Heatwave attribution based on reliable operational weather forecasts

The study addresses how human-induced climate change influenced the unprecedented 2021 Pacific Northwest heatwave. Traditional attribution methods face challenges for such extremes due to limited historical analogues, potential statistical model misspecification, and coarse-resolution climate models that inadequately represent key nonlinear processes (e.g., atmospheric rivers, blocking, soil moisture feedbacks). In contrast, modern numerical weather prediction (NWP) systems have high resolution and demonstrated skill in forecasting extreme events, including the PNW heatwave, at lead times of a week or more. The authors propose leveraging reliable operational forecast models to attribute both the magnitude and probability changes of specific extreme events, aiming to provide decision-relevant, trustworthy estimates that bridge probabilistic and storyline attribution approaches.

Prior work has advanced extreme event attribution but struggles with unprecedented events where statistical models lack physically similar samples and climate models may under-resolve key processes (e.g., atmospheric rivers, omega blocks, land–atmosphere feedbacks). The literature distinguishes probabilistic (event-class) and storyline (event-specific) approaches; the former can misattribute if the event differs from the class, while the latter lacks probabilistic likelihood changes. Previous forecast-based attribution efforts using seasonal/subseasonal systems showed promise but at coarser resolution. Studies of the PNW heatwave identified combined dynamical and thermodynamical drivers, including an omega block, an anomalous warm-season trans-Pacific atmospheric river, dry soils, high insolation, and record mid-tropospheric warmth. Operational forecasts (e.g., ECMWF) successfully anticipated aspects of the event, suggesting forecast systems can be used to attribute events reliably by ensuring the model can reproduce the event and its drivers for the right reasons.

Design: A forecast-based attribution framework using the ECMWF ensemble prediction system (EPS) to generate counterfactual forecasts for the 2021 PNW heatwave. The approach perturbs initial and boundary conditions to create (i) a pre-industrial world and (ii) a warmer future world, then compares these to the operational forecast to quantify anthropogenic impacts on event intensity and likelihood.

Event definition: Regional mean of daily maximum temperature over 45–52°N, 119–123°W. For ERA5 observations, the peak at 00 UTC on 2021-06-29 defines the event. For medium-range forecasts, the peak over 26–30 June 2021 is used; for seasonal forecasts, the peak over JJA 2021.

Experiments and models:

• Medium-range EPS (IFS CY47R2): 18 km atmosphere (Tco639), 137 vertical levels, coupled to 0.25° wave, sea ice (LIM2), and ocean (NEMO v3.4 ORCA025Z75, 75 levels). Ensemble size 51, except 11-day lead expanded to 251.
• Seasonal SEAS5 (IFS CY43R1): 36 km atmosphere (Tco319), 91 levels, coupled to 0.5° wave, LIM2, NEMO v3.4 ORCA025Z75. Ensemble size 51.

Perturbations to create counterfactuals:

1. Atmospheric CO2 set to 285 ppm (pre-industrial) and 615 ppm (future), symmetric about the present-day 420 ppm.
2. Estimated anthropogenic changes in ocean state (3D temperature), sea ice concentration and thickness are subtracted (pre-industrial) or added (future) via ocean restart files, using optimal fingerprinting with the Anthropogenic Warming Index (AWI). Data sources: ORAS5 sea ice (1958–2019), HadISST1.1 SST (1870–2019), WOA18 subsurface temperature (1950–2017). Regression coefficients at each grid point are scaled by the AWI change between 1850–1900 and 2019.
3. Quality control on sea ice fields to maintain physical bounds and consistency between concentration and thickness.
4. Salinity adjustment to preserve in-situ density after 3D temperature perturbations (thermodynamic consistency) using a gradient descent algorithm; no changes to initial circulation, MLD, or horizontal pressure gradients.

Land and atmosphere: Not directly perturbed (due to uncertainties in historical land-state trends and minor role of soil moisture for this event), allowing free adjustment. This leads to partial atmospheric adjustment at short lead times.

Initialisation dates and adjustment handling: Forecasts initialised at multiple leads: 3, 7, 11 days (medium-range) and May 1 for JJA (seasonal, 1–4 months). Predictability and conditioning vary with lead time. To account for incomplete atmospheric adjustment at shorter leads, the local event response is scaled linearly by the coincident global mean land temperature anomaly to a present-day attributable warming of 1.6 °C, supported by a linear relationship observed across leads and a perfect-model slab-ocean experiment.

Bias correction (seasonal): Address SEAS5 climate drift and biases over PNW by (i) removing forced trend via regression on AWI for ERA5 and hindcasts (1981–2020, excluding 2021), (ii) estimating and removing lead-dependent drift by regression versus lead time, and (iii) correcting remaining mean bias in annual maximum temperatures (TXx). Used for estimating seasonal model climatology and probabilities.

Statistical estimation:

• Intensity change: Select ensemble members in the nearest quintile to the event from the operational ensemble and compute mean differences between counterfactual and operational ensembles (upper-tail focused for longer leads).
• Risk change (relative risk): Fit an exponential tail (straight line in return-time space) to the nearest quintile of the operational ensemble (medium-range) or model climatology (seasonal) to estimate tail probabilities at or above the observed threshold. Shift the fitted distribution by the attributable intensity change and compute risk ratio = P_current / P_shifted. Confidence intervals from 10,000-member nonparametric bootstrap (17–83% likely range).

Validation of model reliability: ECMWF forecasts accurately predicted the PNW heatwave’s magnitude and drivers at 7–11 day leads, with seasonal forecasts capturing large-scale anomalies (e.g., elevated 500 hPa heights). Predictability remains stable under perturbations, and attributable signals align with canonical warming responses (thicker lower troposphere, increased water vapor).

• Anthropogenic influence increased the likelihood of an event at least as warm as the 2021 PNW heatwave by a factor of 8 [2, 50] overall (FAR ≈ 0.9 [0.5, 0.98]).
• On seasonal (least conditioned) timescales, relative risk ≈ 5 [2, 9] (FAR ≈ 0.8 [0.6, 0.9]).
• Event intensity attribution (synthesised across leads and scaled to current warming of ~1.25 °C): best-estimate anthropogenic contribution of 1.3 °C [0.7, 1.6]. Seasonal-only attributable warming: 1.2 °C [0.8, 1.6]. Medium-range attributable warming ranges from 0.7 [0.1, 1.3] (11-day lead) to 1.5 [1.3, 1.7] (3-day lead).
• Likelihood doubling time: At the current rate of global land warming, the probability of an event at least as warm as the 2021 PNW heatwave is doubling every 20 [10, 50] years.
• Forecast skill: ECMWF medium-range forecasts captured the event magnitude by 11-day lead and many ensemble members captured key drivers (atmospheric river penetration, low cloud cover, soil moisture anomalies) by 7 days. Seasonal forecasts captured large-scale anomalies (elevated 500 hPa geopotential heights, warmer troposphere) though not the exact sequence leading to record temperatures.
• Relationship to global warming: Log probabilities and local intensity shifts scale approximately linearly with the coincident global mean land warming, suggesting anthropogenic influence primarily modulates development from precursor conditions rather than the occurrence probabilities of the precursors themselves.

The study demonstrates that a reliable, high-resolution operational forecast model that successfully predicted the 2021 PNW heatwave can be used to quantify anthropogenic impacts on the event’s intensity and risk. By conditioning on model skill for the specific event, the approach avoids over-reliance on extrapolating from physically dissimilar events and addresses a key credibility gap in extreme event attribution. The findings indicate that human influence contributed roughly 1.3 °C to the heatwave peak and increased the event probability by factors consistent with rapid risk escalation under ongoing warming. The approximately linear scaling of intensity and log-probability with global land warming across lead times implies that anthropogenic effects act mainly on how precursor states evolve into extremes, rather than strongly changing the precursor occurrence itself. Compared to some previous climate-model-based studies, relative risks are lower, plausibly because the forecast model simulates relevant multi-process dynamics at higher resolution and because the imposed perturbations did not include all anthropogenic influences (e.g., land/atmosphere initial states, aerosols). The results reinforce the value of a forecast-based framework that bridges storyline (highly conditioned, short lead) and probabilistic (less conditioned, longer lead) attribution within a single, operationally viable methodology, providing information relevant to risk management and adaptation planning.

This work introduces and applies a forecast-based attribution framework that leverages a demonstrably reliable operational forecast system (ECMWF EPS) to attribute a specific unprecedented heatwave. Counterfactual forecasts with physically consistent ocean and CO2 perturbations, combined with multi-lead experiments and scaling by global land warming, yield robust estimates of human influence on both intensity and probability. The approach synthesizes storyline and probabilistic attribution in one framework and is directly compatible with operational forecasting, enabling potential real-time or anticipatory attribution services. Future directions include: adding state-consistent perturbations to land and atmospheric initial conditions; incorporating additional forcings (aerosols and non-CO2 greenhouse gases); applying the method across multiple forecast models to assess robustness; generating ensembles of perturbations to represent ocean-state uncertainty; and extending to impact-focused projections by simulating observed events under specified warming levels (e.g., 2 °C worlds) to inform adaptation planning and stress testing.

• Single-model dependence: All results rely on the ECMWF EPS/SEAS5; multi-model application is needed to assess robustness.
• Incomplete anthropogenic perturbations: Land surface and atmospheric initial conditions were not perturbed, potentially underestimating anthropogenic effects and missing nonlinear cross-component interactions.
• Adjustment/spin-up: Short-lead forecasts undergo atmospheric adjustment to perturbed ocean states, requiring post hoc scaling by global land warming; linearity may not hold for all events, especially those dominated by nonlinear dynamical drivers.
• Forcing scope: Only CO2 was altered explicitly; other GHGs and aerosols were not included, possibly underrepresenting total anthropogenic influence.
• Ocean perturbation uncertainty: Subsurface temperature perturbations are uncertain due to sparse pre-ARGO observations; only a single best-estimate perturbation was used (no perturbation ensemble).
• Seasonal model drift and bias: SEAS5 required bias correction; validation of extremes is challenging given the unprecedented 2021 event and sensitivity of higher-order moments.
• Seasonal forecast limitations: Seasonal ensembles did not reproduce the precise event sequence or observed peak magnitude, constraining direct comparison at that lead.
• Event-specificity: The linear scaling relationships and predictability characteristics found here may not generalize to other events or regions without verification.