Plasma Control: PINN + RL

PINN-constrained PPO control agent for BOUT++ plasma simulation

Overview

A reinforcement-learning control agent (PPO) constrained by a physics-informed neural network, with Clifford geometric-algebra layers and a chalcogenide memristor reservoir computing backend, controlling BOUT++ plasma simulations in real time. Source published on GitHub; PyPI packaging in progress.

Architecture

• PPO policy network as the control agent
• Physics-informed neural network layer enforcing plasma-physics constraints on the policy's action space
• Clifford geometric-algebra layers carrying state as multivectors rather than flat vectors
• Chalcogenide memristor crossbar as a physical reservoir computing substrate, supplying nonlinear temporal dynamics in hardware
• BOUT++ as the plasma simulation environment being controlled in real time
• Source published on GitHub as tardigrade_agent; PyPI packaging in progress

Milestones

1. 01Formal benchmark: reservoir-in-hardware vs. simulated-reservoir performance (shared open question with the memristor hardware program)
2. 02Documentation and cleanup pass on tardigrade_agent for wider adoption
3. 03Recorded demo / notebook walkthrough for outreach and applications
4. 04Write up as a standalone paper or portfolio piece for national-lab and fellowship applications

Open questions

• Device-to-device variability across the memristor array (cross-ref hardware/memristor-reservoir)
• Which specific plasma-physics constraints the PINN layer enforces — needs a precise statement for the writeup
• Sim-to-real transfer path if ever run against a physical plasma device