Isaac D. Gerg, Ph.D.

AI Research Scientist: physics-informed generative models for sensing and Earth systems

I have spent 20 years doing machine learning on physical sensor data, the last decade making deep learning work where data is scarce, sensors are unusual, and the physics is non-negotiable. I build the measurement model and domain structure into the network, scattering behavior, coherent-imaging constraints, and signal priors, rather than relying on scale alone. That work spans diffusion and GAN-based image restoration and synthesis, self-supervised representation learning, learned compression, and the signal processing that makes them work.

Much of that career went into one of the hardest sensing domains there is: complex-valued, low-data synthetic aperture sonar. Along the way I published 30+ peer-reviewed papers, earned two best-paper finalists, and won and led $2.5M+ of research programs as PI (DARPA, ONR, Navy), setting their technical direction. I earned my Ph.D. at Penn State in four years while working full time. My current direction is learned, physics-grounded simulators and world models built from real sensor data; ApertureLab, my end-to-end synthetic aperture sonar simulation workbench, is the working example.

Open to remote research roles in generative modeling, physics-based and domain-enriched ML, and multimodal sensing. I do my best work heads-down: developing new methods, evaluating them rigorously, and setting technical direction. US-based, fully remote.

CV Google Scholar ResearchGate GitHub LinkedIn YouTube

How I Work

No problem is unique: I reduce it to its mathematics

Before touching the existing solution base, I strip a problem down to its mathematical underpinnings. Three things follow:

Borrow the math. It has usually been worked out already in some other field, and adopting it is faster than reinventing it.
See the gap. The abstraction shows what a literature's models are missing, and that gap is where the novel work lives.
Solve from first principles. They point to solutions that iterating on an existing approach would not reach.

This is why I port ideas across communities that don't read each other: most SAS researchers come up through acoustics and rarely read the radar literature, but the underlying math doesn't care. Results:

Ported coherent autofocus ideas from the SAR literature into SAS, producing a learned autofocus that sharpens imagery in one shot (Deep Adaptive Phase Learning, IEEE JSTARS 2024).
Brought hierarchical pretraining from mainstream computer vision into low-shot sonar classification (IEEE OCEANS 2025).
Carried weather-model signal into consumer spending prediction (J. Retailing & Consumer Services 2025).

I build the proof before the argument

When I believe in a technical direction, I don't campaign for it; I build the artifact that makes the case:

Advocated GPU computing at Raytheon before it was mainstream, then built the low-SWaP CUDA demonstration on a shoestring budget that turned the argument into an investment.
Wrote ASASIN, real-time synthetic aperture sonar image formation (CUDA/C++), from scratch: 100x speedup, 10x power reduction, the lab's top engineering award out of 1,000+ engineers, and still deployed a decade later.
Built ApertureLab, a fully coherent SAS simulation and beamforming workbench, solo in two months; it now anchors my work toward world models for underwater perception.

I back claims with statistics

Many reported ML gains dissolve under honest evaluation; they turn out to be hyperparameter luck or overfitting to the test set. I design evaluations not to flatter: held-out multi-year validation, fair scoring rules, hypothesis tests, and bootstrap confidence intervals before I claim a win. My work on label noise, metamers, and dataset bias comes from the same instinct: understand how models fail before trusting how they succeed.

Research Interests

Physics-Based & Domain-Enriched ML

Building the measurement model and domain structure into the network instead of relying on scale alone: acoustic scattering, coherent-imaging and phase constraints, rendering forward models, and signal priors. This model-based approach keeps learning sample-efficient, interpretable, and robust in the noisy, low-data regimes where black-box models break down.

Are Magnitude-Only Spectrograms Sufficient? A Phase Retrieval Perspective OCEANS 2025
Deep Adaptive Phase Learning: Learned Coherent Autofocus JSTARS 2024
Resonant Scattering-Inspired Deep Networks for Munition Detection TGRS 2023
Structural Prior Driven Regularized Deep Learning for Sonar TGRS 2022
Coupling Rendering and GANs for Artificial SAS Image Generation OCEANS 2019

Earth-System AI & Earth Observation

Learning from satellite and Earth-observation data toward models of Earth systems: representation, bias, and the building blocks of physics-based world models.

Remote Sensing & AI

Applying machine learning to extract actionable intelligence from sonar, radar, and other sensor modalities.

Prompted, Not Trained: Zero-Shot Classification of SA Imagery with VLMs IGARSS 2026
Low-Shot Learning for SAS Image Classification OCEANS 2025
Exploring the Impact of Erroneous Labels on SAS Classifiers IOA 2023

Signal Processing

Integrating classical signal processing theory with modern deep learning for robust sensor data analysis.

Real-Time, Deep SAS Autofocus IGARSS 2021
Progressive Diffusion Autofocus for SAS IGARSS 2024
GPU Acceleration for SAS Image Reconstruction OCEANS 2020

Robustness & Interpretability

How deep networks represent, generalize, and fail, studied through metamers, perceptual priors, and out-of-distribution behavior, with inspiration from biological perception.