Dynamical Modeling of Spatiotemporal Fields

• Dynamical modeling of spatiotemporal fields is a framework that uses neural, operator, and kernel methods to represent and forecast complex systems in space and time.
• It leverages latent dynamics, neural implicit fields, and variational inference to achieve grid-free prediction and effective uncertainty quantification.
• Applications span climate science, neuroscience, and engineering, offering significant improvements in speed and accuracy for real-world data analysis.

Dynamical Modeling of Spatiotemporally Evolving Fields

Dynamical modeling of spatiotemporally evolving fields addresses the inference, prediction, and analysis of fields that evolve in both space and time under often-complex, potentially nonlinear dynamics. These fields arise in myriad disciplines including climate science, materials, neuroscience, geophysical monitoring, and engineering. The challenge is to learn, represent, and forecast these fields from diverse observational regimes—ranging from gridded sensor arrays to sparse, randomly sampled point process data—while ensuring computational tractability, physical interpretability, and robustness to noise and missing data.

1. Foundations and Motivation

A spatiotemporal field can be formalized as a function $u(\mathbf{x}, t)$ defined over a spatial domain $\mathbf{x} \in \mathbb{R}^d$ and time $t \in [0,T]$. The evolution of $u$ is typically governed by partial differential equations (PDEs), stochastic differential equations (SDEs), or integral operators. Practical modeling must reconcile high-dimensionality, irregular or sparse observations, and partial knowledge of governing laws.

Classical approaches relied on dense-grid numerical solvers, spectral methods, or linear models (e.g., vector autoregressions). Control theory introduced kernel-based methods layered with finite-dimensional linear dynamics for estimation and control under partial observability (Kingravi et al., 2015). However, the last decade has seen an explosion in data-driven and hybrid approaches—neural implicit fields, neural differential operators, dynamic mode decomposition (DMD), operator-valued kernel frameworks, and advanced reduced-order models—enabling mesh-free prediction, uncertainty quantification, and scalability to complex regimes (Iakovlev et al., 2024, Chen et al., 2023, Chen et al., 11 Mar 2026, Kim et al., 25 Nov 2025, Withanachchi, 23 Aug 2025, Szehr et al., 2020, Zhang et al., 1 Dec 2025).

2. Latent Dynamical Representations and Neural Implicit Fields

Recent advances leverage low-dimensional latent representations evolving according to parametric ODEs or SDEs, coupled with meshless neural decoders that reconstruct the full field at arbitrary query points. A canonical paradigm posits an unknown continuous field $u(\mathbf{x}, t)$ represented via

$$u(\mathbf{x}, t) = \phi\bigl(\mathbf{z}(t),\,\mathbf{x}\bigr),$$

where $\mathbf{z}(t) \in \mathbb{R}^{d_z}$ evolves by a neural ODE

$$\frac{d\mathbf{z}}{dt}(t) = f(\mathbf{z}(t)), \qquad \mathbf{z}(t_1)\sim p(\mathbf{z}_1),$$

and $\phi$ is an implicit neural representation, typically an MLP with learnable embeddings for $\mathbf{x}$ (Iakovlev et al., 2024, Regazzoni et al., 2023). This enables continuous, grid-free field evaluation and naturally separates spatial and temporal complexities.

Latent Dynamics Networks (LDNets) further refine this approach by never operating in the high-dimensional field space directly; instead, they define all dynamics on the low-dimensional manifold and reconstruct at any spatial point via a meshless decoder, with the decoder’s weights shared across all space (Regazzoni et al., 2023).

3. Observation Models, Inference, and Learning

Spatiotemporal fields are often observed under severe subsampling, at random (potentially inhomogeneous) spatial locations and irregular or event-driven times. This prohibits standard discretized tensor-based architectures. One solution is to model data acquisition as a nonhomogeneous Poisson process (NHPP) on space-time, with intensity $\mathbf{x} \in \mathbb{R}^d$0, capturing the probabilistic arrival of events (Iakovlev et al., 2024). Given event $\mathbf{x} \in \mathbb{R}^d$1, the field value is modeled as a Gaussian emission centered at $\mathbf{x} \in \mathbb{R}^d$2.

Inference and learning utilize amortized variational methods, introducing a Transformer encoder over observation triplets to produce a posterior over the initial latent state, optimizing an evidence lower bound (ELBO) combining emission errors, point process log-likelihoods, and KL-divergence to the prior (Iakovlev et al., 2024). ODE integration over a sparse auxiliary temporal grid with interpolation dramatically accelerates learning and inference without loss of field accuracy.

Alternatively, stochastic dynamic mode decomposition (NODE-DMD) recovers a field as a time-varying superposition of learned spatial modes under a latent SDE, enabling uncertainty quantification and direct interpretability in terms of modal structure and continuous-time eigenvalues (Kim et al., 25 Nov 2025).

4. Operator Learning and Dynamic Mode Decomposition

Operator-theoretic perspectives encode the dynamics of spatiotemporal fields as evolution under a (potentially nonlinear) infinite-dimensional operator, motivating neural operator architectures and dynamic mode decomposition approaches.

Neural dynamical operators (e.g., Fourier Neural Operator) learn parametric mappings $\mathbf{x} \in \mathbb{R}^d$3 that advance the field in time via integral (or spectral) kernels,

$\mathbf{x} \in \mathbb{R}^d$4

with forward evaluation unconstrained by grid resolution or time step. A key hallmark is "resolution-invariance": the same trained operator generalizes across discretization levels and temporal increments (Chen et al., 2023). Hybrid optimization schemes leverage both gradient-based updates (for short-term error) and derivative-free ensemble Kalman inversion (for long-run statistical property matching).

Dynamic mode decomposition (DMD) and Koopman operator frameworks decouple a field into modes and continuous-time growth rates,

$\mathbf{x} \in \mathbb{R}^d$5

with $\mathbf{x} \in \mathbb{R}^d$6 (complex eigenvalues) explicit, enabling modal-based forecasting and interpretability. Neural architectures, e.g., NeuralDMD, integrate implicit neural representations with DMD structure, supporting reconstruction and stable extrapolation from highly sparse measurements (SaraerToosi et al., 3 Jul 2025, Chen et al., 11 Mar 2026). Probabilistic NODE-DMD augments traditional DMD with a low-dimensional stochastic latent evolving under a residual SDE, providing fully Bayesian inference (Kim et al., 25 Nov 2025).

Kernel–based Koopman methods extend this decomposition to nonparametric settings by constructing operator-valued reproducing kernel Hilbert spaces and learning spectral decompositions from data, with theoretical convergence guarantees for both interpolation and forecast stages (Withanachchi, 23 Aug 2025).

5. Statistical and Kernel Methods for Spatiotemporal Fields

Kernel-based approaches enable mesh-independent, theoretically grounded nonparametric modeling. A prototypical method defines a field expansion in terms of time-dependent Green’s function kernels tied to the underlying PDE’s spectrum,

$\mathbf{x} \in \mathbb{R}^d$7

with $\mathbf{x} \in \mathbb{R}^d$8 the eigenpairs of the spatial operator. The regularized least-squares estimator is then

$\mathbf{x} \in \mathbb{R}^d$9

with coefficients $t \in [0,T]$0 estimated from observed data (Szehr et al., 2020). This provides exact satisfaction of the underlying PDE and enables efficient solution via kernel regression with only minor adjustments for initial/boundary conditions or sparsity.

Operator-valued RKHS (OV-RKHS) and kernel-based Koopman operator learning further generalize this to vector-valued and high-dimensional fields, with Sobolev-type rates and spectral convergence theorems governing the quality of approximation and long-horizon prediction (Withanachchi, 23 Aug 2025).

6. Model Reduction, Control, and Real-World Applications

Surrogate models (reduced-order models, ROMs) constructed by projecting observed fields onto low-dimensional bases (e.g., POD, DMD, kernel bases) serve both as emulators for real-time prediction and as tools for design, optimization, and control (Chang et al., 2018, Chang et al., 2021). Modern approaches (e.g., CKSPOD) address modal phase-alignment—critical for capturing coherent structures—by synthesizing a common Gram matrix shared across parameter variations, subsequently leveraging Gaussian process regression for parameter interpolation and providing full uncertainty quantification (Chang et al., 2021). The result is up to 10^5× speedup over direct simulation with accurate recovery of flow statistics and small-scale structures.

Systems-theoretic kernel controllers retrofit classical estimation and control (Kalman filtering, LQR) onto a kernel-evolved latent space, yielding direct quantification of observability/controllability and provable sensor/actuator placement criteria (Kingravi et al., 2015). Similarly, functional time-series models such as SFDE-driven array regressions with sparse network estimation efficiently recover functional connectivity in neuroscience and generalize to array-valued dynamical fields with delay structure via scalable proximal algorithms (Lund et al., 2018).

Applications span satellite sea surface temperature inference, pollutant transport, neural field imaging, global air quality (e.g., analysis of wildfire events), and video-based 4D scene reconstruction. State-of-the-art frameworks routinely achieve substantial error reductions versus prior mesh-based or black-box machine learning baselines—e.g., 40–80% reduction in MAE, log-likelihood gains of 0.2–1.0 nats/event, and marked robustness to structured data sparsity (Iakovlev et al., 2024, Chen et al., 2023, SaraerToosi et al., 3 Jul 2025, Kim et al., 25 Nov 2025, Regazzoni et al., 2023).

7. Current Limitations and Research Directions

Despite significant progress, open challenges remain. Many current models assume non-overlapping events and Poisson process arrivals, precluding self-exciting dynamics (Iakovlev et al., 2024). Scalability, especially for true online learning or extreme spatiotemporal resolutions (N > 10^5, multi-decade time scales), invites further algorithmic innovation, including randomized or multiscale kernel methods, sparse tensor strategies, and hierarchical neural operator design.

Integrating known physical symmetries or constraints (e.g., incompressibility, invariances), explicit non-Euclidean geometry (spherical manifolds, SE(3) actions for video/graphics), and efficient uncertainty quantification (e.g., distributional outputs, Bayesian neural operators, kernel-based process models) remain active areas of research (Qiao et al., 25 Feb 2026, Zhang et al., 1 Dec 2025, Withanachchi, 23 Aug 2025). There is increasing emphasis on explaining the learned dynamics (interpretability), fusing heterogeneous data types (point process, grid, indirect inverse), and combining short-run trajectory fidelity with long-run statistical and invariant matching (Chen et al., 2023). Limitations of black-box surrogates in chaotic, non-stationary, or extrapolative regimes suggest opportunities for models that blend strict physics with flexible learning.

The field is rapidly converging toward general-purpose, efficient, and physically consistent frameworks that can accommodate real-world complexities—arbitrary sensor placement, multi-physics coupling, and operational forecast requirements—while retaining scientific interpretability and strong theoretical footing.