Research map / Label-efficient and robust learning / Domain adaptation and generalization
Nuclear Fusion Diamond Polishing Dataset
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Releases a public dataset of accelerometer vibration signals paired with surface roughness (Sa) measurements from ICF diamond capsule polishing, with MLP and domain-adaptation baselines for predicting roughness change.
Problem
Surface roughness of high-density carbon ICF capsules must be measured manually and labor-intensively during multi-stage polishing; an automated, sensor-based prediction would remove this bottleneck.
Approach
Provides the first public dataset linking raw accelerometer vibration signals from shell polishing to surface roughness changes, segmented into 6-minute samples with interpolated Sa labels, and applies unsupervised domain adaptation (DANN-style gradient reversal) to an MLP to generalize across polishing batches.
Key results
- Domain adaptation MLP reduced MAE on S233 from 0.11 nm (plain MLP) to 0.0674 nm.
- On S238, domain adaptation MAE was 0.0634 nm versus 0.205 nm for the plain MLP.
- Gradient boosting achieved 100% accuracy in classifying the five polishing batches into domains.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add domain adaptation to the MLP | MLP with 2 hidden layers · S233 | MAE: 0.11nm -> 0.0674nm | Helped |
| Add domain adaptation to the MLP | MLP with 2 hidden layers · S238 | MAE: 0.205nm -> 0.0634nm | Helped |
| Adapt on S233 before adapting on S238 | MLP with domain adaptation · S238 | MAE: 0.0634nm -> 0.0618nm | Helped |
| Use Gradient Boosting instead of the other evaluated classifiers | Polishing domain classification · Polishing batches S179, S173, S211, S233, and S238 | accuracy: 0.99 -> 1 | Helped |
Surprising findings
- Domain classification of polishing batches was perfectly separable (100% accuracy), indicating strong batch-level domain shift.
Limitations
- Sa labels for intermediate 6-minute samples are interpolated under assumed linear/logarithmic decline rather than measured, which may be inaccurate.
- Only 11 hand-crafted time-domain statistical features are used, which may not represent the raw vibration signal well.
- Data is limited to a few batches because raw vibration data collection is expensive.
Experimental setup
Datasets: Diamond Polishing Dataset (accelerometer vibration signals with surface roughness, batches S173, S179, S211, S233, S238)
Benchmarks: Nuclear Fusion Diamond Polishing Dataset
Metrics: MAE accuracy macro F1 weighted F1
Baselines: MLP without domain adaptation Gradient Boosting Decision Tree Random Forest XGBoost Logistic Regression SVM
Builds on: Unsupervised Domain Adaptation by Backpropagation
Closest papers by meaning
- Position: Opportunities Exist for Machine Learning in Magnetic Fusion Energy (ICML 2024 · similarity 0.79)
- Multi-Timescale Dynamics Model Bayesian Optimization for Plasma Stabilization in Tokamaks (ICML 2025 · similarity 0.79)
- Inverse Modeling for Laser Pulse Shape Design in Inertial Confinement Fusion (NeurIPS 2026 · similarity 0.78)
- Diff-PIC: Revolutionizing Particle-In-Cell Nuclear Fusion Simulation with Diffusion Models (ICLR 2025 · similarity 0.77)
- Thin-Shell Object Manipulations With Differentiable Physics Simulations (ICLR 2024 · similarity 0.76)
- ConStellaration: A dataset of QI-like stellarator plasma boundaries and optimization benchmarks (NeurIPS 2025 · similarity 0.75)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.