Physics-based Dataset Augmentation

To address the need for high-quality datasets that are optimized for training Artificial Intelligence (AI) and Machine Learning (ML) models and enable rapid fielding of AI/ML capabilities for sonar systems, ARiA will develop and demonstrate the feasibility of Deep Synthetic Environmental Augmentation algorithms (DeepSEA), a suite of tools able to synthesize realistic physically constrained sonar data for augmenting existing recorded datasets to optimize dataset coverage and completeness across the global range of phenomena including bathymetric, weather, volume-scattering, and contact-density conditions. Leveraging ARiAís experience developing and specializing deep-learning models for Pulsed Active Sonar (PAS) data, DeepSEA combines state-of-the-art conditional generative models with physics-based constraints to synthesize sufficiently realistic data suitable for training both existing and emerging AI/ML models. DeepSEA will be built initially for the datasets and development communities of the Pulsed Active Sonar (PAS) and Continuous Active Sonar (CAS) Functional Segments (PASFS/CASFS) and transitioned to the AN/SQQ-89A(V)15 (SQQ-89) Program of Record (PoR).