arXiv physics.ao-ph · Dec 28, 2025
Synthetic climate data for long-range AI weather forecasts
Markdown · JSON · Original source
Scott A. Martin, Noah Brenowitz, Dale Durran and Michael Pritchard describe “long-range distillation,” a method that trains models to forecast directly at long horizons using synthetic climate data generated by an autoregressive model. They report generating more than 10,000 simulated years. In perfect-model experiments, the student models outperformed climatology and approached the teacher model’s skill; in real-world tests, two approaches with synthetic pre-training achieved subseasonal-to-seasonal skill competitive with the ECMWF IFS ensemble. The abstract also reports improvements beyond ERA5-only training. These are the authors’ reported results, not an independent assessment.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
arXiv physics.ao-ph · Article
Original source date: Dec 28, 2025 (2025-12-28).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-10-16T00:40:25.968Z. First observed: 2026-10-15T04:07:06.285Z. Retrieved: 2026-10-15T23:43:42.123Z.
Limits
The supplied evidence is the paper’s abstract and does not include detailed performance metrics or an independent assessment. The paper distinguishes perfect-model experiments from real-world forecasting results; its reported outcomes should not be treated as a general guarantee of forecast performance.
Open question
How much can synthetic climate data improve long-range forecasts beyond what observational records alone provide?
Source-stated
The paper proposes training long-horizon forecasting models on synthetic data generated by an autoregressive teacher model.
a long-timestep probabilistic ``student'' model is trained to forecast directly at long-range using a huge synthetic dataset generated by an autoregressive ``teacher'' model
Source-stated
The authors say they generated more than 10,000 simulated years of climate data to train models for multiple timescales.
we generate over 10,000 years of simulated climate to train student models across a range of timescales.
Source-stated
In perfect-model experiments, the authors report that distilled models outperformed climatology and approached the teacher’s skill while replacing many autoregressive steps with one step.
In perfect-model experiments, the distilled models outperform climatology and approach their teacher's skill while replacing hundreds of autoregressive steps with a single timestep.
Source-stated
The authors report that two student-model approaches using synthetic pre-training achieved subseasonal-to-seasonal forecast skill competitive with the ECMWF IFS ensemble.
both a diffusion model student fine-tuned on ERA5 and a linear inverse model student with mixed DLESyM-ERA5 training achieve S2S forecast skill competitive with the ECMWF IFS ensemble
Source-stated
The abstract reports that synthetic pre-training improved results beyond ERA5-only training.
synthetic pre-training providing improvements beyond ERA5-only training.
Source and contributor context
arXiv physics.ao-ph · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Pre-training
A model's initial training on a large dataset, before further adaptation for particular uses.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 2. Publication ID: 3da7cfc3-4930-4300-8dac-bc4420224ed9.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-10-16T00:40:25.968Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 77bc7a3b096239103010ed3c71796cd597ab8f55c0ccc7c736f964e41d900492. Source text hash: a6188275dd5429f97d3ca174129e006bd06e9d64d7b85a9f230736ad416233b0.
Correction reason: Corrected the temporal framing: the feed timestamp is 15 October 2026, while the publisher record identifies 28 December 2025 as the first submission and 14 October 2026 as the v2 revision. The substantive reading and its attribution are retained.
Revision and source-check history
Version 2 · 2026-10-16T00:40:25.968Z · Corrected the temporal framing: the feed timestamp is 15 October 2026, while the publisher record identifies 28 December 2025 as the first submission and 14 October 2026 as the v2 revision. The substantive reading and its attribution are retained.
Synthetic climate data for long-range AI weather forecasts
Scott A. Martin, Noah Brenowitz, Dale Durran and Michael Pritchard describe “long-range distillation,” a method that trains models to forecast directly at long horizons using synthetic climate data generated by an autoregressive model. They report generating more than 10,000 simulated years. In perfect-model experiments, the student models outperformed climatology and approached the teacher model’s skill; in real-world tests, two approaches with synthetic pre-training achieved subseasonal-to-seasonal skill competitive with the ECMWF IFS ensemble. The abstract also reports improvements beyond ERA5-only training. These are the authors’ reported results, not an independent assessment.
Original source
arXiv physics.ao-ph · Article
Original source date: Dec 28, 2025 (2025-12-28).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-10-16T00:40:25.968Z. First observed: 2026-10-15T04:07:06.285Z. Retrieved: 2026-10-15T23:43:42.123Z.
Limits
The supplied evidence is the paper’s abstract and does not include detailed performance metrics or an independent assessment. The paper distinguishes perfect-model experiments from real-world forecasting results; its reported outcomes should not be treated as a general guarantee of forecast performance.
Open question
How much can synthetic climate data improve long-range forecasts beyond what observational records alone provide?
Source-stated
The paper proposes training long-horizon forecasting models on synthetic data generated by an autoregressive teacher model.
a long-timestep probabilistic ``student'' model is trained to forecast directly at long-range using a huge synthetic dataset generated by an autoregressive ``teacher'' model
Source-stated
The authors say they generated more than 10,000 simulated years of climate data to train models for multiple timescales.
we generate over 10,000 years of simulated climate to train student models across a range of timescales.
Source-stated
In perfect-model experiments, the authors report that distilled models outperformed climatology and approached the teacher’s skill while replacing many autoregressive steps with one step.
In perfect-model experiments, the distilled models outperform climatology and approach their teacher's skill while replacing hundreds of autoregressive steps with a single timestep.
Source-stated
The authors report that two student-model approaches using synthetic pre-training achieved subseasonal-to-seasonal forecast skill competitive with the ECMWF IFS ensemble.
both a diffusion model student fine-tuned on ERA5 and a linear inverse model student with mixed DLESyM-ERA5 training achieve S2S forecast skill competitive with the ECMWF IFS ensemble
Source-stated
The abstract reports that synthetic pre-training improved results beyond ERA5-only training.
synthetic pre-training providing improvements beyond ERA5-only training.
Source and contributor context
arXiv physics.ao-ph · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Pre-training
A model's initial training on a large dataset, before further adaptation for particular uses.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 2. Publication ID: 3da7cfc3-4930-4300-8dac-bc4420224ed9.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-10-16T00:40:25.968Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 77bc7a3b096239103010ed3c71796cd597ab8f55c0ccc7c736f964e41d900492. Source text hash: a6188275dd5429f97d3ca174129e006bd06e9d64d7b85a9f230736ad416233b0.
Correction reason: Corrected the temporal framing: the feed timestamp is 15 October 2026, while the publisher record identifies 28 December 2025 as the first submission and 14 October 2026 as the v2 revision. The substantive reading and its attribution are retained.
Version 1 · 2026-10-15T23:30:52.368Z
A weather-model study trains long-range forecasts on 10,000 simulated years
Authors Scott A. Martin, Noah Brenowitz, Dale Durran and Michael Pritchard describe “long-range distillation”: training models to forecast directly at longer horizons using synthetic climate data generated by an autoregressive model. They report that their student models approached the teacher’s skill in perfect-model tests, and that two approaches using synthetic pre-training achieved subseasonal-to-seasonal forecast skill competitive with the ECMWF ensemble. The results are the authors’ reported findings; the abstract does not provide detailed metrics or independent validation.
Original source
arXiv physics.ao-ph · Article
Original source date: Oct 15, 2026 (2026-10-15T04:00:00.000Z).
Date basis: Date supplied by the publisher’s feed.
Added to Escape Meridian: 2026-10-15T23:30:52.368Z. First observed: 2026-10-15T04:07:06.285Z. Retrieved: 2026-10-15T23:30:19.577Z.
Limits
These are findings reported in the paper abstract; the supplied evidence does not include detailed performance metrics or independent assessment. The paper distinguishes perfect-model experiments from real-world forecasting results; the reported outcomes should not be treated as a general guarantee of forecast performance.
Open question
How much can synthetic climate data improve long-range forecasts beyond what observational records alone provide?
Source-stated
The paper proposes training long-horizon forecasting models on synthetic data generated by an autoregressive teacher model.
a long-timestep probabilistic ``student'' model is trained to forecast directly at long-range using a huge synthetic dataset generated by an autoregressive ``teacher'' model
Source-stated
The authors say they generated more than 10,000 simulated years of climate data to train models for multiple timescales.
we generate over 10,000 years of simulated climate to train student models across a range of timescales.
Source-stated
In perfect-model experiments, the authors report that distilled models outperformed climatology and came close to the teacher’s skill, replacing many autoregressive forecast steps with one step.
In perfect-model experiments, the distilled models outperform climatology and approach their teacher's skill while replacing hundreds of autoregressive steps with a single timestep.
Source-stated
The authors report that two student-model approaches using synthetic pre-training achieved subseasonal-to-seasonal forecast skill competitive with the ECMWF IFS ensemble.
both a diffusion model student fine-tuned on ERA5 and a linear inverse model student with mixed DLESyM-ERA5 training achieve S2S forecast skill competitive with the ECMWF IFS ensemble
Source-stated
The abstract says synthetic pre-training improved on ERA5-only training and characterizes the work as a first demonstration of real-world skill gains from large volumes of AI-generated training data.
synthetic pre-training providing improvements beyond ERA5-only training. This demonstrates the benefit of synthetic pre-training for both deep learning and classical data-driven methods
Source and contributor context
arXiv physics.ao-ph · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Pre-training
A model's initial training on a large dataset, before further adaptation for particular uses.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: ceae6046-4895-4915-8de9-a90ed512ac83.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-10-15T23:30:52.368Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 77bc7a3b096239103010ed3c71796cd597ab8f55c0ccc7c736f964e41d900492. Source text hash: a6188275dd5429f97d3ca174129e006bd06e9d64d7b85a9f230736ad416233b0.
passed · 2026-10-15T23:30:52.368Z · anthropic/claude-opus-5.5
Publication: ceae6046-4895-4915-8de9-a90ed512ac83
attribution — pass: The summary names the four listed authors correctly and presents the results as the authors' reported findings from an arXiv paper. The paper is cross-listed in physics.ao-ph, which matches the feed publisher label.
Authors:Scott A. Martin, Noah Brenowitz, Dale Durran, Michael Pritchard
summary — pass: The title, summary and question accurately describe the long-range distillation method, the 10,000-plus simulated years, the perfect-model results and the real-world S2S results versus the ECMWF IFS ensemble. The question fits the paper's stated motivation and its claim about data volumes beyond the observational record. The caveats are appropriate.
represents the first demonstration that real-world skill can be improved by training on AI-generated synthetic data in volumes far exceeding the observational record
date — pass: The reading asserts no conflicting date. The feed timestamp of 2026-10-15 is consistent with the v2 revision posted on 14 Oct 2026. v1 was submitted in December 2025.
[Submitted on 28 Dec 2025 (v1), last revised 14 Oct 2026 (this version, v2)]
support — pass: Each claim is backed by a verbatim excerpt from the abstract and attributed to the authors. Claim 5's 'first demonstration' characterization appears in the abstract's final sentence. The reading does not present the claims as independently verified.
synthetic pre-training providing improvements beyond ERA5-only training
media — pass: The selected media is the verified link to the original arXiv abstract page, which is the direct source.
passed · 2026-10-16T00:40:25.968Z · anthropic/claude-opus-5.5
Publication: 3da7cfc3-4930-4300-8dac-bc4420224ed9
attribution — pass: The draft attributes the work to Scott A. Martin, Noah Brenowitz, Dale Durran and Michael Pritchard, matching the listed authors. It frames the results as the authors' reported findings rather than independent verification.
Authors:Scott A. Martin, Noah Brenowitz, Dale Durran, Michael Pritchard
summary — pass: The title, question and summary faithfully reflect the abstract. It covers synthetic teacher-generated data for long-timestep student models, more than 10,000 simulated years, perfect-model results, and real-world S2S skill competitive with the ECMWF IFS ensemble. It also notes the gains beyond ERA5-only training. The draft does not overstate these claims.
We introduce long-range distillation, whereby a long-timestep probabilistic ``student'' model is trained to forecast directly at long-range
date — pass: originalPublished is 2025-12-28, which matches the v1 submission date. It correctly ignores the v2 revision on 14 Oct 2026 and the October 2026 feed timestamp.
[Submitted on 28 Dec 2025 (v1), last revised 14 Oct 2026 (this version, v2)]
support — pass: Each claim's excerpt is an exact substring of the abstract and supports the paraphrased claim. The limitations are appropriately cautious.
In perfect-model experiments, the distilled models outperform climatology and approach their teacher's skill while replacing hundreds of autoregressive steps with a single timestep.
media — pass: The selected media is the verified arXiv abstract page for this paper. It is the direct original source and the only supplied candidate.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.