---
title: 'DeepX-GAN: Simulating Spatial Climate Extremes'
url: https://www.emergentmind.com/topics/deepx-gan
type: topic
---

# DeepX-GAN: Simulating Spatial Climate Extremes

Searching arXiv for the specified paper and closely related background work.
arxiv_search.query({"search_query":"id:2507.09211 OR ti:\"Capturing Unseen Spatial Extremes Through Knowledge-Informed Generative Modeling\"","max_results":5,"sort_by":"submittedDate","sort_order":"descending"})
arxiv_search.query({"search_query":"ti:\"SPATE-GAN\" OR all:\"spatiotemporal data with causal optimal transport and a mixed Sinkhorn loss\"","max_results":10,"sort_by":"relevance","sort_order":"descending"})
arxiv_search.query({"search_query":"all:\"SPATE-GAN\"","max_results":5,"sort_by":"relevance","sort_order":"descending"})
DeepX-GAN, short for Dependence-Enhanced Embedding for Physical eXtremes - Generative Adversarial Network, is a knowledge-informed deep generative framework for simulating spatially compound, historically unseen climate extremes with improved fidelity to the physics of spatial dependence. It was introduced to address two coupled deficiencies in conventional risk assessment: observed records are too short to capture all high-impact extremes, and neglect of spatial dependence undervalues synchronized hazards that amplify impacts. The model augments spatiotemporal climate datasets by embedding an extremal dependence metric, DeepX, into adversarial learning, with the stated aim of enabling zero-shot generalization to rare events that fall outside the historical record yet remain statistically plausible. The framework also formalizes two categories of unrealized extremes—direct-hit “checkmate” events and near-miss “stalemate” events—and applies them to historical and future heat risk in the Middle East and North Africa (MENA) [2507.09211].

## 1. Problem formulation and conceptual scope

DeepX-GAN is motivated by the claim that observed records of climate extremes provide an incomplete picture of risk because unseen extremes can exceed historical bounds, while many regions may have avoided direct hits by chance and thereby acquired a false sense of resilience. The framework is designed around the premise that the critical omission in many generative and statistical approaches is not only the marginal severity of extremes, but their spatial dependence structure, especially when hazards occur synchronously across regions [2507.09211].

Within this framing, the paper contrasts DeepX-GAN with several classes of existing methods. Conventional statistical approaches, including extreme value theory with parametric assumptions and copulas, are described as often missing extremes that exceed theoretical bounds and as inadequately reflecting the spatial dependence structure of synchronized hazards. Physics-based ensemble methods can explore extremes, but they require meticulous setup or large computational resources. Standard AI generative models are characterized as typically neglecting extremal spatial dependence and as producing either univariate time series or static fields rather than full spatiotemporal evolution.

The framework’s stated contributions are fourfold. First, it explicitly encodes spatial tail dependence via a Dependence-Enhanced Embedding, DeepX, that fuses spatiotemporal autocorrelation and extremal correlation into the model’s latent embedding channel. Second, it demonstrates zero-shot generalizability by generating extreme events omitted from training. Third, it introduces the formal distinction between unseen “checkmate” and “stalemate” extremes and quantifies their probabilities with normalization against community-wide unseen extreme probability. Fourth, it reports improved reproduction of tail dependence, extremal angle spectra, and high-dimensional co-occurrence of exceedances relative to a baseline SPATE-GAN, and uses the model to produce risk maps under historical and future scenarios, specifically SSP126 and SSP585.

A plausible implication is that DeepX-GAN is not merely a generator of synthetic extremes, but a framework for stress-testing the consequences of rare spatially coherent events that remain absent from the empirical record. The paper is explicit, however, that such synthetic extremes cannot be guaranteed to occur.

## 2. Formal definitions of unseen extremes

The paper defines unseen extremes using target- and neighborhood-level exceedance thresholds matched to the return level of the highest historically recorded temperature at the target. It distinguishes direct-hit and near-miss scenarios through normalized and unnormalized event probabilities [2507.09211].

The unnormalized unseen “checkmate” probability is written as
$$
P_{\text{target}} = P(I_{\text{target}} \geq \alpha_{\text{target}}).
$$

The unnormalized unseen “stalemate” probability is written as
$$
P(\exists I_{\text{neighbor}} \geq \alpha_{\text{neighbor}} \land I_{\text{target}} < \alpha_{\text{target}}).
$$

The community-wide unseen extreme probability is
$$
P^{\text{community}} = P(\exists I_{\text{neighbor}} \geq \alpha_{\text{neighbor}} \lor I_{\text{target}} \geq \alpha_{\text{target}}).
$$

The normalized probabilities are then defined as
$$
P^{\text{checkmate}} = \frac{P(I_{\text{target}} \geq \alpha_{\text{target}})}{P^{\text{community}}},
$$
and
$$
P^{\text{stalemate}} = \frac{P(\exists I_{\text{neighbor}} \geq \alpha_{\text{neighbor}} \land I_{\text{target}} < \alpha_{\text{target}})}{P^{\text{community}}}.
$$

For comparison, the paper also gives spatially random-process counterparts based on independent Bernoulli exceedances:
$$
P_{\text{random}}^{\text{community}} = 1 - \prod_{i \in N_{\text{target}}} (1 - P_i)(1 - P_{\text{target}}),
$$
$$
P_{\text{random}}^{\text{checkmate}} = \frac{P_{\text{target}}}{P_{\text{random}}^{\text{community}}},
$$
and
$$
P_{\text{random}}^{\text{stalemate}} = \frac{[1 - \prod_{i \in N_{\text{target}}} (1 - P_i)](1 - P_{\text{target}})}{P^{\text{community}}}.
$$

Neighborhoods are defined by spatial adjacency through a binary weight matrix $w_{ij}$. The temporal window for event identification uses 30-day sequences in JJA. For zero-shot experiments, extremes are the top 100 spatially averaged daily Tmax events, corresponding to the top 0.62% of instances. For random-process baselines, the simplified exceedance probability is taken as $1/S$, where $S$ is the record length, with examples of $1/44$ for reanalysis and $1/36$ for the future period.

These definitions are central because the framework does not treat unseen risk as a single scalar. Instead, it separates the probability that an unprecedented community-wide event occurs from the conditional question of whether a target is hit directly or narrowly missed. This suggests a distinction between exposure, direct impact, and warning-like near misses.

## 3. Dependence-Enhanced Embedding and model architecture

DeepX-GAN builds on SPATE-GAN, an adversarial model for spatiotemporal data with causal optimal transport, a mixed Sinkhorn loss, and an autoregressive embedding, SPATE, for detecting spatiotemporal clustering. The distinctive addition is the Dependence-Enhanced Embedding, DeepX, which adds a second embedding component that explicitly weights spatial information by extremal correlations and thereby shapes learning toward tail dependence [2507.09211].

For pixel $i$ at time $t$, the DeepX metric is
$$
\text{DeepX}_{it} = \frac{(n - 1) z_{it}}{\sum_{j=1}^n z_{jt}^2} \sum_{j=1, j \neq i}^n w_{ij} z_{jt},
$$
with deviation from mixed expectation
$$
z_{it} = x_{it} - \mu_{it}^{\text{Mixed}}.
$$

Spatial adjacency is binary:
$$
w_{ij} =
\begin{cases}
1 & \text{if } x_{jt} \text{ is in the neighborhood of } x_{it}, \; t \in \mathbb{Z}^+ \\
0 & \text{elsewise.}
\end{cases}
$$

The mixed space-time expectation is
$$
\mu_{it}^{\text{Mixed}} = \theta_{\text{A}} \mu_{it}^{\text{A}} + \theta_{\text{B}} \mu_{it}^{\text{B}}.
$$

Its spatiotemporal coupling component is
$$
\mu_{it}^{\text{A}} =
\frac{\sum_{j=1}^n x_{jt} \sum_{t' < t} b_{tt'} x_{it'}}
{\sum_{j=1}^n \sum_{t' < t} b_{tt'} x_{jt'}},
$$
where
$$
b_{tt'} = \exp\left(-\frac{|t-t'|}{l}\right).
$$

The tail-dependence-enhanced component is
$$
\mu_{it}^{\text{B}} =
\frac{\sum_{j=1}^n k_{ij,t} \chi_{ij} x_{jt} \sum_{t' < t} b_{tt'} x_{it'}}
{\sum_{j=1}^n \sum_{t' < t} b_{tt'} x_{jt'}}.
$$

The extreme indicator and upper tail dependence coefficient are
$$
k_{ij,t} =
\begin{cases}
1 & \text{if } F_i(x_{it}) > q \text{ and } F_j(x_{jt}) > q \\
0 & \text{else,}
\end{cases}
$$
and
$$
\chi_{ij} = \lim_{q \to 1} P(F_i(x_i) > q \mid F_j(x_j) > q) \in [0,1],
$$
with empirical form
$$
\chi_{ij} = \frac{P(F_i(x_i) > q, F_j(x_j) > q)}{P(F_j(x_j) > q)}.
$$

Architecturally, the sequence generator uses temporal modeling with LSTM and spatial upsampling with Deconv/FCNN to produce 30-day sequences on a $64 \times 32$ grid. The discriminator employs Conv layers over space-time. The DeepX embedding is concatenated to input channels for both real and generated sequences before entering the discriminator, and it is also used to compute an embedding loss that encourages the generator to match observed spatial tail dependence.

The model is unconditional on external climate indices. Conditioning is described as implicit, achieved by fusing DeepX embeddings with the data. Latent vectors are drawn from a Gaussian distribution for generation and are optimized for reconstruction loss evaluation.

The paper does not provide explicit objective-function equations. It states, however, that adversarial training is based on SPATE-GAN and therefore uses an optimal transport or Sinkhorn-style loss together with an embedding loss. It presents the following as an inferred plausible formalization:
$$
\min_G \max_D \; \mathbb{E}_{x \sim p_{\text{data}}}[\log D(x, \text{DeepX}(x))] + \mathbb{E}_{z \sim p(z)}[\log(1 - D(G(z), \text{DeepX}(G(z))))] + \lambda_{\text{OT}} \, \mathcal{L}_{\text{Sinkhorn}}(x, G(z)) + \lambda_{\text{emb}} \, \| \text{DeepX}(x) - \text{DeepX}(G(z)) \|.
$$
This inferred form is explicitly marked as not being stated in the paper as a formal training objective.

## 4. Data, preprocessing, training protocol, and evaluation

The reported experiments use four data sources and one validation process. The LGCP spatiotemporal point process contributes 300 samples for initial validation. Historical climate training uses NCEP Reanalysis 2 daily 2-m Tmax over MENA during JJA from 1979 to 2022, yielding 2,772 samples, each a 30-day sequence on a $64 \times 32$ grid. Future simulations use CMIP6 CMCC-ESM2 daily 2-m Tmax over MENA during JJA from 2065 to 2100, with 2,268 training samples under SSP126 and SSP585. Equity analyses use ND-GAIN vulnerability and readiness indicators for 77 countries [2507.09211].

For unseen experiments, extremes are defined by spatially averaged daily Tmax exceeding a threshold set as the top 100 recorded heat events, again corresponding to the top 0.62%. For pixelwise co-occurrence analysis, local 1-in-100-year thresholds are computed per pixel from real data, and cumulative exceedances are compared with binomial independence. For unseen risk maps, the neighborhood threshold is matched to the return level corresponding to the target’s historical maximum, so that target and neighboring exceedances represent equivalent severity.

The zero-shot strategy uses two training sets: Complete, which contains extremes, and NoExtreme, in which extremes are removed. The paper describes four cases in Table 1, organized around train/test combinations on extremes, complete, or NoExtreme datasets, with evaluation emphasizing reconstruction on ExtremeOnly test sets. This setup is meant to test whether the generator can recover severe events it never observed during training.

Training details are explicit. The implementation uses PyTorch with Adam, a learning rate of $1 \times 10^{-4}$, minibatch size 64, and 20,000 iterations for LGCP or 100,000 iterations for reanalysis, evaluation, and unseen experiments. Hardware consists of two NVIDIA A6000 GPUs. Stopping is based on monitoring convergence to a stable regime in reproducing climate data.

Evaluation spans both conventional distributional similarity and extremes-specific diagnostics. Overall distributional similarity is assessed with statistical moments, marginal distributions, multi-scale sliced Wasserstein distance (MS-SWD), maximum mean discrepancy (MMD), and reconstruction loss via one-to-one matching using latent optimization. The empirical MMD is given as
$$
\text{MMD}^2 = \frac{1}{m(m-1)} \sum_a \sum_{c \neq a} k(x_a, x_c)
- \frac{2}{m^2} \sum_a \sum_c k(x_a, y_c)
+ \frac{1}{m(m-1)} \sum_a \sum_{c \neq a} k(y_a, y_c),
$$
with kernel
$$
\kappa(x_a, x_c) = \exp\left(-\frac{\|x_a - x_c\|^2}{2\sigma^2}\right).
$$

Extremal dependence is evaluated through the upper tail dependence coefficient $\chi_{ij}$ and through the spectral distribution of extremal angles,
$$
H(\omega) = \lim_{u \to \infty} P\left(\frac{\widetilde{X}_i}{\widetilde{X}_i + \widetilde{X}_j} \leq \omega \mid \widetilde{X}_i + \widetilde{X}_j > u\right), \quad \omega \in [0,1],
$$
subject to
$$
\int_0^1 \omega \, dH(\omega) = \frac{1}{2}.
$$

Spatial structure is examined with radially averaged power spectral density, intended to detect multi-scale features and small-scale sharpness. Zero-shot generalizability is evaluated with reconstruction loss on ExtremeOnly test sets and with the distribution of regional mean temperature, both for values above the 90th quantile and for the full distribution. Spatially compound risk is assessed via counts of pixels exceeding local 1-in-100-year thresholds and comparison with binomial independence at $p=0.01$.

## 5. Empirical performance and climate-risk findings

The paper reports consistent improvements in reproducing extremal dependence. DeepX-GAN reduces RMSE of extremal correlation by 22% on LGCP and TX and lowers Wasserstein distance of extremal angle spectra by 25%. It also preserves small-scale spatial features in power spectral density [2507.09211].

For bivariate extremes between countries, the paper highlights geographically distant pairs such as Pakistan–Sudan, where DeepX-GAN is said to match observed dependence much better than the baseline. It notes a “97% improvement in Kendall’s $\tau$ correlation,” consistent with Table S3, in which DeepX-GAN yields 0.24, real data 0.23, and the baseline 0.58. For neighboring pairs such as Pakistan–Afghanistan, both models are described as comparable to observations. Amplification factors for bivariate return periods indicate that the baseline underestimates risk whereas DeepX-GAN captures recurrence.

In high-dimensional compound extremes, DeepX-GAN reproduces the distribution of counts of concurrent exceedances of local 1-in-100-year thresholds and closely matches NCEP reanalysis, while spatially random processes based on binomial independence substantially undercount co-occurrences. The model is also reported to extrapolate to wider co-occurrence extents than observed. This suggests that the learned spatial dependence structure allows the generator to explore extreme configurations beyond the empirical sample without collapsing into purely random exceedance patterns.

The zero-shot experiments are among the paper’s central claims. In reconstruction tests, when trained without extremes, DeepX-GAN significantly outperforms the baseline on ExtremeOnly sets, corresponding to Models M5 versus M6. Its performance is described as robust to omission of extremes during training, summarized as $M1 \sim M5$, while the baseline degrades, summarized as M2 diverging from M6. Generated distributions of regional mean temperatures, including the extreme tail above the 90th quantile and the overall distribution, match empirical data better than the baseline, which the paper interprets as evidence of physically credible generalization rather than overestimation.

Risk mapping over 1979–2022 identifies hotspots in North Africa, West Asia, and Central Asia for community-wide unseen extremes and for checkmate risks. India is singled out as having relatively low community-wide unseen probability but high checkmate probability, implying that any unprecedented regional extreme would likely affect the target directly. The paper further states that over 76% of MENA would have checkmate risks underestimated and stalemate risks overestimated if spatial dependence were ignored.

Under future warming in 2065–2100, the geography of unseen risk shifts. Historical high-risk zones persist over 45% of area under SSP126 and 34% under SSP585. New hotspots account for 54% under SSP126, identified with Indo-Pakistan, and 65% under SSP585, identified with Central Africa and Mediterranean coasts. Areas where checkmate probability exceeds the spatially random baseline expand by 7% under SSP126 and 10% under SSP585, with new-emerging hotspots in Central Africa and southeastern Europe at 18% under SSP126 and 22% under SSP585.

The equity analysis links these hazard patterns to socioeconomic readiness. Among 33 higher-vulnerability and lower-readiness countries, 42% exhibit community-wide unseen risks in the top quartile, versus 9% among lower-vulnerability and higher-readiness countries. Normalized checkmate and stalemate probabilities show no significant socioeconomic correlation, but community-wide unseen risk correlates positively with vulnerability at $p<0.05$. A plausible implication is that the dominant inequity signal lies in exposure to unprecedented regional extremes rather than in the conditional split between direct-hit and near-miss outcomes.

## 6. Baselines, ablations, limitations, and applications

The baseline comparator is SPATE-GAN, characterized in the paper as using causal optimal transport, mixed Sinkhorn loss, and SPATE embedding for autocorrelation. DeepX-GAN is defined as the extension in which an extremal tail dependence component is added, operationally corresponding to $\theta_{\text{B}} > 0$. The paper states that setting $\theta_{\text{A}} = 1$ and $\theta_{\text{B}} = 0$ reduces DeepX-GAN to the baseline, thereby isolating the impact of tail dependence. Under this ablation, the reported improvements include the 22% reduction in extremal-correlation RMSE, the 25% reduction in extremal-angle Wasserstein distance, and correction of small-scale power spectral density features [2507.09211].

Robustness is discussed primarily through the zero-shot setting. DeepX-GAN maintains reconstruction performance when trained without extremes, summarized as $M5 \sim M1$, whereas the baseline fails to generalize, summarized as M6 versus M2. Uncertainty quantification is limited but explicit: reconstruction loss is computed on 100 randomly selected test samples to capture variability, and the reported training regime is said to produce stable convergence.

Several assumptions and limitations are noted. The DeepX metric relies on empirical tail dependence $\chi_{ij}$ estimated from limited records, which may be noisy for very rare events. Neighborhood adjacency is binary and may not capture distance-decay nuance beyond $w_{ij}=1/0$. The simplified exceedance probability $1/S$ for random baselines is heuristic and may deviate from true probabilities because of sampling variability. No explicit physics-constrained losses, such as conservation laws, are imposed; knowledge is incorporated through extremal dependence embedding rather than hard constraints. Code availability is pending publication, and reproducibility depends on careful data preparation and calibration of $\theta_{\text{A}}$, $\theta_{\text{B}}$, $q$, and neighborhood definitions.

The paper also identifies a specific failure mode in the unseen-risk formalism. Because normalized checkmate and stalemate probabilities use “almost unseen” thresholds to avoid division by zero, they may conservatively underestimate true checkmate probabilities. This caveat matters because the framework’s policy interpretation turns on the distinction between direct hits and near misses.

Application guidance is provided in stepwise form. Reproduction begins with acquisition of NCEP TX data for a region of interest, segmentation into 30-day sequences on a consistent grid such as $64 \times 32$, and computation of DeepX embeddings through neighborhood selection, quantile choice for $k_{ij,t}$, empirical estimation of $\chi_{ij}$, and calibration of $\mu^{A}$, $\mu^{B}$, $\mu^{\text{Mixed}}$, and the temporal kernel $b_{tt'}$. Training then proceeds in PyTorch with Adam at learning rate $1 \times 10^{-4}$ and batch size 64 for approximately 100k iterations, monitored through MMD, MS-SWD, reconstruction loss, and extremal metrics. Large ensembles from trained models are used to estimate exceedances beyond local 1-in-100-year thresholds, co-occurrence distributions, and unseen community-wide, checkmate, and stalemate probabilities; future scenarios repeat the process with CMIP6 simulations and can be overlaid with ND-GAIN indicators for equity analysis.

In policy terms, the paper argues that distinguishing checkmate from stalemate reframes preparedness. Near-miss stalemates may be interpreted either as warning signals or as evidence of resilience, depending on risk tolerance, and can therefore foster complacency if misread. The emergence of future hotspots in Indo-Pakistan and Central Africa is presented as evidence of blind spots in hazard planning that extrapolates from history. The broader claim is that high community-wide unseen risks in vulnerable, low-readiness settings underscore climate injustice and support spatially adaptive policies and adaptation finance targeted to emergent hotspots.

Source: https://www.emergentmind.com/topics/deepx-gan