CMWF机器学习天气预报模型AIFS Single最新升级:物理约束、训练策略重构与变量扩展 A selection of new variables available from the revised AIFS Single forecasts: cloud cover (left), surface solar radiation (centre), and 100 m wind speed/direction (right). 24-hour accumulated precipitation forecasts from the AIFS (top row) and IFS (middle row) models, compared with IMERG observational data (bottom row) over northeastern Queens land for 01/02/2025 to 03/02/2025.
阶段 | 数据源 | 时长 | 步数 | LR策略 |
|---|---|---|---|---|
pre-train | ERA5 1979-2022 | 260 k | cosine 5×10⁻⁴→3×10⁻⁷ | |
rollout-finetune | IFS分析 2016-2022 | 7 900 | cosine 1.28×10⁻⁵→3×10⁻⁷ |
约束类型 | 激活函数 | 变量 | 备注 |
|---|---|---|---|
ReLUBounding | ReLU | 总降水、径流、水汽总量、比湿、太阳辐射 | 负值→0,训练阶段即生效 |
HardTanhBounding | HardTanh [0,1] | 总云量、土壤湿度 | 土壤湿度天然0-1,云量同理 |
FractionBounding | HardTanh×母变量 | 对流降水、降雪、分层云量 | 保证cp≤tp,lcc≤tcc … |
以上是对文章的详细解读,如有不当之处欢迎批评指出!也可以私信小编。