{"paper":{"title":"OceanCBM: A Concept Bottleneck Model for Mechanistic Interpretability in Ocean Forecasting","license":"http://creativecommons.org/licenses/by/4.0/","headline":"OceanCBM routes forecasts of ocean heat content through prescribed physical concepts plus one free concept to deliver both skill and mechanistic insight.","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Kieran Ringel, Maike Sonnewald, Sanah Suri","submitted_at":"2026-05-12T18:29:45Z","abstract_excerpt":"Extreme ocean phenomena are challenging not only to predict but to diagnose, as accurate forecasts alone do not reveal the underlying physical drivers. While recent machine learning approaches achieve strong predictive skill, they remain largely opaque and provide limited guarantees of fidelity to ground-truth physics. We introduce OceanCBM, the first concept bottleneck model (CBM) for spatiotemporal prediction and mechanistic interrogation of ocean dynamics. OceanCBM uses mixed supervision to predict mixed layer heat content, a key precursor of marine heatwaves, while routing information thro"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"OceanCBM achieves interpretable, physically grounded representations without sacrificing skill, explicitly characterizing the interpretability-performance trade-off. Across ensemble initializations, mixed supervision yields consistent mechanistic representations, whereas prediction-only and prescription-only baselines learn highly variable latent structures despite similar predictive performance.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the prescribed concepts derived from geophysical fluid dynamics are the right set to capture the key drivers of mixed layer heat content, and that adding one free concept plus mixed supervision will reliably produce consistent mechanistic representations rather than artifacts of the training process.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"OceanCBM is the first concept bottleneck model for spatiotemporal ocean prediction that uses mixed supervision on physical concepts and a free concept to deliver consistent mechanistic representations for mixed layer heat content forecasts.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"OceanCBM routes forecasts of ocean heat content through prescribed physical concepts plus one free concept to deliver both skill and mechanistic insight.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"60ed6d5d82c0a3664aabc184345feee0d73fb8f8663b61423db72b441b649a2c"},"source":{"id":"2605.12639","kind":"arxiv","version":1},"verdict":{"id":"bae51572-73bf-4e09-8554-729fca2eb828","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T21:45:07.876482Z","strongest_claim":"OceanCBM achieves interpretable, physically grounded representations without sacrificing skill, explicitly characterizing the interpretability-performance trade-off. Across ensemble initializations, mixed supervision yields consistent mechanistic representations, whereas prediction-only and prescription-only baselines learn highly variable latent structures despite similar predictive performance.","one_line_summary":"OceanCBM is the first concept bottleneck model for spatiotemporal ocean prediction that uses mixed supervision on physical concepts and a free concept to deliver consistent mechanistic representations for mixed layer heat content forecasts.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the prescribed concepts derived from geophysical fluid dynamics are the right set to capture the key drivers of mixed layer heat content, and that adding one free concept plus mixed supervision will reliably produce consistent mechanistic representations rather than artifacts of the training process.","pith_extraction_headline":"OceanCBM routes forecasts of ocean heat content through prescribed physical concepts plus one free concept to deliver both skill and mechanistic insight."},"references":{"count":46,"sample":[{"doi":"","year":2026,"title":"Process- guided concept bottleneck model.arXiv preprint arXiv:2601.10562, 2026","work_id":"6c85653c-c45f-4952-94a9-540008e95a25","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":1919,"title":"Artificial intelligence for modeling and understanding extreme weather and climate events.Nature Communications, 16(1):1919, 2025","work_id":"0cfb0f05-4d8e-46e0-9035-22c66185e136","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"A global overview of marine heatwaves in a changing climate.Communications Earth & Environment, 5(1):701, 2024","work_id":"75ef718e-a325-48a4-84ba-ba14c7c9bab7","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2012,"title":"Ke Chen, Glen Gawarkiewicz, Young-Oh Kwon, and Weifeng G Zhang. The role of atmospheric forcing versus ocean advection during the extreme warming of the northeast us continental shelf in 2012.Journal ","work_id":"36eb184b-b54e-4821-a6bb-e7de8a91fef9","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2012,"title":"Ke Chen, Glen G Gawarkiewicz, Steven J Lentz, and John M Bane. Diagnosing the warming of the northeastern us coastal ocean in 2012: A linkage between the atmospheric jet stream variability and ocean r","work_id":"6e0c4d36-ee2b-4be1-8a16-4a6251c43b22","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":46,"snapshot_sha256":"2250a44206f99c617f284d5c29dfdbdc0d9e993cc6f26423e2b077c43a66a70d","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"e7fa3b12be9fd28a0f5182662fca6204378d54169efe623181a0f77c08c4aaee"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}