Quantum Machine Learning: A Skeptical View
Quantum machine learning is heavily hyped and thinly proven; Kronos treats it as a research curiosity, not a tool for the digital twin or control.
What QML claims and what it shows
Quantum machine learning (QML) proposes to use quantum circuits as models, quantum kernels, variational classifiers, quantum neural networks, for learning tasks. It attracts strong claims. The honest state of the field is that, for classical data of the kind Kronos actually has (plasma telemetry, diagnostics, simulation outputs), no QML method has shown a robust advantage over classical machine learning.
# Quantum kernel method:
K(x, x') = |<phi(x) | phi(x')>|^2
# phi(x) = quantum feature map (data-encoding circuit)
# used inside a classical SVM. Advantage requires a feature map that is
# (a) hard to compute classically AND (b) actually useful for the task.
# For generic classical data, (a) and (b) rarely hold together.The core obstacles
- Data loading: encoding classical vectors into quantum states can cost as much as the computation itself.
- Barren plateaus: variational QML circuits suffer the same vanishing-gradient problem as VQE ansatze.
- Dequantization: several proposed QML speedups were later matched by classical randomized algorithms.
Where Kronos stands
All production machine learning for the Kronos machines, disruption precursors, surrogates, anomaly detection, digital-twin models, is classical, running in L0 and compiled toward the edge. These methods are mature, verifiable, and already good. There is no identified Kronos ML problem where a quantum model is expected to help.
QML stays a monitored research topic, tracked for the specific case where the data itself is quantum, for example feeding quantum-sensor outputs or quantum-simulation results into a learning step, which is the one setting where a quantum model has a principled reason to help. Absent that, we default to classical and say so. This restraint is the same candor applied throughout the quantum program.