Projects
Selected research projects in scientific machine learning, inverse problems, topology-aware modeling, and biomedical signal analysis. I am especially interested in methods that combine machine learning with physical structure, geometry, topology, or mechanistic interpretability.
Physics-Informed Image Registration for Hyperelastic Material Identification
Context. Soft biological tissue characterization often relies on digital image correlation to extract displacement fields, followed by separate constitutive model fitting. This sequential workflow can propagate errors from image tracking into material-parameter estimation, and for very soft tissues it also depends on estimating a zero-strain reference state that may be distorted by gravity-induced pre-compression.
What I built. I developed a unified physics-informed neural network framework that performs image registration and Ogden hyperelastic parameter identification within a single differentiable inverse problem. The model combines differentiable image similarity, hyperelastic equilibrium constraints, explicit gravity loading, boundary-condition penalties, and multi-step reaction-force losses.
Why it matters: instead of treating displacement extraction and material fitting as disconnected stages, the framework allows image data, mechanics, gravity, boundary conditions, and force measurements to constrain one another during optimization. This provides a proof of concept for end-to-end material characterization directly from image sequences.
Validation: generated FEM-based synthetic tissue image sequences replicating the Gao-Desai porcine liver experiment, including gravity pre-compression, tensile loading, natural texture synthesis, and imaging noise. The model recovered Ogden parameters with relative errors of 0.02% for \(\mu\) and 0.006% for \(\alpha\) on synthetic validation.
Scope. Current validation is synthetic and two-dimensional, but the framework points toward image-based inverse mechanics pipelines for experimental soft-tissue characterization.
PERSIST: Topology-Guided Multiscale Domain Identification in Spatial Transcriptomics
Context. Spatial transcriptomics preserves gene expression together with tissue architecture, but many domain-identification pipelines depend on fixed clustering resolutions, neighborhood sizes, or smoothing assumptions. This makes it difficult to distinguish intrinsic tissue organization from artifacts of a chosen spatial scale.
What I built. I developed PERSIST, an unsupervised topology-guided framework for multiscale spatial domain identification. The method constructs Alpha-complex filtrations over tissue coordinates, uses \(H_1\) persistent homology birth-time distributions to infer intrinsic geometric scales, and then builds a discrete-exterior-calculus-based multiscale stability field over the tissue.
Why it matters: PERSIST uses topology not merely as a descriptor, but as a scale-discovery mechanism. It reframes spatial transcriptomics segmentation as a topology-informed stability problem: stable tissue cores, transition regions, and high-gradient boundaries emerge from the geometry of the tissue rather than from a single clustering resolution.
Validation: applied the framework to breast cancer, human lymph node, and mouse brain spatial transcriptomics datasets. The inferred domains recovered biologically interpretable organization, including tumor microenvironment compartments, lymph-node immune zones, and mouse-brain laminar or regional anatomy, supported by marker enrichment and H&E overlays.
Methodological components: Moran’s-I filtering, PC1 scalar-field construction, Alpha-complex persistent homology, DEC-style Laplacian smoothing and discrete gradients, topology-weighted local variation, bootstrap robust-core estimation, and Morse-like basin decomposition.
Deep Learning for Full Waveform Inversion: Seismic Waveforms to Velocity Fields
Context. Full waveform inversion aims to recover subsurface velocity structure from seismic measurements. Classical FWI is a large-scale nonlinear inverse problem that can be computationally expensive, ill-posed, and vulnerable to cycle skipping, especially in complex geological settings.
What I built. I developed a data-driven FWI pipeline that learns the inverse map from multi-source seismic shot gathers to subsurface velocity models. The model uses a hybrid Swin V2 Transformer encoder and enhanced U-Net decoder, combining long-range wavefield dependency modeling with high-fidelity spatial reconstruction of layers, faults, and heterogeneous structures.
Why it matters: the project explores deep learning as a fast surrogate inverse model for wavefield-to-structure reconstruction. The architecture was designed to combine the global context modeling of self-attention with the spatial precision and inductive bias of U-Net-style decoding.
Scale and result: trained on roughly 470,000 OpenFWI Vel/Fault/Style samples using memory-mapped loading, GPU augmentations, PyTorch Distributed Data Parallel, automatic mixed precision, EMA checkpointing, and H100 GPUs. The best model achieved 63.27 m/s validation MAE and SSIM values between 0.8725 and 0.9188 across dataset families.
What was nontrivial: beyond the model architecture, the project required building a scalable training system for large scientific data: stratified dataset splitting, class-frequency weighting, Huber/gradient/total-variation losses, warm restarts, robust checkpointing, and distributed training on Grid5000.
NeurIPS 2025 EEG Foundation Challenge: Cross-Task and Cross-Subject Modeling
Context. EEG foundation models aim to learn representations that transfer across recording paradigms, prediction tasks, and individuals. The NeurIPS 2025 challenge tested this idea on the Healthy Brain Network EEG dataset through two problems: predicting active-task behavior from passive-task recordings and predicting psychopathology factors for previously unseen subjects.
What I built. I developed DDP/HPC-ready pipelines for both challenge tracks. The cross-task pipeline uses Barlow Twins self-supervised pretraining on passive EEG tasks with an EEGNeX backbone, followed by supervised transfer to contrast-change-detection response-time prediction. The cross-subject pipeline combines masked-autoencoder pretraining, a one-dimensional ConvNeXt-style encoder, subject-level multiple-instance learning, MixStyle, and domain-adversarial training for externalizing-factor prediction.
Why it matters: EEG varies strongly across subjects, developmental stages, recording tasks, and acquisition conditions. The project treats this variability as a representation-learning and domain-generalization problem rather than assuming that window-level features transfer unchanged across people and paradigms.
Technical scope: implemented subject-grouped data splits, multi-recording HBN-EEG loading and caching, distributed training, mixed precision, exponential-moving-average checkpoints, subject-bag sampling, and configurable self-supervised and supervised training stages.
Graph Signal Processing for Characterization of Multipolar Electrograms in AF
Context. Atrial fibrillation is sustained by complex spatiotemporal propagation patterns, including rotors, wavefront collision, conduction block, and re-entry. Many electrogram-analysis methods treat catheter electrodes independently, even though multipolar catheters contain spatial information through their electrode geometry.
What I built. I developed a geometry-aware graph signal processing framework for multipolar intracardiac electrograms. PentaRay catheter electrodes were modeled as graph nodes, and electrogram recordings were analyzed as graph signals over the catheter geometry using graph total variation and joint Short-Time Fourier Transform / Graph Fourier Transform analysis.
Why it matters: rotor-proximal AF activity is not just a local waveform abnormality; it is a spatial propagation phenomenon. By embedding catheter geometry into the analysis, the framework can detect spatial discordance and intermittent wavefront-collision patterns that per-channel temporal features structurally cannot see.
Proof of concept: on a small simulated OpenCARP dataset with close-vs-far rotor catheter positions, AF-band JGFT kurtosis achieved 87.5% accuracy and AUC = 1.00, outperforming conventional time-domain and phase-mapping baselines. Because the dataset contained only eight simulated recordings, the result should be interpreted as a proof of concept rather than clinical validation.
Extension idea: I also proposed edge-flow representations based on voltage, local-activation-time, and phase differences across catheter edges. This reframes electrogram analysis from isolated node signals to propagation-flow patterns, opening a route toward Helmholtz-Hodge-style separation of planar activation, rotor-like circulation, and obstacle-related harmonic flow.