Inverse Reconstruction of Shock Time Series from Shock Response Spectrum Curves using Machine Learning

Adam Watts *, 1 , Andrew Jeon *, 1 Destry Newton *, 1 , Ryan Bowering *, 2 , 1 Los Alamos National Laboratory, P.O. Box 1663, Los Alamos, NM 87545, USA 2 University of Rochester, Rochester, NY, USA {acwatts, andrewjjeon, destryn}@lanl.gov [email protected] · 2026

Read Inverse Reconstruction of Shock Time Series from Shock Response Spectrum Curves using Machine Learning by Adam Watts *, 1 , Andrew Jeon *, 1 Destry Newton *, 1 , Ryan Bowering *, 2 , 1 Los Alamos National Laboratory, P.O. Box 1663, Los Alamos, NM 87545, USA 2 University of Rochester, Rochester, NY, USA {acwatts, andrewjjeon, destryn}@lanl.gov [email protected] (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-03229.txt.

Introduction

Adam Watts, Andrew Jeon, Destry Newton, and Ryan Bowering rebuild a shock's time history from its fingerprint. Engineers describe a shock by its response spectrum, which records how simple oscillators of many frequencies would react. Many different shocks share one spectrum, so reversing the map is ill-posed, and classical searches are slow and rigid.

Their conditional variational autoencoder learns the inverse map from data instead. Once trained, it generates acceleration signals matching a target spectrum with no iterative optimization. Experiments show better spectral fidelity than classical techniques and strong generalization to unseen spectra. Inference runs three to six orders of magnitude faster. The Los Alamos team presents generative modeling as a scalable route to shock reconstruction.

Catalog record

CREATORAdam Watts *, 1 , Andrew Jeon *, 1 Destry Newton *, 1 , Ryan Bowering *, 2 , 1 Los Alamos National Laboratory, P.O. Box 1663, Los Alamos, NM 87545, USA 2 University of Rochester, Rochester, NY, USA {acwatts, andrewjjeon, destryn}@lanl.gov [email protected]
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE87059
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.03229

Browse the full library · About the Tower of Babel

Full text

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

Inverse Reconstruction of Shock Time Series from Shock Response Spectrum Curves using Machine Learning Adam Watts *, 1 , Andrew Jeon *, 1 Destry Newton *, 1 , Ryan Bowering *, 2 , 1 Los Alamos National Laboratory, P.O. Box 1663, Los Alamos, NM 87545, USA 2 University of Rochester, Rochester, NY, USA {acwatts, andrewjjeon, destryn}@lanl.gov [email protected] Abstract The shock response spectrum (SRS) is widely used to characterize the response of single-degree-of-freedom (SDOF) systems to transient accelerations. Because the mapping from acceleration time history to SRS is nonlinear and many-to-one, reconstructing time-domain signals from a target spectrum is inherently ill-posed. Conventional approaches address this problem through iterative optimization, typically representing signals as sums of exponentially decayed sinusoids, but these methods are computationally expensive and constrained by predefined basis functions. We propose a conditional variational autoencoder (CVAE) that learns a data-driven inverse mapping from SRS to acceleration time series. Once trained, the model generates signals consistent with prescribed target spectra without requiring iterative optimization. Experiments demonstrate improved spectral fidelity relative to classical techniques, strong generalization to unseen spectra, and inference speeds three to six orders of magnitude faster. These results establish deep generative modeling as a scalable and efficient approach for inverse SRS reconstruction. 1 1 footnotetext: Alphabetical by first name, these authors contributed equally to this work. Code will be released upon publication. 1 Introduction A mechanical shock is a sudden, short-duration event characterized by a rapid change in velocity and typically accompanied by high acceleration. An event is considered a shock if its duration is significantly shorter than the natural period of the system under consideration [ 12 ] . The foundations of shock analysis were established by Biot, who introduced the study of single-degree-of-freedom (SDOF) systems subjected to transient dynamic loads, showing that the peak response of a system depends on its natural frequency and damping [ 1 ] . Building on this concept, the shock response spectrum (SRS) was developed to characterize shocks by recording the maximum response (displacement, velocity, or acceleration) of a set of SDOF systems, each with a different natural frequency, to a common input. This representation enables engineers to model and quantify the effects of shocks on linear systems across frequency ranges, regardless of the source. The SRS thus provides a standardized means of assessing shock severity, comparing shocks, and establishing equivalence criteria between measured transient environments and laboratory simulations [ 5 ] . The SRS has become a standard tool across engineering disciplines for designing and qualifying assemblies and components to withstand transient shock loads. For example, MIL-STD-810G outlines environmental engineering considerations and laboratory test methods, including the use of SRS in equipment qualification [ 20 ] . When plotted, the SRS displays natural frequency of an SDOF on the horizontal axis (with damping ratio specified by the user) and the peak response of that SDOF on the vertical axis. In this work, the shock input is taken to be a time series representing the acceleration applied to the system by the environment, with the response of interest being acceleration, denoted as \ddot{x}(t) . Of particular relevance is the maximax spectrum, which records the maximum absolute acceleration response across all frequencies. However, the use of absolute values and maxima renders the maximax SRS both nonlinear and many-to-one, meaning that different acceleration time series can yield identical shock response spectra [ 12 ] . In practice, environmental specifications (ES) for shock testing are derived from measured shock response spectra. At each frequency, a one-sided upper normal tolerance interval is calculated to represent the upper bound of observed responses, and repeating this process across frequencies produces a statistically conservative description of the environment [ 14 ] . These tolerance intervals form the foundation for defining shock testing requirements within the ES. Qualification testing then seeks to reproduce the specified shock environment — often using a shaker table — in order to verify system performance, durability, and functionality. However, because shaker control software requires inputs in the form of an acceleration time series rather than SRS curves, a fundamental difficulty arises: the SRS cannot be uniquely inverted to recover the underlying time series. This challenge is not limited to a nominal SRS, but applies even more strongly to statistical envelopes or tolerance bounds, which are abstractions of many different measured shocks. Such bounds describe only the maximum responses expected at each frequency, without preserving timing, phase, or correlation across frequencies. As a result, the underlying time series consistent with a statistical bound is not just unknown but may be physically unrealizable, making the generation of suitable test inputs an inherently approximate process [ 12 ] . Existing reconstruction methods address this problem by casting it as an optimization task, typically representing the acceleration signal as a finite sum of exponentially decayed sinusoids. While effective in certain cases, such methods are computationally demanding, restricted by their choice of basis functions, and limited in their ability to reproduce complex shocks. This paper introduces a novel machine learning (ML) framework that directly learns the mapping from an SRS to an acceleration time series. Unlike traditional approaches, the proposed methodology does not presuppose basis functions, achieves substantially faster computation, and offers improved fidelity in reconstructing time series consistent with target or specification SRS curves. 1.1 Contributions The main contributions of this work are summarized as follows: • A conditional variational autoencoder (CVAE) framework for SRS inversion: We introduce a CVAE architecture that directly maps target shock response spectra (SRS) to corresponding time-domain acceleration signals, eliminating the need for explicit basis functions. The trained model achieves high spectral fidelity while reducing inference time by several orders of magnitude compared to optimization-based synthesis methods such as SDS and SDS+GA. • A GPU-accelerated synthetic shock data generation system with an SRS operator: We develop a PyTorch-based framework that generates synthetic shock data at scale and includes a fully batched, differentiable Shock Response Spectrum (SRS) operator. The system rapidly produces paired time-series signals and their corresponding SRS at arbitrary sampling rates and durations, enabling high-throughput model training, evaluation, and benchmarking. • Two complementary performance metrics: Because no single metric fully captures global spectral agreement, relative performance across individual realizations, and localized frequency-domain fidelity, we evaluate model performance using RMSLE (including a per-sample RMSLE accuracy statistic) and per-frequency dB error. Together, these measures provide a unified evaluation framework that quantifies overall reconstruction quality, probabilistic case-by-case reliability, and localized spectral deviations. • A standardized benchmark dataset for reproducible evaluation: We release a comprehensive SRS–time series benchmark composed of four subsets: (i) two datasets derived from operational shock test recordings normalized to unit amplitude and anonymized to remove identifying test signatures ( \mathcal{D}_{\text{test}}^{\text{A}} and \mathcal{D}_{\text{test}}^{\text{B}} ); (ii) one dataset derived from earthquake recordings ( \mathcal{D}_{\text{test}}^{\text{C}} ); and (iii) one dataset of synthetic shocks generated using the proposed framework, incorporating all four basis function types ( \mathcal{D}_{\text{test}}^{\text{D}} ). These datasets establish a standardized and diverse reference for future research on shock synthesis and SRS inversion . 2 Preliminaries and Problem Formulation 2.1 The Shock Response Spectrum (SRS) The SRS operator is a functional mapping from a time-domain acceleration signal to a vector of spectral magnitudes evaluated over a set of SDOF natural frequencies: \mathrm{SRS}:\ddot{x}(t)\longmapsto\bm{\mathfrak{s}}_{x}, where the mapping

Read the complete text