acceptodds
Under review as a conference paper at ICLR 2027

One Encoding, Continuous Lead Times: Coherent Atmospheric Representations with Compositional Finite-Time Flows

Abstract

Weather forecasts are often required at lead times that do not coincide with the temporal cadence of available data or trained forecasters. Although atmospheric evolution is continuous, most data-driven models learn transitions on a predefined temporal grid. Forecasts away from that grid therefore rely on interpolation, repeated fixed-step rollout, or separately learned horizon mappings, which neither explicitly represent intermediate evolution nor ensure that predictions at different lead times belong to the same atmospheric trajectory. The central challenge is thus not simply to provide time as an additional input, but to couple queried horizons through a common atmospheric representation and evolution rule. We present PhiCast, a coherent latent atmospheric evolution model that addresses this challenge through two coupled mechanisms. PhiCast encodes the observed multivariate atmosphere once as a shared spatial latent, converts forecast duration into a smooth state-aware coordinate, and uses a solver-free finite-time flow to map that same encoding directly to each queried horizon. This structure makes intermediate lead times directly accessible without interpolation, autoregressive feedback, numerical integration, or repeated encoding. Because direct queryability alone does not ensure that separately requested horizons describe one trajectory, composition-aware learning further aligns direct and temporally decomposed paths in both latent and forecast spaces. At lead times without direct forecast supervision, PhiCast attains the best Root Mean Square Error (RMSE) and Anomaly Correlation Coefficient (ACC) for all four evaluated variables, reducing variable-wise RMSE by 4.4–20.4% relative to the strongest competing result for each variable. Representation diagnostics further demonstrate smooth latent trajectories and substantially lower composition inconsistency. By making forecast duration an intrinsic coordinate of one coherent atmospheric representation, PhiCast establishes a path toward weather models whose temporal resolution is decoupled from the observation and supervision cadences that produced their training data.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
data-driven weather forecasting, continuous-time forecasting, latent atmospheric dynamics, intermediate lead-time prediction, finite-time flow, composition-aware learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_XB7QOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes PhiCast, a continuous lead-time weather forecasting framework designed to generate predictions at arbitrary future horizons from a single encoded atmospheric state. The method consists of three main components: (1) a continuous time encoding module that combines Fourier-based temporal features with state-dependent FiLM conditioning; (2) a solver-free finite-time flow that integrates multiscale spectral and convolutional operations to directly map the initial latent representation to a queried future state without numerical integration or autoregressive rollout; and (3) composition-aware learning objectives that encourage consistency between direct and temporally decomposed latent transitions and forecasts.

The authors evaluate PhiCast on WeatherBench using four atmospheric variables—2-meter temperature, 10-meter zonal and meridional winds, and 500-hPa geopotential—at a spatial resolution of 5.625°. The model is trained with supervision at four lead times (6, 12, 18, and 24 hours) and evaluated at all integer-hour lead times from 1 to 24 hours. Experiments against seven baselines, including ClimODE, ClimaX, WeatherGFT, FNO, and SFNO, show that PhiCast achieves the best average RMSE and ACC across all four variables at the 20 lead times without direct forecast supervision. Additional experiments examine temporal composition consistency, component ablations, inference efficiency, and zero-shot transfer to a finer spatial resolution of 2.8125°.

Critical strengths

The paper addresses a practically relevant problem of forecasting at flexible lead times, enabling direct predictions at lead times without explicit training supervision, which offers some practical value in real-world forecasting scenarios.

Critical weaknesses

We believe that this paper has many issues and it is difficult to solve them through rebuttal:

1. Limited Methodological Novelty.: PhiCast utilizes mature technologies, mainly including Fourier Neural Operators (FNO, FourCastNet) for spatial modeling, FiLM conditioning (FiLM, ClimaX, WeatherGFT, Stormer) and identity-preserving mappings and semigroup consistency regularization (Deep-OSG). Even its composition-aware objectives largely rely on existing consistency constraints and conventional forecasting losses. Overall, the framework appears to be an integration of existing methods rather than a fundamentally novel approach, with insufficient evidence demonstrating the necessity or unique contribution of their combination.

2. Outdated and Insufficient Baseline Comparisons. :The experimental comparisons are limited to methods published in or before 2024, omitting several important weather forecasting models, including FourCastNet, Pangu-Weather, Stormer, ARROW, and traditional methods like IFS. Moreover, most baselines are reimplemented or retrained under modified experimental settings rather than evaluated using their official pretrained checkpoints. These omissions and modifications raise substantial concerns about the fairness and representativeness of the comparisons, making the reported performance improvements less convincing.

3. Limited Experimental Settings.:The experiments are limited to a coarse spatial resolution of 5.625°, only four atmospheric variables, and a 24-hour forecasting horizon, which fall short of modern weather forecasting standards. Such settings offer limited practical value and fail to convincingly demonstrate the benefits of forecasting at lead times without direct supervision.

4. Questionable Lead-Time Sensitivity.: Figure 11 raises a serious concern that predictions across different lead times appear nearly identical, suggesting that the model may primarily produce similar atmospheric states regardless of the requested horizon. The increasing forecasting error with lead time does not necessarily demonstrate meaningful temporal modeling, as even a fixed prediction could exhibit this pattern. The authors should evaluate prediction differences across lead times for the same historical input, compare against persistence baselines, and conduct cross-horizon substitution tests (e.g., evaluating a 1-hour prediction against 6-, 12-, and 24-hour targets) to verify whether the model genuinely learns lead-time-dependent atmospheric evolution.

Questions for authors

See Weaknesses.

The authors should standardize the font styles, sizes, and text formatting across Figure 3 and the subsequent code-generated figures to improve visual consistency and overall presentation quality.

Reviewer_5gzVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

PhiCast encodes an atmospheric field once and predicts arbitrary lead times with a duration-conditioned latent map and shared residual decoder. A zero-duration gate enforces the initial condition, while latent and decoded composition objectives encourage agreement between direct and staged queries. On four WeatherBench variables at 5.625° resolution, training uses 6/12/18/24-hour targets and testing covers every integer hour in 1–24. Three-seed comparisons show improved intermediate-lead accuracy; additional experiments assess composition, latency and zero-shot transfer to 2.8125°.

Critical strengths
  1. Tables 1–2 evaluate all 1456 test origins with three training seeds, including intervals strictly between supervised horizons. The 4.4–20.4% variable-wise RMSE gains over the strongest reported competitors are meaningful. The interior-interval evaluation, block-bootstrap analysis and native 2.8125° transfer substantially strengthen the evidence beyond selected examples.
  2. The cached encoding and exact zero-time anchor are sensible design choices. Composition ablations demonstrate a real reduction in path disagreement, and Appendix E states the assumptions and limits of the theory. Appendix F.5 also compares against a cached ClimODE integration, appropriately limiting the efficiency claim.
Critical weaknesses
  1. The central algorithmic increment is insufficiently established. Direct finite-time maps and semigroup regularization have clear precedents in Neural Flows and Deep-OSG [1–2], both cited by the paper. The distinctive contribution therefore needs more than combining these ideas with a spatial backbone. Table 8 shows small, mixed changes in forecast accuracy when both composition losses are removed; the full model slightly worsens both unseen-lead wind RMSEs. Improved internal consistency is useful, but it does not establish more faithful intermediate dynamics. Figure 4 also compares PhiCast's retained latent with a weather-field restart of ClimODE, although the paper acknowledges that native augmented-state continuation is nearly exactly compositional. This comparison cannot establish superior dynamical coherence over native ClimODE.
  2. The most direct alternatives to the claimed benefit are not resolved. The evaluation excludes interpolation altogether, although avoiding interpolation motivates the work. A strong endpoint forecaster with interpolation between its predicted endpoints is an essential practical reference under the same information budget. Appendix C selects checkpoints using all 24 validation leads, so the 20 intermediate leads are absent from training-target supervision, not from model selection. This is not test leakage, but the stronger sparse-temporal-supervision interpretation needs a restricted-validation check.

[1] Biloš et al. Neural Flows: Efficient Alternative to Neural ODEs. NeurIPS 2021. [2] Chen and Wu. Deep-OSG: Deep Learning of Operators in Semigroup. JCP 2023.

Questions for authors
  1. Can you compare against interpolation of the strongest forecaster's predicted 6-hour endpoints, using the observed origin for the first interval and no future ground truth? Report the interior-interval scores of Table 2 and query-suite cost. This would determine whether direct continuous querying adds practical value.
  2. With the same backbone, time encoding and exact origin anchor, how much does the proposed composition scheme add over a standard Deep-OSG-style consistency loss? Please report forecast accuracy and decoded consistency across seeds, and include native augmented-state ClimODE continuation in the composition diagnostic.
  3. Does the intermediate-lead ranking persist when checkpoint selection uses only the four supervised validation horizons? Please also provide parameter counts and training cost for PhiCast and the closest direct baselines, so architectural capacity can be separated from the temporal formulation.
Reviewer_PWFNOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes PhiCast, which reframes weather forecasting as continuous querying of latent embeddings named "atmostphere", which is encoded once and is used throughout inference. Experiments on WeatherBench 5.625° show considerable improvement in RMSE and ACC across baselines.

Critical strengths
  • Caching and creating a queryable latent atmospheric state from data is interesting and novel
  • The semigroup loss is an elegant route for propagating sparse supervision.
Critical weaknesses
  • Cached z₀ cannot be “one atmosphere” because four 2D fields are not an atmospheric state. The input is T2m, u10, v10, and Z500 on a 32×64 grid, three near-surface diagnostics, and one upper-level mass field, with no vertical temperature or wind structure, no humidity, no pressure levels of the atmosphere, and no vertical velocity. The naming is misleading and misguides the paper's premise.
  • Table 8 shows that the composition objectives do not improve accuracy and slightly hurt wind RMSE.
  • Compositional coherence is measured as self-consistency. A constant predictor composes perfectly. Lower disagreement between φ(z0,24)φ(z_0,24) and φ(φ(z0,12),12)φ(φ(z_0,12),12) tells us the model agrees with itself, not that either path is right.
  • Many baselines in the paper are continuously queryable; it cannot be treated as a major contribution.
  • Two of the three “contributions and coupled requirements” are satisfied by many models in the comparison, so the novelty claim is very shallow.
Questions for authors
  • Given that the input is three near-surface diagnostics and one 500 hPa level, in what sense is z0z_0 a sufficient statistic for 1–24 h evolution?
  • Figure 4(b) is a through-origin fit on a map that satisfies φ(z0,0)=z0φ(z_0,0)=z_0 exactly. What does this diagnostic measure beyond the architectural constraint?
  • Since c(h)c(h) is globally average-pooled and γ,β\gamma,\beta are broadcast spatially, the time–state conditioning is one modulation for the entire globe. How can this represent the method's regime-dependent prediction?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.