Climate
Forecasting resilient climate decisions
Fast simulation tools for weather risk, energy demand, and environmental adaptation.
Forecasts have to be actionable, not just accurate
A forecast is only useful if someone can act on it. Our climate work targets the regimes where traditional simulation is too slow and standard ML is too brittle, and it surfaces uncertainty alongside every headline number so decisions are defensible.
accurate
a point forecast that minimises a leaderboard metric, often without saying where it is confident and where it is guessing.
actionable
a headline number paired with a calibrated band an operator can read, defend, and act on under a written policy.
Where traditional simulation is too slow and standard ML is too brittle
energy demand
infrastructure planning
The forecast is honest when the quantiles line up
A 90 % band that covers 90 % of outcomes is calibrated. We report the reliability diagram on every model we ship — if it does not sit on the diagonal, it is not actionable.
A headline number alone is not a decision
Every output ships with the band an operator needs to make the call. The action column is part of the deliverable, not an afterthought.
signal headline band action urgency
peak load 8.4 GW 7.9 – 9.0 GW (P10–P90) arm 1.1 GW reserve high
6-hour gust risk 38 % 24 – 54 % (P25–P75) pre-stage crew medium
next-day temp +1.6 °C +0.9 – +2.4 °C shift maintenance low
flood window 14:00–17:00 70 – 92 % within window open sluice gates high
Faster, calibrated forecasts
Our models target the regimes where traditional simulation is too slow and traditional ML is too brittle: short-horizon weather risk, energy demand, and infrastructure planning.
Decisions, not just numbers
A forecast is only useful if a person can act on it. We design outputs that surface uncertainty alongside the headline number so decisions are easier to defend.
cross-cuts
Calibration is a discipline, not a metric
The same reliability scrutiny we apply here travels into our evaluation library and our alignment posture. A forecast that lies about its uncertainty is treated as a failed system, not a leaderboard outlier.