TOKYO — NEC Corporation in Tokyo, Japan, is developing an AI foundation model for underwater acoustics that could help defense and maritime users analyze large volumes of sonar and hydrophone data without relying on extensive labeled datasets.
Japan's Agency for Defense Equipment awarded NEC the research contract through its Defense Innovation Science and Technology Institute. The project falls under the agency's Innovative Breakthrough Research program and focuses on technology with both defense and commercial applications.
NEC plans to complete the foundation model by fiscal 2027. The company wants to build a general-purpose system that can learn from large datasets and support various analysis tasks, rather than training a separate AI model for each application.
The approach applies some of the principles behind large language models (LLMs) to underwater sound. Instead of learning patterns across large collections of text, the model will learn patterns and relationships within acoustic data.
Underwater sensing creates a different AI problem
Sound plays an important role in underwater sensing because light and radio-frequency signals do not travel through water as effectively over useful distances. Sonar systems and hydrophones can instead collect acoustic signals that carry information about objects, activity and environmental conditions below the surface.
Turning those measurements into useful information takes another step. Analysts use signal-processing techniques and domain expertise to interpret acoustic data, something that can require considerable time and attention.
AI could help with that workload, but training a conventional model generally requires labeled examples that show the system what it should recognize. Producing those datasets can be particularly difficult underwater, where collecting information and determining exactly what produced a signal may require specialized equipment and expertise.
NEC plans to address that limitation with self-supervised learning, which allows an AI system to learn useful representations from data without requiring a human-generated label for every example.
Self-supervised learning reduces the need for labeled data
Rather than beginning with a model built around one narrowly defined task, NEC intends to train a broader foundation model on large amounts of underwater acoustic information. The company will draw on its experience developing sonar technology and its work on cotomi, NEC’s proprietary AI technology. Public-private partnerships will provide the acoustic datasets used for large-scale training.
NEC sees defense and commercial applications
The research contract focuses on dual-use technology. For defense applications, faster interpretation of underwater acoustic information could support a broader understanding of activity and conditions below the surface.
NEC also sees potential outside the defense sector. The company identified marine-life and resource exploration, environmental monitoring and earthquake prediction as possible applications for the technology.
A sufficiently capable model could also contribute to digital representations of the ocean by helping turn continuous acoustic measurements into information about underwater conditions and events.
NEC will work with the Defense Innovation Science and Technology Institute to develop the model and demonstrate its effectiveness through fiscal 2027. If successful, the effort could give engineers a new way to analyze growing volumes of sonar and hydrophone data across defense and commercial maritime applications.
