Research map / Probabilistic and scientific modeling / Dynamical systems

OceanCBM: A Concept Bottleneck Model for Mechanistic Interpretability in Ocean Forecasting

NeurIPS 2026 · Probabilistic and scientific modeling › Dynamical systems · approach: Data assimilation from sparse observations · 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.

Introduces OceanCBM, a U-Net concept bottleneck model with four physically prescribed ocean concepts plus one free concept, trained with mixed supervision to predict mixed layer heat content and diagnose the 2012 Gulf of Maine marine heatwave.

Problem

Ocean forecasting models are accurate but opaque, and skillful forecasts do not reveal the physical drivers of extreme events such as marine heatwaves; existing physics-informed and post-hoc XAI approaches do not guarantee physically meaningful mechanisms.

Approach

A concept bottleneck U-Net that predicts physically derived mixed-layer concepts (vertical shear, buoyancy frequency, heat flux entrainment, MLD tendency) plus one unconstrained free concept, trained with a decaying-weight mixed supervision loss over an ensemble of ORAS5 members to yield more consistent mechanistic pathways than prediction-only or prescription-only models.

Key results

• OceanCBM achieves seasonal MLHC ACC of about 0.91-0.92, essentially matching the prediction-only model (0.91-0.92).
• Heat flux entrainment reaches JJA ACC of 0.77 and buoyancy frequency a maximum ACC of 0.81.
• Vertical shear ACC improves over the prescription-only baseline (e.g., JJA 0.6727 vs 0.5774), which the authors attribute to the free concept regularizing concept predictions.

Surprising findings

• The free concept receives the dominant linear weight in the bottleneck yet does not fully control the prediction, as spatial discrepancies show MLHC is not reducible to it.

Limitations

• The prescribed concept set is non-exhaustive, and the free concept's physical interpretation is unconstrained.
• Only five ensemble members from one reanalysis and monthly-averaged data over the North Atlantic; generalization to other basins or finer timescales is untested.
• Consistency analysis depends on random initialization alone, and baselines share seeds with OceanCBM.

Experimental setup

Datasets: ORAS5

Benchmarks: ORAS5 held-out validation and test periods

Metrics: anomaly correlation coefficient ensemble spread of normalized regression weights

Baselines: prediction-only model prescription-only model

Builds on: concept bottleneck models U-Net process-guided CBM

Closest papers by meaning

• FuXi-Ocean: A Global Ocean Forecasting System with Sub-Daily Resolution (NeurIPS 2025 · similarity 0.86)
• Njord: A Probabilistic Graph Neural Network for Ensemble Ocean Forecasting (NeurIPS 2026 · similarity 0.86)
• Common Task Framework For a Critical Evaluation of Scientific Machine Learning Algorithms (NeurIPS 2025 · similarity 0.85)
• PINP: Physics-Informed Neural Predictor with latent estimation of fluid flows (ICLR 2025 · similarity 0.84)
• ChaosBench: A Multi-Channel, Physics-Based Benchmark for Subseasonal-to-Seasonal Climate Prediction (NeurIPS 2024 · similarity 0.84)
• Geometric and Physical Constraints Synergistically Enhance Neural PDE Surrogates (ICML 2025 · similarity 0.84)