06 // The Digital Twin
LATTICE-BOLTZMANN MHD — THE MODEL THAT KEEPS THE BUBBLE ALIVE
FIG 6.0: THREE-LAYER DIGITAL TWIN — FULL LBM / ROM / SHEATH PATCHES
The digital twin is the computational model of the plasma bubble that runs inside the craft’s control processor, faster than the plasma evolves. It is not solving the full MHD equations from scratch — that would take seconds on any existing hardware, and the bubble would have collapsed in microseconds. It is a hierarchical model stack: three layers of decreasing fidelity and increasing speed, operating simultaneously.[22]
The numerical engine is the Lattice-Boltzmann method (LBM) — a mesoscopic simulation framework that recovers MHD at the macroscopic level through embarrassingly parallel local operations. Instead of discretizing the Navier-Stokes equations on a mesh, LBM tracks a particle distribution function $f_i(\mathbf{x}, t)$ on a discrete lattice. Particles stream to neighboring nodes and collide locally via the BGK relaxation operator. Through Chapman-Enskog analysis, this recovers the full MHD equations including the Lorentz force coupling, with viscosity controlled by the relaxation time. The key advantage: every node computes independently of every other node. No global solves. No matrix inversions. No iterative convergence. Pure arithmetic throughput.[23]
The MHD extension adds a second distribution function $g_i$ for the magnetic field, evolving on the same lattice. The Lorentz force coupling between field and fluid is added as a body force in the collision operator. The complete system — two coupled distributions, BGK relaxation, and Lorentz body force — recovers the full resistive MHD equations at macroscopic scales.
Layer 1: Full MHD-LBM. D3Q19 velocity set (19 discrete velocities in three dimensions). Resolution: 2 cm voxels across the full bubble volume, approximately 1.3 million nodes. Update rate: 55 kHz — one complete bubble state every 18 μs. This is slower than the Alfvén transit time but sufficient to track slow instability evolution. Runs on the dedicated MHD Processing Unit (MPU) — a systolic array processor fabricated by Aetheric Sciences on 3 nm process, 70 TFLOPS dedicated, 8 kW steady state.
Layer 2: Reduced-Order Model (ROM). Spectral decomposition of the bubble into 50–200 dominant eigenmodes. The ROM evolves modal amplitudes via coupled ODEs at 1 MHz update rate. This is the prediction engine: it propagates the bubble state 100 μs forward, identifies growing modes, and passes correction targets to the actuators. Each 1 μs cycle: Kalman state estimation (<100 ns), forward prediction via RK4, instability scan, and control allocation via constrained quadratic optimization (<500 ns). Total cycle: ~700 ns.
Layer 3: Sheath patch models. 128 high-resolution local models of the Debye sheath, each resolving a 1 mm² area at 10 μm spatial resolution. Update rate: 100 MHz. These are the eyes and hands of the THz control system — receiving boundary conditions from Layer 2 and driving the local gyrotron controllers. They run independently and in parallel, tracking sheath dynamics at the surface plasmon timescale.
In-flight calibration. The digital twin is only as accurate as its plasma model. 32 distributed Thomson scattering diagnostics sample electron temperature and density at 10 kHz. Measured profiles are compared with Layer 1 predictions. Discrepancies drive parameter updates through adjoint-based optimization, converging on a timescale of seconds — fast enough to track slow drift, not fast enough for rapid transients. The ROM handles rapid transients using the last converged parameter set with expanding uncertainty bounds.