Source-linked AI summary
EC-EarthFlow: Probabilistic emulation of daily transient global climate model simulations with flow matching
Kirien Whan, Nikolaj T. Mücke, Karin van der Wiel
TL;DR
Climate emulation needs daily transient projections, while prior generative approaches have not yet provided them. EC-EarthFlow uses flow matching to predict EC-Earth3’s next-day temperature from recent fields and annual mean temperature. It reproduces daily variability, annual-cycle features and warming trends, and remains stable through extended-season rollouts.
Problem
Prior generative climate approaches had not yet emulated daily, transient future climate projections.
Source: §1 · pp. 1–2Method
EC-EarthFlow uses generative flow matching to autoregressively predict EC-Earth3’s next-day temperature from recent temperatures and annual mean temperature.
Source: §1 · p. 2Results
EC-EarthFlow reproduces daily variability, global annual-cycle amplitude and phase, warming trend, and eastern-Australia spring–summer variability, remaining stable through extended-season rollouts.
Source: §4 · pp. 8–9Takeaways & Limitations
Temperature-only training reproduced teleconnection-like wave patterns and blocking signatures, suggesting the model learned aspects of atmospheric dynamics.
Source: §4 · p. 10The proof of concept models only 2m temperature, and higher spatial resolution is required for practical use.
Source: §4 · p. 11
Abstract
from arXiv · showhide
We introduce EC-EarthFlow, a generative flow matching model that emulates simulations from the physical climate model EC-Earth3. The model is trained on transient simulations from EC-Earth3 (1950-2166, SSP2-4.5) to predict the day ahead temperature field from the previous days temperature as well as annual mean temperature. Predictions are made auto-regressively with rollout periods of between a month and an extended season. Using only this variable of interest, we are able to reproduce the daily variability, spatial patterns, annual cycle and long-term trend from EC-Earth3 at a substantially lower computational cost than the physical model. We demonstrate that EC-EarthFlow is stable for long inference periods, and that it can learn the physical relationships as simulated in EC-Earth3.
1. Introduction
Large ensembles help climate scientists distinguish forced responses from internal variability, but their computational cost makes them a target for generative emulation. Existing approaches leave a gap in emulating daily, transient future climate projections, which this work addresses for EC-Earth3 temperature.
1. Introduction
- Large ensembles help separate forced responses from internal variability, but are costly to run, motivating generative deep-learning emulation.They provide many realizations of the same future experiment from different initial conditions.Source: §1 · p. 1
- Climate applications of deep learning and generative methods remain less developed than weather forecasting, with long-term stability and climate non-stationarity presenting additional challenges.Deterministic models may lack a sophisticated probabilistic framework and can produce smooth forecasts with reduced high-frequency variability.Source: §1 · p. 1
- Bassetti et al. (2024) generate fixed 30-day temperature or precipitation sequences, representing future-climate extremes well but introducing discontinuities; independent hourly samples cannot simulate multi-day events, and no approach yet emulates daily transient future projections.Source: §1 · pp. 1–2
- The paper aims to emulate EC-Earth3’s 1950–2166 transient daily near-surface temperature, including its forced trend, with autoregressive flow matching.This is presented as a proof of concept focused on one high-impact variable for pragmatic computational and scientific reasons.Source: §1 · p. 2
2. Data and Methods
EC-EarthFlow is trained to predict the next day’s EC-Earth3 temperature field from recent temperatures and annual mean temperature. Flow matching generates conditional samples, which are iteratively fed back to produce ensembles over longer horizons.
2. Data and Methods
- The emulator uses EC-Earth3 p5 data, block-averaged to 128 × 64 grid cells, from 16 transient SSP2-4.5 runs spanning 1950–2166.Data before 1950 are discarded because of unrealistic oceanic oscillations; the final training stages use five ensemble members.Source: a. Data · p. 2, c. Implementation and training · p. 4
- Each prediction models the conditional distribution of next-day 2m temperature given five previous daily fields and the gridded annual-mean temperature.The recent fields carry temporal dynamics, while annual mean temperature fixes the mean climate state.Source: 1) Flow Matching Model · p. 2
- Conditional flow matching transports Gaussian noise to a next-day temperature sample by integrating a neural-network vector field over pseudo-time.Training interpolates linearly between noise and the target along an optimal-transport probability path.Source: 1) Flow Matching Model · p. 2
- A latitude-weighted mean squared error fits the vector field because regular latitude-longitude grid cells represent less physical area toward the poles.An unweighted loss would over-represent polar latitudes.Source: 1) Flow Matching Model · p. 2
- At inference, distinct noise draws generate ensemble members, and each predicted day is added to the context for the next autoregressive step.The flow is numerically integrated to produce each next-day prediction before updating the context window.Source: 1) Flow Matching Model · p. 3, Figure 3 · p. 4
- The vector field uses a PDE-Transformer, and the final model was trained in stages as compute resources became available.Training moved from one to five EC-Earth3 members, and the final model took less than 28 hours; the authors say the benefit of the multistage strategy is unclear.Source: c. Implementation and training · p. 4
3. Results
EC-EarthFlow reproduces daily temperature behavior, global mean temperature patterns, and spatial dependencies in EC-Earth3, while remaining stable across long rollouts. Its ensemble tends to be under-dispersed, and longer rollouts accumulate some biases more than shorter ones.
3. Results
- EC-EarthFlow captures daily variability, the annual cycle, and long-term warming at local grid points, including seasonal variability in De Bilt and Sydney.The model reproduces Sydney's increased October–December variability and De Bilt's increased extended-winter variability using five days of temperatures and annual-mean temperature as inputs.Source: a. Pointwise predictions · p. 5, Figure 4 · p. 6
- Rank histograms show that the EC-EarthFlow ensemble tends to be under-dispersed, with EC-Earth3 temperatures often outside its ensemble range.Rank patterns vary across locations and years; EC-Earth3 is warmer in some cases and cooler in others.Source: b. Ensemble spread · p. 5, Figure 5 · p. 6
- About 2.5°C of Northern Hemisphere summer warming by 2150 is captured alongside the annual cycle in global mean surface temperature.The annual mean temperature from EC-Earth3 is an input to the model.Source: c. Long-term trends · p. 5, Figure 6 · p. 7
- EC-EarthFlow generally preserves inter-annual GMST variability, with generated ensembles following the warming pattern of their conditioning EC-Earth3 member while retaining some spread.Ensembles generated from EC-Earth3 members 15 and 16 overlap.Source: c. Long-term trends · p. 5, Figure 7 · p. 7
- Long rollouts remain stable without errors accumulating into instabilities, and their annual cycle and trends are similar to those from short rollouts.Short rollouts return trajectories toward the EC-Earth3 test member at the start of each inference period; long-rollout spread increases initially and is then consistent through the extended season.Source: c. Long-term trends · pp. 5–6, Figure 8 · p. 8
- Mean differences between EC-EarthFlow and EC-Earth3 are consistent with expected differences, but are smaller than differences between EC-Earth3 members; warm polar winter bias and longer rollouts worsen some issues.The bias is linked to more grid points with positive bias, stemming from a warm bias at the poles in the winter hemisphere.Source: c. Long-term trends · pp. 6–7
- EC-EarthFlow reproduces temperature spatial autocorrelation and teleconnection-like patterns, including wave patterns associated with hot western Europe and cooler regions up- and downstream.Correlations are smoother than EC-Earth3 farther from the point of interest, including in the North Pacific in Sydney-centered plots.Source: d. Large-scale dynamics · pp. 7–8
4. Discussion and Conclusions
EC-EarthFlow reproduces key temperature variability and warming features of EC-Earth3 from a single variable, and remains stable through extended-season rollouts. The discussion also identifies limits in ensemble spread, spatial structure, and model scope.
4. Discussion and Conclusions
- EC-EarthFlow remains stable over extended inference periods, although annual and multi-annual rollout stability and trajectory quality need further study.Source: §4 · p. 9
- The ensemble spread is generally good but slightly under-dispersed, with parameter tuning proposed as a possible improvement.Source: §4 · p. 9
- The emulator reproduces the global mean surface temperature annual cycle and warming trend, alongside eastern Australia’s spring–summer variability and warming trend from 1950 to 2150.Source: §4 · p. 9
- The generated ensemble retains input-member dependence while maintaining internal spread, but that dependence may limit independent exploration of internal variability.Source: §4 · pp. 9–10
- The model produces spatially coherent polar temperature biases, which may reflect the regular latitude–longitude grid; tested normalization methods brought little improvement.Source: §4 · p. 10
- Spatial correlations reproduce teleconnection-like wave patterns and blocking signatures, but are smoother at large distances than in EC-Earth3.Source: §4 · p. 10, Figure 10 · p. 10
- This proof of concept models only 2m temperature; the authors identify additional variables and higher spatial resolution as needed for practical regional climate use.Source: §4 · p. 11