Research + Applied Projects

Tricritical Directed Percolation in Transitional Turbulence

Research on how pipe flow transitions from laminar to turbulent when body forces are applied. The work identifies a tricritical directed-percolation point that enriches the phase diagram of transitional turbulence.

Result: published in Physical Review Letters 135, 104001 (2025), with UC San Diego news coverage. The project connects fluid mechanics with non-equilibrium phase transitions and universal scaling near turbulent onset.

• Physics: laminar-turbulent transition, body-forced shear flow, directed percolation, tricritical scaling.
• Methods: finite-size scaling, stochastic modeling, phase-transition phenomenology, statistical inference.
• Why it matters: gives a compact universality-class explanation for transition behavior that previously looked inconsistent across forcing regimes.

May's Complexity-Stability Hypothesis in Neural Networks

Project asking whether optimized neural networks behave like random complex systems, or whether training produces stabilizing structure analogous to evolved ecological networks.

Goal: test whether gradient descent and selection pressure can move high-dimensional interaction systems away from the random-matrix instability predicted by May's theorem.

• Physics/math: random matrix stability, optimized interaction systems, complex networks, dynamical stability.
• Methods: spectral analysis, trained-network diagnostics, interaction-structure statistics, comparison to ecological stability mechanisms.
• Why it matters: reframes neural networks as optimized complex systems whose stability may be governed by statistical-physics structure, not only architecture or loss curves.

Universality Class of Stratified Shear-Flow Transitions

Project on whether stable density stratification is a relevant perturbation to directed-percolation universality at turbulent onset in shear flows.

Goal: use finite-size scaling and Binder cumulants to separate finite-domain artifacts from genuine stratification-induced changes to critical behavior.

• Physics: stratified Waleffe flow, geophysical shear turbulence, laminar-turbulent onset, directed-percolation scaling.
• Methods: large-domain simulation constraints, finite-size scaling, Binder cumulants, universality-class diagnostics.
• Why it matters: connects fundamental turbulence transition theory to stratified flows relevant to oceans, thermoclines, and atmospheric boundary layers.

Structural Health Monitoring from Sensor Time Series

End-to-end pipeline on a 34.8M-entry multi-sensor dataset to detect structural damage in buildings. Multivariate LSTM predicts 4-floor acceleration response from excitation input; prediction residuals serve as damage signal.

Result: test MSE 0.00278; on heavily damaged state error rose to MSE 0.0347 (~12.5× separation), enabling clean anomaly classification. Welch PSD / FFT feature engineering + multinomial classifier achieved 85.29% accuracy across 17 structural states.

• Physics/data: vibration response, excitation forcing, and frequency-domain damage signatures.
• ML role: forecast healthy dynamics, then use residuals and spectral features as damage signals.
• Analysis: residual-based anomaly detection and spectral classification separate healthy and damaged structural states.

Graph-RNN for High-Dimensional ECG Forecasting

NSF HDR ML hackathon ($4k prize pool). Forecasts ECG signals from multi-channel neural activity. 3-layer GRU + graph convolutions with scheduled sampling; tackles out-of-distribution generalization from session-to-session distribution shift.

Result: in-distribution MSE 32,925 with R² = 0.80; beat ARIMA/ETS baselines; currently 17th on global leaderboard.

• Physics/data: high-dimensional coupled physiological time series with session-level distribution shift.
• ML role: recurrent forecasting with graph structure and scheduled sampling for stable rollouts.
• Evaluation: baseline comparison, leaderboard-style benchmarking, and tests of session-to-session generalization.

Physics-Informed ML for Laser-Plasma Accelerators

AI/ML Intern at TAU Systems (Carlsbad). Surrogate modeling and physics-informed prediction for electron-bunch properties from laser/plasma parameters. Sanitized writeup will follow after the internship.

• Physics/data: laser-plasma accelerator inputs and electron-bunch output properties.
• ML role: physics-informed prediction and surrogate modeling for expensive experimental/simulation regimes.
• Focus: scientific ML at the interface of accelerator physics, surrogate modeling, and experimental tuning.

Golden Handcuffs: Mortgage Rate Lock-In and Housing Supply

A step outside physics: applying the same modeling discipline to a macroeconomic question. Automated pipeline quantifying how mortgage rate lock-in affects housing supply, pulling 24 FRED macroeconomic series into a 67-variable national/state-level panel dataset.

Result: XGBoost outperformed SARIMAX by 30%+ on national RMSE and 70%+ at the state level. Lock-in effects are amplified by weak labor markets, while industry structure shapes listing volumes nonlinearly in ways tree-based models detect but SARIMAX misses.

• Data: FRED API macro series, national housing indicators, and state-level economic panels.
• ML role: PCA, LASSO, and K-means to identify latent macro structure; XGBoost and SHAP for nonlinear forecasting and interpretability.
• Modeling focus: automated data ingestion, panel-data feature engineering, econometric baselines, model comparison, and interpretable ML.