A million dollars per fusion shot: a diffusion language model cuts neutron-peak timing error from 11.6 to 9.2 steps

The neutron-rate waveform is no longer written out as 512 equivalent numbers from left to right; it is first written as the three quantities fusion physicists read off: total yield, the time the peak occurs, and the local waveform around that peak. Earlier autoregressive language models and deterministic regressors emit the whole waveform directly and commit to a peak location before the surrounding waveform exists. On the 50-shot held-out experimental test set of ICFBench, ICF-DLM reaches a peak-timing error of 9.2 steps, where the error is counted in time steps and lower is better; an autoregressive LLaMA-3-8B under the same SFT plus PPO pipeline reaches 11.6 steps.

The model can revise a peak position it has already written during generation instead of fixing it once. Autoregressive decoding fixes the peak first and fills in the surrounding waveform afterwards, so small early deviations are locked in by later generation. Tracing denoising trajectories on the 50 test shots, 60% of shots show some peak revision, and revisions lower peak-timing error by 2.48 steps on average; a control that commits the peak at step 0 and freezes it for the remaining 127 steps raises the error by 2.48 steps.

The second training stage scores predictions with the physical quantities themselves, restoring the numeric proximity that a discrete number vocabulary discards. Token-level cross-entropy treats adjacent numeric tokens as fully distinct classes, which mismatches the error scales of yield and peak timing. On the test set, removing the waveform reward term raises MSE from 0.064 to 0.360 and yield error from 35% to 186%; removing the magnitude reward term raises peak-timing error from 9.2 to 17.8 steps.

The demonstrated scope is a single fusion facility and a single diagnostic, the neutron-rate waveform, with 232 quality-controlled experimental shots split into 182 training and 50 held-out test shots; the authors leave generalization to other diagnostics such as X-ray imaging and neutron time-of-flight, and to facilities with different driver geometries, to future work.

The authors state the recipe has not been empirically validated on domains such as tokamak disruption prediction, supernova light curves or shock-induced phase transitions; diffusion generation needs multiple forward passes per prediction, so per-sample latency exceeds single-pass predictors and real-time settings would need distillation or few-step variants. Prediction quality degrades in the tails of the yield distribution, with the largest relative error on shots with yield below 0.5 or above 5.0, both thinly covered by the training data.