Deon Fernando

Deon Fernando

Biography

Deon Fernando is a rising junior majoring in Physics and Mathematics at The University
of Alabama. In Dr. Jeremy Bailin’s galaxy formation group, he recognized a major computational
bottleneck in circumgalactic medium analysis and built a denoising autoencoder to improve
MCMC analysis speed, conditioning its latent space on the underlying physics. At the Alabama
Water Institute, he developed a framework to detect when hydrology surrogate models are
operating outside their learned experience, work he presented at the 2025 American Geophysical
Union conference and is now developing into a publication. This summer, as an MSRP intern with
Dr. Phillip Harris and Dr. Christina Reissel, Deon is building a contrastive Gaussian embedding
framework that uses negative log-likelihood in latent space for statistical inference and anomaly
detection in gravitational waves. He aspires to pursue a PhD in computational physics, focusing on
building tools that help discoveries in physics move faster and scale further.

Witness-Augmented Contrastive Learning for Gravitational Wave
Transient Classification
Deon Fernando1,2, Christina Reissel2 and Philip Harris2
1Department of Physics and Astronomy, The University of Alabama
2Department of Physics, Massachusetts Institute of Technology

Gravitational waves have opened a new window into our understanding of the universe, and
their detection relies on advanced laser interferometry to measure minute ripples in spacetime.
However, these signals are frequently masked by environmental and instrumental transients,
known as “glitches,” which complicate signal identification. Current gravitational-wave
astronomy pipelines mitigate this noise by analyzing over 200,000 auxiliary detector channels;
yet, this rich diagnostic information is traditionally restricted to post-detection veto analysis.
While modern neural networks have revolutionized rapid signal classification, they typically
rely on primary strain data alone, leaving this extensive auxiliary metadata unused during the
critical real-time detection window. We bridge this gap by introducing a multi-modal jointembedding
framework with dual-encoder architecture that integrates auxiliary witness
channels directly into the classification process. Our architecture utilizes state-space models
(Mamba) regularized by a Supervised Sketched Isotopic Gaussian Regularization
(SuperSIGReg) framework. This method enforces a geometrically structured embedding space,
enabling robust out-of-distribution detection and statistical inference. We demonstrate that
incorporating auxiliary channels improves the network’s ability to distinguish true
gravitational-wave signals from noise. We expect to integrate this framework into existing
online parameter estimation and anomaly detection pipelines to provide a robust foundation for
real-time gravitational-wave inference.