Research map / Label-efficient and robust learning / OOD and anomaly detection

IDOL: Meeting Diverse Distribution Shifts with Prior Physics for Tropical Cyclone Multi-Task Estimation

NeurIPS 2025 · Label-efficient and robust learning › OOD and anomaly detection · approach: Robust fine-tuning for OOD generalization · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes IDOL, a physics-guided identity-token framework (Holland wind model plus dark correlation graph) that regularizes TC feature distributions to estimate wind speed, pressure, and inner/outer-core size under distribution shift.

Problem

Tropical cyclone multi-task estimation (wind speed, pressure, inner-core and outer-core size) degrades under covariate, label, and concept distribution shifts caused by varying geography and seasons, so models generalize poorly to unseen TCs.

Approach

Instead of regularizing raw features, model Gaussian-sampled task-specific and task-shared identity tokens constrained by prior physics: the Holland wind-field relation drives a Prior-aware Approximator for inter-task dependencies, and a dark knowledge graph of TC correlation factors drives a GMM-based task-shared token, with an identity-oriented loss aligning intra-sample correlations with labels

Key results

• On PDTC, IDOL reaches wind speed MAE 5.93 vs 8.84 for DeepTCNet (best prior), and pressure MAE 5.77 vs 6.99 for PeRCNN.
• On Digital Typhoon, IDOL achieves MAE 5.82 for wind speed and 24.74 for outer-core size, best across all four tasks.
• Ablation shows adding task-specific identity tokens reduces wind speed MAE from 10.13 to 7.24, and adding the task-shared token further lowers it to 5.93.

Measured effects reported in the paper

Surprising findings

• Mutual information between task-specific identity tokens exceeds that between the ground-truth attributes themselves.
• A lightweight CNN backbone (12.88M params) gives comparable accuracy to the VGG13 version (275.89M), with slightly better inner-core MAE.

Limitations

• No error bars or statistical significance reported, following prior TC estimation work.
• Inner-core size RMSE and STD for IDOL are not uniformly best on PDTC, and the model's advantage is partly tied to hand-specified prior physics.
• Evaluation is limited to Western North Pacific TC data from 2015-2023 with a single temporal split.

Experimental setup

Datasets: Physical Dynamic TC (PDTC) dataset built from IBTrACS and Himawari-8 data Digital Typhoon dataset MGTCF/TC-Diffuser forecasting dataset (1950-2021 WNP)

Benchmarks: PDTC Digital Typhoon

Metrics: MAE RMSE STD

Baselines: ADT MTCSWA STIA NS TC-MTLNet DeepTCNet PeRCNN IRM V-REx SADE DirMixE

Builds on: Holland wind field model LoCa dark knowledge PeRCNN IRM DirMixE

Closest papers by meaning

• The Perception–Physics Paradox: Probing Scientific Alignment with TC-Bench (ICML 2026 · similarity 0.84)
• TC-Diffuser: Bi-Condition Multi-Modal Diffusion for Tropical Cyclone Forecasting (AAAI 2025 · similarity 0.84)
• VQLTI: Long-Term Tropical Cyclone Intensity Forecasting with Physical Constraints (AAAI 2025 · similarity 0.84)
• TCP-Diffusion: A Multi-modal Diffusion Model for Global Tropical Cyclone Precipitation Forecasting with Change Awareness (ICML 2025 · similarity 0.82)
• OceanCBM: A Concept Bottleneck Model for Mechanistic Interpretability in Ocean Forecasting (NeurIPS 2026 · similarity 0.81)
• Common Task Framework For a Critical Evaluation of Scientific Machine Learning Algorithms (NeurIPS 2025 · similarity 0.81)