Weather, Later

The problem

CMIP6 climate projections are scientifically rich and practically inaccessible: the data sits in specialist archives, the bias correction and downscaling that make a raw model cell trustworthy at a point is its own research problem, and the products that package it for a general audience tend to charge for a PDF report. Weather, Later asks what a climate forecast looks like built with the same rigour as a weather app: search a place, see how its heat, water, winter, and energy exposure shifts out to 2100, and get a scientifically grounded answer to where your future climate already exists today, for free, with no server standing between the model and the browser.

How it works

All the heavy computation runs once, offline, on the DKRZ Levante supercomputer: xclim derives 23 from daily CMIP6 output, xsdba quantile-delta-mapping bias-corrects every absolute-threshold metric against ERA5, and a delta-change step downscales the ensemble to a fine point grid. The output is not a live model: it is a set of small, pre-computed static artefacts (quantised grids, coarse globe textures, a Climate Twin lookup table) that Cloudflare Pages and R2 serve straight to the browser, so the site itself carries no backend and no per-visit compute cost.

Every result carries its ensemble spread rather than a bare mean, and the flagship feature is the Climate Twin: for any point, a sigma-dissimilarity match against a pool of 3,979 finds the place whose current climate looks like this one’s projected future, so the answer to “what will it feel like here” is another city’s today rather than an abstract number.

> ENGINE_ROOM — Bias correction, downscaling, and the honest framing

Bias correction is deliberately narrow: only the metrics defined on an absolute threshold, such as hot days or heating and cooling degree-days, are quantile-mapped against ERA5, because a threshold count is exactly the statistic a raw model bias distorts. The Climate Twin match runs on each location’s full climate-normal vector rather than a single variable, so the analogue reflects the same multivariate comparison a climatologist would make. Every plotted field keeps its ensemble median alongside its p10 to p90 band rather than collapsing to a mean, against a fixed 1995 to 2014 baseline and three emissions scenarios (SSP1-2.6, SSP2-4.5, SSP5-8.5). The framing throughout stays honest about what CMIP6 can and cannot say: it is climate, not weather, the resolution is a broad area rather than a street, and the site says so rather than implying a precision the data doesn’t have.

What this demonstrates