Production-grade machine learning systems bridging computational physics and artificial intelligence. Interactive demos, open-source code, and deployable models.
Each project includes source code, interactive demos, and documentation.
Physics-Informed Deep Learning for Topological Phononic Crystal Design
An end-to-end ML pipeline that predicts topological bandgap properties, edge-state frequencies, and wave propagation directionality in 2D phononic crystals using physics-informed neural networks (PINNs) that enforce the acoustic wave equation during training. Reduces design iteration time from hours to milliseconds while maintaining <3% error vs. full-physics simulations.
Real-Time Acoustic Metamaterial Simulator with Physics-Informed Neural Surrogates
A production-grade Python framework for acoustic metamaterial design combining rigorous Plane Wave Expansion (PWE) solvers, 2D FDTD wave propagation, and SSH topological models with a physics-informed neural surrogate that achieves 100× speedup. Includes 80+ tests, CI/CD, Docker support, and an interactive Streamlit app for real-time design exploration.
Conditional GAN for Acoustic Metamaterial Inverse Design
Solves the acoustic metamaterial inverse design problem: given a target absorption or transmission spectrum, generates the unit cell geometry that produces it. Combines a conditional GAN with a differentiable Transfer Matrix Method (TMM) forward solver implementing Helmholtz resonators, quarter-wavelength tubes, and micro-perforated panels with physics-augmented discriminator loss for guaranteed physical consistency.
Deep Reinforcement Learning for Aerodynamic Shape Optimization
Trains a PPO agent to iteratively modify airfoil shapes to maximize aerodynamic efficiency (L/D ratio) while satisfying geometric constraints. The agent operates in the CST (Class-Shape Transformation) parameter space — a Bernstein polynomial basis that guarantees smooth, valid airfoils — evaluated using a Hess-Smith panel method with Kutta condition and Schlichting skin friction.
Neural Radiance Fields for 3D Fluid Flow Visualization
Adapts the exact NeRF architecture (Mildenhall et al., ECCV 2020) for scientific volume rendering of 3D fluid flow data. Training data generated from exact analytical solutions of Navier-Stokes/Euler equations. Supports novel view synthesis, Q-criterion vortex identification, and Beer-Lambert volume rendering — replacing GB-scale CFD volume data with a compact neural representation.
ML-Driven Bandgap Prediction for Rhombic Phononic Crystals
Predicts phononic bandgap presence and width in rhombic unit cells with cylindrical air hole inclusions using gradient boosting regression and random forest classification trained on 3,000 COMSOL FEA band structure simulations. Bandgaps extracted directly from raw eigenfrequency data using a density-of-states approach with band-major data reshaping.