Multi-Modal Sensor Fusion

The fabric fuses complementary diagnostics into a single state estimate, so the whole constellation constrains one coherent picture with honest uncertainty.

Why fuse

Each diagnostic family is strong in one regime and blind in another. Interferometry is fast but line-integrated; Thomson is local but pulsed; Mirnov is fast but only magnetic; flux loops are absolute but slow. Fusing them yields a state estimate better than any single sensor and, crucially, an uncertainty on that estimate the twin can weight.

A Bayesian frame

Fusion is naturally Bayesian: each diagnostic contributes a likelihood over the plasma state given its measurement, and the fabric combines them with the physics prior encoded by the equilibrium. The result is a posterior estimate with covariance. In fast paths, this is approximated by a calibrated weighted combination whose weights come from each channel's quality score and known variance.

# inverse-variance fusion of two density estimates
def fuse(x1, var1, x2, var2):
    w1, w2 = 1/var1, 1/var2
    x = (w1*x1 + w2*x2) / (w1 + w2)
    var = 1.0 / (w1 + w2)          # fused uncertainty
    return x, var

# Thomson (local, noisier) with interferometry (robust)
fuse(1.02e20, 4e37, 1.00e20, 1e37)

Robustness

• A dead channel drops out of the fusion cleanly, its weight going to zero.
• Disagreement between modalities raises the fused uncertainty and triggers validation.
• Imputed values enter with reduced weight, reflecting their lower quality score.

Both machines

The fusion framework is machine-agnostic: it combines whatever diagnostics a machine carries into the coordinate frame the twin uses. For the breeder it centers on equilibrium and profiles; for the burner on density, potential, and DEC-train state. The physics priors differ; the fusion machinery is shared.