EnScale-t: Temporal Generative Downscaling

• The paper introduces EnScale-t, a generative framework that extends EnScale by conditioning on the previous day to produce temporally coherent high-resolution climate fields.
• It factorizes the downscaling process into a temporal coarse model using dense MLPs and a spatial super-resolution cascade that refines low-resolution fields.
• Empirical results show that EnScale-t nearly eliminates lag-1 autocorrelation bias, thereby enhancing the simulation of multiday events while remaining computationally efficient.

Searching arXiv for the specified paper and closely related downscaling work to ground the article. EnScale-t is the temporally consistent variant of EnScale, a generative machine-learning framework for statistical downscaling of climate data that emulates the conditional distribution of high-resolution regional climate model (RCM) fields given coarse global circulation model (GCM) fields. In the formulation studied for Central Europe, the predictors are daily coarse-resolution GCM fields $X$ over Europe, comprising tas, pr, sfcWind, rsds, and GCM psl, while the targets are corresponding high-resolution RCM fields $Y$ over Central Europe for tas, pr, sfcWind, and rsds. EnScale-t extends EnScale’s two-step stochastic map $X_t \to Z_t \to Y_t$ by conditioning the coarse state on the previous day, thereby generating multivariate daily high-resolution fields with realistic temporal dependence while retaining the spatial and multivariate generative design of the base framework (Schillinger et al., 30 Sep 2025).

1. Problem formulation and relation to EnScale

The underlying downscaling problem is to approximate the conditional distribution

$p_{Y\mid X}(y\mid x),$

interpreted as the distribution of possible RCM realizations given one GCM boundary-condition state. In this setting, the variability of $Y$ conditional on $X$ reflects internal variability, so the objective is not a deterministic predictor but a stochastic emulator of the full GCM-to-RCM map. EnScale addresses this by introducing coarsened RCM fields $Z$, obtained by average pooling $Y$ over $16\times 16$ patches, reducing resolution from $128\times 128$ to $Y$0. It then factorizes generation into a coarse model $Y$1, which captures large-scale GCM–RCM mismatch, and a super-resolution model $Y$2, which reconstructs full-resolution fields from coarsened RCM states (Schillinger et al., 30 Sep 2025).

Under the conditional independence approximation

$Y$3

the overall conditional law is approximated by

$Y$4

This decomposition separates large-scale bias correction and variability adjustment from the high-dimensional spatial refinement problem. A common misconception is that EnScale-t replaces this factorization with a fully temporal high-resolution model. It does not. The temporal extension leaves the super-resolution stage structurally unchanged and introduces temporal dependence only in the coarse latent dynamics.

Both EnScale and EnScale-t are trained as stochastic neural networks with Gaussian noise injected as a latent source. Their optimization target is the multivariate energy score, used here as a proper scoring rule for conditional generative modeling. In the conditional setting, with model law $Y$5, the loss is

$Y$6

Its key property is strict propriety: the loss is minimized if and only if $Y$7 for $Y$8-almost every $Y$9.

2. Temporal extension and probabilistic assumptions

The suffix “-t” denotes temporal. Whereas EnScale generates each day independently, EnScale-t conditions on the previous day and aims to emulate the conditional process $X_t \to Z_t \to Y_t$0. In the paper’s terminology, this makes EnScale-t effectively a first-order Markov emulator in time at the RCM scale, akin to autoregressive AI weather forecasting but driven by daily GCM fields (Schillinger et al., 30 Sep 2025).

Temporal consistency is defined operationally through properties of generated sequences rather than through a separate dynamical model at full resolution. The relevant properties include lag-1 autocorrelation of daily fields at each grid point, realistic evolution of multi-day events such as heatwaves or wet spells, consistent temporal co-evolution of tas, pr, rsds, and sfcWind, and temporal coherence of spatial patterns such as moving synoptic systems. The paper quantifies this primarily with the lag-1 autocorrelation

$X_t \to Z_t \to Y_t$1

computed separately at each grid cell and for each variable.

Directly modeling $X_t \to Z_t \to Y_t$2 is treated as impractical because $X_t \to Z_t \to Y_t$3 is high-dimensional. EnScale-t therefore introduces a temporal coarse model

$X_t \to Z_t \to Y_t$4

and retains the same decoder $X_t \to Z_t \to Y_t$5 used in base EnScale. The temporal construction rests on two explicit assumptions: $X_t \to Z_t \to Y_t$6 and

$X_t \to Z_t \to Y_t$7

Under these assumptions, generation proceeds through a stochastic transition from $X_t \to Z_t \to Y_t$8 to $X_t \to Z_t \to Y_t$9, followed by conditional super-resolution from $p_{Y\mid X}(y\mid x),$0 to $p_{Y\mid X}(y\mid x),$1.

This design localizes temporal memory to the low-dimensional large-scale state. The resulting high-resolution time series is therefore approximated by a Markov chain on $p_{Y\mid X}(y\mid x),$2 driven by $p_{Y\mid X}(y\mid x),$3, together with conditional draws $p_{Y\mid X}(y\mid x),$4. The authors numerically test the key approximation $p_{Y\mid X}(y\mid x),$5 by comparing energy-score losses with and without conditioning on $p_{Y\mid X}(y\mid x),$6: for precipitation there is no change, while for surface wind there is an improvement of about $p_{Y\mid X}(y\mid x),$7 when including $p_{Y\mid X}(y\mid x),$8. They conclude that the approximation is acceptable.

3. Architecture and multivariate generative design

EnScale-t reuses the spatial architecture of EnScale and augments only the coarse model with temporal conditioning. The coarse components are dense MLPs with about five hidden layers of sizes $p_{Y\mid X}(y\mid x),$9. Random Gaussian noise channels are concatenated at each layer as latent input. The models also receive one-hot encodings of the GCM–RCM pair index $Y$0 and seasonal predictors including day-of-year and sine/cosine terms (Schillinger et al., 30 Sep 2025).

The super-resolution model $Y$1 is a cascade of four upscaling steps,

$Y$2

Each step models $Y$3, where $Y$4 denotes a pooled version of the target field at resolution $Y$5. The architecture at each stage combines deterministic local upsampling and stochastic refinement. In the deterministic component, each variable is interpolated separately from nearest low-resolution neighbors using location-specific learnable linear weights. In the stochastic component, the upsampled variables are concatenated with Gaussian noise channels; local neighbors are aggregated through learnable location-specific interpolation into a latent vector of dimension $Y$6; and a small shared MLP maps this representation to final high-resolution values for the four variables.

The pair index is used at least in the first linear upsampling step to allow pair-specific large-scale patterns, while stochastic refinement is shared. The architecture is jointly multivariate and spatial because the energy score is computed over all variables and grid points simultaneously. This encourages correct inter-variable and spatial dependence rather than marginal fidelity alone.

Temporal modeling is deliberately not implemented with RNNs, temporal CNNs, or transformers. The paper states that temporal modeling is entirely through the coarse MLP applied in an autoregressive roll-out. This is an important architectural distinction: EnScale-t does not impose a learned sequence model directly at $Y$7 resolution. A plausible implication is that the method trades explicit fine-scale temporal recurrence for lower computational cost and a simpler inductive decomposition between large-scale memory and local spatial refinement.

4. Training objective, data regime, and inference pipeline

The temporal coarse model is trained with the same conditional energy-score objective as the base method, but with $Y$8 replaced by $Y$9. For each training sample, the inputs are $X$0, the target is $X$1, and the output distribution is implicit, sampled by the stochastic MLP. The super-resolution component remains unchanged and models $X$2 with the same proper scoring rule (Schillinger et al., 30 Sep 2025).

Training uses daily samples rather than long sequences. For the temporal coarse model, each example is a triple $X$3; for super-resolution, each example is $X$4. The method therefore learns one-step transitions and does not backpropagate through long temporal sequences. The paper characterizes this as making training efficient and stable. No time-specific augmentation is described beyond seasonality and the preceding day’s coarse RCM state.

The implementation details given for EnScale-t specify the same MLP structure as the marginal coarse model, with $X$5 Gaussian channels per hidden layer concatenated to activations. The super-resolution model uses four sparse local stochastic stages, $X$6 Gaussian noise channels per pixel at each stochastic refinement step, local neighborhoods of $X$7 low-resolution neighbors in deterministic upsampling and $X$8 high-resolution neighbors in stochastic refinement, and a shared MLP with $X$9 hidden layers of $Z$0 units. Optimization uses Adam with learning rate $Z$1. The coarse and temporal coarse models are trained for $Z$2 epochs; super-resolution stages $Z$3–$Z$4 are trained for $Z$5 epochs each, and the final stage for $Z$6 epochs. Training data cover 1971–2029 and 2040–2089; testing is performed on 2030–39 as interpolation and 2090–99 as mild extrapolation.

At inference time, EnScale-t initializes the process by sampling

$Z$7

from the marginal non-temporal coarse model. For $Z$8, it performs an autoregressive coarse roll-out via

$Z$9

forming a Markov chain in latent coarse space. Each $Y$0 is then decoded independently in time through the super-resolution cascade,

$Y$1

The model is therefore autoregressive in time at coarse scale, but conditionally independent in time at high resolution given $Y$2. The paper reports stable sequences over multiple years, and ensemble generation is obtained by varying random seeds in the coarse and super-resolution models.

5. Evaluation framework and empirical behavior

The evaluation framework spans calibration, spatial structure, extremes, and multivariate dependencies, with temporal consistency assessed primarily by lag-1 autocorrelation at each grid point. For each location and variable, the empirical lag-1 autocorrelation of generated series is compared with that of the RCM, using both signed and absolute errors aggregated over locations (Schillinger et al., 30 Sep 2025).

The principal temporal result is that EnScale underestimates persistence, whereas EnScale-t reduces this bias almost to zero. Averaged signed lag-1 autocorrelation errors are $Y$3, $Y$4, $Y$5, and $Y$6 for EnScale on tas, pr, sfcWind, and rsds, respectively, compared with $Y$7, $Y$8, $Y$9, and $16\times 16$0 for EnScale-t. The appendix absolute-error table likewise shows EnScale-t to have the lowest absolute ACF-1 error across variables: $16\times 16$1, $16\times 16$2, $16\times 16$3, and $16\times 16$4 for tas, pr, sfcWind, and rsds, versus $16\times 16$5, $16\times 16$6, $16\times 16$7, and $16\times 16$8 for EnScale. CorrDiff, by contrast, is reported to overshoot autocorrelation for pr and sfcWind, with signed errors of $16\times 16$9 and $128\times 128$0.

For single-day metrics such as energy score, CRPS, spatial CRPS with pooling, and PSD, EnScale and CorrDiff are described as best overall, with EnScale-t very close to EnScale but sometimes slightly worse. On the interpolation period 2030–39, the reported energy scores are: for tas, EnScale $128\times 128$1, EnScale-t $128\times 128$2, CorrDiff $128\times 128$3; for pr, EnScale $128\times 128$4, EnScale-t $128\times 128$5, CorrDiff $128\times 128$6; for sfcWind, EnScale $128\times 128$7, EnScale-t $128\times 128$8, CorrDiff $128\times 128$9; and for rsds, EnScale $Y$00, EnScale-t $Y$01, CorrDiff $Y$02. The interpretation given in the paper is that EnScale-t maintains almost all of EnScale’s spatial and marginal performance while adding temporal structure.

Qualitative time-series examples show that EnScale captures seasonality and mean behavior but produces daily fluctuations that are rougher, while EnScale-t yields smoother trajectories closer to the RCM. The uncertainty bands, defined as min–max over $Y$03 samples, remain stable over time and often contain the RCM. The paper does not define formal metrics for multi-day extremes such as spell lengths or heatwave duration. This is significant because an improvement in lag-1 persistence does not by itself establish fidelity for all multi-day event statistics, although the reported smoother trajectories and improved ACFs suggest better multi-day persistence than the base method.

6. Computational characteristics, limitations, and uses

The temporal extension adds mainly a second coarse MLP, while leaving the high-resolution super-resolution machinery unchanged. The paper therefore characterizes the cost increase from EnScale to EnScale-t as modest relative to diffusion-model baselines. For the interpolation setup on a single NVIDIA A100, the reported comparison between EnScale and CorrDiff gives training times of $Y$04 h and $Y$05 h, respectively, making CorrDiff about $Y$06 slower, and inference times of $Y$07 h and $Y$08 h, respectively, for a 100-year high-resolution ensemble with $Y$09 samples per day for a single GCM–RCM pair, making CorrDiff about $Y$10 slower. EnScale-t is stated to share almost all of this efficiency because the extra temporal coarse model is cheap relative to the high-resolution super-resolution stage and inference only adds one MLP call per day. The EnScale architecture is reported as parameter-efficient, with about $Y$11M parameters versus about $Y$12M for CorrDiff, and the code, including EnScale-t, is released at https://github.com/m-schillinger/enscale (Schillinger et al., 30 Sep 2025).

The method’s main limitations follow directly from its temporal factorization. Temporal structure relies on a 1-day Markov assumption at coarse scale, no explicit temporal link at high resolution beyond what is carried by $Y$13, and the approximation $Y$14. Empirically, these approximations cause at most about $Y$15 loss of energy-score performance for some variables such as wind, but they may limit the representation of longer-memory local processes, including soil-moisture feedbacks, multi-week anomalies, and fine-scale temporal clustering of rare extremes. The model is trained on $Y$16 EURO-CORDEX GCM–RCM pairs, with pair-specific conditioning in the coarse model, and assumes approximate stationarity across 1971–2089 within each RCM aside from climate trend; performance in the far future or under other scenarios is not studied.

The paper also notes that temporal evaluation focuses on daily lag-1 autocorrelation and does not directly assess wet or dry spell lengths, duration of heatwaves or cold spells, or temporal clustering of extremes. A common overstatement would therefore be to treat EnScale-t as a complete solution to multi-day extreme-event fidelity. The documented claim is narrower: it substantially improves temporal autocorrelation with negligible loss elsewhere and produces stable, realistic multi-year time series. Within that scope, the suggested use cases are impact models that depend critically on temporal structure, including hydrology, energy systems, agriculture, and risk assessment for compound events over multiple days. When only single-day statistics or climatological distributions are needed, the paper states that EnScale without the temporal extension may suffice and is slightly better on some single-day metrics; when multi-day behavior matters, EnScale-t is recommended.