Operationalizing ML-Based S2S Forecasts at Salient

Salient is a private-sector startup focused on providing maximally skillful and usable sub-seasonal to seasonal (S2S) weather forecasts. Beginning with an initial seed of academic research, the team shifted to operational forecast products and a commercial model, and now has nearly 4 years of experience serving customers.

We describe the journey, challenges, and lessons learned in building an ML-based product in this domain. For S2S applications, ML approaches have many of the same benefits (low cost of inference, flexible model outputs, conduciveness to probabilistic approaches) and challenges (complex data pipelines, high training cost, "black box" properties) as in other weather domains.

There are also unique considerations in the science of S2S forecasting, and the logistics of delivering the associated data and insights. Operational uses of S2S forecasts are less mature than short and medium-range forecasts, primarily due to low accuracy of existing models. This means that there is both pre-existing skepticism about the value of forecasts, and hunger for new solutions. ML approaches offer a direction for real improvement, and user trust can be built through transparency and rigorous validation frameworks just as in other domains. S2S information is also fundamentally probabilistic, requiring sophisticated application layers to help users make decisions from this complex data that create real long-term value.

Ultimately, operationalizing a new model presents myriad challenges beyond any initial proof of concept in a research context. Because of this cost, development of ML solutions is best targeted in areas where existing NWP products have significant shortcomings. Fortunately, ML models lend themselves well to cloud-native compute frameworks and mature open-source packages, and we discuss approaches used to minimize initial development costs with these tools.