acceptodds
Withdrawn from ICLR 2027

PG-MoE: Physically Grouped Mixture-of-Experts with Soft Track Generation for Multi-Horizon Typhoon Track Prediction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Accurate multi-horizon typhoon track forecasting is crucial for disaster prevention and mitigation, yet existing deep learning methods for ensemble forecast post-correction suffer from two key limitations: they discard probabilistic uncertainty from ensemble members, leading to error accumulation, and single networks lack capacity for diverse typhoon dynamics, while conventional MoE lacks physical priors. We propose PG-MoE, integrating soft-track probabilistic encoding and physically grouped mixture-of-experts. The soft-track module converts multi-source ensemble forecasts into standardized probabilistic representations via a downsampling-free fully convolutional network and confidence ellipse fitting, preserving uncertainty evolution. The MoE partitions experts into 23 homogeneous groups based on six physical attributes (e.g., lead time, intensity, moving distance) with a hierarchical sparse router and two-stage decoupled training. On a Northwest Pacific typhoon test dataset (2024–2025), PG-MoE significantly outperforms mainstream deep learning baselines and ECMWF deterministic forecasts within the critical 12–72H window. Ablations validate the core components. Our method offers a general probabilistic input pipeline and a physics-informed MoE paradigm, balancing accuracy and interpretability. Current framework supports MLP experts and track data only, with future extensions to multimodal observations and adaptive partitioning.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Mixture-of-Experts, Sparse Hierarchical Routing, Typhoon Track Forecasting, Ensemble Forecast Post-processing, Probabilistic Uncertainty Modeling, Physically Informed Machine Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_wGRyOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PG-MoE framework merges and corrects ECMWF and NCEP GEFS ensemble outputs to predict typhoon tracks across multiple horizons. It relies on a Soft-Track module to translate ensemble paths into probability heatmaps and elliptical features, which are subsequently handled by frozen, physically grouped MLP experts under a hierarchical sparse router. In benchmarks covering 2024–2025 Northwest Pacific typhoons, the model beats evaluated baselines in 18 of 20 year–horizon pairs (e.g., cutting 48-hour errors from ECMWF's 91.59 km and 90.83 km down to 86.07 km and 86.79 km), with ablation studies supporting each architectural component. The results demonstrate improved typhoon track accuracy by combining ensemble-derived uncertainty descriptors with physically grouped experts.

Critical strengths
  1. The evaluation uses a subsequent-year test period, operational forecast baselines, and a broad set of component and architectural ablations. Table 2 reports improvements over ECMWF deterministic forecasts at several relevant horizons; for example, the 2024 48-hour error decreases from 91.59 km to 86.07 km. Together, the forecast comparisons and ablations provide useful evidence of the pipeline's predictive value and allow scrutiny of its individual design choices.
Critical weaknesses
  1. The experiments do not establish a practical advantage for hierarchical routing or separate expert specialization from increased model capacity.

Table 6 shows that Flat-R has lower mean track error than PG-MoE at 9 of 10 horizons in 2024 and 6 of 10 in 2025. These comparisons concern reported means, not established statistical significance, but they do not support an accuracy advantage for hierarchical routing. Appendix B.2 instead emphasizes more balanced activation and more stable routing accuracy. Balanced expert utilization is not itself a forecasting objective, and no corresponding practical benefit is established. The claim that hierarchical routing is essential for robust long-term generalization is therefore insufficiently supported.

The Single-E comparison in Table 4 does not resolve the capacity question. Section 4.4.1 claims matching "activation compute", but does not provide Single-E's exact architecture or numerical parameter counts. A single MLP matched to the total parameter count of the expert bank and routers is an important missing baseline. The practical benefits for enabling MoE for such small networks are hardly justifiable (lower VRAM requirement, lower latency etc.). Thus a comparison at this total capacity is important for testing whether physical specialization adds value beyond a larger prediction network.

  1. The Soft-Track representation and physical partitions are insufficiently specified and validated.

Table 10 provides limited evidence that preserving the full ellipse geometry improves prediction. Despite the information loss discussed in Appendix B.6, four boundary points remain competitive with six and eight, and achieve the lowest 120-hour error in both test years. Comparisons using only the center, as well as the center and covariance of the Gaussian (5 DoF in total), would help establish how much the uncertainty representation contributes beyond the estimated center.

Section 3.2 motivates the expert boundaries using physical guidelines, but their effect on prediction accuracy is not directly tested. Appendix B.3 compares using buffers with removing them, without varying the boundaries or buffer widths. Since physical partitioning is a central contribution, comparisons with nearby thresholds or alternative partitions would help establish whether the chosen divisions are beneficial and how sensitive the results are to these choices.

Questions for authors
  1. What practical benefit justifies choosing hierarchical routing over the simpler flat router? Flat-R achieves lower mean errors in 15 of the 20 comparisons in Table 6. More balanced expert activation alone does not establish a useful advantage for this small forecasting model. What evidence shows that the added routing complexity benefits the forecasting task?

  2. Capacity and ablation controls: What is Single-E's exact architecture, and how is "activation compute" matched? Please report parameter counts and compare forecast accuracy against a single prediction network matched to the total parameters of the experts and routers. For Dense-MoE, what exactly is the setup of this experiment? For Table 7, do Top1 and Top2 retain the same expert bank and change only the number selected per group?

  3. Soft-Track input, supervision, and representation: How are the input and target of the FCN constructed? Please clarify whether the FCN is trained separately and held fixed for subsequent MoE training. Please also compare sampled ellipse points with direct center-covariance features mentioned in the weakness.

  4. Physical partition validation: How sensitive are the results to changes in the Table 1 boundaries and buffer widths? Do the physically motivated boundaries improve accuracy over nearby thresholds or alternative partitions, such as equal-frequency bins, with the same expert architecture?

Reviewer_CTbKOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper tackles post-correction of ensemble typhoon track forecasts in the Northwest Pacific. It has two components: (1) Soft-Track, which renders raw ECMWF/NCEP-GEFS ensemble tracks into a 0.25° probability heatmap via a downsampling-free FCN (trained with a KL loss), fits a 75% bivariate-normal confidence ellipse, and compresses it into 14-dim features (weighted center + 6 boundary points per lead time); and (2) PG-MoE, a bank of 23 small MLP experts partitioned by six physical attributes (lead time/“delay”, intensity, move distance, turn angle, coast distance, area) with overlapping buffers, a two-level sparse router (global top-3 groups, sub-router top-2 experts), and a decoupled two-stage training scheme (experts pre-trained with Haversine loss, frozen, then router trained). On 2024–2025 test typhoons, PG-MoE reports the lowest average track error at most lead times vs. BiLSTM/ConvLSTM/TMPI and ECMWF/NCEP operational forecasts.

Critical strengths

The paper addresses a practically important problem in ensemble typhoon-track post-processing. Converting variable-size ensemble forecasts into a fixed-size representation is useful, and the physically organized MoE provides an intuitive way to inspect routing behavior. The paper also includes a broad set of architectural and hyperparameter ablations.

Critical weaknesses
  1. The main comparison is not conducted on equal footing. The appendix states that NCEP baselines “can only be evaluated on a reduced subset with complete NCEP records,” yet Table 2 places them in the same ranking as PG-MoE, which is evaluated on all samples. The headline claim of “only two instances of second-best results” therefore rests on cross-subset comparisons — most consequentially at 2024-120H, where NCEP_D (314.52 km) beats PG-MoE (348.11 km) on a different, possibly easier, sample set. Additionally, PG-MoE consumes the combined ECMWF+GEFS ensemble, while the ECMWF_A/NCEP_A baselines are single-source; a combined-ensemble-mean baseline is the fair “trivial” reference and is missing. All main results should be re-reported on the intersection of samples, with per-sample paired comparisons.

  2. The central claim — that physical grouping beats unsupervised grouping — is never directly tested. The introduction criticizes unsupervised clustering-based MoE partitioning, but no ablation instantiates the same architecture with k-means/random/hash-based expert groups under identical sparse routing and two-stage training. The closest variants keep the physical partition (NoPrior-R only removes the router’s inputs). Meanwhile, Dense-MoE matches or beats PG-MoE at several cells (2024-72H: 135.60 vs. 138.23; 2024-120H: 339.27 vs. 348.11; 2025-6H: 39.07 vs. 40.24), so the accuracy benefit of physical grouping is unproven — its defensible value is interpretability and routing stability, and the paper should be re-framed accordingly.

Questions for authors

1.How exactly is the KL target heatmap YY constructed (kernel type, width)? Is the fitted ellipse’s empirical coverage close to 75%?

2.Please give explicit computational definitions of all six attributes and certify that none uses post-initialization best-track information.

3.Can you report standard deviations over seeds and paired significance versus ECMWF_D at 12/24/48/72H?

4.Does the method still beat ECMWF_D and NCEP_D when all methods are restricted to the common complete-sample subset?

5.How does the proposed physical grouping compare with k-means, random, or hash-based expert grouping under the same architecture, sparse routing mechanism, two-stage training procedure, and computational budget?

Reviewer_tokgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes PG-MoE for multi-horizon typhoon track prediction. SoftTrack converts ECMWF and NCEP ensemble tracks into a spatial heatmap and confidence ellipse, and a hierarchical router combines experts grouped by hand-designed physical attributes such as intensity, motion, turning angle, and coastal distance. The method is evaluated on chronological 2024 and 2025 test periods with several architecture and grouping ablations.

Critical strengths
  1. The problem is operationally important, and the attempt to combine ensemble guidance with storm-specific routing is well motivated. Using future calendar years for testing is also preferable to a random temporal split.
  2. The paper includes comparisons with a dense MoE, expert-removal studies, lead-time breakdowns, and case analyses rather than relying only on one aggregate error number.
Critical weaknesses
  1. The reported means over three seeds are not accompanied by standard deviations, confidence intervals, or paired event-level tests, and the paper does not state the number of storms, initialization cycles, or valid samples at each horizon. Forecast cycles from the same storm are correlated, so seed variation alone would not establish the claimed significance. Missing NCEP forecasts and the fact that SoftTrack retains samples that some baselines discard also make it unclear whether every comparison uses the same cases.
  2. The experiments do not clearly isolate the physically grouped MoE. Dense-MoE matches or exceeds PG-MoE in several year–horizon cells, and removing individual physical groups sometimes improves performance. SoftTrack also compresses an ensemble to a heatmap-derived ellipse and seven points, which cannot preserve multimodality or tail structure despite the stronger wording in the paper. Equal-data comparisons and storm-level uncertainty are needed before attributing gains to the physical grouping.
Questions for authors
  1. How many distinct storms, initialization cycles, and valid samples are used at each horizon, and are all methods evaluated on exactly the same cases when NCEP members are missing?
  2. Can the authors report paired storm-level bootstrap intervals for PG-MoE versus Dense-MoE and the strongest non-MoE baseline?
  3. What exactly are the FCN input and KL target used by SoftTrack, and is every routing attribute and target-derived quantity available without leakage at inference time?

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

Related papers

Discussion (0)

Sign in to comment.