Columbia Fusion Research Center
Fusion ML research
This page tracks the arc of my work on the RMP Databases project at the Columbia Fusion Research Center, from onboarding and manual labeling in early 2025 through the current push to turn the classifier into a forecaster that can run inside the DIII-D plasma control system.
Spring 2025
Built the project's first end-to-end ML pipeline
Joined the Columbia Fusion Research Center's RMP Databases project and ramped up on the DIII-D dataset, the reviewplus and OMFIT toolchain, and the prior student work in the lab. Built the project's first labeled database from scratch by hand-tagging shots into four plasma states (Suppressed, Mitigated, ELMing, Dithering), then shipped an end-to-end training and evaluation pipeline for a Random Forest classifier.
Summer 2025
Scaled to sequence and hybrid deep-learning models
Expanded the labeled database and replaced ad hoc tuning with systematic 1D and 2D hyperparameter sweeps and forward feature selection. Migrated training to an A100 GPU and designed three new architectures in PyTorch: a CNN, a Temporal Convolutional Network, and a hybrid CNN plus BiLSTM with attention. Mapped how the four states sit in feature space using class histograms alongside PCA and t-SNE cluster plots to pinpoint where the classes overlap and where labeling itself breaks down.
Fall 2025
Hardened the pipeline with causal hygiene and label propagation
Rebuilt the labeling methodology with a biLSTM-assisted tool, eliminated causal leakage by dropping post-event variables, and reframed labels so the model could not lean on the suppressed phase itself as a predictor. Designed a k-nearest-neighbor plus hypersphere novelty detector to flag shots sitting outside the training distribution. Engineered a label-propagation pipeline that extended manual labels to every shot passing an H-mode filter, expanding the training set by orders of magnitude.
Spring 2026 (present)
Forecasters toward the DIII-D plasma control system
Pivoted from offline classification to real-time forecasting: added plasma actuators (neutral-beam power and torque, electron-cyclotron heating and current drive, shaping, fueling, RMP amplitude) at time t + δt as inputs so the model could anticipate the transition from suppressed to ELMing before it happens. Slimmed the forecaster to close the train-versus-validation gap, ran experiments across 2, 3, and 4 state formulations, and started prototyping LSVM variants. Curating a permanent, hand-labeled test set inside the synthetic-shot range so evaluation stays honest. The target is a forecaster fast and stable enough to run inside the DIII-D Plasma Control System in real time.
Related project writeups: four-state classification, Random Forest search, and RMP threshold modeling.