Peer-reviewed
Machine learning · Arnav Garg · 2026 · Earth Systems and Environment
Seven-Day Landslide Forecasting from PCA-Derived InSAR Data with a Random Forest Classifier
Combines PCA-reduced InSAR deformation histories with a random forest to forecast instability seven days ahead. The study asks whether satellite displacement signals can support short-horizon prediction, rather than only retrospective detection.
Machine learning · Arnav Garg, Aksh Garg, Dominique Duncan · 2025 · Electronics 14(13):2571
2.5 CNN: Leveraging 2D CNNs to Pretrain 3D Models in Low-Data Regimes for COVID-19 Diagnosis
Uses 2D CNN pretraining to give a 3D CT model a stronger starting point when labeled volumetric data are scarce. The method was developed for COVID-19 diagnosis, where full 3D training sets are relatively small.
Current projects
Manuscripts currently under peer review.
Scientific machine learning
Broadening Training Dictionaries in State Space for Composed Neural PDE Surrogates
Composed neural PDE operators fail when intermediate states leave the support each component saw during training. Expanding that state-space support cuts composition error by up to 30× and generalizes across pairs, resolutions, horizons, and architectures.
Machine learning
Weak-Form Physics-Informed Neural Networks for Noise-Robust Inverse Problems
Weak-form constraints transfer derivatives from noisy data to smooth test functions, stabilizing inverse parameter recovery on scientific imaging problems where pointwise physics-informed losses amplify noise.
Machine learning
Covert-Channel Forensics of Deceptive Language Models: When Internal-State Monitoring Adds Admissible Evidence
Internal activations add evidence when deceptive or collusive information is hidden from the transcript. When visible behavior already reveals it, white-box probes add much less beyond the behavioral evidence.
Computational science
Modified-Equation-Informed Solution-Discrepancy Signatures Detect Hidden Stabilization Changes in Convective Heat-Transfer Solvers
Small stabilization changes can leave heat-transfer outputs nearly unchanged even when the solver has changed. Modified-equation signatures reveal those hidden numerical changes directly from the computed solution fields.
Computational science
Emergence of a Deterministic Butterfly Cone in Structural Glasses
A localized perturbation spreads through a structural glass inside a deterministic butterfly cone. Its speed changes little as structural relaxation slows dramatically, separating information spreading from glassy dynamics.
Computational science
Predicting and Relocating the Shear Band in a Two-Dimensional Model Glass
The undeformed glass contains enough structural information to predict where its shear band will form. Softening a targeted region before loading can then move the eventual band to a different location.