In-Theatre Probabilistic Forecasting with Deep Learning-Powered Edge Computing

Numerical weather prediction (NWP) is executed in two processes: data assimilation (DA) and forecasting. Weather forecast models must be fed by analyses derived from DA. DA is a process in which many historic forecasts within an assimilation window of time are sequenced to inform a ōbest guessö of the current state of the atmosphere. This sequencing requires complex modeling and computation to fill in gaps within the data, or space between weather sensors. Depending on the span of the assimilation window, atmospheric data may be run through various weather models tens to hundreds of times to generate a single DA analysis which can then be fed into forecasting models to create a weather forecast. The requirements for this type of computation surpass the bandwidth and equipment capabilities that are available to the warfighter in operation. In other words, operational forecasts cannot be locally updated and accurate weather forecasting data must be relayed from command centers to the field. Recent advancements in artificial intelligence, machine learning, and deep learning have dramatically shifted the world in which the DAF operates and the realm of what is possible. We have arrived at a point where we can use deep learning to execute hundreds of forecasting model runs in the same time and using the same computational power as a single model run on a legacy system. China is currently at the forefront of AI development and weather modeling, and already HUAWEI CLOUDÆs Pangu-Weather AI model enables the second step of NWP,Āforecasting, at a 10,000x improvement over legacy speeds. This technology is already in the hands of our adversaries, and it is imperative that the DAF seek to procure similar or better solutions. WindBorne Systems is developing a capability which will apply deep learning AI to not only forecasting, but also to DA. Our engineers are already building a prototype which will ensure that the DAF can maintain a superior environmental advantage over our adversaries while strengthening the operative capabilities of the warfighter. Through SBIR phases I, II, and III, we will build an edge computing system powered by deep learning that can enable the warfighter to run advanced NWP, from DA to forecasting, within minutes for the purposes of mission planning, mission readiness, go/no-go decision-making, and augmenting operational execution. It is this type of technical advancement that will ensure that DAF remains dominant in environmental and battlespace awareness in the age of AI. Ā