[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7ASSPSA2Z9PJAS45Q1A2BWV":3,"skill-same:01M7ASSPSA2Z9PJAS45Q1A2BWV":48},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":24,"commitSha":25,"skillDirectory":26,"forksCount":27,"repositoryHtmlUrl":28,"repositoryHomepageUrl":29,"ownerHtmlUrl":30,"ownerBlogUrl":31,"ownerDescription":32,"updatedAt":33,"install":34,"resources":47},"01M7ASSPSA2Z9PJAS45Q1A2BWV","weather-observation-nwp-verification-and-machine-learning",null,"Use when working on forecasting systems: the observation network, numerical weather prediction including the governing system, data assimilation as the underrated half, and ensembles, forecast verification and the scoring rules, predictability and the intrinsic chaos limit, machine-learning weather prediction with its limits, the live scientific dispute and the metrics problem underneath it, and the weather-versus-climate distinction.","the-vibey-project\u002Fvibey","the-vibey-project","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F329298758?v=4","01M7AJHGZ9722CTJ4JNQZVRH0Q","the vibey project","01M7AJHGZVMG2K7TGXQ6XFHVG3",5,"MIT","https:\u002F\u002Fgithub.com\u002Fthe-vibey-project\u002Fvibey\u002Fblob\u002Fdevelop\u002Fsrc\u002Fvibey_tools\u002Fskills\u002Fplugins\u002Fweather-science\u002Fskills\u002Fweather-observation-nwp-verification-and-machine-learning\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: weather-observation-nwp-verification-and-machine-learning\ndescription: \"Use when working on forecasting systems: the observation network, numerical weather prediction including the governing system, data assimilation as the underrated half, and ensembles, forecast verification and the scoring rules, predictability and the intrinsic chaos limit, machine-learning weather prediction with its limits, the live scientific dispute and the metrics problem underneath it, and the weather-versus-climate distinction.\"\n---\n\n# Weather Science: Observation, Numerical Weather Prediction, Verification, Predictability, and Machine Learning\n\n> **Part 4 of 5** of the *Weather Science* reference (plugin `weather-science`), covering §11–§16. Sibling skills: `weather-atmosphere-radiation-thermodynamics-and-moisture` (§0–§4), `weather-dynamics-circulation-and-synoptic` (§5–§7), `weather-severe-storms-cyclones-and-boundary-layer` (§8–§10), `weather-reference` (§17–§21). Section numbers are shared across the set; a reference written as §N → `skill` points into that sibling skill.\n>\n> **Currency:** Atmospheric physics is settled — hydrostatic balance, geostrophy, Lorenz's 1963 chaos work. One area moved dramatically. See §15 below for machine-learning weather prediction and the live scientific dispute around it.\n\n> **Scope.** Complements a Newtonian-mechanics reference (§10 there covers Coriolis and\n> rotating frames properly) and a fundamental-physics reference (radiation, thermodynamics).\n> ⚠️ **This is the atmosphere specifically.**\n>\n> **⚠️ GOTCHA** boxes mark genuine misconceptions and the places where intuition\n> systematically misleads.\n>\n> **The three ideas that organize the whole field:**\n> 1. **⚠️ Weather is a heat engine driven by differential solar heating.** The tropics\n>    receive more energy than they radiate, the poles the reverse, and **essentially all\n>    weather is the atmosphere and ocean moving that surplus poleward** (§2.3 → `weather-atmosphere-radiation-thermodynamics-and-moisture`).\n> 2. **⚠️ Rotation changes everything.** On a non-rotating planet air would flow directly\n>    from high to low pressure. **The Coriolis effect makes it flow *along* isobars\n>    instead**, and that single fact produces jet streams, cyclones, and the entire\n>    structure of mid-latitude weather (§5 → `weather-dynamics-circulation-and-synoptic`).\n> 3. **⚠️ The atmosphere is chaotic, and this is a mathematical property, not a\n>    measurement problem.** Lorenz (1963) showed the predictability limit is intrinsic.\n>    **No observing system and no model — physics-based or learned — removes it** (§14.1).\n\n---\n\n## §11. Observation\n\n| System | Provides | ⚠️ Notes |\n|---|---|---|\n| **Surface stations \u002F METAR** | T, Td, P, wind, visibility | Sparse and land-biased |\n| **Radiosonde** | ⚠️ **The vertical profile** | ~2×\u002Fday, ~800 sites; **the backbone of upper-air truth** |\n| **Weather radar** | Precipitation, ⚠️ **Doppler velocity** | **Dual-pol** distinguishes hydrometeor type |\n| **Satellite — geostationary** | Continuous imagery | ⚠️ **Poor at high latitudes** |\n| **Satellite — polar orbiting** | Sounding, global coverage | ⚠️ **The dominant data volume in assimilation** |\n| **Aircraft (AMDAR)** | Upper-air en route | ⚠️ **Dropped sharply during COVID and measurably degraded forecasts** |\n| **GNSS radio occultation** | ⚠️ **Bias-free temperature\u002Fhumidity profiles** | High-value, growing |\n| **Buoys, ships, profilers, lightning networks** | | |\n\n**⚠️ Radar caveats worth knowing**: the beam rises with distance so distant echoes sample\nhigher altitudes; **ground clutter, anomalous propagation, bright band at the melting\nlevel, and beam blockage** all produce artifacts. ⚠️ **Doppler measures only the radial\ncomponent of velocity** — motion across the beam is invisible.\n\n---\n\n## §12. Numerical Weather Prediction\n\n### 12.1 The governing system\n**The primitive equations**: momentum (Navier-Stokes on a rotating sphere), continuity,\nthermodynamic energy, the ideal gas law, and moisture conservation.\n⚠️ **A coupled nonlinear PDE system with no analytic solution — so you discretize.**\n\n**Discretization**: spectral (⚠️ **spherical harmonics — historically dominant for global\nmodels**) or grid-point (finite difference\u002Fvolume, icosahedral and cubed-sphere grids).\n**Vertical coordinates**: sigma, pressure, hybrid, isentropic.\n**⚠️ CFL condition constrains the timestep given grid spacing and wave speed** — **which is\nwhy resolution is expensive: halving grid spacing roughly costs 8–16×.**\n**Hydrostatic** for coarse global models; ⚠️ **non-hydrostatic required below ~10 km grid\nspacing, where convection begins to be resolved** (§1 → `weather-atmosphere-radiation-thermodynamics-and-moisture`).\n\n**⚠️ Parameterization is where the physics that can't be resolved lives**: convection,\ncloud microphysics, radiation, boundary layer turbulence, gravity wave drag, land surface.\n⚠️ **This is the largest source of model error and the hardest part of NWP.** **The\n\"grey zone\" — grid spacings around 1–10 km where convection is partly resolved and partly\nparameterized — is genuinely awkward.**\n\n### 12.2 ⚠️ Data assimilation — the underrated half\n**The forecast is only as good as its initial conditions, and observations are sparse,\nirregular and noisy.** **DA combines a short-range forecast (the \"background\") with\nobservations, weighted by their respective error statistics**, to produce an **analysis**.\n\n**Methods**: 3D-Var, **4D-Var** (⚠️ **assimilates over a time window, using the model\nitself as a constraint — ECMWF's long-standing strength**), **EnKF** (flow-dependent\nerror covariances from an ensemble), and **hybrid** approaches which now dominate.\n\n> **⚠️ GOTCHA — data assimilation is arguably a larger contributor to modern forecast\n> skill than model improvements, and it's almost invisible outside the field.**\n> ⚠️ **It's also the reason §15's ML models are not yet a full replacement: they mostly\n> consume an analysis produced by conventional physics-based assimilation.**\n\n**Reanalysis** — ⚠️ **rerunning a fixed modern DA system over the historical record to\nproduce a physically consistent gridded dataset.** **ERA5 is the standard**, and it is\n**the training data for essentially every ML weather model** (§15).\n\n### 12.3 Ensembles\n**⚠️ Since initial conditions and the model are both uncertain, run many forecasts.**\nPerturb initial conditions (singular vectors, bred modes, EDA) and represent model\nuncertainty (stochastic physics, multi-model, multi-physics).\n**⚠️ Ensemble prediction arrived operationally in 1992 and was the previous methodological\nbreak in forecasting** — before §15. **The ensemble mean is more skillful than any member;\nthe spread is the uncertainty estimate.**\n⚠️ **A well-calibrated ensemble's spread should match its error. Under-dispersion is the\ncommon failure and it makes forecasts overconfident.**\n\n---\n\n## §13. Forecast Verification\n\n**⚠️ \"Was the forecast good?\" is a harder question than it looks, and the metric you choose\ndetermines the answer** (§15.3).\n```\nRMSE \u002F MAE                deterministic error  ⚠️ REWARDS SMOOTHING — see §15.3\nACC (anomaly correlation) ⚠️ the standard synoptic skill score; 0.6 conventionally\n                          taken as the limit of useful deterministic skill\nBrier score               probabilistic, binary events\nCRPS                      ⚠️ probabilistic, continuous — the standard ensemble metric\nReliability diagram       ⚠️ do 30% forecasts verify 30% of the time?\nPOD \u002F FAR \u002F CSI           categorical events\n```\n**⚠️ The double penalty problem**: a sharp forecast of a feature in slightly the wrong place\nis penalized twice — **once for predicting it where it wasn't, once for missing it where it\nwas.** ⚠️ **A blurry forecast scores better on RMSE while being less useful.** **This is\nnot a technicality; it distorts model development toward smoothness** (§15.3).\n\n**⚠️ Skill must beat a reference** — persistence, climatology, or a previous model.\n**\"85% accurate\" means nothing without one.**\n\n---\n\n## §14. Predictability and Chaos\n\n### 14.1 The intrinsic limit\n**Lorenz (1963)** — ⚠️ **deterministic nonlinear systems exhibit sensitive dependence on\ninitial conditions.** Errors grow, and **small-scale errors propagate upscale.**\n> **⚠️ GOTCHA — the predictability limit is a property of the atmosphere, not of our\n> instruments or models.** ⚠️ **Roughly two weeks for synoptic-scale deterministic\n> forecasting is a mathematical bound, not an engineering target.** **Perfect observations\n> and a perfect model would not remove it** — see a Newtonian-mechanics reference §13 for\n> why better data buys only logarithmic improvement.\n>\n> **⚠️ And predictability is flow-dependent.** Some regimes are far more predictable than\n> others, which is exactly what the ensemble spread is telling you (§12.3).\n\n**Scale matters**: convective cells minutes to hours; mesoscale systems hours; synoptic\nsystems days; **planetary waves and teleconnections longer.** ⚠️ **Seasonal forecasting\nworks not by predicting weather but by predicting boundary-condition-driven shifts in the\ndistribution** — chiefly ENSO (§6 → `weather-dynamics-circulation-and-synoptic`).\n\n---\n\n## §15. Machine Learning Weather Prediction\n\n**⚠️ This is the section that moved, and it moved fast enough that most general knowledge\nabout it is out of date.** **Verified August 2026.**\n\n### 15.1 What happened\n**⚠️ One assessment frames it as the biggest methodological break in forecasting since\nensemble prediction arrived in 1992** — and that seems right. **Between November 2023 and\nmid-2026, data-driven models trained on reanalysis went from research curiosities to\noperational systems running alongside the physics engines they were built to challenge.**\n\n**The models**: **FourCastNet** (NVIDIA, Fourier neural operator; later SFNO),\n**Pangu-Weather** (Huawei, 3D Earth-Specific Transformer), **GraphCast** (Google DeepMind,\nGNN on a multi-scale icosahedral mesh), **GenCast** (⚠️ **DeepMind, a conditional diffusion\nmodel producing ensembles**), **AIFS** (ECMWF, GNN-transformer hybrid), **Aurora**\n(Microsoft), **NeuralGCM** (hybrid), **FuXi \u002F FengWu**, **NVIDIA Atlas**, **FGN**.\n\n**⚠️ Operational status is the important part:**\n- **ECMWF's AIFS Single has run operationally since 25 February 2025** — ⚠️ **the first\n  operational ML weather system**, with a **51-member probabilistic version (AIFS-CRPS)\n  following.**\n- **NOAA\u002FNWS made AI\u002FML models available in mid-December 2025**, including **AIGFS** and\n  **HyGEFS**.\n- **Google's WeatherNext** family operationalized GenCast (as WeatherNext Gen).\n\n**⚠️ The skill claims, and they're substantiated:** GraphCast **matches or exceeds IFS HRES\non global benchmarks**; **GenCast was the first probabilistic MLWP model to significantly\noutperform ECMWF's ENS at high resolution** (Nature, 2024); and by 2026 ⚠️ **ensemble\nsystems including GenCast and FGN have surpassed the skill of the ECMWF ensemble — the\ngold standard in operational meteorology — at a small fraction of the computational\ncost.**\n\n**⚠️ The computational asymmetry is the genuinely disruptive part**: training is expensive\nand one-off; **inference is seconds on a single GPU versus hours on a supercomputer.**\n\n### 15.2 ⚠️ The limits — and there is a live scientific dispute here\n> **⚠️ GOTCHA — the headline \"AI beats physics\" and the extremes evidence genuinely\n> conflict, and you should know both sides.**\n>\n> **A Science Advances paper (2026, Zhang et al.) found that for record-breaking weather\n> extremes, ECMWF's HRES still consistently outperforms GraphCast, Pangu-Weather and FuXi**\n> — ⚠️ **with AI errors larger for record-breaking heat, cold and wind across nearly all\n> lead times, systematic underestimation of both frequency and intensity of records, and\n> growing bias the further the record is exceeded.** ⚠️ **The performance gap was widest\n> at SHORT lead times.**\n>\n> **The proposed mechanism is important**: ⚠️ **models trained on 1979–2017 tend to be\n> limited to extreme values already observed, \"as if they had an implicit ceiling,\" while\n> physics-based models are not so constrained and can in principle represent unprecedented\n> situations.**\n\n**Other documented limitations:**\n- **⚠️ Smoothing.** Models trained to minimize average error produce **overly smooth\n  fields, blurring small-scale structure and systematically underrepresenting extremes,\n  worsening with lead time.** ⚠️ **This is §13's double-penalty problem expressed as a\n  training objective.** **Diffusion-based ensembles (GenCast) substantially address it** —\n  which is a large part of why they work.\n- **⚠️ Dependence on physics-based assimilation.** **Every current ML model relies on a\n  conventional analysis for initial conditions** (§12.2). ⚠️ **The replacement narrative is\n  oversold and the dependency is underreported** — and **GraphCast initialized on GFS\n  rather than ERA5 produces systematic inconsistencies.** **ML-based data assimilation is\n  the live frontier that would close this.**\n- **Physical consistency** — ⚠️ **outputs can violate known physical laws or expected\n  statistical consistency**, and one assessment concluded AI models **do not properly\n  reproduce sub-synoptic and mesoscale phenomena.**\n- **Precipitation** — ⚠️ **persistent difficulty with light precipitation across GraphCast,\n  Pangu, FuXi and CREDIT, with a positive frequency bias in the drizzle regime**, attributed\n  to symmetric regression losses on a non-negative intermittent variable.\n- **Temporal resolution** — ⚠️ **AIFS and AIGFS step every 6 hours where GFS gives hourly\n  output; that matters for hurricanes and sharp fronts.**\n- **Tropical cyclones** — track prediction is good; ⚠️ **intensity is uneven, with earlier\n  models underestimating peak intensity and both tested models overestimating inner-core\n  size. And a subtle trap: ERA5 itself underestimates peak intensity, so agreement with\n  reanalysis does not imply accuracy.**\n- **Climate shift** — ⚠️ **behaviour in novel climate states not represented in training\n  is an open question.**\n- **⚠️ Nowcasting (0–12 h) is still behind high-resolution NWP.**\n\n**⚠️ The counterargument deserves stating fairly too**: one 2026 review argues it is\n**\"certainly an over-statement to say that they can only predict what has been seen locally\nin their training dataset,\"** notes that ML models have **overcome long-standing physics\nmodel biases** such as slow tropical cyclone track bias, and points to **next-generation\nmodels incorporating observations directly** — potentially removing the reanalysis\ndependency entirely.\n\n### 15.3 ⚠️ The metrics problem underneath the dispute\n**Part of why the two camps disagree is that they're measuring different things.**\n⚠️ **RMSE rewards smoothing (§13), so a model that hedges scores well on the headline\nmetric while underrepresenting exactly the extremes that matter operationally.** **Work on\nfair comparison for extremes — weighted potential CRPS and similar — exists precisely\nbecause the standard scores flatter the smoothing problem.**\n**⚠️ When you read \"AI outperforms on 90% of metrics,\" ask which metrics, and on what\ndistribution of events.**\n\n### 15.4 The state of play\n**⚠️ As of 2026, no major meteorological agency has decommissioned its NWP system.**\n**ECMWF, NOAA and the Met Office all run AI models *alongside* traditional ones, not\ninstead of them**, with AI as one guidance product among several.\n**⚠️ That is the correct read**: this is an additional, extraordinarily cheap, often more\nskillful source of guidance — **not a replacement for a physics-based system that still\nprovides the analysis, the extremes, the physical consistency, and the fallback.**\n\n---\n\n## §16. Weather vs Climate\n\n**⚠️ \"Weather is what you get, climate is what you expect.\"** More usefully: **weather is\nan initial-value problem; climate is a boundary-value problem.**\n⚠️ **This is why a two-week predictability limit (§14) does not imply climate projection is\nimpossible** — you are not predicting the trajectory, you are predicting the statistics of\nthe attractor under changed forcing. **These are different mathematical problems.**\n\n**Attribution** — ⚠️ **modern extreme event attribution asks how much a specific event's\nprobability or intensity changed, not whether climate change \"caused\" it.** The\neither\u002For framing is the wrong question.\n**Climate sensitivity** — ⚠️ **cloud feedback remains the dominant uncertainty** (§2 → `weather-atmosphere-radiation-thermodynamics-and-moisture`).\n","# Weather Science: Observation, Numerical Weather Prediction, Verification, Predictability, and Machine Learning\n\n> **Part 4 of 5** of the *Weather Science* reference (plugin `weather-science`), covering §11–§16. Sibling skills: `weather-atmosphere-radiation-thermodynamics-and-moisture` (§0–§4), `weather-dynamics-circulation-and-synoptic` (§5–§7), `weather-severe-storms-cyclones-and-boundary-layer` (§8–§10), `weather-reference` (§17–§21). Section numbers are shared across the set; a reference written as §N → `skill` points into that sibling skill.\n>\n> **Currency:** Atmospheric physics is settled — hydrostatic balance, geostrophy, Lorenz's 1963 chaos work. One area moved dramatically. See §15 below for machine-learning weather prediction and the live scientific dispute around it.\n\n> **Scope.** Complements a Newtonian-mechanics reference (§10 there covers Coriolis and\n> rotating frames properly) and a fundamental-physics reference (radiation, thermodynamics).\n> ⚠️ **This is the atmosphere specifically.**\n>\n> **⚠️ GOTCHA** boxes mark genuine misconceptions and the places where intuition\n> systematically misleads.\n>\n> **The three ideas that organize the whole field:**\n> 1. **⚠️ Weather is a heat engine driven by differential solar heating.** The tropics\n>    receive more energy than they radiate, the poles the reverse, and **essentially all\n>    weather is the atmosphere and ocean moving that surplus poleward** (§2.3 → `weather-atmosphere-radiation-thermodynamics-and-moisture`).\n> 2. **⚠️ Rotation changes everything.** On a non-rotating planet air would flow directly\n>    from high to low pressure. **The Coriolis effect makes it flow *along* isobars\n>    instead**, and that single fact produces jet streams, cyclones, and the entire\n>    structure of mid-latitude weather (§5 → `weather-dynamics-circulation-and-synoptic`).\n> 3. **⚠️ The atmosphere is chaotic, and this is a mathematical property, not a\n>    measurement problem.** Lorenz (1963) showed the predictability limit is intrinsic.\n>    **No observing system and no model — physics-based or learned — removes it** (§14.1).\n\n---\n\n## §11. Observation\n\n| System | Provides | ⚠️ Notes |\n|---|---|---|\n| **Surface stations \u002F METAR** | T, Td, P, wind, visibility | Sparse and land-biased |\n| **Radiosonde** | ⚠️ **The vertical profile** | ~2×\u002Fday, ~800 sites; **the backbone of upper-air truth** |\n| **Weather radar** | Precipitation, ⚠️ **Doppler velocity** | **Dual-pol** distinguishes hydrometeor type |\n| **Satellite — geostationary** | Continuous imagery | ⚠️ **Poor at high latitudes** |\n| **Satellite — polar orbiting** | Sounding, global coverage | ⚠️ **The dominant data volume in assimilation** |\n| **Aircraft (AMDAR)** | Upper-air en route | ⚠️ **Dropped sharply during COVID and measurably degraded forecasts** |\n| **GNSS radio occultation** | ⚠️ **Bias-free temperature\u002Fhumidity profiles** | High-value, growing |\n| **Buoys, ships, profilers, lightning networks** | | |\n\n**⚠️ Radar caveats worth knowing**: the beam rises with distance so distant echoes sample\nhigher altitudes; **ground clutter, anomalous propagation, bright band at the melting\nlevel, and beam blockage** all produce artifacts. ⚠️ **Doppler measures only the radial\ncomponent of velocity** — motion across the beam is invisible.\n\n---\n\n## §12. Numerical Weather Prediction\n\n### 12.1 The governing system\n**The primitive equations**: momentum (Navier-Stokes on a rotating sphere), continuity,\nthermodynamic energy, the ideal gas law, and moisture conservation.\n⚠️ **A coupled nonlinear PDE system with no analytic solution — so you discretize.**\n\n**Discretization**: spectral (⚠️ **spherical harmonics — historically dominant for global\nmodels**) or grid-point (finite difference\u002Fvolume, icosahedral and cubed-sphere grids).\n**Vertical coordinates**: sigma, pressure, hybrid, isentropic.\n**⚠️ CFL condition constrains the timestep given grid spacing and wave speed** — **which is\nwhy resolution is expensive: halving grid spacing roughly costs 8–16×.**\n**Hydrostatic** for coarse global models; ⚠️ **non-hydrostatic required below ~10 km grid\nspacing, where convection begins to be resolved** (§1 → `weather-atmosphere-radiation-thermodynamics-and-moisture`).\n\n**⚠️ Parameterization is where the physics that can't be resolved lives**: convection,\ncloud microphysics, radiation, boundary layer turbulence, gravity wave drag, land surface.\n⚠️ **This is the largest source of model error and the hardest part of NWP.** **The\n\"grey zone\" — grid spacings around 1–10 km where convection is partly resolved and partly\nparameterized — is genuinely awkward.**\n\n### 12.2 ⚠️ Data assimilation — the underrated half\n**The forecast is only as good as its initial conditions, and observations are sparse,\nirregular and noisy.** **DA combines a short-range forecast (the \"background\") with\nobservations, weighted by their respective error statistics**, to produce an **analysis**.\n\n**Methods**: 3D-Var, **4D-Var** (⚠️ **assimilates over a time window, using the model\nitself as a constraint — ECMWF's long-standing strength**), **EnKF** (flow-dependent\nerror covariances from an ensemble), and **hybrid** approaches which now dominate.\n\n> **⚠️ GOTCHA — data assimilation is arguably a larger contributor to modern forecast\n> skill than model improvements, and it's almost invisible outside the field.**\n> ⚠️ **It's also the reason §15's ML models are not yet a full replacement: they mostly\n> consume an analysis produced by conventional physics-based assimilation.**\n\n**Reanalysis** — ⚠️ **rerunning a fixed modern DA system over the historical record to\nproduce a physically consistent gridded dataset.** **ERA5 is the standard**, and it is\n**the training data for essentially every ML weather model** (§15).\n\n### 12.3 Ensembles\n**⚠️ Since initial conditions and the model are both uncertain, run many forecasts.**\nPerturb initial conditions (singular vectors, bred modes, EDA) and represent model\nuncertainty (stochastic physics, multi-model, multi-physics).\n**⚠️ Ensemble prediction arrived operationally in 1992 and was the previous methodological\nbreak in forecasting** — before §15. **The ensemble mean is more skillful than any member;\nthe spread is the uncertainty estimate.**\n⚠️ **A well-calibrated ensemble's spread should match its error. Under-dispersion is the\ncommon failure and it makes forecasts overconfident.**\n\n---\n\n## §13. Forecast Verification\n\n**⚠️ \"Was the forecast good?\" is a harder question than it looks, and the metric you choose\ndetermines the answer** (§15.3).\n```\nRMSE \u002F MAE                deterministic error  ⚠️ REWARDS SMOOTHING — see §15.3\nACC (anomaly correlation) ⚠️ the standard synoptic skill score; 0.6 conventionally\n                          taken as the limit of useful deterministic skill\nBrier score               probabilistic, binary events\nCRPS                      ⚠️ probabilistic, continuous — the standard ensemble metric\nReliability diagram       ⚠️ do 30% forecasts verify 30% of the time?\nPOD \u002F FAR \u002F CSI           categorical events\n```\n**⚠️ The double penalty problem**: a sharp forecast of a feature in slightly the wrong place\nis penalized twice — **once for predicting it where it wasn't, once for missing it where it\nwas.** ⚠️ **A blurry forecast scores better on RMSE while being less useful.** **This is\nnot a technicality; it distorts model development toward smoothness** (§15.3).\n\n**⚠️ Skill must beat a reference** — persistence, climatology, or a previous model.\n**\"85% accurate\" means nothing without one.**\n\n---\n\n## §14. Predictability and Chaos\n\n### 14.1 The intrinsic limit\n**Lorenz (1963)** — ⚠️ **deterministic nonlinear systems exhibit sensitive dependence on\ninitial conditions.** Errors grow, and **small-scale errors propagate upscale.**\n> **⚠️ GOTCHA — the predictability limit is a property of the atmosphere, not of our\n> instruments or models.** ⚠️ **Roughly two weeks for synoptic-scale deterministic\n> forecasting is a mathematical bound, not an engineering target.** **Perfect observations\n> and a perfect model would not remove it** — see a Newtonian-mechanics reference §13 for\n> why better data buys only logarithmic improvement.\n>\n> **⚠️ And predictability is flow-dependent.** Some regimes are far more predictable than\n> others, which is exactly what the ensemble spread is telling you (§12.3).\n\n**Scale matters**: convective cells minutes to hours; mesoscale systems hours; synoptic\nsystems days; **planetary waves and teleconnections longer.** ⚠️ **Seasonal forecasting\nworks not by predicting weather but by predicting boundary-condition-driven shifts in the\ndistribution** — chiefly ENSO (§6 → `weather-dynamics-circulation-and-synoptic`).\n\n---\n\n## §15. Machine Learning Weather Prediction\n\n**⚠️ This is the section that moved, and it moved fast enough that most general knowledge\nabout it is out of date.** **Verified August 2026.**\n\n### 15.1 What happened\n**⚠️ One assessment frames it as the biggest methodological break in forecasting since\nensemble prediction arrived in 1992** — and that seems right. **Between November 2023 and\nmid-2026, data-driven models trained on reanalysis went from research curiosities to\noperational systems running alongside the physics engines they were built to challenge.**\n\n**The models**: **FourCastNet** (NVIDIA, Fourier neural operator; later SFNO),\n**Pangu-Weather** (Huawei, 3D Earth-Specific Transformer), **GraphCast** (Google DeepMind,\nGNN on a multi-scale icosahedral mesh), **GenCast** (⚠️ **DeepMind, a conditional diffusion\nmodel producing ensembles**), **AIFS** (ECMWF, GNN-transformer hybrid), **Aurora**\n(Microsoft), **NeuralGCM** (hybrid), **FuXi \u002F FengWu**, **NVIDIA Atlas**, **FGN**.\n\n**⚠️ Operational status is the important part:**\n- **ECMWF's AIFS Single has run operationally since 25 February 2025** — ⚠️ **the first\n  operational ML weather system**, with a **51-member probabilistic version (AIFS-CRPS)\n  following.**\n- **NOAA\u002FNWS made AI\u002FML models available in mid-December 2025**, including **AIGFS** and\n  **HyGEFS**.\n- **Google's WeatherNext** family operationalized GenCast (as WeatherNext Gen).\n\n**⚠️ The skill claims, and they're substantiated:** GraphCast **matches or exceeds IFS HRES\non global benchmarks**; **GenCast was the first probabilistic MLWP model to significantly\noutperform ECMWF's ENS at high resolution** (Nature, 2024); and by 2026 ⚠️ **ensemble\nsystems including GenCast and FGN have surpassed the skill of the ECMWF ensemble — the\ngold standard in operational meteorology — at a small fraction of the computational\ncost.**\n\n**⚠️ The computational asymmetry is the genuinely disruptive part**: training is expensive\nand one-off; **inference is seconds on a single GPU versus hours on a supercomputer.**\n\n### 15.2 ⚠️ The limits — and there is a live scientific dispute here\n> **⚠️ GOTCHA — the headline \"AI beats physics\" and the extremes evidence genuinely\n> conflict, and you should know both sides.**\n>\n> **A Science Advances paper (2026, Zhang et al.) found that for record-breaking weather\n> extremes, ECMWF's HRES still consistently outperforms GraphCast, Pangu-Weather and FuXi**\n> — ⚠️ **with AI errors larger for record-breaking heat, cold and wind across nearly all\n> lead times, systematic underestimation of both frequency and intensity of records, and\n> growing bias the further the record is exceeded.** ⚠️ **The performance gap was widest\n> at SHORT lead times.**\n>\n> **The proposed mechanism is important**: ⚠️ **models trained on 1979–2017 tend to be\n> limited to extreme values already observed, \"as if they had an implicit ceiling,\" while\n> physics-based models are not so constrained and can in principle represent unprecedented\n> situations.**\n\n**Other documented limitations:**\n- **⚠️ Smoothing.** Models trained to minimize average error produce **overly smooth\n  fields, blurring small-scale structure and systematically underrepresenting extremes,\n  worsening with lead time.** ⚠️ **This is §13's double-penalty problem expressed as a\n  training objective.** **Diffusion-based ensembles (GenCast) substantially address it** —\n  which is a large part of why they work.\n- **⚠️ Dependence on physics-based assimilation.** **Every current ML model relies on a\n  conventional analysis for initial conditions** (§12.2). ⚠️ **The replacement narrative is\n  oversold and the dependency is underreported** — and **GraphCast initialized on GFS\n  rather than ERA5 produces systematic inconsistencies.** **ML-based data assimilation is\n  the live frontier that would close this.**\n- **Physical consistency** — ⚠️ **outputs can violate known physical laws or expected\n  statistical consistency**, and one assessment concluded AI models **do not properly\n  reproduce sub-synoptic and mesoscale phenomena.**\n- **Precipitation** — ⚠️ **persistent difficulty with light precipitation across GraphCast,\n  Pangu, FuXi and CREDIT, with a positive frequency bias in the drizzle regime**, attributed\n  to symmetric regression losses on a non-negative intermittent variable.\n- **Temporal resolution** — ⚠️ **AIFS and AIGFS step every 6 hours where GFS gives hourly\n  output; that matters for hurricanes and sharp fronts.**\n- **Tropical cyclones** — track prediction is good; ⚠️ **intensity is uneven, with earlier\n  models underestimating peak intensity and both tested models overestimating inner-core\n  size. And a subtle trap: ERA5 itself underestimates peak intensity, so agreement with\n  reanalysis does not imply accuracy.**\n- **Climate shift** — ⚠️ **behaviour in novel climate states not represented in training\n  is an open question.**\n- **⚠️ Nowcasting (0–12 h) is still behind high-resolution NWP.**\n\n**⚠️ The counterargument deserves stating fairly too**: one 2026 review argues it is\n**\"certainly an over-statement to say that they can only predict what has been seen locally\nin their training dataset,\"** notes that ML models have **overcome long-standing physics\nmodel biases** such as slow tropical cyclone track bias, and points to **next-generation\nmodels incorporating observations directly** — potentially removing the reanalysis\ndependency entirely.\n\n### 15.3 ⚠️ The metrics problem underneath the dispute\n**Part of why the two camps disagree is that they're measuring different things.**\n⚠️ **RMSE rewards smoothing (§13), so a model that hedges scores well on the headline\nmetric while underrepresenting exactly the extremes that matter operationally.** **Work on\nfair comparison for extremes — weighted potential CRPS and similar — exists precisely\nbecause the standard scores flatter the smoothing problem.**\n**⚠️ When you read \"AI outperforms on 90% of metrics,\" ask which metrics, and on what\ndistribution of events.**\n\n### 15.4 The state of play\n**⚠️ As of 2026, no major meteorological agency has decommissioned its NWP system.**\n**ECMWF, NOAA and the Met Office all run AI models *alongside* traditional ones, not\ninstead of them**, with AI as one guidance product among several.\n**⚠️ That is the correct read**: this is an additional, extraordinarily cheap, often more\nskillful source of guidance — **not a replacement for a physics-based system that still\nprovides the analysis, the extremes, the physical consistency, and the fallback.**\n\n---\n\n## §16. Weather vs Climate\n\n**⚠️ \"Weather is what you get, climate is what you expect.\"** More usefully: **weather is\nan initial-value problem; climate is a boundary-value problem.**\n⚠️ **This is why a two-week predictability limit (§14) does not imply climate projection is\nimpossible** — you are not predicting the trajectory, you are predicting the statistics of\nthe attractor under changed forcing. **These are different mathematical problems.**\n\n**Attribution** — ⚠️ **modern extreme event attribution asks how much a specific event's\nprobability or intensity changed, not whether climate change \"caused\" it.** The\neither\u002For framing is the wrong question.\n**Climate sensitivity** — ⚠️ **cloud feedback remains the dominant uncertainty** (§2 → `weather-atmosphere-radiation-thermodynamics-and-moisture`).\n",{"name":5,"description":7},"develop","ae32736a1aabddabca5d2c96d665aa0ebc7d8700","src\u002Fvibey_tools\u002Fskills\u002Fplugins\u002Fweather-science\u002Fskills\u002Fweather-observation-nwp-verification-and-machine-learning",3,"https:\u002F\u002Fgithub.com\u002Fthe-vibey-project\u002Fvibey","https:\u002F\u002Fthe-vibey-project.github.io\u002Fvibey\u002F","https:\u002F\u002Fgithub.com\u002Fthe-vibey-project","https:\u002F\u002Fvibewithadam.matthewsteinberger.com","freedom for software engineers","2026-11-15 05:47:42",{"source":8,"skillName":5,"commands":35,"prompt":45,"zipAvailable":46},[36,39,42],{"tool":37,"command":38},"npx","npx skills add the-vibey-project\u002Fvibey --skill weather-observation-nwp-verification-and-machine-learning",{"tool":40,"command":41},"bunx","bunx skills add the-vibey-project\u002Fvibey --skill weather-observation-nwp-verification-and-machine-learning",{"tool":43,"command":44},"pnpm","pnpm dlx skills add the-vibey-project\u002Fvibey --skill weather-observation-nwp-verification-and-machine-learning","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fthe-vibey-project\u002Fvibey\u002Fblob\u002Fdevelop\u002Fsrc\u002Fvibey_tools\u002Fskills\u002Fplugins\u002Fweather-science\u002Fskills\u002Fweather-observation-nwp-verification-and-machine-learning\u002FSKILL.md",true,[],[49,56,63,70,77,84],{"skillNo":50,"name":51,"nameZh":6,"description":52,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":53,"categories":54,"keywords":55,"publishedAt":19},"01M7AT1FQJAE1FQYAMDTF490Y4","white-paper-writing","Use when writing white papers, technical reports, policy documents, thought leadership pieces, or any long-form authoritative document. Triggers on requests to write, structure, outline, or review white papers, technical reports, position papers, policy briefs, or evidence-driven arguments. Also use for B2B content marketing strategy, technical documentation requiring authority, academic-style reports, or any document that must educate while persuading. Use when someone asks \"how should I structure this\" for a long-form document.","https:\u002F\u002Fgithub.com\u002Fthe-vibey-project\u002Fvibey\u002Fblob\u002Fdevelop\u002Fsrc\u002Fvibey_tools\u002Fskills\u002Fplugins\u002Fwriting-craft\u002Fskills\u002Fwhite-paper-writing\u002FSKILL.md",[],[],{"skillNo":57,"name":58,"nameZh":6,"description":59,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":60,"categories":61,"keywords":62,"publishedAt":19},"01M7AT16BJRJJKK2MASXRZS76G","prose-technical-mechanics","Use when editing or writing prose at a technical\u002Fmechanical level: choosing between Anglo-Saxon and Latinate diction, placing information in topic vs stress positions, writing periodic vs cumulative sentences, handling sentence rhythm, punctuation as logic, transitions, paragraph structure, or phonesthetics. Use when someone wants to understand WHY prose works or doesn't, not just surface-level style advice. Triggers on word choice, sentence structure, emphasis, rhythm, diction, or any deep writing mechanics question.","https:\u002F\u002Fgithub.com\u002Fthe-vibey-project\u002Fvibey\u002Fblob\u002Fdevelop\u002Fsrc\u002Fvibey_tools\u002Fskills\u002Fplugins\u002Fwriting-craft\u002Fskills\u002Fprose-technical-mechanics\u002FSKILL.md",[],[],{"skillNo":64,"name":65,"nameZh":6,"description":66,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":67,"categories":68,"keywords":69,"publishedAt":19},"01M7AT0W9HT0ZZ9Y287NW99MR3","prose-mechanics","Use when writing, editing, or analyzing prose style; when trying to avoid generic AI-sounding writing; when applying specific prose styles such as journalistic, academic, narrative, KJV biblical, Hemingway plain, Latinate\u002FJamesian, Orwellian, Augustan, Victorian, Early Modern (Bacon, Browne, Milton), or free indirect discourse; or when coaching someone on writing voice and mechanics. Triggers on any request to improve, analyze, or write in a specific prose style, to fix \"AI-sounding\" writing, to make prose more vivid or concrete, to vary sentence rhythm, or to understand how a particular writer achieves their effects.","https:\u002F\u002Fgithub.com\u002Fthe-vibey-project\u002Fvibey\u002Fblob\u002Fdevelop\u002Fsrc\u002Fvibey_tools\u002Fskills\u002Fplugins\u002Fwriting-craft\u002Fskills\u002Fprose-mechanics\u002FSKILL.md",[],[],{"skillNo":71,"name":72,"nameZh":6,"description":73,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":74,"categories":75,"keywords":76,"publishedAt":19},"01M7AT0KR635JX66X2X6F3E2CP","narrative-structure","Use when structuring stories, pitches, presentations, white papers, or any persuasive narrative; analyzing the hero's journey in fiction or non-fiction; applying StoryBrand for business communication; designing narrative arcs for thought leadership; or understanding Vogler, Snyder, Propp, or Campbell frameworks. Triggers on any narrative structure, storytelling, pitch structure, StoryBrand, hero's journey, three-act structure, Freytag, Save the Cat, narrative arc, or \"how should I structure this story\u002Fpitch\u002Fpresentation\" request.","https:\u002F\u002Fgithub.com\u002Fthe-vibey-project\u002Fvibey\u002Fblob\u002Fdevelop\u002Fsrc\u002Fvibey_tools\u002Fskills\u002Fplugins\u002Fwriting-craft\u002Fskills\u002Fnarrative-structure\u002FSKILL.md",[],[],{"skillNo":78,"name":79,"nameZh":6,"description":80,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":81,"categories":82,"keywords":83,"publishedAt":19},"01M7AT0AKGW09NKKZDYBT8VRDE","legal-brief-writing","Use when writing legal briefs, motions, or any court filing; when analyzing judicial cognition and persuasion; when applying elite brief-writing techniques from Roberts, Clement, Kagan, Garner; or when structuring legal arguments for maximum persuasion. Also use for autistic\u002Fneurodivergent advocates who need to understand how their natural communication style maps to elite legal advocacy. Triggers on any legal writing, brief drafting, motion preparation, or legal argument structuring request.","https:\u002F\u002Fgithub.com\u002Fthe-vibey-project\u002Fvibey\u002Fblob\u002Fdevelop\u002Fsrc\u002Fvibey_tools\u002Fskills\u002Fplugins\u002Fwriting-craft\u002Fskills\u002Flegal-brief-writing\u002FSKILL.md",[],[],{"skillNo":85,"name":86,"nameZh":6,"description":87,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":88,"categories":89,"keywords":90,"publishedAt":19},"01M7AT00VWPS4Q8B5VEMXX9NQT","journal-publishability-criteria","Use whenever a research paper, article or preprint is being written, revised, reviewed or published, and specifically before every publication of docs\u002Fpaper.md. Covers what legitimate journals actually gate on, in order of gating — the editor's desk screen (scope, novelty, package problems), what reviewers judge (methodological soundness, claims the evidence can carry, statistical and qualitative reporting), EQUATOR reporting guidelines, the ICMJE authorship criteria, COPE's view of text recycling, the near-universal generative-AI policy (no AI author, disclosure, author responsibility), TOP 2025 transparency standards, how to verify a journal is legitimate, submitting without an institution, and a pre-submission checklist. Triggers on requests to evaluate, review, prepare or submit a paper, to judge publishability, to write a declarations or data-availability section, to disclose AI use, or to check a journal.","https:\u002F\u002Fgithub.com\u002Fthe-vibey-project\u002Fvibey\u002Fblob\u002Fdevelop\u002Fsrc\u002Fvibey_tools\u002Fskills\u002Fplugins\u002Fwriting-craft\u002Fskills\u002Fjournal-publishability-criteria\u002FSKILL.md",[],[]]