Semaphore: A Software Platform for Operational Hosting of Coastal AI Models

Semaphore is a Python application designed to semi-automate the operationalization of AI models. Semaphore’s focus is on operationalizing coastal prediction models, such as models that predict coastal inundation, coastal fog, or water temperature. An aim of the water temperature models is to help predict sea turtle and fisheries cold stunning events. Having operational AI models predicting important coastal processes helps support and enhance coastal decision-making ahead of impactful events. The AI models are for specific locations and can be associated with threshold action levels for the likelihood of inundation or cold stunning events.

Semaphore is designed to work specifically with models that can be saved as .H5 files. Leveraging an agile design, specialized data objects, and an easily expandable data ingestion module, Semaphore supports a wide variety of data series. It employs a combination of .H5 files and Semaphore-specific data specification files to incorporate AI models, regardless of data or design. Data is standardized to a unique set of keys, enabling Semaphore to comprehend and manage any data series. This standardization facilitates seamless communication, storage, and querying of series information. Semaphore is designed with scalability and extensibility in mind, allowing it to adapt to future advancements in AI modeling with minimal effort. Semaphore also incorporates a data migration technique for its database, storing both the input data for the model and the model's predictions. This ensures the database remains organized and ready for new models without losing any historical data.

Simplifying the process of bringing research models into operation (R2O) is highly beneficial but challenging, given the rapid evolution of technologies and modeling approaches. Operationalizing these models enhances the deployment and communication of predictions to stakeholders and the public, generating valuable exposure and feedback. Semaphore is currently running 55 AI prediction models on the Semaphore production server. This number includes models created to predict water level for Virginia Key, Florida, inundation models for Aransas Pass Texas, and models to predict surge across various locations along the Texas coast and with increasing lead times. A model to predict water level for Magnolia Beach, Texas, a location with a recently installed low-cost water level sensor but with the model taking advantage of surrounding TCOON stations with decade-long datasets was also operationalized during the summer of 2024.

Rapidly transitioning to at least the prototype and providing real-time outputs to local users is critical for obtaining feedback and co-producing operational models, an essential part of convergent approaches to AI model development. Semaphore eases the burden of operationalizing models for researchers, accelerating the process, and improving the overall efficiency of model deployment. Providing an efficient pipeline to operational status not only provides more information to forecasters and the general public, but also allows researchers to test their models live, including during active events, allowing them to iteratively develop their design according to the results saved in the Semaphore database. Developing efficient AI model deployment platforms such as Semaphore is an enabling component for the co-production with stakeholders and an overall convergent approach to AI model development.