PaperlayerView exact revision on arXiv

Source-linked AI summary

Interpretable AI predicts a 2026 summer dry anomaly in central China

Anran Wang, Wen Shi, Yong Luo, Jianbin Huang, Lijuan Chen, Junhu Zhao, Weixin Jin, Huihui Yuan

arXiv:2608.19163v1physics.ao-phcs.AI

TL;DR

Seasonal precipitation forecasts have modest and variable skill, motivating approaches that exploit the more reliable prediction of atmospheric circulation. This study bridges circulation predictions to precipitation estimates and finds a consistently predicted summer 2026 dry anomaly over central China, with interpretable evidence implicating northerly winds.

  • Problem

    Seasonal precipitation prediction skill is generally modest and varies substantially, despite the value of early information for China’s water management.

  • Method

    A circulation-to-precipitation bridging model generates real-time precipitation predictions, which are evaluated against historical analogue years and interpreted using LRP.

  • Results

    The model consistently predicts a summer 2026 dry anomaly over central China, while historical analogue years show higher skill and northerly anomalies dominate the model attribution.

  • Takeaways & Limitations

    The framework supports physically interpretable, case-specific assessment of regional climate projections before observational data become available.

  • Takeaways & Limitations

    LRP identifies input features associated with the model output but does not establish causal relationships.

Abstract

from arXiv · show

Seasonal precipitation anomalies are largely regulated by atmospheric circulation, which dynamical models predict with greater reliability than precipitation itself. Here, we employ a deep learning model that translates dynamical circulation predictions into precipitation estimates. Predictions initialized from March to May consistently indicate a dry anomaly over central China in summer 2026. Retrospective evaluations revealed higher predictive skill in the analogue years, which also tended to feature central equatorial Pacific warming persisting from the preceding winter into summer. This warming favors an anomalous cyclonic circulation over the western North Pacific-South China Sea-South China region, which induces northerly winds and moisture divergence that jointly suppress rainfall over central China. Supporting this mechanism, layer-wise relevance propagation (LRP) independently identifies these northerly winds as the dominant driver of the prediction among all model inputs. Perturbation tests supported this attribution: removing LRP-identified features effectively eliminates the dry anomaly. Our framework thus provides physically interpretable explanations for AI-derived regional climate projections, facilitating evidence-based assessment before observational data become available.

Introduction

Seasonal precipitation prediction over China is difficult because skill varies with climate state and complex atmospheric, oceanic, and land-surface interactions. This study predicts a summer 2026 dry anomaly over central China and assesses its credibility using historical analogues, physical diagnostics, feature attribution, and perturbation tests.

  • Motivation: Regional summer precipitation is challenging to predict, with modest skill that varies substantially across prevailing climate states.Averaging hindcast skill can obscure state-dependent predictability and provide limited guidance on individual real-time forecasts.
  • Motivation: Credibility assessment requires tracing individual predictions to physically meaningful signals supported by historical diagnostics and model behavior.Convergent evidence can strengthen the basis for issuing an early warning before observations become available.
  • Contribution: The proposed framework assesses AI-based seasonal precipitation predictions through convergent historical, physical, and model-centered evidence available before the target season.It combines feature-attribution methods with LRP-guided perturbation tests to assess whether identified circulation features are relevant to the prediction.
  • Framework: A circulation-to-precipitation bridging model links predictable large-scale dynamical predictions with regional precipitation and enables interpretable attribution of circulation signals.Deep-learning models extend this framework by learning nonlinear relationships between multivariable circulation fields and regional precipitation.
  • Study objective: Predictions initialized in March, April, and May consistently indicate a dry anomaly over central China in summer 2026.The study evaluates historical analogue-year skill, checks whether associated climate conditions are evident in 2026, and tests attribution stability and feature relevance.

Results · 1. Consistent predictions point to a central China dry anomaly in summer 2026 · 2. Higher prediction skill in historical dry-anomaly years

March–May initialized predictions consistently indicate a stable, prominent central China dry anomaly for summer 2026, alongside above-normal precipitation in South and parts of North China. Historical analogue years showed higher median predictive skill, suggesting a recurrent circulation pattern may underlie the predicted drying.

  • 1. Consistent predictions point to a central China dry anomaly in summer 2026: March, April and May initializations consistently predicted below-normal summer precipitation across central China, with above-normal precipitation over South China and parts of North China.The central China dry anomaly was the most prominent regional feature.
  • 1. Consistent predictions point to a central China dry anomaly in summer 2026: The central China anomaly’s location, spatial extent and magnitude remained relatively stable as initialization approached summer 2026.Its inter-initialization range was comparatively small despite the largest anomaly magnitude.
  • 1. Consistent predictions point to a central China dry anomaly in summer 2026: The bridging-model MME achieved ACC values of 0.140–0.161 and Ps of 73.27–74.25% across March–May initializations.Its largest gains relative to the dynamical MME occurred for March, while the MMEs performed similarly in May.
  • 2. Higher prediction skill in historical dry-anomaly years: ERA5-driven outputs attained markedly higher ACC and Ps than dynamical-circulation-driven predictions across all initializations and both historical year groups.This indicates that errors in predicted circulation partly constrained real-time prediction skill.
  • 2. Higher prediction skill in historical dry-anomaly years: Median skill was higher in dry-anomaly analogue years in every comparison except May-initialized dynamically driven ACC.For that exception, the dry-year median was slightly lower and the distributions overlapped substantially.
  • 2. Higher prediction skill in historical dry-anomaly years: Higher skill in dry-anomaly analogue years suggests the model may perform better for precipitation states resembling the predicted 2026 anomaly than average historical skill implies.Because predictions derive entirely from atmospheric circulation, the pattern also suggests a recurrent circulation pattern may underlie central China drying.

3. Climate conditions conducive to a dry anomaly over central China

The predicted central China dry anomaly is associated with a recurrent circulation pattern featuring a poleward-shifted jet, continental high pressure, and western North Pacific–South China Sea cyclonic flow. Central equatorial Pacific warming, especially rising Niño-4 SST, supports this circulation and the associated moisture divergence.

  • March-, April-, and May-initialized predictions showed similar spatial circulation patterns linked to the central China dry anomaly.
  • A western North Pacific–South China Sea–South China cyclonic anomaly strengthened northeasterly flow over central China and opposed moisture transport into the target region.Central China lay on the circulation’s northwestern flank, where anomalous moisture-flux divergence occurred; the stronger 2026 winds may intensify drying relative to the seven-year composite.

4. Physically coherent prediction signals identified by LRP

LRP identified a physically coherent circulation pathway underlying the predicted summer 2026 dry anomaly over central China. Northerly meridional-wind anomalies were the dominant signal, while attribution also implicated upstream warmth and geopotential-height gradients.

  • Attribution setup: LRP targeted the regional-mean precipitation anomaly over central China from the May-initialized summer 2026 prediction.The predicted principal components were propagated through EOF reconstruction and regional averaging to attribute the regional anomaly directly.
  • Circulation signals: Northerly anomalies along the northwestern flank of the western North Pacific–South China Sea–South China cyclone dominated dry-supporting relevance across 500, 700 and 850 hPa.The three meridional-wind predictors accounted for 37.5% of area-weighted dry-supporting relevance, with the strongest signal at 700 hPa.
  • Physical mechanism: The model linked the rainfall deficit primarily to circulation-induced weakening of moisture transport into central China rather than limited moisture availability.Positive specific-humidity anomalies received little dry-supporting relevance, whereas the northerly anomalies opposed climatological moisture transport.
  • Additional signals and limitation: Dry-supporting relevance also appeared over upstream anomalous warmth in 2-m air temperature and along strong geopotential-height gradients on the WNPSH periphery.Some geopotential-height attribution structure may reflect alignment with the ViT patch grid and tokenization.
  • Cross-method consistency: Integrated Gradients and Guided Integrated Gradients broadly reproduced the May-initialized wind-field attribution patterns, with high cross-method agreement in predictor rankings across initializations.These comparisons are reported in the supplementary figures and Table 2.

5. Perturbation tests support the faithfulness of LRP explanations

LRP-guided perturbation tests showed that removing the highlighted circulation features reversed the predicted dry signal, whereas retaining them preserved or strengthened it. Random masks did not reproduce these responses, supporting the faithfulness of the LRP attribution.

  • Removal tests: LRP-guided Removal reversed ΦƼ from negative to positive for 2026 and all seven analog years, while random Removal stayed near baseline values.ΦƼ is the mean standardized precipitation anomaly across central China stations, with negative values indicating dry anomalies.
  • Retention tests: LRP-guided Retention preserved negative ΦƼ and produced more negative values than baselines in every case, whereas random Retention clustered near zero.The LRP-guided masks retained features in the top 5% of dry-supporting relevance and their neighboring grid cells.
  • Random-mask comparisons: None of the 100 random perturbations produced a response more extreme than the LRP-guided result in the prespecified direction for any of the eight cases.The one-sided empirical tests yielded ƺ < 0.01.
  • Interpretation: The disappearance of the dry signal after Removal and its persistence under Retention indicate that LRP-highlighted circulation features concentrated the model’s dependence supporting the predicted dry anomaly.The tests used case-specific masks for 2026 and seven analog years in the May initialization.

Discussion

The study supports the predicted summer 2026 dry anomaly over central China through historical skill, circulation diagnostics, and interpretable attribution analyses. The framework makes the evidential basis of seasonal AI predictions explicit while distinguishing supportive evidence from causal proof and event-level certainty.

  • Evidence for the prediction: Historical evaluations showed generally higher predictive skill in selected analogue years, while predicted circulation reproduced circulation and moisture features associated with suppressed central-China precipitation.These features included a lower-tropospheric cyclone over the western North Pacific–South China Sea–South China region, northerly anomalies, weakened monsoon moisture transport, and moisture-flux divergence.
  • Interpretable attribution: LRP identified mid-to-lower-tropospheric northerly anomalies as the strongest evidence-supported signal, consistent with climate diagnostics and broadly reproduced by IG and Guided IG.Perturbation tests confirmed the joint importance of selected features: removing them reversed the predicted anomaly from dry to wet, whereas retaining them alone intensified dryness.
  • Interpretable attribution: The attribution target reconstructed the regional precipitation anomaly from all 512 EOF modes, integrating reduced-order outputs into a region-specific map directly comparable with circulation patterns.The formulation may also support regional interpretation of other spatial prediction models with reduced-order outputs.
  • Implications: The framework provides physically grounded explanations for potentially high-impact seasonal climate anomalies and supports transparent, evidence-based AI use in pre-season climate-risk assessment.It evaluates predictability, physical plausibility, and model reliance on pathway-associated features before observations are available.
  • Implications and limitations: For summer 2026, historical-case, physical-pathway, and feature-reliance analyses all provided supportive evidence, but the framework does not guarantee verification after observations become available.The predicted anomaly denotes a seasonal tendency toward drier conditions and does not exclude heavy rainfall that could substantially affect the observed summer total or resolve individual events.

Method · 1. Data · 2. Bridging-model development and evaluation

The study combines long-term Chinese station precipitation observations with atmospheric circulation and seasonal prediction datasets to develop and evaluate a ViT-based circulation-to-precipitation bridging model. The model is pretrained on dynamical-model pairs, transferred using ERA5 and observations, and assessed through blocked cross-validation.

  • 1. Data: C3S hindcasts for 1993–2025 comprised 27 forecast-system versions from eight forecasting centers and supplied paired circulation inputs and precipitation targets for pretraining.The same 13 atmospheric variables and corresponding precipitation formed the paired inputs and targets.
  • 2. Bridging-model development and evaluation: The bridging model maps 13 seasonal circulation-anomaly fields to station precipitation anomalies using a ViT architecture that replaces the original convolutional backbone.Multichannel fields become spatial tokens, while the target season is represented by a CLS token conditioning feature extraction.
  • 2. Bridging-model development and evaluation: The model predicts principal-component coefficients for the leading 512 EOF modes, then reconstructs station precipitation anomalies from those coefficients and the EOF basis.Six ViT blocks process the token sequence before the output layers predict the coefficients.
  • 2. Bridging-model development and evaluation: Because each three-month window yields one seasonal sample, the model was pretrained on dynamical-model anomaly pairs and transferred by fitting its final two linear layers with ridge regression.Transfer learning used ERA5 circulation anomalies and observed precipitation anomalies.
  • 2. Bridging-model development and evaluation: Blocked cross-validation for 1993–2025 used six contiguous year blocks, holding one block for testing, the preceding cyclic block for validation, and four for training.The same splits were applied during pretraining and transfer learning, and the six test-block predictions formed the complete evaluation.

3. Identification of historical analogue years

Historical analogue years were identified by comparing observed JJA precipitation anomalies from 1993–2025 with the corresponding 2026 prediction across four metrics. Years were ranked for each metric and combined, retaining the seven closest matches.

  • Metric-based analogue identification: Observed JJA precipitation anomalies during 1993–2025 were compared with the corresponding 2026 prediction for each initialization.The comparison used observations from the 33-year period 1993–2025.
  • Metric-based analogue identification: Four metrics assessed similarity: all-station ACC (ACCall), Central China ACC (ACCCC), dry coverage (DC), and dry-intensity distance (DID).These metrics jointly evaluated spatial agreement, regional agreement, dry-area coverage, and dry intensity.
  • Metric-based analogue identification: Rank 1 indicated the closest match, with larger values preferred for ACCall, ACCCC, and DC, but smaller values preferred for DID.Years were ranked separately for each metric before combining the rankings.
  • Metric-based analogue identification: The overall rank sum combined the individual ranks for ACCall, ACCCC, DC, and DID.The rank sum was defined as the sum of the four metric-specific ranks.
  • Metric-based analogue identification: The seven years with the smallest overall rank sums were retained as historical analogues, corresponding to the top 20% of the 33 ranked years.Ties were resolved by smaller DID, followed by higher Central China and all-station ACCs.

4. Layer-wise relevance propagation and perturbation tests

The study used CP-LRP to trace circulation signals supporting the predicted central China dry anomaly and designed LRP-guided perturbation tests to assess their attribution. The method defined a regional precipitation target, propagated relevance through the ViT model, and compared targeted masks with random controls.

  • Layer-wise relevance propagation: CP-LRP traced circulation signals contributing to the predicted central China dry anomaly through the ViT architecture.Attention weights were fixed, and relevance propagated through the value path without redistribution through query–key scores or softmax operations.
  • Layer-wise relevance propagation: The scalar LRP target was the predicted mean standardized precipitation anomaly over the central China region.A fixed linear reconstruction layer propagated relevance from the regional target through precipitation PC outputs to circulation inputs.
  • Layer-wise relevance propagation: Only negative relevance values were retained for visualization because they contributed toward a more negative regional precipitation anomaly.Predictor contributions were calculated using cosine-latitude-weighted integration and normalized across the 13 predictors.
  • Perturbation tests: LRP-guided perturbation tests examined the May-initialized 2026 prediction and seven analogue years using the strongest 5% of dry-supporting relevance values.The selected mask was expanded by one grid cell in each spatial direction.
  • Perturbation tests: Removal set selected input elements to zero, whereas Retention preserved only those elements for comparison with random masks of equal size.The random-mask comparisons used Ƭ = 100 masks and one-sided empirical ƺ values, with a smallest attainable value of 1/101 (<0.01).

5. Integrated Gradients and Guided Integrated Gradients · Reference · central China

IG and Guided IG were applied with a common climatological baseline to test whether LRP’s dry-supporting signals were reproduced by different attribution principles. Both methods broadly recovered the identified meridional-wind signals, with strong cross-method agreement in predictor rankings and attribution patterns.

  • 5. Integrated Gradients and Guided Integrated Gradients: IG and Guided IG produce signed feature-level attributions for a specified scalar output, enabling comparison with LRP’s dry-supporting signals.IG integrates gradients along a straight path, whereas Guided IG uses an adaptive path to reduce noisy-gradient accumulation.
  • 5. Integrated Gradients and Guided Integrated Gradients: Both methods were applied to the frozen ViT using the regional attribution target ΦƼ for March-, April-, and May-initialized predictions.The same attribution procedures were used across all three initialization months.
  • Reference: Zero standardized-anomaly fields represented predictor climatology and served as the common baseline for assessing how input anomalies changed the regional prediction.Negative attribution values indicate dry-supporting contributions.
  • 5. Integrated Gradients and Guided Integrated Gradients: The IG and Guided IG attribution fields were smoothed for visualization consistently with the LRP fields, and cross-method agreement was assessed in rankings and attribution patterns.The comparison is presented in the supplementary attribution figures and agreement analysis.
  • central China: Analogue composites and the March-initialized 2026 prediction both showed western North Pacific–South China Sea–South China cyclonic circulation and central-China moisture-flux divergence.The supplementary circulation figure documents this correspondence for the March initialization.
  • central China: IG broadly reproduced LRP’s dry-supporting meridional-wind signals at 500, 700, and 850 hPa, while T2m received the largest predictor contribution.This result was reported for the May-initialized prediction.
  • central China: Guided IG likewise recovered dry-supporting meridional-wind signals at 500, 700, and 850 hPa, while assigning the largest predictor contribution to T2m.This result was also reported for the May-initialized prediction.
  • central China: Spearman rank correlations of 0.88–0.98 indicate strong agreement in the hierarchy of area-weighted dry-supporting contributions across the 13 predictors.The supplementary table also evaluates full-tensor cosine similarity of unsmoothed attribution patterns across predictors and grid cells.
Loading 2608.19163v1…