Research vision
How can plasma diagnostics help us infer uncertain physical parameters and design more energy-efficient discharges? My proposed research connects differentiable simulation with parameter inference and neural surrogate models, with an immediate focus on semiconductor plasma processing and longer-term relevance to preionization for fusion startup.
This page describes planned work. The solver, inference pipeline, and design results outlined below are research objectives.
Three connected directions
1. Build a verified differentiable solver
Develop a one-dimensional discharge solver in JAX, coupling drift-diffusion transport, Poisson electrostatics, electrode emission, and an external circuit. PASCHEN-1D benchmarks would provide a reference. Mesh and timestep refinement, finite-difference gradient checks, and GPU runtime and memory measurements would establish numerical reliability before optimization.
2. Learn physical parameters from measurements
Use discharge-current and gap-voltage traces to estimate effective secondary-electron emission and an ionization-rate correction. Synthetic observations would test parameter recovery; experimental measurements would support calibration and validation. Sensitivity analysis would help identify which operating conditions distinguish parameters that produce similar signals.
3. Design energy-efficient plasma generation
Optimize driving-voltage waveforms and gas pressure, and compare characterized electrode surfaces, to reduce source energy while reaching a target electron density. Neural surrogates, including a Fourier neural operator developed with NVIDIA PhysicsNeMo, would screen candidate designs. The verified discharge solver would then check the selected candidates and their operating constraints.
Foundations and planned outputs
This direction builds on my contributions to PASCHEN-1D and my studies of short-pulse electron transport and emission. Independent numerical comparisons and uncertainty analysis are central to the proposed approach.
Planned outputs include reproducible solver and gradient benchmarks, parameter-recovery studies, and verified design comparisons. My currently available codes and results cover particle-in-cell and multiple-sheet examples.
A small working example
The Langmuir oscillations example demonstrates density inference through a differentiable JAX fluid solver, with analytical and gradient checks. It is a compact educational example; the discharge solver and broader research aims above remain proposed work.