UCSD🔱 ECE228: Machine Learning for Physical Applications (Spring 2025)

English Version | 中文版

Welcome to the official course repository for ECE228: Machine Learning for Physical Applications at UCSD, Spring 2025. This repo contains all the completed assignments and labs demonstrating fundamental concepts of ML in physical sciences, such as regression, sparsity, and deep learning for PDEs and dynamical systems.

Explore each folder for detailed implementation, report writeups, and source code.

📐 Assignment 1: Sparse Linear Regression

• Implemented linear regression and LASSO from scratch
• Developed ISTA algorithm to solve the L1-regularized problem
• Compared results with scikit-learn implementation
• Conducted experiments on synthetic and real-world datasets

🧠 Assignment 2: Transformers & Attention

• Built Transformer modules from scratch using PyTorch
• Implemented scaled dot-product attention, multi-head attention, and positional encoding
• Applied Transformer encoders to classification tasks
• Trained and evaluated models on real-world data

🌊 Assignment 3: Neural Operators

• Introduced the concept of learning mappings between function spaces
• Implemented FNO (Fourier Neural Operator) using spectral convolution
• Modeled physical systems like fluid dynamics with operator learning
• Compared generalization across resolutions and domains

🚦 Final Project – Adaptive Spatial-Temporal Transformer Network (ASTTN-Weather)

Goal: Accurate short-term traffic-flow forecasting by capturing road-network topology and exogenous weather effects.

Aspect	Implementation Highlights
Model	Extended ASTTN [paper] with: • Static adjacency from GIS road graph • Joint spatial & temporal self-attention in a single block • Multi-modal fusion of weather features (temperature, precipitation, wind) • Lightweight output head with weighted multi-horizon loss
Variants	Baseline ASTTN, ASTTN-Weather (ours), and ablations w/-wo weather or adaptive edges
Dataset	1-min loop-detector volume & speed + NOAA hourly weather, aligned to 15-min intervals
Results	AGSTTN-Weather ↓MAE by 9 % and ↑R² by 6 % over vanilla AGSTTN on 15-min horizon
Reproducibility	train.py supports CPU-only training (PyTorch 2.7.1), configurable via YAML; logs, checkpoints, and tensorboard summaries auto-saved
Visualization	Scripts for attention heat-maps and per-edge importance under different weather scenarios

Enjoy your journey into scientific machine learning! 🚀