Merge branch 'zed/models'

Author: Flowm

CONTEXT.md

@@ -1,21 +1,21 @@
 # MeteoCompare
 
-Multi-model weather forecast comparison. The app pulls deterministic numerical-weather-prediction (NWP) forecasts from many models via open-meteo, weights them, and produces a single aggregate forecast plus a per-timestep confidence score. A secondary verification surface compares past forecasts against a reference reanalysis field to expose which models (and the aggregate) were actually right.
+Multi-model weather forecast comparison. The app pulls operational forecast products from many models via open-meteo, weights them, and produces a single aggregate forecast plus a per-timestep confidence score. A secondary verification surface compares past forecasts against a reference reanalysis field to expose which models (and the aggregate) were actually right.
 
 ## Language
 
 ### Models & runs
 
 **Model**:
-A specific NWP product exposed by open-meteo, e.g. ECMWF IFS or DWD ICON-EU. One logical model per row in `src/domain/models.ts`.
+A specific forecast product exposed by open-meteo, e.g. ECMWF IFS HRES, DWD ICON-EU, or ECMWF AIFS. One logical model per row in `src/domain/models.ts`.
 _Avoid_: provider, source, dataset.
 
 **Probability graft**:
 open-meteo derives `precipitation_probability` only from ensembles, so deterministic models return null for it. Where a model has an ensemble-backed _seamless_ sibling, we fetch that sibling solely for the probability variable and read its series under the registered model's id — the _graft_. Today `icon_seamless` is grafted onto DWD ICON (`icon_global`). The **graft source** (`icon_seamless`) is never a Model: no registry row, no chip, no aggregate vote of its own; it supplies one variable under the host model's identity and weight. Note we consume open-meteo's _derived_ probability, never raw ensemble members (see "Aggregate").
 _Avoid_: registering a graft source as a Model.
 
 **Model class**:
-Resolution-and-scope tier: `global`, `regional-mid`, or `regional-cam` (convection-allowing). Drives lead-time decay and precipitation boost.
+Resolution/product-and-scope tier: `global`, `regional-mid`, `regional-cam` (convection-allowing), `ai` (machine-learned single forecast product), or `ensemble-mean` (mean of ensemble members exposed as one forecast product). Drives lead-time decay, precipitation boost, and how strongly a model contributes relative to deterministic NWP products.
 
 **Home region**:
 The geographic bounding box where a regional model has a structural advantage. Drives the region bonus inside the weighting recipe. Global models have no home region.

README.md

@@ -6,7 +6,7 @@ Frontend-only (Vue 3 + Vite). Forecasts come straight from [open-meteo.com](http
 
 ## Features
 
-- **11 NWP models**, automatically dropped in/out based on geographic coverage and forecast horizon.
+- **21 forecast models/products**, automatically dropped in/out based on geographic coverage and forecast horizon.
 - **Aggregate-first UI**: temperature + ±1σ confidence band, precipitation bars, daily strip with weather icon / high / low / precip prob / wind.
 - **Confidence score** per timestep — derived from inter-model spread normalised against typical seasonal spread, a model-count penalty, and lead-time decay encoded in the model weights.
 - **Multi-model breakdown** (opt-in) — spaghetti chart of every contributing model with per-model toggles, switchable between temperature, precipitation, precipitation probability, wind speed, and cloud cover.
@@ -22,7 +22,7 @@ Per timestep and per variable:
 2. **Weight them.**
    - Base weight = 1.
    - Region bonus of +0.2 (mid-resolution) or +0.3 (convection-allowing) when the location is inside the model's home region.
-   - Lead-time decay per model class: convection-allowing models fade out by 60 h, mid-resolution regionals by 120 h, globals decay gently from 72 h → 0.4× by 240 h.
+   - Lead-time decay per model class: convection-allowing models fade out by 60 h, mid-resolution regionals by 120 h, globals decay gently from 72 h → 0.4× by 240 h, and AI plus ensemble-mean products follow global decay with a smaller vote.
    - Variable boost: CAMs get ×1.3 for precipitation, since they explicitly resolve convection.
 3. **Aggregate**:
    - **Temperature / precip / cloud cover / wind speed** → weighted mean + weighted standard deviation.
@@ -50,28 +50,36 @@ Weather codes have no meaningful stdDev, so they use severity-group agreement in
 `confidence = clamp(weightShare(same severity group) × modelFactor, 0, 1)`.
 
 Lead-time decay is handled entirely in the model weighting layer (not as a separate
-multiplier here): CAMs fade out by 60 h, regionals by 120 h, globals decay past 72 h.
+multiplier here): CAMs fade out by 60 h, regionals by 120 h, globals decay past 72 h,
+and AI plus ensemble-mean products follow global decay with a smaller vote.
 
 The badge maps the result to one of three tiers — high (≥70 %, emerald), mid (≥40 %, amber), low (rose).
 
 ## Models
 
-| Open-meteo id          | Provider           | Resolution / scope          | Class        | Max lead |
-| ---------------------- | ------------------ | --------------------------- | ------------ | -------- |
-| `ecmwf_ifs025`         | ECMWF              | 25 km global                | global       | 240 h    |
-| `gfs_global`           | NOAA               | 13–25 km global             | global       | 384 h    |
-| `gfs_hrrr`             | NOAA               | 3 km CONUS CAM              | regional-cam | 48 h     |
-| `icon_global`          | DWD                | 11 km global                | global       | 180 h    |
-| `icon_eu`              | DWD                | 7 km Europe                 | regional-mid | 120 h    |
-| `icon_d2`              | DWD                | 2 km central Europe CAM     | regional-cam | 48 h     |
-| `gem_seamless`         | Environment Canada | 2.5–15 km, NA focus         | regional-mid | 240 h    |
-| `meteofrance_seamless` | Météo-France       | 1.3 km AROME / 25 km ARPEGE | regional-cam | 102 h    |
-| `ukmo_seamless`        | UK Met Office      | 2 km UKV / 10 km global     | regional-mid | 168 h    |
-| `knmi_seamless`        | KNMI               | 2.5 km Benelux              | regional-cam | 60 h     |
-| `metno_seamless`       | MET Norway         | 2.5 km Nordics              | regional-cam | 60 h     |
-| `jma_seamless`         | JMA                | 5 km Japan / 55 km global   | regional-mid | 264 h    |
-| `kma_seamless`         | KMA                | 1.5–13 km, Korea focus      | regional-mid | 288 h    |
-| `bom_access_global`    | BOM                | 15 km global, Aus. focus    | global       | 240 h    |
+| Open-meteo id                 | Provider           | Resolution / scope                 | Class         | Max lead |
+| ----------------------------- | ------------------ | ---------------------------------- | ------------- | -------- |
+| `ecmwf_ifs`                   | ECMWF              | 9 km HRES global                   | global        | 240 h    |
+| `gfs_seamless`                | NOAA               | seamless NOAA global/U.S. coverage | global        | 384 h    |
+| `gem_seamless`                | Environment Canada | 2.5–15 km, NA focus                | regional-mid  | 240 h    |
+| `ukmo_seamless`               | UK Met Office      | 2 km UKV / 10 km global            | regional-mid  | 168 h    |
+| `meteofrance_seamless`        | Météo-France       | 1.3 km AROME / 25 km ARPEGE        | regional-cam  | 102 h    |
+| `cma_grapes_global`           | CMA                | 15 km global, East Asia focus      | global        | 240 h    |
+| `bom_access_global`           | BOM                | 15 km global, Aus. focus           | global        | 240 h    |
+| `jma_seamless`                | JMA                | 5 km Japan / 55 km global          | regional-mid  | 264 h    |
+| `kma_seamless`                | KMA                | 1.5–13 km, Korea focus             | regional-mid  | 288 h    |
+| `icon_global`                 | DWD                | 11 km global                       | global        | 180 h    |
+| `icon_eu`                     | DWD                | 7 km Europe                        | regional-mid  | 120 h    |
+| `icon_d2`                     | DWD                | 2 km central Europe CAM            | regional-cam  | 48 h     |
+| `knmi_harmonie_arome_europe`  | KNMI               | 2 km Harmonie AROME Europe         | regional-cam  | 60 h     |
+| `dmi_harmonie_arome_europe`   | DMI                | 2 km Harmonie AROME Europe         | regional-cam  | 60 h     |
+| `metno_nordic`                | MET Norway         | 2.5 km Nordics                     | regional-cam  | 60 h     |
+| `meteoswiss_icon_seamless`    | MeteoSwiss         | 1–2 km ICON Switzerland seamless   | regional-cam  | 120 h    |
+| `geosphere_arome_austria`     | GeoSphere Austria  | AROME Austria                      | regional-cam  | 60 h     |
+| `ecmwf_aifs025_single`        | ECMWF              | 0.25° AI forecast                  | ai            | 360 h    |
+| `gfs_graphcast025`            | NOAA               | 0.25° GraphCast forecast           | ai            | 384 h    |
+| `ncep_aigfs025`               | NOAA               | 0.25° AI-enhanced GFS              | ai            | 384 h    |
+| `ncep_hgefs025_ensemble_mean` | NOAA               | 0.25° ensemble mean                | ensemble-mean | 384 h    |
 
 ## Tech
 

src/api/omForecast.ts

@@ -100,7 +100,7 @@ export async function fetchForecast(req: ForecastRequest, signal?: AbortSignal):
 }
 
 /** Pull a per-model hourly series for one base variable.
- *  open-meteo suffixes each variable with the model id (e.g. temperature_2m_ecmwf_ifs025). */
+ *  open-meteo suffixes each variable with the model id (e.g. temperature_2m_ecmwf_ifs). */
 export function extractHourlyByModel(resp: ForecastResponse, variable: HourlyVar, modelIds: string[]): Record<string, (number | null)[]> {
   const out: Record<string, (number | null)[]> = {};
   for (const id of modelIds) {

src/components/chartOption.test.ts

@@ -33,7 +33,7 @@ const DATA: HourlySeries = {
     cloud_cover: [pt(50), pt(60), pt(70), pt(80), pt(90)],
   },
   perModel: {
-    temperature_2m: { ecmwf_ifs025: [10, 11, 12, 13, 14], gfs_global: [9, 10, 11, 12, 13] },
+    temperature_2m: { ecmwf_ifs: [10, 11, 12, 13, 14], gfs_seamless: [9, 10, 11, 12, 13] },
   },
 };
 
@@ -87,12 +87,12 @@ describe("buildHourlyChartOption — axis pinning", () => {
 });
 
 describe("buildHourlyChartOption — spaghetti", () => {
-  const models = [getModel("ecmwf_ifs025")!, getModel("gfs_global")!];
+  const models = [getModel("ecmwf_ifs")!, getModel("gfs_seamless")!];
   it("builds one per-model series per available model when showModels is on", () => {
     const ids = seriesOf({ ...base, models, showModels: true })
       .map((s) => s.id)
       .filter((id): id is string => !!id?.startsWith("s-"));
-    expect(ids.toSorted()).toEqual(["s-ecmwf_ifs025", "s-gfs_global"]);
+    expect(ids.toSorted()).toEqual(["s-ecmwf_ifs", "s-gfs_seamless"]);
   });
   it("builds no per-model series when showModels is off", () => {
     const ids = seriesOf({ ...base, models, showModels: false }).filter((s) => s.id?.startsWith("s-"));
@@ -128,7 +128,7 @@ describe("buildHourlyChartOption — night/now marks", () => {
 });
 
 describe("buildHourlyChartOption — visibility toggles", () => {
-  const models = [getModel("ecmwf_ifs025")!, getModel("gfs_global")!];
+  const models = [getModel("ecmwf_ifs")!, getModel("gfs_seamless")!];
 
   it("emits aggregate + band toggles for a line view", () => {
     const t = togglesOf(base);
@@ -155,7 +155,7 @@ describe("buildHourlyChartOption — visibility toggles", () => {
 
   it("emits one model toggle per spaghetti series and none when models are off", () => {
     const on = togglesOf({ ...base, models, showModels: true }).filter((x) => x.group === "model");
-    expect(on.map((x) => x.id).toSorted()).toEqual(["s-ecmwf_ifs025", "s-gfs_global"]);
+    expect(on.map((x) => x.id).toSorted()).toEqual(["s-ecmwf_ifs", "s-gfs_seamless"]);
     expect(on.every((x) => !!x.modelId && x.props.includes("lineStyle"))).toBe(true);
     expect(togglesOf({ ...base, models, showModels: false }).some((x) => x.group === "model")).toBe(false);
   });

src/domain/aggregate.test.ts

@@ -4,7 +4,7 @@ import { aggregateSeries } from "./aggregate";
 import { getModel } from "./models";
 
 const PARIS = { lat: 48.85, lon: 2.35 };
-const subset = [getModel("ecmwf_ifs025")!, getModel("gfs_global")!, getModel("icon_global")!, getModel("meteofrance_seamless")!];
+const subset = [getModel("ecmwf_ifs")!, getModel("gfs_seamless")!, getModel("icon_global")!, getModel("meteofrance_seamless")!];
 
 function makeTimes(n: number, baseISO: string): string[] {
   const base = new Date(baseISO).getTime();
@@ -16,8 +16,8 @@ describe("aggregateSeries (temperature)", () => {
     const baseTime = new Date("2026-05-20T00:00:00Z");
     const times = makeTimes(4, "2026-05-20T00:00:00Z");
     const series = {
-      ecmwf_ifs025: [10, 11, 12, 13],
-      gfs_global: [10, 11, 12, 13],
+      ecmwf_ifs: [10, 11, 12, 13],
+      gfs_seamless: [10, 11, 12, 13],
       icon_global: [10, 11, 12, 13],
       meteofrance_seamless: [10, 11, 12, 13],
     };
@@ -38,8 +38,8 @@ describe("aggregateSeries (temperature)", () => {
     const baseTime = new Date("2026-05-20T00:00:00Z");
     const times = makeTimes(1, "2026-05-20T00:00:00Z");
     const series = {
-      ecmwf_ifs025: [10],
-      gfs_global: [14],
+      ecmwf_ifs: [10],
+      gfs_seamless: [14],
       icon_global: [12],
       meteofrance_seamless: [11],
     };
@@ -59,8 +59,8 @@ describe("aggregateSeries (temperature)", () => {
     const baseTime = new Date("2026-05-20T00:00:00Z");
     const times = makeTimes(1, "2026-05-20T00:00:00Z");
     const series = {
-      ecmwf_ifs025: [10],
-      gfs_global: [null],
+      ecmwf_ifs: [10],
+      gfs_seamless: [null],
       icon_global: [null],
       meteofrance_seamless: [null],
     };
@@ -84,8 +84,8 @@ describe("aggregateSeries (wind_direction_10m)", () => {
     const out = aggregateSeries(
       times,
       {
-        ecmwf_ifs025: [350],
-        gfs_global: [10],
+        ecmwf_ifs: [350],
+        gfs_seamless: [10],
         icon_global: [355],
         meteofrance_seamless: [5],
       },
@@ -110,8 +110,8 @@ describe("aggregateSeries (wind_direction_10m)", () => {
     const out = aggregateSeries(
       times,
       {
-        ecmwf_ifs025: [180],
-        gfs_global: [185],
+        ecmwf_ifs: [180],
+        gfs_seamless: [185],
         icon_global: [175],
         meteofrance_seamless: [180],
       },
@@ -131,8 +131,8 @@ describe("aggregateSeries (wind_direction_10m)", () => {
     const out = aggregateSeries(
       times,
       {
-        ecmwf_ifs025: [0],
-        gfs_global: [180],
+        ecmwf_ifs: [0],
+        gfs_seamless: [180],
         icon_global: [0],
         meteofrance_seamless: [180],
       },
@@ -155,8 +155,8 @@ describe("aggregateSeries (weather_code)", () => {
     const times = makeTimes(1, "2026-05-20T00:00:00Z");
     // 3 models say "rain" (61, 63, 80), 1 says "clear" (0).
     const series = {
-      ecmwf_ifs025: [61],
-      gfs_global: [63],
+      ecmwf_ifs: [61],
+      gfs_seamless: [63],
       icon_global: [80],
       meteofrance_seamless: [0],
     };

src/domain/aggregateVariables.test.ts

@@ -6,7 +6,7 @@ import { confidenceFor } from "./confidence";
 import { getModel } from "./models";
 
 const PARIS = { lat: 48.85, lon: 2.35 };
-const subset = [getModel("ecmwf_ifs025")!, getModel("gfs_global")!, getModel("icon_global")!, getModel("meteofrance_seamless")!];
+const subset = [getModel("ecmwf_ifs")!, getModel("gfs_seamless")!, getModel("icon_global")!, getModel("meteofrance_seamless")!];
 
 function makeTimes(n: number, baseISO: string): string[] {
   const base = new Date(baseISO).getTime();
@@ -24,8 +24,8 @@ describe("aggregateVariables — index→lead-hours convention", () => {
     // (no clamping) — the typicalSpread band differs between lead 1 and lead 36.
     const perModel = {
       temperature_2m: {
-        ecmwf_ifs025: [10, 10],
-        gfs_global: [10.6, 10.6],
+        ecmwf_ifs: [10, 10],
+        gfs_seamless: [10.6, 10.6],
         icon_global: [10.3, 10.3],
         meteofrance_seamless: [10.1, 10.1],
       },
@@ -48,7 +48,7 @@ describe("aggregateVariables — index→lead-hours convention", () => {
 describe("aggregateVariables — key vs family", () => {
   it("keys the result by `key` but weights/scores by `family`", () => {
     const times = makeTimes(1, "2026-05-20T00:00:00Z");
-    const byModel = { ecmwf_ifs025: [18], gfs_global: [22], icon_global: [20], meteofrance_seamless: [19] };
+    const byModel = { ecmwf_ifs: [18], gfs_seamless: [22], icon_global: [20], meteofrance_seamless: [19] };
     const out = aggregateVariables({
       times,
       perModel: { temperature_2m_max: byModel },
@@ -72,8 +72,8 @@ describe("aggregateVariables — shape", () => {
   it("echoes perModel and keys aggregate/confidence per variable, lengths match times", () => {
     const times = makeTimes(3, "2026-05-20T00:00:00Z");
     const perModel = {
-      temperature_2m: { ecmwf_ifs025: [10, 11, 12], gfs_global: [10, 11, 12], icon_global: [10, 11, 12], meteofrance_seamless: [10, 11, 12] },
-      precipitation: { ecmwf_ifs025: [0, 1, 0], gfs_global: [0, 1, 0], icon_global: [0, 2, 0], meteofrance_seamless: [0, 1, 0] },
+      temperature_2m: { ecmwf_ifs: [10, 11, 12], gfs_seamless: [10, 11, 12], icon_global: [10, 11, 12], meteofrance_seamless: [10, 11, 12] },
+      precipitation: { ecmwf_ifs: [0, 1, 0], gfs_seamless: [0, 1, 0], icon_global: [0, 2, 0], meteofrance_seamless: [0, 1, 0] },
     };
     const out = aggregateVariables({
       times,
@@ -100,7 +100,7 @@ describe("aggregateVariables — shape", () => {
 describe("aggregateVariables — weather_code is lead-independent", () => {
   it("scores weather_code by agreement, so cadence does not change confidence", () => {
     const times = makeTimes(1, "2026-05-20T00:00:00Z");
-    const perModel = { weather_code: { ecmwf_ifs025: [61], gfs_global: [63], icon_global: [80], meteofrance_seamless: [0] } };
+    const perModel = { weather_code: { ecmwf_ifs: [61], gfs_seamless: [63], icon_global: [80], meteofrance_seamless: [0] } };
     const vars = [{ key: "weather_code", family: "weather_code" as const }];
     const base = { times, perModel, vars, models: subset, lat: PARIS.lat, lon: PARIS.lon, baseTime };
 
@@ -115,7 +115,7 @@ describe("aggregateVariables — weather_code is lead-independent", () => {
 describe("aggregateVariables — degenerate input", () => {
   it("yields a NaN aggregate value and zero confidence when every model is null", () => {
     const times = makeTimes(1, "2026-05-20T00:00:00Z");
-    const perModel = { temperature_2m: { ecmwf_ifs025: [null], gfs_global: [null], icon_global: [null], meteofrance_seamless: [null] } };
+    const perModel = { temperature_2m: { ecmwf_ifs: [null], gfs_seamless: [null], icon_global: [null], meteofrance_seamless: [null] } };
     const out = aggregateVariables({
       times,
       perModel,