PickMicrosoft uses ML for longer-range weather forecasts
Microsoft researchers say a cost-effective machine-learning method improves forecasts beyond the usual two-week reliability limit.
Microsoft researchers say they have found a cheaper, more accurate way to forecast weather two to six weeks ahead, a range where both traditional and AI forecasts usually lose reliability.
The problem
Physics-based weather models simulate the atmosphere on supercomputers. Newer AI weather models learn patterns from historical data and typically run on GPUs. Both approaches start to struggle beyond about two weeks, partly because they depend on approximations to fill gaps, for example in regions with few weather stations.
Microsoft's approach
Instead of building another standalone forecasting model, the Microsoft team used machine learning to correct the systematic biases in existing models. The results were published in the journal Nature, and the tool has been released as open-source software.
A key selling point is efficiency. Lester Mackey, a senior principal researcher at Microsoft Research New England, says the method can run on ordinary CPUs rather than requiring supercomputers or large GPU clusters. That makes it far more accessible to national weather services and agencies with limited budgets. The researchers say better forecasts at this range could help farmers plan planting and harvests and help emergency responders prepare for extreme weather.
Microsoft joins Google, NVIDIA and others that are developing AI tools for weather and climate.
Why it matters for builders
There are two lessons here beyond meteorology. First, "fix the existing model" can beat "build a new model". Many real products improve fastest by adding a learned correction layer on top of a trusted baseline, whether that baseline is a physics simulation, a rules engine or a large language model.
Second, CPU-friendly machine learning is underrated. If a method can deliver most of the value without GPUs, it can be deployed in far more places, from small agencies to edge devices, and it costs much less to run.
For developers in agriculture, logistics, energy or insurance, an open-source subseasonal forecasting tool is also a concrete building block. Better two-to-six-week outlooks could feed scheduling, pricing and risk features in your own applications.
You can find model hubs and AI platforms for experimenting with open-source models in our APIs & Model Platforms category.
Source: The Verge. This article is PickAI’s own summary and analysis based on the original reporting.
