Spatiotemporal Engression Models

• Spatiotemporal engression processes are a family of models that learn non-deterministic, conditional distributions of space-time fields using stochastic noise injection.
• They integrate neural architectures such as LSTMs, GRUs, graph convolutions, and attention mechanisms to capture dynamic dependencies in complex systems.
• These models employ scalable inference methods like energy score minimization and invariant embeddings to ensure robust, distributional predictions across varied applications.

Spatiotemporal engression processes denote a family of models for representing uncertainty, dependence, and emergence in systems that evolve over space and time. In the most explicit usage, engression is neural distributional regression: instead of learning only a point predictor, the model learns a conditional sampler for the full law of a spatiotemporal field given observed inputs, typically by injecting stochastic noise before or within the forward map. Closely related literatures formulate the same objective as posterior predictive inference over stochastic processes, conditional intensity modeling of events, or event-centric network simulation of emergent dynamics. The terminology is therefore heterogeneous: some papers define engression directly, whereas others state that the term does not appear and interpret the topic through spatiotemporal point processes or event-based spatiotemporal networks (Gao et al., 1 Apr 2026, Pathak et al., 7 Mar 2026, Jenson et al., 10 Jun 2025, Mukherjee et al., 13 Feb 2025, Romeijnders et al., 15 May 2026).

1. Terminological scope and conceptual variants

Within uncertainty-aware sparse sensing, engression is defined as neural distributional regression that learns the full conditional law of the output given the input rather than a single point estimate. UQ-SHRED formulates this as a conditional sampler

$Y = g_\theta(X,\epsilon),$

where $\epsilon \sim P_\epsilon$ and $\epsilon \perp X$, so the predictive distribution at $X=x$ is induced by the map $\epsilon \mapsto g_\theta(x,\epsilon)$. In that setting, the target is a high-dimensional spatial field $x_t \in \mathbb{R}^n$ conditioned on a sensor-history lag window $h_t \in \mathbb{R}^{L \times p}$, and engression is the mechanism that turns a recurrent decoder into a distributional learner (Gao et al., 1 Apr 2026).

In epidemic forecasting, engression is defined more specifically through a pre-additive noise model. The paper contrasts this with post-additive noise models and states: for regression problems, noise $\eta$ is added directly to the covariates before they are passed through a nonlinear transformation, represented as $Y=g(X+\eta)$, whereas conventional post-additive formulations use $Y=g(X)+\eta$. The resulting network acts as a “distributional lens”: sampling the input-side noise generates samples from the learned conditional distribution without requiring explicit parametric likelihoods or MCMC (Pathak et al., 7 Mar 2026).

Other adjacent papers align with the same inferential goal but do not use the term. BSA-TNP describes spatiotemporal regression as inference of the posterior predictive distribution of a stochastic process over space and time, $\epsilon \sim P_\epsilon$0, within the Neural Process framework (Jenson et al., 10 Jun 2025). The STPP review states explicitly that the established concept matching the phrase is spatiotemporal point processes, where the central object is the conditional intensity of events in continuous space-time (Mukherjee et al., 13 Feb 2025). EBSTNs go further in a different direction: the paper states that the term “engression” does not appear, and interprets the query in context as emergence from micro-events, with event-based spatiotemporal networks providing a computational framework in which discrete events, rather than static states, are the primitive units of macroscopic dynamics (Romeijnders et al., 15 May 2026).

2. Mathematical formulations

A common mathematical thread is that uncertainty is represented through a conditional distribution rather than a deterministic trajectory. In UQ-SHRED, the modeling target is

$\epsilon \sim P_\epsilon$1

with forward map

$\epsilon \sim P_\epsilon$2

where $\epsilon \sim P_\epsilon$3 is an LSTM or GRU operating on the augmented lag window and $\epsilon \sim P_\epsilon$4 is a shallow decoder. The same $\epsilon \sim P_\epsilon$5 is concatenated to each time step in the lag window, and resampling $\epsilon \sim P_\epsilon$6 at test time yields predictive samples from $\epsilon \sim P_\epsilon$7 (Gao et al., 1 Apr 2026).

The epidemic engression models define a closed-loop stochastic recurrence on a state vector $\epsilon \sim P_\epsilon$8:

$\epsilon \sim P_\epsilon$9

with $\epsilon \perp X$0 in the general theorem and, for GCEN/STEN, an extended state containing outputs and LSTM hidden and cell states. This induces the conditional law $\epsilon \perp X$1 through the deterministic network map and the distribution of the pre-additive noise. Because the noises are i.i.d. across time, the process is homogeneous Markovian (Pathak et al., 7 Mar 2026).

In BSA-TNP, the central object is the posterior predictive map of a stochastic process observed through context data $\epsilon \perp X$2 and queried at targets $\epsilon \perp X$3. The model directly parameterizes $\epsilon \perp X$4, and the paper proves that a process is $\epsilon \perp X$5-stationary if and only if the posterior predictive map $\epsilon \perp X$6 is $\epsilon \perp X$7-invariant in its inputs:

$\epsilon \perp X$8

This theorem turns invariance from a modeling intuition into an architectural prescription (Jenson et al., 10 Jun 2025).

The STPP formalism uses a different but closely related object: the conditional intensity function

$\epsilon \perp X$9

with likelihood

$X=x$0

In Hawkes-type models,

$X=x$1

so previous events modulate future event risk through a triggering kernel (Mukherjee et al., 13 Feb 2025).

EBSTNs replace intensity-based event prediction with an explicit event graph. Each event is encoded as

$X=x$2

optionally with duration $X=x$3, where $X=x$4 is a spatial coordinate or compartment, $X=x$5 a start time, $X=x$6 an event type or class, and $X=x$7 event attributes. Event-to-event influence is written generically as

$X=x$8

with domain-specific realizations such as per-actor infection hazard or max-plus delay propagation (Romeijnders et al., 15 May 2026).

3. Principal architectural realizations

UQ-SHRED extends SHRED by concatenating random noise to sparse sensor inputs and processing the augmented sequence with a recurrent encoder and shallow decoder. The model remains single-architecture: no retraining, ensembling, or auxiliary networks are required, because uncertainty is produced entirely by resampling the input noise $X=x$9. The paper emphasizes minimal computational overhead, with training using two forward passes per batch under a single-sample energy-score approximation and inference cost scaling linearly in the number of resamples $\epsilon \mapsto g_\theta(x,\epsilon)$0 (Gao et al., 1 Apr 2026).

Deep epidemic engression is instantiated through three lightweight generative architectures. MVEN is a multivariate temporal LSTM-engression model in which noise is injected directly into the input tensor. GCEN combines a graph convolutional spatial encoder with an LSTM temporal module; the graph is built from Haversine distances through a thresholded Gaussian adjacency. STEN replaces graph convolution with a STAR-inspired differentiable spatial layer based on powers of inverse-distance spatial weights $\epsilon \mapsto g_\theta(x,\epsilon)$1, followed by an LSTM and dense multi-horizon head. In GCEN and STEN, the pre-additive noise is injected into spatial embeddings before temporal processing, either additively or by concatenation (Pathak et al., 7 Mar 2026).

BSA-TNP belongs to the Neural Process family and is specialized for scalable spatiotemporal inference under full or partial translation invariance and more general group invariances. Its key components are Kernel Regression Blocks, group-invariant attention biases, and Biased Scan Attention. The attention rule is

$\epsilon \mapsto g_\theta(x,\epsilon)$2

with bias

$\epsilon \mapsto g_\theta(x,\epsilon)$3

For spatiotemporal translation invariance, the spatial and temporal biases depend only on relative positions. The scan-based implementation computes attention chunk-by-chunk without materializing the full bias matrix, allowing arbitrary bias functions together with constant memory with respect to sequence length (Jenson et al., 10 Jun 2025).

These realizations differ in what is treated as primitive. UQ-SHRED and the epidemic models are explicit conditional samplers driven by injected noise. BSA-TNP is a posterior predictive model whose uncertainty is encoded in the predictive distribution head and invariant attention structure. A plausible implication is that “spatiotemporal engression” is best understood as a modeling objective—distributional inference over space-time—rather than as a single fixed architecture.

4. Event-based and point-process perspectives

The event-centered interpretation broadens the concept beyond field reconstruction and direct forecasting. STPPs model realizations as finite ordered sets of events $\epsilon \mapsto g_\theta(x,\epsilon)$4, possibly with marks, and use conditional intensity functions to describe the instantaneous rate of event occurrence. Neural STPPs then parameterize the intensity through recurrent encoders, transformers, continuous-time flows, diffusion models, CNN-based covariate encoders, GNNs on networks, or neural Hawkes-type kernels. The review highlights both continuous-space-time modeling and grid-based approximations, together with simulation by spatiotemporal Ogata thinning and calibration diagnostics based on time rescaling (Mukherjee et al., 13 Feb 2025).

EBSTNs offer a distinct but related alternative. Events become nodes, and actor trajectories linking events become edges. The paper presents this as a dual representation to conventional temporal networks, one in which causal interdependencies among processes become explicit and simulation proceeds event-to-event rather than by iterating over all actor states at uniform time steps. In epidemiology, an event may be pathogen transmission in a room during a particular hour; in transport, a train occupying a blocking section for a scheduled run. Actor mobility constraints and event durations are encoded explicitly, so only feasible interactions occur, and fragmentation into fully connected components at each step avoids the $\epsilon \mapsto g_\theta(x,\epsilon)$5 contact search cost of agent-based models (Romeijnders et al., 15 May 2026).

The same paper explicitly extends this interpretation to developmental biology. It states that, for domain-specific ingression in developmental biology, EBSTNs provide a natural way to encode discrete topological transitions and their mechanical and chemical consequences as events. The suggested event vocabulary includes neighbor exchange/T1, T2, cell delamination/ingression, mitotic rounding, apoptosis, focal adhesion formation, and localized contractility increases. This is presented as a way to propagate their effects across space and time and to model tissue-level shape changes as emergence from micro-scale transitions (Romeijnders et al., 15 May 2026).

A common misconception is that event-based models and engression models are interchangeable. The papers do not support that equivalence. Engression papers center conditional distribution learning through stochastic forward maps, whereas STPPs center intensities and EBSTNs center event graphs and influence rules. The alignment lies in their shared concern with uncertainty, dependence, and emergence in space-time, not in a single formalism.

5. Training, inference, and empirical evaluation

The principal training criterion for the explicit engression models is the energy score. In UQ-SHRED, for predictive law $\epsilon \mapsto g_\theta(x,\epsilon)$6 induced by $\epsilon \mapsto g_\theta(x,\epsilon)$7, the objective minimizes

$\epsilon \mapsto g_\theta(x,\epsilon)$8

with $\epsilon \mapsto g_\theta(x,\epsilon)$9. The empirical loss uses two independent noise draws per sample. Evaluation additionally reports coordinate-wise predictive intervals, empirical coverage, CRPS, RMSE, and calibration diagrams. Reported calibration behavior includes, for SST, observed coverage percentages of $x_t \in \mathbb{R}^n$0 at nominal levels $x_t \in \mathbb{R}^n$1, together with $x_t \in \mathbb{R}^n$2 and $x_t \in \mathbb{R}^n$3; analogous calibration summaries are reported for isotropic turbulence, neural activity, solar activity, and a rotating detonation engine transient (Gao et al., 1 Apr 2026).

The epidemic engression models also optimize an energy score, with $x_t \in \mathbb{R}^n$4 and ensemble size $x_t \in \mathbb{R}^n$5 during training, and generate out-of-sample probabilistic forecasts by repeated forward passes with independently sampled noise tensors. Point forecasts are ensemble medians and prediction intervals are empirical quantiles. Across six epidemiological datasets over three forecast horizons, the proposed models achieved best CRPS in $x_t \in \mathbb{R}^n$6 of cases. The paper further states that proposed models yield sharp, informative prediction intervals with strong CRPS and Winkler scores, while PIT Q-Q plots exhibit slight S-shaped curves indicating mild overconfidence but remain close to the diagonal across horizons and datasets (Pathak et al., 7 Mar 2026).

BSA-TNP uses conditional Neural Process training by maximizing the log likelihood of targets given contexts. For regression tasks, the decoder often outputs Gaussian parameters so that

$x_t \in \mathbb{R}^n$7

The empirical emphasis is on accuracy, calibration, and scalability. On off-grid 2D Gaussian Processes, BSA-TNP achieves best NLL and fastest training and inference; on epidemiological SIR simulations it matches the best NLL on the training domain and remains stable under shifts and scales; and the paper reports inference with more than $x_t \in \mathbb{R}^n$8 million test points and $x_t \in \mathbb{R}^n$9 thousand context points in under a minute on a single $h_t \in \mathbb{R}^{L \times p}$0GB GPU (Jenson et al., 10 Jun 2025).

These studies collectively treat probabilistic quality as a first-class criterion. Point accuracy remains important, but calibration, interval width, CRPS, NLL, and coverage are repeatedly used to judge whether the learned stochastic mechanism is informative rather than merely noisy.

6. Theoretical properties, applications, and limitations

The strongest formal results concern distributional correctness, invariance, and stochastic stability. UQ-SHRED states that under mild regularity, any population minimizer of the energy-score risk recovers the true conditional law almost everywhere; under finite second moments, predictive mean and covariance match the true conditional moments; and empirical quantiles converge almost surely as the number of Monte Carlo samples grows (Gao et al., 1 Apr 2026). BSA-TNP proves the equivalence between $h_t \in \mathbb{R}^{L \times p}$1-stationarity of a process and $h_t \in \mathbb{R}^{L \times p}$2-invariance of the posterior predictive map, thereby justifying invariant embeddings and attention biases as exact design principles rather than heuristic inductive biases (Jenson et al., 10 Jun 2025).

The epidemic engression paper develops explicit Markov-chain stability theory. It defines irreducibility and geometric ergodicity, uses Tweedie’s drift criterion, and proves that if the pre-additive recurrent map satisfies a contraction bound $h_t \in \mathbb{R}^{L \times p}$3 with $h_t \in \mathbb{R}^{L \times p}$4, then the process is geometrically ergodic. Under additional assumptions on bounded spatial output, bounded dense output, and forget-gate contraction,

$h_t \in \mathbb{R}^{L \times p}$5

the GCEN, STEN, and MVEN processes are geometrically ergodic and asymptotically stationary (Pathak et al., 7 Mar 2026).

The application range is correspondingly broad. UQ-SHRED targets sparse sensing in turbulent flow, atmospheric dynamics, neuroscience, astrophysics, and a rotating detonation engine transient (Gao et al., 1 Apr 2026). BSA-TNP is evaluated on geology-style Gaussian Processes, epidemiology, climate forecasting, air quality, and generic spatial modeling against NUTS MCMC (Jenson et al., 10 Jun 2025). Neural STPPs span crime, earthquakes, epidemiology, transportation, and wildfires (Mukherjee et al., 13 Feb 2025). EBSTNs are demonstrated on pathogen transmission in the Netherlands and delay propagation in the Zürich S-bahn, and are discussed in relation to developmental biology and community ecology (Romeijnders et al., 15 May 2026).

The limitations are equally diverse. UQ-SHRED notes that mild miscalibration can appear under finite data or model misspecification, that too small a noise dimension leads to underdispersed predictions, and that no finite-sample coverage guarantees are provided (Gao et al., 1 Apr 2026). The epidemic engression models note extreme peak underprediction, sensitivity to the spatial graph or weights, empirical choice of Uniform versus Gaussian noise, and the absence of exogenous covariates in the reported study (Pathak et al., 7 Mar 2026). BSA-TNP still has quadratic time in the number of context points, and translation invariance may mis-specify strongly nonstationary or anisotropic processes if suitable fixed effects or bias functions are not supplied (Jenson et al., 10 Jun 2025). STPPs continue to face difficult continuous-space-time integration, robustness to missingness and covariate shift, limited Bayesian uncertainty quantification, and fairness concerns in high-stakes domains (Mukherjee et al., 13 Feb 2025). EBSTNs identify identifiability and attribution problems, sensitivity to event definition, non-stationarity of future event schedules, simultaneity and interference among concurrent events, missing data, and unresolved questions about formal coarse-graining and renormalization of event kernels (Romeijnders et al., 15 May 2026).

Taken together, the literature supports a precise but plural conclusion: spatiotemporal engression processes are not a single standardized model class, but a convergent set of distributional, event-centric, and posterior-predictive approaches for learning how micro-level uncertainty and interaction generate macro-level behavior across space and time.