✨ Page info

<aside> 📢 This page is the repository for introducing my research progress and resources for fusion plasma research in PLARE.

• Academic CV: [Jinsu Kim, CV]
• Github: https://github.com/ZINZINBIN
• E-mail : [email protected] or [email protected] </aside>

Research topics

🔎 Disruption physics

<aside> 🔥 Disruption prediction and its analysis using multimodal data in KSTAR via deep learning

• Disruption: Collapse of the plasma carrying large amount of energy loss and causing harmful damage on the device

• Non-linear dynamics → hard to predict in advance to the occurrence of the disruption

• A neural network-based predictor can detect the precursor and show high performance with various modalities of the data

  • IVIS video data: spatial-temporal information including time-varying position and shape of plasma → Computer vision model will detect the precursor
  • 0D parameters: time-series data including the state of the plasma

• Multi-modal learning can handle low precision of disruption alarms and enhance the model performance

  • ViViT model recognition enhancement via multimodal learning

    

    

  • Low precision issue handled by multimodal learning (left : ViViT, right : multimodal)

    

• Presentation resource

  Disruption prediction and its analysis using multimodal data in KSTAR via deep learning.pdf

• code : https://github.com/ZINZINBIN/research-predict-disruption </aside>

<aside> 📘 Bayesian neural networks for predicting disruption using 0D parameters and 1D profiles in KSTAR

• Bayesian neural network: the neuron is now stochastic so the value of the neuron depends on the prior and posterior distribution → computation of the uncertainty is now available.

  • BNN can allow computing epistemic uncertainty and this makes BNNs very data-efficient since they can now learn from a small dataset without overfitting.
  • Weight parameters as random variables → Marginalization of the output for disruption prediction = Ensemble of the disruption prediction output → enhancement of the model precision

• Based on BayesBackProps algorithm, we can construct the stochastic networks by approximating the MCMC sampling procedure.

• Not only 0D parameters but also 1D profiles obtained from Tompson data and magnetic diagnostic signals are used.

• Simulation results

  

  

</aside>

🔎 Reinforcement learning application on fusion plasma

<aside> ⚙ Tokamak plasmas operation control using reinforcement learning in KSTAR

• We can approximate the steady-state tokamak plasma operation by using neural networks as a function approximator (simulator)

• The virtual KSTAR environment based on Transformer model are used as an environment for reinforcement learning.

• The aim of this research is to find out the optimal way for dynamic decision-making → feasible way to approach the 0D parameters to the target value.

• The agent will learn how to control the variables to achieve the multi-objectives.

  • 0D parameter control: High beta-n
  • Shape parameter control: Elongation, Aspect ratio

• presentation resource

  PLARE_JSKim_230407_v1.pdf

• code : https://github.com/ZINZINBIN/plasma-shape-control

• Simulation results

  

</aside>

<aside> 💡 Fusion reactor design: reinforcement learning for reactor design optimization

• An inverse design problem with constraints: operation limits + cost-efficiency

• Using a single-step Proximal Policy optimization algorithm, the reactor design optimization can be handled.

• RL-based design optimization algorithm can find out the optimal configuration which satisfies both operation limits and cost-efficient solutions.

• presentation resource

  Fusion reactor design, reinforcement learning for reactor design optimization.pdf

• code: https://github.com/ZINZINBIN/Fusion-Reactor-Design-Project

</aside>

🔎 Physics-informed neural networks

<aside> 🎮 PINN-based Free boundary Grad-Sharfanov solver

• Neural network-based Grad-Shafanov equation solver
• Using Physics-informed loss, the network can satisfy the physical consistency and compensate for the value error simultaneously
• code: https://github.com/ZINZINBIN/Tokamak-GS-solver

</aside>