{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"indexed":{"date-parts":[[2026,8,24]],"date-time":"2026-08-24T19:24:52Z","timestamp":1787599492040,"version":"build-2736575974"},"reference-count":39,"publisher":"National Academy of Sciences","issue":"35","license":[{"start":{"date-parts":[[2026,8,24]],"date-time":"2026-08-24T00:00:00Z","timestamp":1787529600000},"content-version":"vor","delay-in-days":0,"URL":"https:\/\/creativecommons.org\/licenses\/by\/4.0\/"}],"funder":[{"DOI":"10.13039\/501100000266","name":"UKRI | Engineering and Physical Sciences Research Council","doi-asserted-by":"publisher","award":["EP\/X01259X\/1"],"award-info":[{"award-number":["EP\/X01259X\/1"]}],"id":[{"id":"10.13039\/501100000266","id-type":"DOI","asserted-by":"publisher"}]}],"content-domain":{"domain":["www.pnas.org"],"crossmark-restriction":true},"short-container-title":["Proc. Natl. Acad. Sci. U.S.A."],"published-print":{"date-parts":[[2026,9,1]]},"abstract":"<jats:p>Coarse-grained models of chaotic systems neglect unresolved degrees of freedom, inducing structured model error that limits predictability and distorts long-term statistics. Typical data-driven closures are trained to minimize prediction error over a single time step, implicitly assuming Markovian dynamics and often failing to capture long-term behavior. Recent approaches instead optimize losses over finite trajectories. However, when such trajectory-based training is carried out with deterministic pointwise losses, it introduces a fundamental mathematical degeneracy. We prove that optimizing pointwise deterministic losses, including but not limited to mean squared error, over chaotic trajectories suppresses predictive variance, with a corresponding loss of physical variability in long integrations. In contrast, strictly proper scoring rules avoid this degeneracy. By targeting forecast distributions rather than realized trajectories, they remove the penalty against predictive spread and align the long-lead optimum with the invariant measure. Using quasi-geostrophic turbulence as a canonical chaotic system, we validate this theory: Closures trained with one-step losses fail to capture stable coarse-grained dynamics, while deterministic closures optimized over trajectories exhibit the variance-loss tendency predicted by our analysis. Stochastic closures calibrated over trajectories using the energy score, however, overcome both structural limitations, yielding skillful ensemble forecasts and realistic long-term statistics. Our results establish that both stochastic modeling and trajectory-based calibration are essential for faithfully representing the dynamics of coarse-grained systems.<\/jats:p>","DOI":"10.1073\/pnas.2609143123","type":"journal-article","created":{"date-parts":[[2026,8,24]],"date-time":"2026-08-24T19:02:04Z","timestamp":1787598124000},"update-policy":"https:\/\/doi.org\/10.1073\/pnas.cm10313","source":"Crossref","is-referenced-by-count":0,"title":["Stochasticity and probabilistic trajectory scoring are essential for data-driven closures of chaotic systems"],"prefix":"10.1073","volume":"123","author":[{"ORCID":"https:\/\/orcid.org\/0000-0002-8508-2847","authenticated-orcid":false,"given":"Martin T.","family":"Brolly","sequence":"first","affiliation":[{"id":[{"id":"https:\/\/ror.org\/01nrxwf90","id-type":"ROR","asserted-by":"publisher"}],"name":"School of Mathematics, Maxwell Institute for Mathematical Sciences, University of Edinburgh","place":["Edinburgh, United Kingdom"]}],"role":[{"vocabulary":"crossref","role":"author"}]}],"member":"341","published-online":{"date-parts":[[2026,8,24]]},"reference":[{"key":"e_1_3_4_1_2","doi-asserted-by":"publisher","DOI":"10.1017\/CBO9780511840531"},{"key":"e_1_3_4_2_2","doi-asserted-by":"publisher","DOI":"10.1256\/qj.04.03"},{"key":"e_1_3_4_3_2","doi-asserted-by":"publisher","DOI":"10.1002\/cpa.1014"},{"key":"e_1_3_4_4_2","doi-asserted-by":"publisher","DOI":"10.1088\/0951-7715\/19\/4\/001"},{"key":"e_1_3_4_5_2","doi-asserted-by":"crossref","unstructured":"H. Christensen L. Zanna \u201cParametrization in weather and climate models\u201d in Climate Science Oxford Research Encyclopedias H. V. Storch Ed. (Oxford University Press 2022). vol. 12.","DOI":"10.1093\/acrefore\/9780190228620.013.826"},{"key":"e_1_3_4_6_2","doi-asserted-by":"publisher","DOI":"10.1175\/BAMS-D-15-00268.1"},{"key":"e_1_3_4_7_2","doi-asserted-by":"publisher","DOI":"10.1017\/S0022112059000362"},{"key":"e_1_3_4_8_2","doi-asserted-by":"publisher","DOI":"10.1140\/epjp\/s13360-022-03360-9"},{"key":"e_1_3_4_9_2","doi-asserted-by":"publisher","DOI":"10.1175\/JAS-D-23-0146.1"},{"key":"e_1_3_4_10_2","doi-asserted-by":"publisher","DOI":"10.1073\/pnas.97.7.2968"},{"key":"e_1_3_4_11_2","doi-asserted-by":"publisher","DOI":"10.1093\/oso\/9780195140187.001.0001"},{"key":"e_1_3_4_12_2","doi-asserted-by":"publisher","DOI":"10.1088\/1742-5468\/2012\/03\/P03003"},{"key":"e_1_3_4_13_2","doi-asserted-by":"publisher","DOI":"10.1007\/s10955-013-0726-8"},{"key":"e_1_3_4_14_2","doi-asserted-by":"publisher","DOI":"10.1126\/science.adi2336"},{"key":"e_1_3_4_15_2","unstructured":"J. Pathak et al. A global data-driven high-resolution weather model using adaptive Fourier neural operators. arXiv [Preprint] (2022). https:\/\/arxiv.org\/abs\/2202.11214 (Accessed 30 March 2026)."},{"key":"e_1_3_4_16_2","unstructured":"I. Price et al. Diffusion-based ensemble forecasting for medium-range weather. arXiv [Preprint] (2023). https:\/\/arxiv.org\/abs\/2312.15796 (Accessed 30 March 2026)."},{"key":"e_1_3_4_17_2","doi-asserted-by":"publisher","DOI":"10.1038\/s41586-024-07744-y"},{"key":"e_1_3_4_18_2","doi-asserted-by":"publisher","DOI":"10.1073\/pnas.1810286115"},{"key":"e_1_3_4_19_2","doi-asserted-by":"publisher","DOI":"10.1029\/2019MS001711"},{"key":"e_1_3_4_20_2","doi-asserted-by":"publisher","DOI":"10.1038\/s41467-020-17142-3"},{"key":"e_1_3_4_21_2","doi-asserted-by":"publisher","DOI":"10.1029\/2019MS001896"},{"key":"e_1_3_4_22_2","doi-asserted-by":"publisher","DOI":"10.1029\/2020GL088376"},{"key":"e_1_3_4_23_2","doi-asserted-by":"publisher","DOI":"10.1029\/2021MS002534"},{"key":"e_1_3_4_24_2","doi-asserted-by":"publisher","DOI":"10.1029\/2023MS003681"},{"key":"e_1_3_4_25_2","doi-asserted-by":"publisher","DOI":"10.1029\/2025MS005223"},{"key":"e_1_3_4_26_2","doi-asserted-by":"crossref","unstructured":"S Rasp Coupled online learning as a way to tackle instabilities and biases in neural network parameterizations: General algorithms and Lorenz 96 case study (v1. 0). Geosci. Model. Dev. 13 2185\u20132196 (2020).","DOI":"10.5194\/gmd-13-2185-2020"},{"key":"e_1_3_4_27_2","doi-asserted-by":"publisher","DOI":"10.1073\/pnas.2101784118"},{"key":"e_1_3_4_28_2","doi-asserted-by":"publisher","DOI":"10.1029\/2022MS003124"},{"key":"e_1_3_4_29_2","doi-asserted-by":"publisher","DOI":"10.1029\/2024MS004883"},{"key":"e_1_3_4_30_2","doi-asserted-by":"publisher","DOI":"10.1198\/016214506000001437"},{"key":"e_1_3_4_31_2","doi-asserted-by":"publisher","DOI":"10.1111\/j.2517-6161.1952.tb00104.x"},{"key":"e_1_3_4_32_2","doi-asserted-by":"publisher","DOI":"10.1175\/WAF966.1"},{"key":"e_1_3_4_33_2","doi-asserted-by":"publisher","DOI":"10.1175\/WAF-D-19-0205.1"},{"key":"e_1_3_4_34_2","doi-asserted-by":"publisher","DOI":"10.1002\/qj.2270"},{"key":"e_1_3_4_35_2","doi-asserted-by":"publisher","DOI":"10.1038\/s41586-023-06185-3"},{"key":"e_1_3_4_36_2","unstructured":"M. Chantry et al. \u201cAIFS \u2013 ECMWF\u2019s data-driven forecasting system\u201d in 105th Annual AMS Meeting 2025 (American Meteorological Society 2025) vol. 105 p. 449087."},{"key":"e_1_3_4_37_2","doi-asserted-by":"publisher","DOI":"10.1002\/qj.4647"},{"key":"e_1_3_4_38_2","unstructured":"M. T. Brolly mtbrolly\/qg_closure: Version at article submission. Zenodo. https:\/\/doi.org\/10.5281\/zenodo.19335604. Deposited 30 March 2026."},{"key":"e_1_3_4_39_2","unstructured":"M. T. Brolly Trained neural networks for the paper \u201cStochasticity and probabilistic trajectory scoring are essential for data-driven closures of chaotic systems\u201d. Zenodo. https:\/\/doi.org\/10.5281\/zenodo.19335706. Deposited 30 March 2026."}],"container-title":["Proceedings of the National Academy of Sciences"],"original-title":[],"language":"en","link":[{"URL":"https:\/\/pnas.org\/doi\/pdf\/10.1073\/pnas.2609143123","content-type":"unspecified","content-version":"vor","intended-application":"similarity-checking"}],"deposited":{"date-parts":[[2026,8,24]],"date-time":"2026-08-24T19:03:04Z","timestamp":1787598184000},"score":1,"resource":{"primary":{"URL":"https:\/\/pnas.org\/doi\/10.1073\/pnas.2609143123"}},"subtitle":[],"short-title":[],"issued":{"date-parts":[[2026,8,24]]},"references-count":39,"journal-issue":{"issue":"35","published-print":{"date-parts":[[2026,9,1]]}},"alternative-id":["10.1073\/pnas.2609143123"],"URL":"https:\/\/doi.org\/10.1073\/pnas.2609143123","relation":{},"ISSN":["0027-8424","1091-6490"],"issn-type":[{"value":"0027-8424","type":"print"},{"value":"1091-6490","type":"electronic"}],"subject":[],"published":{"date-parts":[[2026,8,24]]},"assertion":[{"value":"2026-03-30","order":0,"name":"received","label":"Received","group":{"name":"publication_history","label":"Publication History"}},{"value":"2026-07-14","order":2,"name":"accepted","label":"Accepted","group":{"name":"publication_history","label":"Publication History"}},{"value":"2026-08-24","order":3,"name":"published","label":"Published","group":{"name":"publication_history","label":"Publication History"}}],"article-number":"e2609143123"}}