Ashesh Chattopadhyay
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Ashesh Chattopadhyay has a total of 51 co-authors. Here are the top co-authors by number of shared publications:
- Pedram Hassanzadeh
The University of Chicago Physical Sciences Division, Unknown, Chicago, IL, United States [US
7 shared publications Moein Darman
Department of Applied Mathematics University of California Santa Cruz CA USA
5 shared publicationsTianning Wu
North Carolina State University Marine, Earth, and Atmospheric Sciences Raleigh NC USA
4 shared publicationsLeonard Lupin-Jimenez
Department of Applied Mathematics, University of California, Santa Cruz, CA 95064, United States of America
3 shared publicationsMichael Gray
North Carolina State University Marine, Earth, and Atmospheric Sciences Raleigh NC USA
3 shared publicationsAnthony Wong
Converging Technologies Laboratory Fujitsu Research of America Santa Clara CA USA
2 shared publications
Latest Publications
- Open Access
- 13 November 2026
- Published by Springer Nature in Research Square
The Dynamics of Spectral Bias in Operator Learning
Spectral bias is the phenomenon where neural networks learn the low-wavenumber content of a target before its high-wavenumber content. We analyze the training dynamics of neural operators on the discrete Poisson problem and find that the bias is a property of the eigenbasis of the kernel that governs the training update, not of the wavenumber basis in which it is usually reported. A model learning a discretized operator retains its error in the eigendirections of this kernel with small eigenvalues, which are the rates at which training contracts the error along those directions. Under smooth forcing, low variance is placed at high wavenumbers, and any model, linear or nonlinear, has small eigenvalues along the eigenvectors with high-wavenumber content. Under white forcing, the linear models learn the operator, while the nonlinear models retain error in slow directions that depend on the coupling between training samples in the kernel; for the GELU dense network they disappear entirely when that coupling is removed. Modes that a model cannot represent at all contribute zero eigenvalues. The slow eigenvectors decide which wavenumbers carry this error, through their projection onto the sine modes. For the spectral operator and the linear Fourier neural operator these eigenvectors are close to sine modes, so each wavenumber has its own rate and the error collects at high wavenumbers. Delocalized models, such as the dense networks under either forcing, have error of nearly the same magnitude at every wavenumber. The spectrum of the true solution decays with $k$ and falls below this floor at some wavenumber, past which the relative error grows. No mode is learned more slowly than another in this case; the growth comes from the decay of the target spectrum. Spectral bias is usually reported by looking at this retained error in the sine basis, and we use its projection onto the kernel's eigendirections to explain the shape it has there.- 13 November 2026
- Published by Springer Nature in Research Square
Open AccessThe Dynamics of Spectral Bias in Operator Learning
Spectral bias is the phenomenon where neural networks learn the low-wavenumber content of a target before its high-wavenumber content. We analyze the training dynamics of neural operators on the discrete Poisson problem and find that the bias is a property of the eigenbasis of the kernel that governs the training update, not of the wavenumber basis in which it is usually reported. A model learning a discretized operator retains its error in the eigendirections of this kernel with small eigenvalues, which are the rates at which training contracts the error along those directions. Under smooth forcing, low variance is placed at high wavenumbers, and any model, linear or nonlinear, has small eigenvalues along the eigenvectors with high-wavenumber content. Under white forcing, the linear models learn the operator, while the nonlinear models retain error in slow directions that depend on the coupling between training samples in the kernel; for the GELU dense network they disappear entirely when that coupling is removed. Modes that a model cannot represent at all contribute zero eigenvalues. The slow eigenvectors decide which wavenumbers carry this error, through their projection onto the sine modes. For the spectral operator and the linear Fourier neural operator these eigenvectors are close to sine modes, so each wavenumber has its own rate and the error collects at high wavenumbers. Delocalized models, such as the dense networks under either forcing, have error of nearly the same magnitude at every wavenumber. The spectrum of the true solution decays with $k$ and falls below this floor at some wavenumber, past which the relative error grows. No mode is learned more slowly than another in this case; the growth comes from the decay of the target spectrum. Spectral bias is usually reported by looking at this retained error in the sine basis, and we use its projection onto the kernel's eigendirections to explain the shape it has there. - 1 November 2026
- Published by AIP Publishing in Chaos: An Interdisciplinary Journal of Nonlinear Science
- Vol. 36 (11)
Lazy diffusion: Mitigating spectral collapse in generative diffusion-based stable autoregressive emulation of turbulent flows
Turbulent flows possess broadband, power-law spectra in which multiscale interactions couple high-wavenumber fluctuations to large-scale dynamics. Although diffusion-based generative models offer a principled probabilistic forecasting framework, we show that linear-schedule denoising diffusion probabilistic models exhibit spectral collapse: a Fourier-space analysis of the forward stochastic differential equation reveals a closed-form, mode-wise signal-to-noise ratio that decays monotonically in wavenumber |k|. For spectra S(k) ∝ |k|−λ, this renders high-wavenumber modes noise dominated earlier and produces an intrinsic spectral bias. We reinterpret the noise schedule as a spectral regularizer and study power-law schedules β(τ)∝τγ that preserve fine-scale structure deeper into diffusion time, along with lazy diffusion, a one-step retraining method that reuses the pretrained score network to bypass long reverse-time trajectories. Applied to high-Reynolds-number 2D Kolmogorov turbulence and 1/12° Gulf of Mexico ocean reanalysis, these methods mitigate spectral collapse, stabilize long-horizon autoregression, and restore physically realistic inertial-range scaling. Together, they show that naïve Gaussian scheduling can impair long-term spectral fidelity in turbulent flows and that physics-aware diffusion processes can yield accurate, efficient, and fully probabilistic surrogates for multiscale dynamical systems.
- 1 November 2026
- Published by AIP Publishing in Chaos: An Interdisciplinary Journal of Nonlinear Science
- Vol. 36 (11)
Lazy diffusion: Mitigating spectral collapse in generative diffusion-based stable autoregressive emulation of turbulent flows
Turbulent flows possess broadband, power-law spectra in which multiscale interactions couple high-wavenumber fluctuations to large-scale dynamics. Although diffusion-based generative models offer a principled probabilistic forecasting framework, we show that linear-schedule denoising diffusion probabilistic models exhibit spectral collapse: a Fourier-space analysis of the forward stochastic differential equation reveals a closed-form, mode-wise signal-to-noise ratio that decays monotonically in wavenumber |k|. For spectra S(k) ∝ |k|−λ, this renders high-wavenumber modes noise dominated earlier and produces an intrinsic spectral bias. We reinterpret the noise schedule as a spectral regularizer and study power-law schedules β(τ)∝τγ that preserve fine-scale structure deeper into diffusion time, along with lazy diffusion, a one-step retraining method that reuses the pretrained score network to bypass long reverse-time trajectories. Applied to high-Reynolds-number 2D Kolmogorov turbulence and 1/12° Gulf of Mexico ocean reanalysis, these methods mitigate spectral collapse, stabilize long-horizon autoregression, and restore physically realistic inertial-range scaling. Together, they show that naïve Gaussian scheduling can impair long-term spectral fidelity in turbulent flows and that physics-aware diffusion processes can yield accurate, efficient, and fully probabilistic surrogates for multiscale dynamical systems.
- Open Access
- 10 October 2026
- Published by American Geophysical Union (AGU) in Journal of Geophysical Research: Machine Learning and Computation
- Vol. 3 (5)
High‐Resolution Downscaling of a Lightweight Climate Emulator Using Diffusion‐Based Super‐Resolution
High‐resolution climate fields are essential for the assessment of their regional and local effects. Current global climate models, along with being computationally expensive, are often too coarse for applications that depend on local extremes and orographic gradients. Machine learning offers a computationally efficient route for both climate emulation and super‐resolution from coarse simulations to finer spatial scales. Therefore, we develop a data‐driven downscaling framework for the Lightweight Uncoupled ClImate Emulator (LUCIE), a lightweight Spherical Fourier Neural Operator climate emulator introduced by Guan et al. (2025, https://doi.org/10.1029/2025ms005152 ). We compare deterministic super‐resolution baselines, including a U‐Net and Spherical Fourier Neural Operator (SFNO), with two probabilistic diffusion formulations: a conditional Elucidated Diffusion Model and a diffusion‐posterior‐sampling approach. The models are trained on 6‐hourly ERA5 reanalysis fields from 2000 to 2009 and evaluated on out‐of‐sample LUCIE predictions from 2010 to 2018, downscaling coarse emulator states to km resolution. Performance is assessed against ERA5 using latitude‐weighted root mean square error (RMSE), power spectra, probability density functions, and the leading Empirical Orthogonal Function of zonal wind. These diagnostics test whether deterministic and probabilistic super‐resolution models can recover fine‐scale climatological statistics, high‐wavenumber variability, and dominant modes with an imperfect coarse emulator. Our results characterize the strengths and limitations of emulator‐driven super‐resolution, including its ability to reproduce ERA5‐like fine‐scale statistics while preserving explicit coarse‐scale consistency with LUCIE.
- 10 October 2026
- Published by American Geophysical Union (AGU) in Journal of Geophysical Research: Machine Learning and Computation
- Vol. 3 (5)
Open AccessHigh‐Resolution Downscaling of a Lightweight Climate Emulator Using Diffusion‐Based Super‐Resolution
High‐resolution climate fields are essential for the assessment of their regional and local effects. Current global climate models, along with being computationally expensive, are often too coarse for applications that depend on local extremes and orographic gradients. Machine learning offers a computationally efficient route for both climate emulation and super‐resolution from coarse simulations to finer spatial scales. Therefore, we develop a data‐driven downscaling framework for the Lightweight Uncoupled ClImate Emulator (LUCIE), a lightweight Spherical Fourier Neural Operator climate emulator introduced by Guan et al. (2025, https://doi.org/10.1029/2025ms005152 ). We compare deterministic super‐resolution baselines, including a U‐Net and Spherical Fourier Neural Operator (SFNO), with two probabilistic diffusion formulations: a conditional Elucidated Diffusion Model and a diffusion‐posterior‐sampling approach. The models are trained on 6‐hourly ERA5 reanalysis fields from 2000 to 2009 and evaluated on out‐of‐sample LUCIE predictions from 2010 to 2018, downscaling coarse emulator states to km resolution. Performance is assessed against ERA5 using latitude‐weighted root mean square error (RMSE), power spectra, probability density functions, and the leading Empirical Orthogonal Function of zonal wind. These diagnostics test whether deterministic and probabilistic super‐resolution models can recover fine‐scale climatological statistics, high‐wavenumber variability, and dominant modes with an imperfect coarse emulator. Our results characterize the strengths and limitations of emulator‐driven super‐resolution, including its ability to reproduce ERA5‐like fine‐scale statistics while preserving explicit coarse‐scale consistency with LUCIE.
- Open Access
- 26 September 2026
- Published by American Geophysical Union (AGU) in Journal of Geophysical Research: Machine Learning and Computation
- Vol. 3 (5)
AI Emulation of Stochastic Sudden Stratospheric Warming With Interpretable Latent Structure
Rare weather regime transitions pose a challenge for data‐driven modeling due to class imbalance. In this study, we develop a probabilistic deep learning emulator for a prototypical system with regime transitions, the stochastic Holton–Mass (HM) model of stratospheric variability, and analyze the structure of its learned latent space. The HM model exhibits two metastable regimes, a strong and a weak polar vortex, maintained by nonlinear wave–mean flow interactions, with weak stochastic forcing intermittently triggering rare transitions between these regimes that qualitatively represent sudden stratospheric warming (SSW) events. We employ a ResNet‐inspired Conditional Variational Autoencoder with six‐layer encoder and decoder layers and explicit current‐state conditioning to model the distribution of the system's state at the next time step (1 day). The emulator accurately reproduces short‐term dynamics, steady‐state probability distributions, regime persistence statistics, rare transition rates, the transition committor function, and the transition expected lead time of the physical model. Beyond emulation fidelity, we interrogate the learned latent representation to understand how the model internalizes the underlying metastable structure of the dynamics. Principal component analysis of the 32‐dimensional latent space reveals a clear and unsupervised separation into four physically interpretable clusters corresponding to strong versus weak vortex regimes and stable versus transition‐prone configurations. Such emergent regime separation in latent space is hard to identify for deep generative models applied to high‐dimensional stochastic systems. Our results show that carefully designed probabilistic emulators can uncover physically meaningful manifolds governing extreme‐event dynamics, potentially aiding the development of improved operational advanced warning systems.
- 26 September 2026
- Published by American Geophysical Union (AGU) in Journal of Geophysical Research: Machine Learning and Computation
- Vol. 3 (5)
Open AccessAI Emulation of Stochastic Sudden Stratospheric Warming With Interpretable Latent Structure
Rare weather regime transitions pose a challenge for data‐driven modeling due to class imbalance. In this study, we develop a probabilistic deep learning emulator for a prototypical system with regime transitions, the stochastic Holton–Mass (HM) model of stratospheric variability, and analyze the structure of its learned latent space. The HM model exhibits two metastable regimes, a strong and a weak polar vortex, maintained by nonlinear wave–mean flow interactions, with weak stochastic forcing intermittently triggering rare transitions between these regimes that qualitatively represent sudden stratospheric warming (SSW) events. We employ a ResNet‐inspired Conditional Variational Autoencoder with six‐layer encoder and decoder layers and explicit current‐state conditioning to model the distribution of the system's state at the next time step (1 day). The emulator accurately reproduces short‐term dynamics, steady‐state probability distributions, regime persistence statistics, rare transition rates, the transition committor function, and the transition expected lead time of the physical model. Beyond emulation fidelity, we interrogate the learned latent representation to understand how the model internalizes the underlying metastable structure of the dynamics. Principal component analysis of the 32‐dimensional latent space reveals a clear and unsupervised separation into four physically interpretable clusters corresponding to strong versus weak vortex regimes and stable versus transition‐prone configurations. Such emergent regime separation in latent space is hard to identify for deep generative models applied to high‐dimensional stochastic systems. Our results show that carefully designed probabilistic emulators can uncover physically meaningful manifolds governing extreme‐event dynamics, potentially aiding the development of improved operational advanced warning systems.
- Open Access
- 1 June 2026
- Published by IOP Publishing in Journal of Physics: Conference Series
- Vol. 3252 (1)
Stability of Coupling Learned Diffusion to Shock Capturing Schemes
We propose a hybrid ML-numerical solver for mixed hyperbolic–parabolic PDEs in which a neural operator replaces only the parabolic (diffusive) contribution while a standard shock-capturing finite-volume method advances the hyperbolic (advective) fluxes. This selective coupling targets the parabolic timestep bottleneck without discarding the robustness of classical advection discretizations. The central challenge is long-horizon stability since small spectral errors in the learned diffusion update can accumulate under repeated timesteps and drive drift or instability in long rollouts. We investigate this in a controlled setting using the one-dimensional viscous Burgers equation, where a U-shaped neural operator (UNO) predicts the cell-centered diffusive contribution and is coupled to a finite-volume hyperbolic update. We find that (i) continuation error is strongly correlated with spectral error, indicating that high-frequency drift is the dominant failure mode, (ii) multi-step (substep) consistency losses provide effective implicit spectral regularization and substantially improve long-time stability compared with single-step MSE and explicit spectral penalties, and (iii) cross-scheme rollouts (MUSCL ↔ WENO5) demonstrate that the learned diffusion surrogate adapts to the training-time hyperbolic discretization, producing rapid instability when numerical diffusion in the hyperbolic flux is reduced and stable but over-diffusive behavior when it is increased. Finally, we show that an eigenvalue-based Jacobian diagnostic can detect unstable regimes.
- 1 June 2026
- Published by IOP Publishing in Journal of Physics: Conference Series
- Vol. 3252 (1)
Open AccessStability of Coupling Learned Diffusion to Shock Capturing Schemes
We propose a hybrid ML-numerical solver for mixed hyperbolic–parabolic PDEs in which a neural operator replaces only the parabolic (diffusive) contribution while a standard shock-capturing finite-volume method advances the hyperbolic (advective) fluxes. This selective coupling targets the parabolic timestep bottleneck without discarding the robustness of classical advection discretizations. The central challenge is long-horizon stability since small spectral errors in the learned diffusion update can accumulate under repeated timesteps and drive drift or instability in long rollouts. We investigate this in a controlled setting using the one-dimensional viscous Burgers equation, where a U-shaped neural operator (UNO) predicts the cell-centered diffusive contribution and is coupled to a finite-volume hyperbolic update. We find that (i) continuation error is strongly correlated with spectral error, indicating that high-frequency drift is the dominant failure mode, (ii) multi-step (substep) consistency losses provide effective implicit spectral regularization and substantially improve long-time stability compared with single-step MSE and explicit spectral penalties, and (iii) cross-scheme rollouts (MUSCL ↔ WENO5) demonstrate that the learned diffusion surrogate adapts to the training-time hyperbolic discretization, producing rapid instability when numerical diffusion in the hyperbolic flux is reduced and stable but over-diffusive behavior when it is increased. Finally, we show that an eigenvalue-based Jacobian diagnostic can detect unstable regimes.
- Open Access
- 10 April 2026
- Published by IOP Publishing in Machine Learning: Earth
- Vol. 2 (1) , 015013
Fourier analysis of the physics of transfer learning for data-driven subgrid-scale models of ocean turbulence
Transfer learning (TL) is a powerful tool for enhancing the performance of neural networks (NNs) in applications such as weather and climate prediction and turbulence modeling. TL enables models to generalize to out-of-distribution data with minimal training data from the new system. In this study, we employ a 9-layer convolutional NN to predict the subgrid forcing in a two-layer ocean quasi-geostrophic system and examine which metrics best describe its performance and generalizability to unseen dynamical regimes. Fourier analysis of the NN kernels reveals that they learn low-pass, Gabor, and high-pass filters, regardless of whether the training data are isotropic or anisotropic. By analyzing the activation spectra, we identify why NNs fail to generalize without TL and how TL can overcome these limitations: the learned weights and biases from one dataset underestimate the out-of-distribution sample spectra as they pass through the network, leading to an underestimation of output spectra. By re-training only one layer with data from the target system, this underestimation is corrected, enabling the NN to produce predictions that match the target spectra. These findings are broadly applicable to data-driven parameterization of dynamical systems.
- 10 April 2026
- Published by IOP Publishing in Machine Learning: Earth
- Vol. 2 (1) , 015013
Open AccessFourier analysis of the physics of transfer learning for data-driven subgrid-scale models of ocean turbulence
Transfer learning (TL) is a powerful tool for enhancing the performance of neural networks (NNs) in applications such as weather and climate prediction and turbulence modeling. TL enables models to generalize to out-of-distribution data with minimal training data from the new system. In this study, we employ a 9-layer convolutional NN to predict the subgrid forcing in a two-layer ocean quasi-geostrophic system and examine which metrics best describe its performance and generalizability to unseen dynamical regimes. Fourier analysis of the NN kernels reveals that they learn low-pass, Gabor, and high-pass filters, regardless of whether the training data are isotropic or anisotropic. By analyzing the activation spectra, we identify why NNs fail to generalize without TL and how TL can overcome these limitations: the learned weights and biases from one dataset underestimate the out-of-distribution sample spectra as they pass through the network, leading to an underestimation of output spectra. By re-training only one layer with data from the target system, this underestimation is corrected, enabling the NN to produce predictions that match the target spectra. These findings are broadly applicable to data-driven parameterization of dynamical systems.
- Open Access
Generative forecasting with joint probability models
Chaotic dynamical systems exhibit strong sensitivity to initial conditions and often contain unresolved multiscale processes, making deterministic forecasting fundamentally limited. Generative models offer an appealing alternative by learning distributions over plausible system evolutions; yet, most existing approaches focus on next-step conditional prediction rather than the structure of the underlying dynamics. In this work, we reframe forecasting as a fully generative problem by learning the joint probability distribution of lagged system states over short temporal windows and obtaining forecasts through marginalization. This new perspective allows the model to capture nonlinear temporal dependencies, represent multistep trajectory segments, and produce next-step predictions consistent with the learned joint distribution. We also introduce a general, model-agnostic training and inference framework for joint generative forecasting and show how it enables assessment of forecast robustness and reliability using three complementary uncertainty quantification metrics (ensemble variance, short-horizon autocorrelation, and cumulative Wasserstein drift), without access to ground truth. We evaluate the performance of the proposed method on two canonical chaotic dynamical systems, the Lorenz-63 system and the Kuramoto-Sivashinsky equation, and show that joint generative models yield improved short-term predictive skill, preserve attractor geometry, and achieve substantially more accurate long-range statistical behaviour than conventional conditional next-step models.
Open AccessGenerative forecasting with joint probability models
Chaotic dynamical systems exhibit strong sensitivity to initial conditions and often contain unresolved multiscale processes, making deterministic forecasting fundamentally limited. Generative models offer an appealing alternative by learning distributions over plausible system evolutions; yet, most existing approaches focus on next-step conditional prediction rather than the structure of the underlying dynamics. In this work, we reframe forecasting as a fully generative problem by learning the joint probability distribution of lagged system states over short temporal windows and obtaining forecasts through marginalization. This new perspective allows the model to capture nonlinear temporal dependencies, represent multistep trajectory segments, and produce next-step predictions consistent with the learned joint distribution. We also introduce a general, model-agnostic training and inference framework for joint generative forecasting and show how it enables assessment of forecast robustness and reliability using three complementary uncertainty quantification metrics (ensemble variance, short-horizon autocorrelation, and cumulative Wasserstein drift), without access to ground truth. We evaluate the performance of the proposed method on two canonical chaotic dynamical systems, the Lorenz-63 system and the Kuramoto-Sivashinsky equation, and show that joint generative models yield improved short-term predictive skill, preserve attractor geometry, and achieve substantially more accurate long-range statistical behaviour than conventional conditional next-step models.
- Open Access
- 4 December 2025
- Published by Wiley
An AI Emulator of Holton-Mass Stochastic Sudden Stratospheric Warming with Faithful Dynamical Representation and an Interpretable Latent Structure
Extreme climate events like sudden stratospheric warmings (SSWs) are rare yet impactful, posing significant modeling challenges due to their infrequent occurrence in historical data. AI-based emulators offer a fast and data-efficient alternative to traditional numerical models, so long as they can reliably represent rare transitions and internal variability. In this study, we develop a deep learning architecture tailored to emulate the stochastic Holton-Mass stratospheric model and investigate its latent space. The Holton-Mass model is capable of simulating both a standard polar vortex state and a disrupted vortex, warm stratosphere state, as well as rare transitions between the two states. The architecture we use is a ResNet-inspired Conditional Variational Autoencoder (CVAE) with 6-layer encoder/decoder stacks and state conditioning. The model is trained on 300,000 days of simulated real and imaginary streamfunction (ψ) fields, with the goal of forecasting ψ and mean zonal wind at the next time step. To handle the system’s inherent stochasticity, we use a KL-annealed training procedure and a specially weighted loss function that balances reconstruction and latent regularization. To interpret the learned representation, we apply Principal Component Analysis (PCA) to the latent mean vectors (μ) and discover a striking result: the encoded representations self-organize into well-separated clusters corresponding to four physical regimes (standard or warm stratosphere and stable or about to transition ). This degree of unsupervised regime separation in latent space is rare for deep generative models, particularly in high-dimensional, stochastic systems. This work contributes a scalable, interpretable emulator architecture for stochastic climate dynamics and introduces a latent space probing framework for diagnosing what AI models internalize about rare events. Our findings suggest that effectively designed deep emulators can not only accelerate simulation but may also uncover physically meaningful manifolds of variability through latent space interrogation. Future directions include incorporating rare-event sampling and developing disentangled latent models to further enhance interpretability and control.
- 4 December 2025
- Published by Wiley
Open AccessAn AI Emulator of Holton-Mass Stochastic Sudden Stratospheric Warming with Faithful Dynamical Representation and an Interpretable Latent Structure
Extreme climate events like sudden stratospheric warmings (SSWs) are rare yet impactful, posing significant modeling challenges due to their infrequent occurrence in historical data. AI-based emulators offer a fast and data-efficient alternative to traditional numerical models, so long as they can reliably represent rare transitions and internal variability. In this study, we develop a deep learning architecture tailored to emulate the stochastic Holton-Mass stratospheric model and investigate its latent space. The Holton-Mass model is capable of simulating both a standard polar vortex state and a disrupted vortex, warm stratosphere state, as well as rare transitions between the two states. The architecture we use is a ResNet-inspired Conditional Variational Autoencoder (CVAE) with 6-layer encoder/decoder stacks and state conditioning. The model is trained on 300,000 days of simulated real and imaginary streamfunction (ψ) fields, with the goal of forecasting ψ and mean zonal wind at the next time step. To handle the system’s inherent stochasticity, we use a KL-annealed training procedure and a specially weighted loss function that balances reconstruction and latent regularization. To interpret the learned representation, we apply Principal Component Analysis (PCA) to the latent mean vectors (μ) and discover a striking result: the encoded representations self-organize into well-separated clusters corresponding to four physical regimes (standard or warm stratosphere and stable or about to transition ). This degree of unsupervised regime separation in latent space is rare for deep generative models, particularly in high-dimensional, stochastic systems. This work contributes a scalable, interpretable emulator architecture for stochastic climate dynamics and introduces a latent space probing framework for diagnosing what AI models internalize about rare events. Our findings suggest that effectively designed deep emulators can not only accelerate simulation but may also uncover physically meaningful manifolds of variability through latent space interrogation. Future directions include incorporating rare-event sampling and developing disentangled latent models to further enhance interpretability and control.
- Open Access
- 11 November 2025
- Published by IOP Publishing in Machine Learning: Earth
- Vol. 1 (1) , 015009
Generative Lagrangian data assimilation for ocean dynamics under extreme sparsity
Reconstructing ocean dynamics from observational data is fundamentally limited by the sparse, irregular, and Lagrangian nature of spatial sampling, particularly in subsurface and remote regions. This sparsity poses significant challenges for forecasting key phenomena such as eddy shedding and rogue waves. Traditional data assimilation methods and deep learning models often struggle to recover mesoscale turbulence under such constraints. We leverage a deep learning framework that combines neural operators with denoising diffusion probabilistic models to reconstruct high-resolution ocean states from extremely sparse Lagrangian observations. By conditioning the generative model on neural operator outputs, the framework accurately captures small-scale, high-wavenumber dynamics even at 99% sparsity (for synthetic data) and 99.9% sparsity (for real satellite observations). We validate our method on benchmark systems, synthetic float observations, and real satellite data, demonstrating robust performance under severe spatial sampling limitations as compared to other deep learning baselines.
- 11 November 2025
- Published by IOP Publishing in Machine Learning: Earth
- Vol. 1 (1) , 015009
Open AccessGenerative Lagrangian data assimilation for ocean dynamics under extreme sparsity
Reconstructing ocean dynamics from observational data is fundamentally limited by the sparse, irregular, and Lagrangian nature of spatial sampling, particularly in subsurface and remote regions. This sparsity poses significant challenges for forecasting key phenomena such as eddy shedding and rogue waves. Traditional data assimilation methods and deep learning models often struggle to recover mesoscale turbulence under such constraints. We leverage a deep learning framework that combines neural operators with denoising diffusion probabilistic models to reconstruct high-resolution ocean states from extremely sparse Lagrangian observations. By conditioning the generative model on neural operator outputs, the framework accurately captures small-scale, high-wavenumber dynamics even at 99% sparsity (for synthetic data) and 99.9% sparsity (for real satellite observations). We validate our method on benchmark systems, synthetic float observations, and real satellite data, demonstrating robust performance under severe spatial sampling limitations as compared to other deep learning baselines.
- Open Access
- 6 November 2025
- Published by American Geophysical Union (AGU) in Journal of Advances in Modeling Earth Systems
- Vol. 17 (11)
LUCIE: A Lightweight Uncoupled Climate Emulator With Long‐Term Stability and Physical Consistency
We present a lightweight, easy‐to‐train, low‐resolution, fully data‐driven climate emulator, LUCIE, that can be trained on as low as 2 years of 6‐hourly ERA5 data. Unlike most state‐of‐the‐art AI weather models, LUCIE remains stable and physically consistent for 100 years of autoregressive simulation with 100 ensemble members. Long‐term mean climatology from LUCIE's simulation of temperature, wind, precipitation, and humidity matches that of ERA5 data, along with the variability. We further demonstrate how well extreme weather events and their return periods can be estimated from a large ensemble of long‐term simulations. We further discuss an improved training strategy with a hard‐constrained first‐order integrator to suppress autoregressive error growth, a novel spectral regularization strategy to better capture fine‐scale dynamics, and finally an optimization algorithm that enables data‐limited (as low as 2 years of 6‐hourly data) training of the emulator without losing stability and physical consistency. Finally, we provide a scaling experiment to compare the long‐term bias of LUCIE with respect to the number of training samples. Importantly, LUCIE is an easy to use model that can be trained in just 2.4 hr on a single A‐100 GPU, allowing for multiple experiments that can explore important scientific questions that could be answered with large ensembles of long‐term simulations, for example, the impact of different variables on the simulation, dynamic response to external forcing, and estimation of extreme weather events, amongst others.
- 6 November 2025
- Published by American Geophysical Union (AGU) in Journal of Advances in Modeling Earth Systems
- Vol. 17 (11)
Open AccessLUCIE: A Lightweight Uncoupled Climate Emulator With Long‐Term Stability and Physical Consistency
We present a lightweight, easy‐to‐train, low‐resolution, fully data‐driven climate emulator, LUCIE, that can be trained on as low as 2 years of 6‐hourly ERA5 data. Unlike most state‐of‐the‐art AI weather models, LUCIE remains stable and physically consistent for 100 years of autoregressive simulation with 100 ensemble members. Long‐term mean climatology from LUCIE's simulation of temperature, wind, precipitation, and humidity matches that of ERA5 data, along with the variability. We further demonstrate how well extreme weather events and their return periods can be estimated from a large ensemble of long‐term simulations. We further discuss an improved training strategy with a hard‐constrained first‐order integrator to suppress autoregressive error growth, a novel spectral regularization strategy to better capture fine‐scale dynamics, and finally an optimization algorithm that enables data‐limited (as low as 2 years of 6‐hourly data) training of the emulator without losing stability and physical consistency. Finally, we provide a scaling experiment to compare the long‐term bias of LUCIE with respect to the number of training samples. Importantly, LUCIE is an easy to use model that can be trained in just 2.4 hr on a single A‐100 GPU, allowing for multiple experiments that can explore important scientific questions that could be answered with large ensembles of long‐term simulations, for example, the impact of different variables on the simulation, dynamic response to external forcing, and estimation of extreme weather events, amongst others.