ApertureLab is not a first attempt at this problem; it is the current state of a line of work that runs from acoustic wave scattering through to the models that read the resulting imagery. The record below is 29 peer-reviewed publications across 14 years, and it covers the whole chain rather than one stage of it.
ApertureLab's physics engine is grounded in peer-reviewed sonar science: GPU-accelerated TDBP beamforming, precision interpolation kernels for sub-wavelength micronavigation, and validated sub-bottom acoustic simulation.
GPU Acceleration for Synthetic Aperture Sonar Image Reconstruction
GPU-accelerated TDBP image reconstruction suite enabling real-time SAS imaging on unmanned underwater vehicles.
Read paper →
Interpolation Kernels for Synthetic Aperture Sonar Along-Track Motion Estimation
Kernel design for sub-wavelength along-track displacement estimation underpinning ApertureLab's micronavigation pipeline.
Read paper →
Simulation and Testing Results for a Sub-Bottom Imaging Sonar
Validated simulation of sub-bottom acoustic scattering; the same physics model that underlies ApertureLab's seafloor rendering.
Read paper →Physics-accurate synthetic data is the missing ingredient for underwater AI; ApertureLab generates it. The research arc below runs from synthesizing scarce training data with a physics-coupled GAN, to few-shot classifiers trained on synthetic augmentation, to zero-shot VLMs that classify sonar imagery with only a text prompt.
Coupling Rendering and Generative Adversarial Networks for Artificial SAS Image Generation
A physics renderer coupled with a GAN produces realistic SAS target imagery when labeled field data is too scarce to train from alone.
Read paper →
Low-Shot Learning for Synthetic Aperture Sonar Image Classification Using Hierarchical Pretraining & AirSAS
Hierarchical pretraining on AirSAS synthetic data reduces labeled-data requirements to a handful of real examples per class.
Prompted, Not Trained: On Zero-Shot Classification of Synthetic Aperture Imagery with Vision-Language Models
VLMs classify SAS targets at 0.946 AUC using only text prompts describing highlight-shadow geometry; zero domain-specific training required.
To appear.The cards above are highlights. The complete record is 29 publications across 14 years, and it spans the entire chain: physics-based simulation and synthetic data, micronavigation, GPU image formation, learned autofocus, compression, perceptual image quality, and the models that read the imagery. The diagram below places each paper on the stage of the processing chain it advances.
29 publications, 2012 to 2026, mapped onto the SAS processing chain; each line ties a paper to the stage it advances.