening parameter aha_h or the residual-strength parameter rr. This ranking should not be interpreted as a conventional forward-sensitivity measure of the structural model. Rather, it quantifies how distinctly variations in each parameter are expressed in the median IDA responses used to construct the manifold. The stronger recoverability of aca_c suggests that post-capping degradation leaves a more distinguishable signature over the response range sampled by the IDA curves, whereas the effects of aha_h and rr may be more localized or partially confounded with variations in other parameters. The result therefore highlights the distinction between physical influence on the nonlinear system and identifiability from the selected response observable.

Generative learning of the compact state space

The score-based diffusion modeling framework (see § 2.1) was applied directly to the 12-dimensional joint generative state formed by the six standardized structural parameters and the six retained Diffusion Maps response coordinates. The probability-flow Ordinary Differential Equation (ODE) associated with the forward diffusion process was integrated backward from t=1t=1 to t=0t = 0 using 1000 uniform explicit-Euler steps. At each step, the empirical score was evaluated using the fixed training reference ensemble. A total of 10,000 standard normal samples were transported through this probability-flow ODE. Generated realizations were retained when all six structural parameters remained within their prescribed bounds and all response variables were finite and non-negative. This produced 9,850 admissible realizations.

To assess the accuracy of the learned joint distribution, we compare the generated structural parameters and seismic responses with those in the original simulation dataset. In addition to examining their marginal distributions, we investigate whether the generated samples reproduce the dependence between structural properties and seismic performance. We then evaluate the model’s ability to generate representative IDA responses consistent with the held-out validation cases and to capture the changes in response distributions induced by conditioning on specific structural parameters. Figure 8 shows that the generated samples reproduce the training marginals closely for all six structural parameters. As an additional diagnostic, particular attention was given to the lower tail of the capping ductility parameter μc\mu_c, which was found during model development to be especially sensitive to generative smoothing. The 5th percentile is 1.378 in the training ensemble and 1.364 in the admissible generated ensemble. The corresponding lower-tail probabilities at μc=1.50\mu_c = 1.50, 1.60, 1.80, and 2.00 are, respectively, 0.086, 0.106, 0.166, and 0.212 for the training data and 0.090, 0.111, 0.168, and 0.216 for the generated data. These results indicate that the generative model preserves this sensitive lower-tail behavior closely.

Marginal distributions of the six structural parameters

Figure 8. Marginal distributions of the six structural parameters. Solid curves correspond to the training samples, dashed curves correspond to the admissible generated samples, and vertical marks denote the nine held-out validation cases.

The generated response ensemble is summarized in Figure 9a, where gray curves denote generated median IDA responses, the dashed red line denotes their mean, and the black solid line denotes the mean response of the training structural realizations. The close agreement between the two mean curves indicates that the generative model reproduces the central trend of the training response ensemble. Figure 9b provides an additional held-out validation by comparing the nine validation IDA curves with the unconditional 90% envelope of the generated response distribution. Across all held-out response ordinates, 92.0% lie within this envelope. One held-out response extends partially beyond the envelope, which is not unexpected because the shaded region represents the central 90% of the unconditional generated distribution rather than a calibrated case-specific prediction interval. Realizations associated with less probable combinations of structural properties may therefore fall outside this range.

Figure 10 compares the probability densities of the 5%-damped first-mode spectral acceleration Sa(T1,5%)S_a(T_1, 5\%) at Dmax⁡=0.05D_{\max} = 0.05, 0.100.10, and 0.200.20. The generated and reference densities remain close across the selected response regimes, including the increasingly non-Gaussian structure observed at larger drift. To assess whether the generated samples preserve the joint dependence between structural properties and seismic response, we compared the parameter–response cross-correlation matrices of the training and generated ensembles. The generated cross-correlation structure closely reproduces that of the training data, with a relative Frobenius error of 0.0696, a mean absolute entry-wise error of 0.0156, and a correlation coefficient of 0.9965 between the vectorized training and generated cross-correlation matrices. These results indicate that the generative model preserves the dominant first-order parameter–response dependence in addition to matching the marginal distributions.

Probability densities of spectral acceleration at three values of maximum interstory drift

Figure 10. Probability densities of the 5%-damped first-mode spectral acceleration Sa(T1,5%)S_a(T_1, 5\%) at three values of the maximum interstory drift Dmax⁡D_{\max}. Solid lines correspond to the training distributions, while dashed lines correspond to the admissible generated samples.

Forward Conditional Sampling

An important advantage of learning the joint probability distribution fxs,xr(⋅)f_{x_s,x_r}(\cdot) is that the resulting generative representation can be used for conditional inference, rather than being used only for unconditional generation of synthetic structural-response realizations. In particular, the joint model enables forward conditional inference of the form

Generated representative IDA responses

(a) Generated representative IDA responses. Gray lines denote 1,600 admissible generated responses, the red dashed line denotes their mean, and the black solid line denotes the mean response of the training structural realizations.

Nine held-out representative IDA responses

(b) Nine held-out representative IDA responses together with the 5–95% interval of the admissible generated response distribution.

Figure 9. Unconditional generative performance and held-out validation. (a) Representative IDA responses sampled from the learned generative model compared with the training mean response. (b) Held-out representative IDA responses relative to the generated 5–95% response envelope.

fxr∣xs(xr∣xs),(24)f_{x_r\mid x_s}(x_r \mid x_s), \tag*{(24)}

thereby allowing the distribution of seismic response to be evaluated for prescribed structural properties without requiring additional nonlinear responsehistory analyses.

Because conditioning on the complete structural-parameter vector xs\mathbf{x}_s would substantially restrict the remaining response variability, we demonstrate the conditional capability by prescribing a single physically meaningful structural parameter while allowing the remaining uncertain parameters to vary. Specifically, we consider the capping ductility parameter μc\mu_c and evaluate

fxr∣μc(xr∣μc=c),(25)f_{\mathbf{x}_r\mid\mu_c}\left(\mathbf{x}_r \mid\mu_c = c\right), \tag*{(25)}

for representative low, nominal, and high values of μc\mu_c. The selected conditioning values, μc=1.924\mu_c = 1.924, μc=3.068\mu_c = 3.068, μc=4.114\mu_c = 4.114, correspond to low, nominal, and high regions of the sampled parameter distribution, respectively.

Conditional statistics are estimated directly from samples generated by the learned joint model. For a target conditioning value cc, each generated realization is assigned a Gaussian kernel weight

wi(c)=exp⁡[−12(μc,i−chc)2],(26)w_i(c) = \exp\left[-\frac{1}{2}\left(\frac{\mu_{c,i}-c}{h_c}\right)^2\right], \tag*{(26)}

where hch_c denotes the conditioning bandwidth. In the present analysis, hc=0.20×std⁡train(μc)=0.2085h_c = 0.20 \times\operatorname{std}_{\mathrm{train}}(\mu_c) = 0.2085. The normalized weights are then used to estimate the conditional median and the corresponding response quantiles at each value of maximum interstory drift Dmax⁡D_{\max}. The effective sample size associated with each conditional estimate is computed as

Neff=(∑i=1N∗wi)2∑i=1N∗wi2,(27)N_{\mathrm{eff}} = \frac{\left(\sum_{i=1}^{N^*} w_i\right)^2}{\sum_{i=1}^{N^*} w_i^2}, \tag*{(27)}

providing a measure of the number of generated samples that contribute effectively to the conditional estimate.

Figure 11 presents the conditional representative IDA responses together with the corresponding 5–95% intervals. The unconditional median response is included as a reference. A clear systematic dependence on μc\mu_c is observed. At relatively small drift levels, the conditional responses remain close to one another, indicating that the influence of the capping ductility parameter is limited in the predominantly elastic response regime. As Dmax⁡D_{\max} increases and the response enters the nonlinear range, the conditional distributions separate progressively. Larger values of μc\mu_c are associated with higher spectral-acceleration capacity, whereas smaller values of μc\mu_c produce lower conditional median responses. The 5–95% intervals reflect variability in the learned response induced by the remaining five uncertain structural parameters; they do not represent record-to-record ground-motion variability.

Forward conditional sampling of the representative IDA response

Figure 11. Forward conditional sampling of the representative IDA response for low (μc=1.924\mu_c = 1.924), nominal (μc=3.068\mu_c = 3.068), and high (μc=4.114\mu_c = 4.114) values of the capping ductility parameter. Solid colored lines denote the conditional median responses, shaded regions denote the corresponding 5–95% intervals, and the dashed black line denotes the unconditional median response. Conditional statistics are evaluated from samples of the learned joint distribution without additional nonlinear response-history analyses.

These results illustrate a capability that is not directly available from a conventional deterministic surrogate model. Once the joint distribution fxs,xr(⋅)f_{x_s,x_r}(\cdot) has been learned, the model can be queried to characterize response distributions associated with prescribed structural-property values while accounting for variability in the remaining uncertain parameters. In this way, the generative representation provides a computationally efficient mechanism for probabilistic interrogation of the structural parameter–response relationship across the seismic-performance range.

Inverse Conditional Sampling

The learned joint distribution can also be inferred in the inverse direction, by conditioning the structural parameters on an observed response state. We consider

fxs∣Sa(xs∣Sa(Dmax⁡=0.10)=0.35 g).(28)f_{x_s \mid S_a}\left(x_s \mid S_a(D_{\max} = 0.10) = 0.35\ \mathrm{g}\right). \tag*{(28)}

For each generated realization, the Gaussian conditioning weights are

wi=exp⁡[−12(Sa,i(0.10)−0.35hSa)2],(29)w_i = \exp\left[-\frac{1}{2}\left(\frac{S_{a,i}(0.10)-0.35}{h_{S_a}}\right)^2\right], \tag*{(29)}

with hSa=0.10 std [Sa(Dmax⁡=0.10)]=0.01414 gh_{S_a} = 0.10\,\mathrm{std}\,[S_a(D_{\max} = 0.10)] = 0.01414\,\mathrm{g}. The corresponding effective sample size is Neff=138N_{\mathrm{eff}} = 138. Figure 12 compares the unconditional structural-parameter distributions with the response-conditioned distributions. Conditioning on the single response state produces substantial changes in several parameters, including strong concentration toward low μc\mu_c and high μu\mu_u, together with multimodal structure in a number of the remaining parameters.

Unconditional and response-conditioned structural-parameter distributions

Figure 12. Unconditional and response-conditioned structural-parameter distributions for the observation Sa=0.35 gS_a = 0.35\,\mathrm{g} at Dmax⁡=0.10D_{\max} = 0.10.

The multimodal one-dimensional conditional distributions indicate that the inverse problem does not admit a unique structural-parameter estimate. A single response ordinate constrains six uncertain structural variables, and distinct combinations of these variables may therefore remain compatible with the same observed performance state. To examine whether the apparent multimodality reflects coherent structure in the full parameter space, the response-conditioned samples were further analyzed in the six-dimensional space θ=[aMy,ah,μc,ac,r,μu]\boldsymbol{\theta} = [a_{M_y}, a_h, \mu_c, a_c, r, \mu_u].

The six-dimensional conditional sample cloud was standardized and represented using a Gaussian mixture model, with the number of mixture components selected by the Bayesian information criterion (BIC) over the candidate models considered. Several mixture components occupied closely related regions of parameter space and were therefore grouped into two dominant posterior regimes, while the remaining low-weight components were classified as secondary modes. Together, the two dominant regimes contained approximately 95% of the conditional probability mass. The resulting joint structure is illustrated through representative two-dimensional projections in Figure 14. The two dominant regimes remain clearly separated across these projections, together with several smaller secondary regions, and correspond to substantially different combinations of aMyaM_y, aca_c, rr, and μu\mu_u, while μc\mu_c remains comparatively similar between them.

Joint posterior regimes identified in the six-dimensional parameter space

Figure 14. Representative projections of the joint response-conditioned distribution for Sa=0.35S_a = 0.35 g at Dmax⁡=0.10D_{\max} = 0.10. Colors denote posterior regimes identified in the full six-dimensional parameter space. Two dominant modes contain the majority of the posterior probability mass, while the remaining disconnected regions correspond to lower-probability secondary modes.

The joint posterior structure can be interpreted more directly by referring to the post-capping branch of the backbone in Figure 2. Let McM_c denote the normalized moment at the capping point. For a realization with parameters (aMy,ah,μc,ac,r,μu)(aM_y, a_h, \mu_c, a_c, r, \mu_u),

Figure 2. (a) The LA9 steel moment-resisting frame and (b) The moment-rotation beam-hinge backbone (adopted from [19])

Mc=aMy+ah(μc−1).(30)M_c = aM_y + a_h(\mu_c - 1). \tag*{(30)}

The degrading branch reaches the residual-strength plateau rr at a rotation μr\mu_r satisfying

Mc+ac(μr−μc)=r,(31)M_c + a_c(\mu_r - \mu_c) = r, \tag*{(31)}

and therefore

μr=μc+r−Mcac.(32)\mu_r = \mu_c + \frac{r - M_c}{a_c}. \tag*{(32)}

The deformation interval associated with post-capping degradation is then

Δμdeg=μr−μc,(33)\Delta\mu_{\mathrm{deg}} = \mu_r - \mu_c, \tag*{(33)}

while the area under the degrading branch is

Adeg=Mc+r2Δμdeg.(34)A_{\mathrm{deg}} = \frac{M_c + r}{2}\Delta\mu_{\mathrm{deg}}. \tag*{(34)}

We also define the area above the residual plateau,

Aexcess=Mc−r2Δμdeg,(35)A_{\mathrm{excess}} = \frac{M_c - r}{2}\Delta\mu_{\mathrm{deg}}, \tag*{(35)}

which isolates the additional resistance carried during degradation above the residual level.

When these quantities are evaluated over the response-conditioned posterior, the two dominant regimes remain clearly distinct in terms of post-capping capacity. Primary mode I has a median Adeg≈0.634A_{\mathrm{deg}} \approx0.634, whereas Primary mode II has Adeg≈0.855A_{\mathrm{deg}} \approx0.855. Thus, the same conditioning state, (Sa,Dmax⁡)=(0.35 g,0.10)(S_a, D_{\max}) = (0.35\ \mathrm{g}, 0.10), is compatible with two distinct ranges of integrated post-capping resistance. At the same time, the rotations at which the degrading branch reaches the residual plateau are nearly identical, with median values μr≈2.46\mu_r \approx2.46 and μr≈2.47\mu_r \approx2.47 for Primary modes I and II, respectively. The corresponding degradation intervals are also similar, Δμdeg≈1.01\Delta\mu_{\mathrm{deg}} \approx1.01 and 0.98. These results suggest that the same observed maximum-drift state can arise from different post-capping strength-capacity regimes while preserving a similar deformation extent over which degradation develops.

To examine which aspects of this post-capping behavior are actually encoded in the learned response manifold, the same nested five-fold cross-validation procedure used for the individual backbone parameters was applied to μr\mu_r, Δμdeg\Delta\mu_{\mathrm{deg}}, AdegA_{\mathrm{deg}}, and AexcessA_{\mathrm{excess}}. The results are summarized in Figure 13. The rotation μr\mu_r is by far the most recoverable quantity, with R2≈0.865R^2 \approx0.865. The integrated post-capping quantities are also strongly recoverable, with R2≈0.640R^2 \approx0.640 for AdegA_{\mathrm{deg}} and R2≈0.569R^2 \approx0.569 for AexcessA_{\mathrm{excess}}, while Δμdeg\Delta\mu_{\mathrm{deg}} gives R2≈0.513R^2 \approx0.513. By comparison, the individual post-capping slope aca_c has only modest recoverability (R2≈0.245R^2 \approx0.245), and the residual-strength parameter rr is essentially unrecoverable on its own (R2≈−0.061R^2 \approx-0.061).

Bar chart showing recoverability of post-capping backbone quantities

Figure 13. Recoverability of individual and derived post-capping backbone quantities from the six retained response Diffusion Maps coordinates. The derived rotation μr\mu_r, corresponding to the point at which the degrading branch reaches the residual-strength plateau, is the most strongly encoded quantity. The integrated post-capping measures AdegA_{\mathrm{deg}} and AexcessA_{\mathrm{excess}} and the degradation interval Δμdeg\Delta\mu_{\mathrm{deg}} are also substantially more recoverable than the individual parameters aca_c and rr. The reported values are out-of-sample coefficients of determination obtained using the same nested cross-validation procedure as in Figure 7.

Taken together, these results indicate that the response manifold is not primarily encoding the residual-strength level rr as an isolated parameter. Instead, the maximum-drift IDA response is substantially more informative about the location and evolution of post-capping degradation and about integrated measures of the resistance retained during this process. In particular, the strong recoverability of μr\mu_r suggests that the observable response is sensitive to when the degrading branch reaches the residual regime, whereas the precise height of the idealized residual plateau is only weakly identifiable. This also explains how rr can participate in the joint posterior structure despite being individually unrecoverable: different combinations of degradation slope and residual strength can produce distinct post-capping capacity profiles while remaining compatible with the same global response state.

This observation is consistent with the stronger recoverability of μc\mu_c found previously and suggests that the observable IDA response constrains some parameter directions more strongly than others. The joint posterior structure also helps reconcile the parameter-recoverability results of Figure 7 with the multimodal inverse distributions. In particular, the bimodality observed in aca_c is accompanied by a systematic shift in the residual-strength parameter rr between the two dominant posterior regimes. Thus, different combinations of post-capping slope and residual strength can produce comparable response at the conditioning state. This does not contradict the low individual recoverability of rr in Figure 7. The recoverability analysis asks whether rr can be inferred uniquely from the response manifold, whereas the joint inverse analysis reveals that rr participates in compensating parameter combinations. Hence, rr may be weakly identifiable individually while still playing an important role in the joint conditional structure.

Referring to the backbone response in Figure 2(b), aca_c controls the rate of post-capping strength degradation, whereas rr defines the residual-strength plateau reached after substantial degradation. The stronger recoverability of aca_c therefore suggests that the maximum-drift IDA response is more informative about the onset and progression of post-capping degradation than about the precise residual strength attained at later stages. This is physically plausible because rr represents an idealized description of severely degraded behavior and may contribute relatively little additional information to the global maximum interstory drift response. At the same time, the joint posterior indicates that changes in aca_c are accompanied by systematic changes in rr, so that different combinations of degradation rate and residual strength can produce similar observable IDA response. This provides a physical interpretation of the inverse non-uniqueness: rr is weakly identifiable on its own, but variations in residual strength can compensate for changes in the better-constrained post-capping slope.

The separation of the posterior modes alone does not establish that all of these regions correspond to physically admissible structural states, since some disconnected branches could arise from the generative approximation. To assess this directly, five samples were selected from the conditional distribution: two representative realizations from each dominant mode and one realization from a secondary low-probability mode. Each parameter vector was then propagated independently through the original OpenSees model using the same 30-record adaptive IDA procedure employed to construct the training dataset. The resulting IDA curves were postprocessed using the same common-drift interpolation and record-wise median procedure as the original data. Figure 15 compares the resulting median IDA curves with the conditioning point (Dmax⁡,Sa)=(0.10,0.35(D_{\max}, S_a) = (0.10, 0.35 g). The two realizations drawn from Primary mode II give Sa(Dmax⁡=0.10)=0.347S_a(D_{\max} = 0.10) = 0.347 and 0.3660.366 g, corresponding to relative errors of approximately 0.85%0.85\% and 4.44%4.44\%, respectively. The two realizations from Primary mode I give 0.4100.410 and 0.3290.329 g, showing larger within-mode variability but remaining in the vicinity of the conditioning response. In contrast, the representative secondary-mode realization gives Sa(Dmax⁡=0.10)=0.465S_a(D_{\max} = 0.10) = 0.465 g, approximately 33%33\% above the conditioning value. These forward-model evaluations provide evidence that the two dominant disconnected regions of the conditional distribution contain distinct structural parameter combinations capable of producing comparable observable IDA behavior. In this sense, the multimodality is consistent with non-uniqueness of the inverse mapping: substantially different parameter vectors can correspond to similar response states. The poorer agreement of the tested secondary mode further suggests that not all low-probability branches are equally consistent with the original forward model. Because only a small number of representative realizations are examined here, these results are interpreted as a validation of the existence of alternative admissible parameter regimes.

Forward-model validation of representative parameter realizations

Figure 15. Forward-model validation of representative parameter realizations selected from the response-conditioned posterior at Sa=0.35S_a = 0.35 g and Dmax⁡=0.10D_{\max} = 0.10. Each curve is the median IDA response over the same 30 ground-motion records used in the original dataset. The two dominant posterior modes contain realizations that reproduce the conditioning response with varying accuracy, whereas the tested secondary low-probability mode exhibits a larger deviation. The black marker denotes the conditioning point.

The inverse results were also examined for sensitivity to the conditioning bandwidth over the range hSa/σSa=0.075h_{S_a}/\sigma_{S_a} = 0.075–0.200.20. The dominant concentration toward low μc\mu_c and high μu\mu_u remains, while some marginal modes and quantiles broaden as the bandwidth increases. Taken together with the multimodal joint structure and the forward-model validation above, these results support interpreting the inverse solution as a response-conditioned probability distribution over admissible structural states rather than as a unique parameter estimate.

Discussion

The results demonstrate that manifold-based generative learning can complement high-fidelity nonlinear earthquake simulations by providing a probabilistic representation of the relationship between uncertain structural properties and seismic performance. Rather than approximating a deterministic mapping from structural parameters to selected response quantities, the proposed framework learns their joint probability distribution. This distinction enables the generation of additional parameter–response realizations and supports probabilistic inference in both directions: predicting seismic response for prescribed structural properties and inferring plausible structural configurations from observed performance. By combining nonlinear manifold learning, score-based diffusion modeling, and Geometric Harmonics, these capabilities are obtained from a limited ensemble of nonlinear structural simulations, without requiring additional nonlinear analyses during generation.

The numerical results for the nine-story steel moment-resisting frame demonstrate that the learned representation reproduces important characteristics of the original simulation ensemble. The six-dimensional response manifold reconstructs the nine held-out representative IDA curves with an overall root-mean-square error of 0.0106 g. Furthermore, 92.0% of the held-out responses fall within the unconditional 5–95% envelope of the generated response distribution. The generative model also reproduces the marginal distributions of the structural parameters, including the lower tail of the capping ductility parameter μc\mu_c, and closely preserves the parameter–response cross-correlation structure. These results indicate that the learned model captures important features of both the variability in seismic performance and its dependence on uncertain structural properties. Nevertheless, agreement in marginal distributions and low-order dependence measures does not guarantee that all aspects of the joint distribution are accurately represented, particularly in regions of parameter space sparsely covered by the training data.

An important finding concerns the information about structural properties that can be recovered from the representative IDA responses. Although the response manifold provides an accurate low-dimensional description of seismic performance, its coordinates do not encode all structural parameters equally. The capping ductility parameter μc\mu_c is the most recoverable individual parameter, whereas the post-capping slope aca_c is only modestly recoverable and the residual-strength parameter rr is essentially unrecoverable from the selected response representation. These differences should not be interpreted as conventional forward sensitivities of the structural model. Rather, they characterize how distinctly variations in the structural parameters are expressed in the observable IDA responses. More importantly, the results show that physically meaningful combinations of structural parameters can be substantially more recoverable than the individual parameters themselves. The rotation μr\mu_r at which the degrading backbone reaches the residual-strength plateau achieves an out-of-sample coefficient of determination of R2≈0.865R^2 \approx0.865. The degradation interval Δμdeg\Delta\mu_{\mathrm{deg}} and the integrated post-capping resistance measures AdegA_{\mathrm{deg}} and AexcessA_{\mathrm{excess}} are also more strongly recoverable than the individual post-capping slope and residual-strength parameters. This suggests that the representative IDA relation Sa(Dmax⁡)S_a(D_{\max}) contains more information about the location, extent, and integrated resistance characteristics of post-capping degradation than about the precise values of the parameters defining the degrading branch. Consequently, poor recoverability of an individual parameter does not necessarily imply that the associated physical mechanism has little influence on seismic performance. Instead, the observable response may constrain particular combinations of parameters while remaining relatively insensitive to their individual values.

This distinction is central to the interpretation of structural identifiability. From a mathematical perspective, identifiability concerns whether distinct parameter configurations produce distinguishable model responses. The present recoverability analysis does not establish formal structural identifiability, but it provides evidence that the information available in the selected seismic-performance representation is distributed unevenly among the physical parameters and their combinations. In particular, the greater recoverability of derived post-capping quantities suggests that alternative parameterizations based on observable characteristics of the degrading backbone may be more informative for structural model updating than the original backbone parameters considered individually.

The conditional inference results further illustrate the implications of learning the joint distribution. Forward conditioning on μc\mu_c produces systematic changes in the representative IDA response. At relatively small drift levels, the conditional responses remain close, whereas differences become more pronounced as the structure enters the nonlinear regime. Larger capping ductility is associated with increased spectral-acceleration capacity at large drift, consistent with the physical role of μc\mu_c in delaying the onset of post-capping degradation. Thus, the learned joint distribution captures physically meaningful relationships between structural properties and seismic performance while accounting for variability in the remaining uncertain parameters.

Inverse conditioning provides a complementary perspective by revealing the ambiguity of structural parameter inference from observed performance. For the conditioning state (Sa,Dmax⁡)=(0.35 g,0.10)(S_a, D_{\max}) = (0.35\ \mathrm{g}, 0.10), the inferred parameter distribution contains two dominant, separated regions in the six-dimensional structural-parameter space. These regions correspond to substantially different combinations of post-capping slope and residual strength, yet both remain compatible with the prescribed response state. Interestingly, they exhibit different levels of integrated post-capping resistance while reaching the residual-strength plateau at nearly the same rotation and over similar degradation intervals. This demonstrates that a single IDA performance state may constrain certain characteristics of post-capping behavior without uniquely determining the underlying strength-degradation mechanism.

To further assess the inferred parameter configurations, representative samples from the conditional distribution were evaluated using the numerical model. Samples from both dominant modes produced IDA responses close to the prescribed conditioning state, although the agreement varied among realizations. In contrast, the sample from the secondary mode exhibited a larger deviation from the target response. These results provide independent evidence that different combinations of structural parameters can produce similar seismic performance, while also revealing discrepancies between the learned distribution and the original nonlinear model. The latter highlights the need for physics-based validation of generated parameter configurations, particularly when interpreting low-probability regions of the conditional distribution.

The choice of generative representation also has important methodological implications. By applying Diffusion Maps only to the seismic responses, the framework reduces the dimensionality of the response trajectories while retaining the six structural parameters explicitly. Consequently, the structural parameters can be sampled and conditioned upon directly, without requiring their reconstruction from the learned response manifold. This distinction is particularly relevant when the mapping from structural properties to observable performance is non-unique. Nevertheless, the reliability of the generated responses remains dependent on the accuracy and support of the learned manifold and its reconstruction through Geometric Harmonics.

Several limitations should be considered when interpreting these findings. First, the analysis is restricted to a single nine-story steel moment-resisting frame and six uncertain nonlinear backbone parameters. Although this benchmark provides a physically meaningful setting for evaluating the methodology, the findings cannot automatically be generalized to other structural systems, modeling assumptions, or earthquake environments. Second, the generative model is learned from a relatively small dataset, which may limit the accuracy of the estimated joint distribution in sparsely sampled regions. Third, the inverse conditional distributions depend on the conditioning bandwidth, and the independent forward-model validation of the identified conditional regimes is based on a limited number of representative parameter configurations. More extensive validation is therefore required to establish the reliability of the inferred distributions throughout parameter space. An additional limitation concerns the representation of ground-motion uncertainty. Although each representative IDA response is constructed from 30 ground-motion records, record-to-record variability is aggregated through the pointwise median rather than represented explicitly in the learned generative state. The resulting model therefore characterizes uncertainty in structural properties and their representative seismic-performance responses, but does not provide a complete probabilistic description of both structural uncertainty and ground-motion variability. In particular, the conditional response intervals should not be interpreted as capturing the full variability of individual earthquake response histories.

Future developments should extend the joint generative representation to include ground-motion descriptors and record-level responses, allowing a more complete characterization of the interaction between structural uncertainty and earthquake excitation. A further direction is the systematic investigation of structural parameter identifiability using complementary analytical and data-driven approaches. In particular, Jacobian-based sensitivity analysis and the associated Fisher information matrix could help identify poorly constrained parameter directions and establish how the identifiable combinations inferred from the response manifold relate to the underlying structural model. Incorporating additional engineering demand measures and response observations could further improve parameter recoverability and help distinguish structural configurations that produce similar maximum-interstory-drift responses. Finally, embedding physical constraints into the generative process could reduce inadmissible samples, while validation across different structural typologies and larger independent datasets would help establish the generalizability and reliability of the proposed framework.

Overall, this study demonstrates that manifold-based generative learning can extend conventional seismic surrogate modeling from forward response prediction to probabilistic characterization of structural behavior and inverse inference. Beyond reproducing the distribution of seismic responses, the framework reveals which aspects of nonlinear structural behavior are recoverable from the selected performance observables and which remain ambiguous. These capabilities provide a basis for uncertainty-aware structural model updating and seismic performance assessment, while emphasizing that generative predictions must remain grounded in the information content of the available observations and validated against the underlying physics-based simulations.

Diffusion Maps

Diffusion Maps for Dimensionality Reduction

Diffusion Maps [26] is a nonlinear dimensionality reduction technique that captures the intrinsic geometry of a dataset X={x1,x2,…,xN}\mathcal{X} = \{\mathbf{x}_1,\mathbf{x}_2,\ldots,\mathbf{x}_N\} by constructing a local similarity graph among the data points. A kernel-based affinity matrix K∈RN×N\mathbf{K} \in\mathbb{R}^{N \times N} is defined using the following diffusion kernel:

K(xi,xj)=exp⁡(−∥xi−xj∥22ϵ),(36)K(\mathbf{x}_i,\mathbf{x}_j) = \exp\left(-\frac{\lVert\mathbf{x}_i-\mathbf{x}_j\rVert^2}{2\epsilon}\right), \tag*{(36)}

where ϵ\epsilon is a kernel bandwidth parameter [42]. Data points xi\mathbf{x}_i and xj\mathbf{x}_j are considered connected if their squared distance is less than ϵ\epsilon. To correct for sampling density and recover the Laplace-Beltrami operator Δ\Delta on the underlying manifold M\mathcal{M}, normalization is applied:

K~=P−αKP−α,(37)\widetilde{\mathbf{K}} = \mathbf{P}^{-\alpha}\mathbf{K}\mathbf{P}^{-\alpha}, \tag*{(37)}

where P\mathbf{P} is a diagonal matrix with Pii=∑j=1NKijP_{ii} = \sum_{j=1}^{N} K_{ij}, and α\alpha adjusts the influence of sampling density. A further normalization produces a row-stochastic matrix:

M=D−1K~,Dii=∑j=1NK~ij.(38)\mathbf{M} = \mathbf{D}^{-1}\widetilde{\mathbf{K}}, \qquad D_{ii} = \sum_{j=1}^{N}\widetilde{K}_{ij}. \tag*{(38)}

The eigendecomposition of M\mathbf{M} yields eigenvectors ϕi\phi_i and eigenvalues λi\lambda_i via:

Mϕi=λiϕi.(39)\mathbf{M}\phi_i = \lambda_i\phi_i. \tag*{(39)}

A reduced representation is constructed from non-harmonic eigenvectors, excluding those that are harmonic or redundant. These selected eigenvectors Φ={ϕi1,…,ϕik}\mathbf{\Phi} = \{\phi_{i_1}, \ldots, \phi_{i_k}\} provide a parsimonious parametrization of the data when k≪dk \ll d.

To assess the sensitivity of the response representation to the Diffusion Maps kernel scale, we considered

ϵr=c×median⁡i<j∥xr,i(std)−xr,j(std)∥2,c∈{0.5,1,2,5,10}.\epsilon_r = c \times\operatorname{median}_{i<j} \left\| \mathbf{x}_{r,i}^{(\mathrm{std})} - \mathbf{x}_{r,j}^{(\mathrm{std})} \right\|^2, \qquad c \in\{0.5, 1, 2, 5, 10\}.

For each value of cc, six non-harmonic coordinates were retained and the response was reconstructed using Geometric Harmonics. The corresponding held-out reconstruction RMSEs were 0.0254, 0.0196, 0.0235, 0.0159, and 0.0106 g, respectively. Thus, the reconstruction quality improves for the broader kernels considered here, with the best performance obtained for c=10c = 10. Because the selected value corresponds to a kernel (Kij≈0.95K_{ij} \approx0.95 at the median pairwise distance), we also compared the 6-dimensional Diffusion Maps representation with a 6-component PCA baseline. PCA yielded a held-out reconstruction RMSE of 0.0104 g, which is comparable to the 0.0106 g obtained with Diffusion Maps. The present results therefore do not indicate a clear reconstruction advantage of the nonlinear embedding over PCA. Rather, Diffusion Maps provides a compact nonlinear coordinate system that is used within the manifold-based generative framework, while the subsequent generative ablation evaluates whether this representation provides advantages beyond reconstruction alone.

Diffusion Maps for Function Lifting (Double Diffusion Maps)

The Double Diffusion Maps method [30] extends the use of Diffusion Maps to both dimensionality reduction and function extension. Starting from a reduced set of non-harmonic coordinates Φ\mathbf{\Phi}, a new basis Ψ\mathbf{\Psi}, known as Latent Harmonics, is constructed by applying a second Diffusion Maps algorithm:

K∗(ϕi,ϕj)=exp⁡(−∥ϕi−ϕj∥22ϵ2).(40)K^*(\phi_i, \phi_j) = \exp\left(-\frac{\lVert\phi_i - \phi_j \rVert^2}{2\epsilon_2}\right). \tag*{(40)}

The eigenvectors Ψ={ψ0,…,ψm−1}\mathbf{\Psi} = \{\psi_0, \ldots, \psi_{m-1}\} and eigenvalues σ\boldsymbol{\sigma} are computed. A function hh defined on the dataset is projected as:

Pδh=∑j=1m⟨h,ψj⟩ψj.(41)P_{\delta}h = \sum_{j=1}^{m} \langle h, \psi_j \rangle\psi_j. \tag*{(41)}

Latent Harmonics at a new point ϕnew\phi_{\mathrm{new}} are evaluated via:

Ψj(ϕnew)=σj−1∑i=1mK∗(ϕnew,ϕi)ψj(ϕi).(42)\mathbf{\Psi}_j(\phi_{\mathrm{new}})=\sigma_j^{-1}\sum_{i=1}^{m}K^{*}(\phi_{\mathrm{new}},\phi_i)\psi_j(\phi_i). \tag*{(42)}

The extended function value at ϕnew\phi_{\mathrm{new}} is then estimated as:

(Ef)(ϕnew)=∑j=1m⟨h,ψj⟩Ψj(ϕnew).(43)(Ef)(\phi_{\mathrm{new}})=\sum_{j=1}^{m}\langle h,\psi_j\rangle\mathbf{\Psi}_j(\phi_{\mathrm{new}}). \tag*{(43)}

Algorithm 1 Double Diffusion Maps

Require: Dataset X={xi}i=1N⊂Rd∼fX(⋅)\mathcal{X}=\{\mathbf{x}_i\}_{i=1}^{N}\subset\mathbb{R}^{d}\sim f_X(\cdot)

  1. Step 1: Dimensionality Reduction with Diffusion Maps

    • Compute kernel K(xi,xj)K(\mathbf{x}_i,\mathbf{x}_j)

    • Normalize: K~=P−αKP−α\widetilde{K}=P^{-\alpha}KP^{-\alpha}, with α=1\alpha=1

    • Row-normalize: M=D−1K~M=D^{-1}\widetilde{K}

    • Eigendecompose MM and select Φ∈RN×k\Phi\in\mathbb{R}^{N\times k}

  1. Step 2: Latent Harmonics via Double Diffusion Maps

    • Construct kernel K∗(ϕi,ϕj)K^{*}(\phi_i,\phi_j)

    • Compute top ll eigenvectors Ψ∈RN×l\mathbf{\Psi}\in\mathbb{R}^{N\times l}

    • Project f→Pδff\to P_{\delta}f

  1. Step 3: Function Extension

    • Evaluate Ψj(ϕnew)\mathbf{\Psi}_j(\phi_{\mathrm{new}})

    • Compute (Ef)(ϕnew)(Ef)(\phi_{\mathrm{new}})

Geometric Harmonics for Function Extension

Geometric Harmonics (GH) extends a function hh defined on a dataset X\mathcal{X} to new points xnew∉X\mathbf{x}_{\mathrm{new}}\notin\mathcal{X} using a kernel-based interpolation inspired by the Nyström method [43]. The mapping xnew↦ϕnew\mathbf{x}_{\mathrm{new}}\mapsto\phi_{\mathrm{new}} (restriction) is computed as:

K(xnew,xj)=exp⁡(−∥xnew−xj∥22ϵ),(44)K(x_{\mathrm{new}}, x_j) = \exp\left(-\frac{\lVert x_{\mathrm{new}} - x_j\rVert^2}{2\epsilon}\right), \tag*{(44)}
K~(xnew,xj)=K(xnew,xj)p(xnew)αp(xj)α,(45)\widetilde{K}(x_{\mathrm{new}}, x_j) = \frac{K(x_{\mathrm{new}}, x_j)}{p(x_{\mathrm{new}})^\alpha p(x_j)^\alpha}, \tag*{(45)}

where p(xnew)=∑jK(xnew,xj)p(x_{\mathrm{new}}) = \sum_j K(x_{\mathrm{new}}, x_j) and p(xj)=∑kK(xj,xk)p(x_j) = \sum_k K(x_j, x_k). The interpolation weights are defined as:

A(xnew,xi)=K~(xnew,xi)∑j=1NK~(xnew,xj).(46)A(x_{\mathrm{new}}, x_i) = \frac{\widetilde{K}(x_{\mathrm{new}}, x_i)}{\sum_{j=1}^{N} \widetilde{K}(x_{\mathrm{new}}, x_j)}. \tag*{(46)}

The extended eigenfunction value for component β\beta is given by:

ϕβ(xnew)=1λβ∑i=1NA(xnew,xi)ϕβ(xi).(47)\phi_\beta(x_{\mathrm{new}}) = \frac{1}{\lambda_\beta}\sum_{i=1}^{N} A(x_{\mathrm{new}}, x_i)\phi_\beta(x_i). \tag*{(47)}

References

  1. [1]B. Stojadinovići, The case for systemic resilience: Urban communities in natural disasters, Domains of resilience for complex interconnected systems. (2018) 141.
  2. [2]Y. Li, J. P. Conte, P. E. Gill, Probabilistic performance-based optimum seismic design framework: illustration and validation, Computer Modeling in Engineering & Sciences 120 (3) (2019) 517–543.DOI
  3. [3]D. Vamvatsikos, M. Fragiadakis, Incremental dynamic analysis for estimating seismic performance sensitivity and uncertainty, Earthquake engineering & structural dynamics 39 (2) (2010) 141–163.DOI
  4. [4]D. Vamvatsikos, Seismic performance uncertainty estimation via ida with progressive accelerogram-wise latin hypercube sampling, Journal of Structural Engineering 140 (8) (2014) A4014015.DOI
  5. [5]D. Patsialis, A. Taflanidis, Multi-fidelity monte carlo for seismic risk assessment applications, Structural Safety 93 (2021) 102129.DOI
  6. [6]D. Patsialis, A. Taflanidis, Reduced order modeling of hysteretic structural response and applications to seismic risk assessment, Engineering Structures 209 (2020) 110135.DOI
  7. [7]F. Jalayer, H. Ebrahimian, Seismic reliability assessment and the nonergodicity in the modelling parameter uncertainties, Earthquake Engineering & Structural Dynamics 49 (5) (2020) 434–457.DOI
  8. [8]J. Chen, J. Li, Stochastic seismic response analysis of structures exhibiting high nonlinearity, Computers & structures 88 (7-8) (2010) 395–412.DOI
  9. [9]K. dos Santos, I. Kougioumtzoglou, A. Beck, Incremental dynamic analysis: a nonlinear stochastic dynamics perspective, Journal of Engineering Mechanics 142 (10) (2016) 06016007.DOI
  10. [10]I. Gidaris, A. Taflanidis, G. Mavroeidis, Kriging metamodeling in seismic risk assessment based on stochastic ground motion models, Earthquake Engineering & Structural Dynamics 44 (14) (2015) 2377–2399.DOI
  11. [11]J. Zou, D. Welch, A. Zsarnoczay, A. Taflanidis, G. Deierlein, Surrogate modeling for the seismic response estimation of residential wood frame structures, in: Proceedings of the 17th World Conference on Earthquake Engineering, Japan, 2020.
  12. [12]G. Abbiati, M. Broccardo, I. Abdallah, S. Marelli, F. Paolacci, Seismic fragility analysis based on artificial ground motions and surrogate modeling of validated structural simulators, Earthquake Engineering & Structural Dynamics 50 (9) (2021) 2314–2333.
  13. [13]A. Kyprioti, A. Taflanidis, Kriging metamodeling for seismic response distribution estimation, Earthquake Engineering & Structural Dynamics 50 (13) (2021) 3550–3576.DOI
  14. [14]D. Giovanis, M. Fragiadakis, V. Papadopoulos, Epistemic uncertainty assessment using incremental dynamic analysis and neural networks, Bulletin of Earthquake Engineering 14 (2016) 529–547.DOI
  15. [15]Y. Pang, X. Dang, W. Yuan, An artificial neural network based method for seismic fragility analysis of highway bridges, Advances in Structural Engineering 17 (3) (2014) 413–428.DOI
  16. [16]K. Bakalis, D. Vamyatsikos, M. Fragiadakis, Seismic risk assessment of liquid storage tanks via a nonlinear surrogate model, Earthquake Engineering & Structural Dynamics 46 (15) (2017) 2851–2868.DOI
  17. [17]X. Zhu, M. Broccardo, B. Sudret, Seismic fragility analysis using stochastic polynomial chaos expansions, Probabilistic Engineering Mechanics (2023) 103413.DOI
  18. [18]S. Yi, A. Taflanidis, Stochastic emulation with enhanced partial-and no-replication strategies for seismic response distribution estimation, Earthquake Engineering & Structural Dynamics 53 (7) (2024) 2354–2381.DOI
  19. [19]D. G. Giovanis, A. Taflanidis, M. D. Shields, Accelerating uncertainty quantification in incremental dynamic analysis using dimension reduction-based surrogate modeling, Bulletin of Earthquake Engineering 23 (1) (2025) 391–410.DOI
  20. [20]Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, B. Poole, Score-based generative modeling through stochastic differential equations, arXiv preprint arXiv:2011.13456 (2020).arxiv.org/abs/2011.13456
  21. [21]Y. Song, S. Ermon, Generative modeling by estimating gradients of the data distribution, Advances in neural information processing systems 32 (2019).arxiv.org/abs/1907.05600
  22. [22]Y. Song, S. Ermon, Improved techniques for training score-based generative models, Advances in neural information processing systems 33 (2020) 12438–12448.arxiv.org/abs/2006.09011
  23. [23]J. Ho, A. Jain, P. Abbeel, Denoising diffusion probabilistic models, Advances in neural information processing systems 33 (2020) 6840–6851.arxiv.org/abs/2006.11239
  24. [24]C. Soize, R. Ghanem, Data-driven probability concentration and sampling on manifold, Journal of Computational Physics 321 (2016) 242–258.DOI
  25. [25]C. Soize, R. Ghanem, Probabilistic learning on manifolds (plom) with partition, International Journal for Numerical Methods in Engineering 123 (1) (2022) 268–290.arxiv.org/abs/2102.10894
  26. [26]R. Coifman, S. Lafon, Diffusion maps, Applied and Computational Harmonic Analysis 21 (1) (2006) 5–30.DOI
  27. [27]D. G. Giovanis, N. Evangelou, I. G. Kevrekidis, R. G. Ghanem, Enabling probabilistic learning on manifolds through double diffusion maps, Journal of Computational Physics 551 (2026) 114663. doi: 10.1016/j.jcp.2026.114663.DOI
  28. [28]D. G. Giovanis, E. Crabtree, R. G. Ghanem, I. G. Kevrekidis, Generative learning of densities on manifolds, Computer Methods in Applied Mechanics and Engineering 446 (2025) 118266. doi:10.1016/j.cma.2025.118266.DOI
  29. [29]E. R. Crabtree, D. G. Giovanis, N. Evangelou, J. M. Bello-Rivas, I. G. Kevrekidis, Generative learning for slow manifolds and bifurcation diagrams, Computers & Chemical Engineering 207 (2026) 109544. doi:10.1016/j.compchemeng.2025.109544.DOI
  30. [30]N. Evangelou, F. Dietrich, E. Chiavazzo, D. Lehmberg, M. Meila, I. G. Kevrekidis, Double diffusion maps and their latent harmonics for scientific computations in latent space, Journal of Computational Physics 485 (2023) 112072.DOI
  31. [31]R. Coifman, S. Lafon, Geometric harmonics: a novel tool for multiscale out-of-sample extension of empirical functions, Applied and Computational Harmonic Analysis 21 (1) (2006) 31–52.DOI
  32. [32]D. A. Foutch, S.-Y. Yun, Modeling of steel moment frames for seismic loads, Journal of Constructional Steel Research 58 (5-8) (2002) 529–564.DOI
  33. [33]P. Vincent, A connection between score matching and denoising autoencoders, Neural computation 23 (7) (2011) 1661–1674.DOI
  34. [34]Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, B. Poole, Score-based generative modeling through stochastic differential equations (2020). doi:10.48550/ARXIV.2011.13456. URL https://arxiv.org/abs/2011.13456
  35. [35]Y. Liu, M. Yang, Z. Zhang, F. Bao, Y. Cao, G. Zhang, Diffusion-model-assisted supervised learning of generative models for density estimation, Journal of Machine Learning for Modeling and Computing 5 (1) (2024).arxiv.org/abs/2310.14458
  36. [36]M. Fragiadakis, D. Vamvatsikos, M. Papadrakakis, Evaluation of the influence of vertical irregularities on the seismic performance of a 9-storey steel frame, Earthquake Engineering & Structural Dynamics 35 (12) (2006) 1489–1509.
  37. [37]R. L. Iman, W.-J. Conover, A distribution-free approach to inducing rank correlation among input variables, Communications in Statistics-Simulation and computation 11 (3) (1982) 311–334.DOI
  38. [38]D. Vamvatsikos, C. A. Cornell, Applied incremental dynamic analysis, Earthquake spectra 20 (2) (2004) 523–553.
  39. [39]C. J. Dsilva, R. Talmon, R. R. Coifman, I. G. Kevrekidis, Parsimonious representation of nonlinear dynamical systems through manifold learning: A chemotaxis case study, Applied and Computational Harmonic Analysis 44 (3) (2018) 759–773.
  40. [40]T. J. Rothenberg, Identification in parametric models, Econometrica 39 (3) (1971) 577–591. doi:10.2307/1913267.DOI
  41. [41]A. F. Brouwer, M. C. Eisenberg, The underlying connections between identifiability, active subspaces, and parameter space dimension reduction (2018). arXiv:1802.05641, doi:10.48550/arXiv.1802.05641. URL https://arxiv.org/abs/1802.05641
  42. [42]C. Gear, E. Chiavazzo, Y. G. Kevrekidis, Manifolds defined by points: Parameterizing and boundary detection, in: AIP Conference Proceedings, 2016.DOI
  43. [43]C. Fowlkes, S. Belongie, J. Malik, Efficient spatiotemporal grouping using the nystrom method, in: Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. CVPR 2001, Vol. 1, IEEE, 2001, pp. I–I.

Paper details

vlnc:W_q7jma2vg
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents