{"resourceId":"nvidia-stormscope-nowcasting-paper-limits-20260124","versions":[{"version":"external-2e25ae3c3ae5fad97d0af82934cabc4c9b4abb89134d561506e6eaa463c43947","resource":{"id":"nvidia-stormscope-nowcasting-paper-limits-20260124","title":"StormScope paper by NVIDIA authors: generative nowcasting beats HRRR up to 4 hours but authors do not claim state of the art for 0 to 2 hour radar-only nowcasting","organization":"NVIDIA (paper authors)","sector":"Weather and climate AI research","geography":"Continental United States (evaluation domain)","publishedAt":"2026-01-24 (arXiv v1)","publicationDate":"2026-01-24","eventDate":"2026-01-24","sourceName":"arXiv","sourceLabel":"Preprint","sourceUrl":"https://arxiv.org/html/2601.17268v1","evidenceClass":"academic-research","outcomeClass":"mixed","topics":["infrastructure"],"finding":"The StormScope preprint, by NVIDIA authors, reports that the generative nowcasting model outperforms the HRRR operational model for forecasts up to about 4 hours and is comparable around 6 hours, at 6 km resolution. The authors state they do not claim state-of-the-art performance on the 0 to 2 hour timescale for radar-only nowcasting. The paper does not evaluate hail. It is included now because CSU's Oct 26 hail work builds on this model.","sledRelevance":"Gives agencies and universities the base model's documented limits when assessing CSU's hail adaptation.","evidence":"Preprint with quantitative comparison against HRRR; not yet peer reviewed as accessed; authored by the vendor's researchers.","architectureImplications":"Generative model at 6 km; CSU's reported 1.3 km models differ markedly from the evaluated configuration.","governanceImplications":"Agencies should distinguish base-model skill from adapted-model skill when reviewing claims.","securityPrivacyImplications":"None specific.","caveats":"Vendor-authored preprint; no hail evaluation; resolution differs from CSU target.","streamIds":["nvidia"],"roles":{"sales":"Interpretation: this paper is the evidence base behind CSU's hail work, and it is candid about limits. Account teams discussing weather AI with state agencies should cite both the reported improvement over HRRR up to about 4 hours and the authors' statement that they do not claim state-of-the-art results for 0 to 2 hour radar-only nowcasting. That balance matters because hail warnings often depend on the very short horizon.","engineering":"Interpretation: the evaluated configuration is 6 km over the continental United States with comparison against HRRR, and hail is not assessed. CSU's stated resolution of about 1.3 km and a hail-specific output are therefore new work. Engineers reviewing any proposed use should ask for verification at the target resolution and for hail specifically, and should treat the base paper's skill numbers as not transferable.","delivery":"Interpretation: for delivery teams this is background, not a deliverable. If an agency later pilots a StormScope-derived product, delivery plans should require verification at the resolution and horizon actually used, with a proposed acceptance criterion that the product's skill on 0 to 2 hour forecasts is compared with the agency's current radar nowcasting before it informs any public warnings or emergency decisions."},"roleDetails":{"sales":{"customerProblem":"Agencies need to know what the base model behind new hail work can and cannot do.","stakeholders":["State weather and emergency agencies","University researchers"],"discoveryQuestions":["Which forecast horizon matters most for your hail decisions?"],"valueHypothesis":"Interpretation: the model may add value at 2 to 6 hours; short-horizon value is not claimed.","potentialEngagement":"Briefing on documented limits during weather AI discussions.","unsupportedClaims":["Hail skill","Skill at 1.3 km"]},"engineering":{"fit":"Research use for multi-hour convective nowcasting.","architectureIntegration":"Generative nowcasting model compared with HRRR at 6 km.","prerequisites":["GPU compute","Radar and satellite inputs"],"constraints":["No state-of-the-art claim for 0 to 2 hours radar-only","No hail evaluation","6 km resolution"],"security":"None specific.","proofOfValue":["Verify at target resolution","Evaluate hail-specific output"]},"delivery":{"work":"Use as background for evaluating derived products.","dependencies":["Derived product verification"],"ownership":"Agency forecasting lead.","skillsAdoption":"Forecasters need to understand horizon-dependent skill.","governanceCheckpoints":["Require verification at used resolution and horizon"],"acceptanceCriteria":["Proposed: 0 to 2 hour skill compared with current radar nowcasting before operational use"],"risks":["Transferring base-model skill claims to adapted products"]}},"retrievedAt":"2026-10-27T02:56:53Z","enrichedAt":"2026-10-27T03:07:35Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"None specific.","procurementImplications":"No procurement implication on its own.","operatingModelImplications":"Supports using AI nowcasts as one input alongside operational models, not a replacement for 0 to 2 hour radar nowcasting.","sourceVerification":{"openedUrl":"https://arxiv.org/html/2601.17268v1","referenceExcerpt":"As such, we do not claim our proposed method is state-of-the-art on the 0-2 hour timescale for radar-only nowcasting.","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}