Scenario-conditioned flow matching for probabilistic generation of three-component ground-motion waveforms
summary
The gist
Scenario-conditioned flow matching for probabilistic generation of three-component ground-motion waveforms presents WaveFlowGMM…
In short
WaveFlowGMM is a two-stage model that generates three-component ground motions compatible with specific earthquake scenarios. It separates amplitude estimation using physics-informed learning from waveform generation using AlphaFlow. This allows for probabilistic modeling of seismic risk by ensuring generated motions match source, path, and site conditions while preserving crucial physical details.
Key concepts
- Two-stage Probabilistic Framework
- The model splits the complex task into two simpler steps: first estimating the overall earthquake amplitude (PGA) using a physics-informed symbolic learning approach, and second generating the detailed waveform shape based on that amplitude. This separation makes the generation process more manageable and physically grounded.
- Physics-Informed Symbolic Learning (PISL)
- This technique is used in the first stage to estimate component-wise Peak Ground Acceleration (PGA) medians and their covariance. It uses a library of interpretable terms like magnitude, distance, and site conditions to select the most relevant physical factors for scaling the motion.
- AlphaFlow
- This is a generative model used in the second stage to create the actual waveform shape. It operates in an invertible wavelet-packet coefficient space, allowing it to generate normalized three-component histories that are then scaled by the sampled PGA, ensuring realistic wave characteristics.
Terminology used across episodes
This episode discusses
- Scenario-conditioned flow matching for probabilistic generation of three-component ground-motion waveforms · Paper Radio
- Site-Specific Ground Motion Generative Model for Crustal Earthquakes in Japan Based on Generative Adversarial Networks
- Learning Physics for Unveiling Hidden Earthquake Ground Motions via Conditional Generative Modeling
- High Resolution Seismic Waveform Generation using Denoising Diffusion
- Broadband Ground Motion Synthesis by Diffusion Model with Minimal Condition
- Waveform-Based Probabilistic Seismic Hazard Analysis Using Ground-Motion Generative Models
- Elucidating the Design Space of Diffusion-Based Generative Models
- High-Resolution Image Synthesis with Latent Diffusion Models
- Scalable Diffusion Models with Transformers
- Progressive Distillation for Fast Sampling of Diffusion Models
- Adversarial Diffusion Distillation
- Consistency Models
- Improved Techniques for Training Consistency Models
- Simplifying, Stabilizing and Scaling Continuous-Time Consistency Models
- Flow Matching for Generative Modeling
- Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow
- Mean Flows for One-step Generative Modeling
- Wavelet Packet-Based Diffusion Model for Ground Motion Generation with Multi-Conditional Energy and Spectral Matching · Paper Radio
- Denoising Diffusion Implicit Models
The paper
Scenario-conditioned flow matching for probabilistic generation of three-component ground-motion waveforms · Read on arXiv
Yi Ding, Jinjun Hu, Su Chen, Xianwei Liu, Zhongxiang Zhang, Zongchao Li, Xiaojun Li, Lili Xie
State Key Laboratory of Precision Blasting, Jianghan University, Wuhan, China · Key Laboratory of Earthquake Engineering and Engineering Vibration, Institute of Engineering Mechanics, China Earthquake Administration · Key Laboratory of Urban Security and Disaster Engineering of the Ministry of Education, Beijing University of Technology · School of Civil Engineering and Architecture, East China Jiaotong University · Key Laboratory of Roads and Railway Engineering Safety Control, Ministry of Education, Shijiazhuang Tiedao University · Institute of Geophysics, China Earthquake Administration
Performance-based seismic risk assessment requires three-component acceleration histories compatible with specified source, path, and site conditions. Conventional ground-motion prediction equations provide scalar intensity measures, while many generative waveform models learn amplitude and waveform shape within a single high-dimensional target. We present WaveFlowGMM, a two-stage probabilistic ground-motion model that uses peak ground acceleration (PGA) as an amplitude interface between scenario conditioning and waveform generation. The amplitude stage uses physics-informed symbolic learning to estimate component-wise PGA medians and a full cross-component covariance. The waveform stage uses few-step AlphaFlow in an invertible wavelet-packet coefficient space to generate normalised three-component histories that are rescaled by sampled PGA. Tests on an event-level NGA-West2 holdout set show that the generated motions recover the main magnitude, distance, and site scaling, keep peak and spectral residuals close to zero, preserve three-component amplitude dependence, and yield velocity and displacement histories without systematic drift after integration of the generated three-component acceleration histories. The framework provides an interpretable and computationally efficient candidate component for waveform-level seismic hazard and risk analysis.
Transcript
Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.
Kai: We're looking at the paper: "Scenario-conditioned flow matching for probabilistic generation of three-component ground-motion waveforms".
Mira: Here's the gist of the paper. Scenario-conditioned flow matching for probabilistic generation of three-component ground-motion waveforms presents WaveFlowGMM, a two-stage probabilistic ground-motion model that uses peak ground acceleration (PGA) as an amplitude interface between scenario conditioning and waveform generation The framework is designed to generate three-component acceleration histories that are compatible with specified source, path, and site conditions while preserving phase, duration, envelope, and component dependence Framework Overview The core of WaveFlowGMM is a two-stage probabilistic framework that separates amplitude sampling from waveform-shape generation The first stage, the PGA amplitude model, uses physics-informed symbolic learning (PISL) to estimate component-wise PGA medians and a full cross-component covariance The second stage employs AlphaFlow in an invertible wavelet-packet coefficient space to generate normalised three-component histories that are rescaled by sampled PGA This design allows the low-dimensional PGA model to carry the absolute scale and three-component amplitude dependence, while AlphaFlow models the conditional distribution of normalised waveform shape PGA Amplitude Modeling The PISL PGA amplitude model describes the conditional distribution of the three-component PGA target q from the scenario condition x This model has two parts: a conditional mean mu(x) that sets the median PGA scaling, and an inter-event and intra-event covariance that sets the joint three-component dispersion The conditional mean is obtained with physics-informed symbolic learning (PISL), which selects a small number of interpretable terms from a candidate library of magnitude, distance, site, and faulting mechanism terms The resulting PGA model is defined as p PISL(q x) = (q; μ(x), Στ + Σf) This formulation ensures that the sampled H1, H2, and V values preserve the observed cross-component dependence by using full covariance matrices for the inter-event and intra-event terms Waveform Generation AlphaFlow models PGA-normalised three-component wavelet-packet coefficients The training objective defines a linear path between the data distribution and a prior noise distribution in the wavelet-packet space The neural network is conditioned on the scenario condition x and the PGA condition r, which carries the horizontal geometric-mean amplitude, relative contrast, and vertical-to-horizontal ratio The generator backbone is a two-dimensional convolutional U-Net because the wavelet-packet coefficient array has an explicit time-frequency grid structure Performance and Efficiency Tests on the event-level NGA-West2 holdout set show that the generated motions recover the main magnitude, distance, and site scaling, keep peak and spectral residuals close to zero, preserve three-component amplitude dependence, and yield velocity and displacement histories without systematic drift after integration The full covariance sampling reproduces the three-component residual correlations in the test set AlphaFlow reduces the per-record generation cost to the order of one percent of the diffusion-sampler cost compared to other generative models The results show that peak amplitude agreement alone is not enough to select a sampler; duration, Arias energy build-up, velocity, displacement, and long-period response are the quantities that reveal whether a generated time history remains useful after integration and structural filtering Limitations The main limitations arise from data coverage and from the simplified conditioning set The present training data are limited to shallow crustal earthquakes in the western United States, and the model does not include coordinates, regional terms, rupture directivity, basin structure, or path-specific attenuation The PISL site term is a linear saturated Vs30 ratio and does not model nonlinear soil response that couples Vs30 with input intensity The model generates aligned waveform windows, not absolute arrival times, so network-level timing applications remain outside the present scope The NGA-West2 remains sparse in the large-magnitude, near-source, and long-period regimes where both empirical GMMs and data-driven generators are hardest to constrain WaveFlowGMM is supported as a candidate component for waveform-level seismic hazard and risk workflows The study introduces WaveFlowGMM, a two-stage ground-motion model for scenario-conditioned generation of three-component acceleration histories Its current scope remains bounded by NGA-West2 coverage, aligned waveform windows, and the conditioning variables used here Larger regional databases and richer rupture-path-site descriptors are the next step towards site-specific waveform-level analysis Acknowledgements The authors thank the Pacific Earthquake Engineering Research Center for providing the NGA-West2 database used in this study Conflict of Interest Statement The authors declare no conflicts of interest Data Availability Statement The NGA-West2 database used in this study is available from the Pacific Earthquake Engineering Research Center References
Paper summary: Mira: To wrap up, the paper "Scenario-conditioned flow matching for probabilistic generation of three-component ground-motion waveforms" introduces WaveFlowGMM, a two-stage ground-motion model that separates amplitude sampling from waveform shape generation. It uses peak ground acceleration as an amplitude interface to generate three-component histories compatible with specified source, path, and site conditions.
Kai: The core claim is that the PGA amplitude model uses physics-informed symbolic learning to estimate a full cross-component covariance, which helps preserve three-component amplitude dependence in the sampled H1, H2, and V values.
Lev: And they then use AlphaFlow in an invertible wavelet-packet coefficient space to generate normalized three-component histories that are rescaled by those sampled PGA values. This design allows the low-dimensional PGA model to carry the absolute scale and three-component amplitude dependence while AlphaFlow models the conditional distribution of the normalized waveform shape.
Mira: The performance tests on NGA-West2 show that these generated motions recover main magnitude, distance, and site scaling, keep peak and spectral residuals close to zero, and yield velocity and displacement histories without systematic drift after integration.
Kai: It means this framework provides an interpretable and computationally efficient candidate for waveform-level seismic hazard analysis. The limitations they state are clear: the model is currently limited by NGA-West2 coverage, aligned waveform windows, and the conditioning variables used here.
Lev: The next step for this research involves using larger regional databases and richer rupture-path-site descriptors to move towards more site-specific waveform-level analysis.
Conclusion: Kai: So, we're wrapping up this look at WaveFlowGMM, this two-stage model for generating three-component ground motion waveforms that uses peak ground acceleration as an interface between the input scenario and the actual waveform shape.
Mira: Yeah, it’s about taking those complex source and site conditions—source path site—and turning them into a probabilistic set of three-component histories that actually look like real shaking on paper.
Lev: From where I sit with hardware, it’s cool that they managed to get the amplitude sampling and the waveform generation separated, because trying to learn everything in one big model is just too much overhead for what we're building.
Kai: Right, and their way of doing it seems pretty clever—they use physics-informed symbolic learning for the PGA part to get those component medians and covariance, which then feeds into AlphaFlow for the shape generation.
Mira: The implication is that you don't need one giant model that tries to predict everything at once; you can break it down into a simpler amplitude estimation and a more complex shape modeling step.
Lev: That separation is what makes it feasible for real hardware, because if the amplitude part is too messy, the waveform part gets impossibly high dimensional.
Kai: The authors are essentially saying that for seismic hazard analysis, this two-stage approach gives you an interpretable component to work with—something you can actually trust.
Mira: It shifts the focus from just predicting a single output to creating a whole distribution of possible motions based on the inputs.
Lev: So, if you’re building something for real hardware, it suggests that this kind of structured approach could be useful for error correction or perhaps even generating synthetic training data for those systems.
Kai: It opens up the door to asking what kind of seismic input really needs to be sampled and how much fidelity we actually need at each stage.
Mira: And that leads us right into the limitations they pointed out—the current data set is small, limited mostly to shallow crustal earthquakes in western United States.
Lev: Exactly, so the real challenge now is scaling up that database and making those physics-informed symbolic learning terms more robust when we apply them to different geological settings.
More episodes
- 2611.14082-Two atoms with strong light-induced dipolar interactions
- 2611.14040-Characterization of BNLs 20 and 30 micrometer thick LGADs for the Electron Ion Collider
- 2611.13399-Lift-element-based dynamic network of vortex gust-airfoil interactions
- 2611.13476-Hopfion Superlattices in 3D Twisted Moir'e Light Fields
- 2611.13642-Investigation of the timing performance of double-sided p-type 3D silicon sensors produced by CNM
- 2611.13830-SiPM optical readout of a PICOSEC Micromegas precise timing detector
- 2611.14152-Where are the shortest-path nodes in spatial networks?
- 2611.14219-Sub-Ice: semi-automated mapping of basal channels from ice shelf surface elevation data
- 2611.16341-SUMNO: Spatially Unrolled Multigrid Neural Operator for Fluid Flow Surrogates
- 2611.16090-Multi-UTE Myelin Fraction Mapping: Repeatability and Cross-Field-Strength Reproducibility at 1.5T and 3T