AI for fusion: reconstructing a plasma is not controlling one.

Two results with different roles

PPPL reported Diag2Diag in October 2025 as a way to reconstruct missing diagnostic information from other measurements. In September 2026, it reported PACMAN, a prediction-and-control framework tested in five real-world experiments. The latter is an October editorial update to this May-June coverage window. Reconstruction and control are connected tasks, but they are not the same form of evidence.

A reconstruction estimates a state or measurement. A controller uses information to choose an action. A protection system constrains operation against independently defined limits. An article that calls all three AI control hides their different acceptance tests. The reported experiments are encouraging within their stated boundaries, not a guarantee that an arbitrary model can safely manage an arbitrary reactor.

Sources: [1] PPPL: Diag2Diag diagnostic reconstruction, October 2025 · [2] PPPL: PACMAN experimental control framework, September 2026

Start with the provenance of the input

The model sees a data stream, not the physical plasma directly. Sensor calibration, timing, clipping, missing channels and preprocessing determine what reaches it. A model can perform well on a curated dataset while failing when those inputs change. The acceptance question must therefore include the whole measurement route rather than stopping at an offline accuracy score.

We would preserve a record of raw signals, transformations, model version and inferred output. An inferred value should remain distinguishable from a direct measurement even when both appear in one display. This is especially important when reconstruction is used to replace a failed channel: the system should expose that change of evidence, not conceal it behind a smooth curve.

Average accuracy can miss the event that matters

Suppose a model is accurate through most of a discharge but produces a misleading estimate during a rare disturbance. A low average error does not tell us whether that failure is acceptable. Evaluation should therefore be organised around operating situations and failure costs. The event that is rare in the training set may be the event the protection system most needs to recognise.

A practical evaluation matrix would include missing inputs, shifted calibration, previously unseen trajectories and disagreement between independent diagnostics. It would examine confidence calibration as well as point accuracy. A useful model should identify when its confidence falls, or at least operate within a clearly bounded envelope where an external mechanism determines when its output may be used.

Latency is a chain, not a benchmark number

For control, the relevant latency runs from the physical event through acquisition, preprocessing, inference, communications and actuation. A one-millisecond inference result does not establish a one-millisecond loop. Variable latency also matters: occasional long delays can alter behaviour even when the average is acceptable. A timing claim needs a distribution under the actual operating conditions.

We would measure end-to-end timing with synchronized clocks and preserve late or dropped decisions in the event log. The controller should have a defined response when an input arrives too late or a model exceeds its deadline. Reusing an old prediction without an explicit age limit can make the system appear responsive while its actions are based on stale information.

Validation should cross an evidence ladder

Offline replay is useful for finding obvious errors. Simulation allows controlled disturbances but inherits modelling assumptions. Hardware-in-the-loop tests include acquisition and actuation interfaces. Bounded experiments reveal interactions the earlier stages omitted. These stages are complementary; none should be described as automatic proof of all the others.

A serious release record identifies the model version, training provenance, validation cases, operating envelope and rollback method. It keeps optimisation goals separate from protection rules and states how the system returns to a known operating policy. A change in model weights or preprocessing is a functional change, even when the user interface looks identical.

COTT assessment

The public results support a focused role for AI in diagnostics, prediction and bounded control. Our conclusion is that the most valuable engineering advance is not simply faster inference. It is a complete evidence chain that connects input quality, uncertainty, timing and machine constraints.

This approach also makes negative results useful. A model that fails under a documented disturbance teaches the team where its envelope ends; a model whose failures are hidden behind an aggregate score does not. COTT evaluates these methods as enabling computing technologies. This article does not claim that we deployed Diag2Diag, PACMAN or a fusion machine controller.