Calibrated Reconstruction of Picosecond Kilotesla Fields from Slow Diagnostics

Lasers can now drive magnetic fields of ten thousand tesla or more inside tiny plasmas, but these fields live for about a picosecond. The detectors used to observe them average over tens of picoseconds or longer. Recovering the field from such a blurred record is an ill-posed inverse problem, meaning many different field histories produce nearly the same measurement. Image restoration has already shown what happens when such problems are solved by minimizing mean squared error: the answer is the average of all plausible solutions. Sharp peaks are flattened, and two equally likely but opposite answers cancel.

For one experiment that matters now, this failure is decisive. Simulations predict that in laser-driven microtube implosions, the amplified core field can flip direction relative to the seed field. An error-minimizing reconstruction of a shot where both directions are plausible will report a weak field near zero. That answer is physically wrong, yet it scores well on pixel-wise metrics. We propose to train a reconstruction model on plasma simulations that outputs a calibrated distribution over field histories instead of a single field. We will judge it by the quantities a physicist must decide: the field's sign, its peak strength, and its decay time. We will also audit whether the measurement, rather than the learned prior, supports each answer.

Core Claim

Let the measurement be $y = \int K(t)\, \mathcal{A}[B(\cdot,t)]\, dt + \epsilon$. Here $B(\cdot,t)$ is the magnetic field at time $t$, $\mathcal{A}$ is the physical probe operator (for example, how a field deflects a proton beam), $K(t)$ is the detector's time window, and $\epsilon$ is noise.

The estimator that minimizes squared error is the posterior mean $\mathbb{E}[B \mid y]$. Suppose the posterior puts probability $p$ on a positive core field of size $B^+$ and $1-p$ on a reversed field of size $B^-$. Then the estimate is roughly $pB^+ - (1-p)|B^-|$. When $p \approx 1/2$, this is close to zero, a value neither scenario predicts. This is the double-penalty and blur effect documented for images by Bayesian Tensor Inference for Sketch-Based Facial Photo Hallucination, which reports that a lower-RMSE method gave visibly blurrier faces. PIPAL found the same pattern at scale, with PSNR and SSIM anti-correlated with human judgment on GAN-based super-resolution. Here the damage is not a blurry texture but a wrong physical conclusion.

Picking the most probable answer is not enough either. Mode-seeking MAP estimation with diffusion priors, as in Inverse problems with diffusion models: MAP estimation via mode-seeking loss, returns one sharp field. On a near-50/50 shot it would report a confident sign that the data do not support.

Our claim has two parts. First, both kinds of reconstruction can look acceptable under field-wise error metrics while being badly wrong on sign, peak, and decay time. Second, a simulation-trained posterior sampler with an explicit, time-resolved forward model gives calibrated answers to those questions. Where the measurement truly cannot tell, it says so, and the provenance tests we describe below verify that claim.

Research Program

Training data. We will generate an ensemble of EPOCH particle-in-cell simulations (codes that track individual charged particles and the fields they produce) in two families.

The first family is microtube implosions with an applied seed field, following Generation of megatesla magnetic fields by intense-laser-driven microtube implosions and Sign reversal in magnetic field amplification by relativistic laser-driven microtube implosions. We will vary laser intensity, wall perturbation amplitude, and seed strength across the boundary between amplification and reversal. That paper motivates the work: it reports more than 40-fold amplification, identifies reversal as a sub-picosecond race between ion collapse and filament growth, and names the lack of time-resolved diagnostics as the barrier. It also constrains the work, because its results are two-dimensional only.

The second family is the four-beam twisted-pointing configuration of Generation of 10 kT axial magnetic fields using multiple conventional laser beams. It reports roughly 1 ps exponential decay, tolerance to 250 fs inter-beam delays, and about a 40% drop in peak field at 1 ps delay. Sweeping delay and polarization disorder gives ground-truth decay curves of known shape. Budget dictates the split: most runs will be 2D, with a smaller set of held-out 3D runs reserved for testing transfer across dimensionality.

Synthetic diagnostic. We will build a differentiable model of proton deflectometry, which records how a field bends a beam of protons on a stack of radiochromic film layers. This is the charged-particle forward operator that the VML-MAP paper leaves out.

The temporal-resolution literature supplies the key idea. Protons of different energies cross the target at different times, so each film layer records a different time window. Energy maps to delay, much as Spatial Phase-Sweep maps 3 mm of source displacement to 10 ps of delay to beat a 100 ps electronic limit. The width of each window, set by the energy band and the proton source duration, plays the role of the detector's response function. We will vary that width from "nearly instantaneous" to "integrates over the whole field lifetime." This controls how much timing information reaches the measurement. Deflection direction depends on the sign of the field, so sign information exists in principle. Whether it survives the time averaging is the empirical question.

Reconstruction model. We will train a conditional diffusion model, a generator that learns to turn random noise into samples by gradual denoising. It will produce space-time field histories conditioned on the film data, with the forward model enforced during sampling. We will represent the field through a vector potential, so it has zero divergence by construction (a law any magnetic field must obey). From many samples per shot we read off a sign probability, a peak-field distribution, and a decay-time distribution.

Provenance audit. We adopt the audit from Brain-Measurable Diffusion Decoding. That work showed generative priors can fabricate plausible detail with no support in the measurement. It controls for this with a mismatch test: feed the wrong scan and check that performance collapses to chance. Our versions are:

  • a mismatched-shot test;
  • a prior-only control that samples without any data;
  • a comparison of the model's sign probabilities against the Fisher information of the forward model, a calculation of how much sign information the films can carry at all.

If the model is confident about sign where the films carry no information, the confidence came from the prior, and the result is reported as a failure.

Evaluation. The metrics are decision quantities, not field-wise error:

  • sign-call AUROC (how well the model ranks reversed versus non-reversed shots);
  • Brier score and expected calibration error on sign;
  • relative bias in peak field;
  • error in decay time.

Field-wise MSE is reported alongside them so the gap between the two kinds of metric is visible directly. The image-quality literature shows this gap is real and measurable. Perceptual Fidelity Aware Mean Squared Error raises rank correlation with human scores on TID2008 from 0.7718 to 0.9162 just by filtering before comparing. We expect a similar or larger reordering of methods once physics-relevant quantities replace field-wise error.

Baselines

Naive reference. Treat the time-integrated film data as if the field were static, and invert each layer with standard regularized least squares. This is what is done when time is ignored. We expect it to underestimate peak fields by roughly the ratio of field lifetime to integration window, and it has no way to report sign uncertainty. Every method must beat it on peak-field bias and sign calibration.

Parametric Bayesian fit. This is the strongest competitor. We fit a low-dimensional physical model to the data: an exponential decay with amplitude, decay time, onset time, and a sign variable, possibly with a mixture over sign. The forward model and response kernels are the same as ours, and the fit uses Markov chain Monte Carlo sampling. It is strong because the twisted-beam fields really are close to single-exponential decays, according to the paper that reports them, and because it is calibrated by construction when its model is correct. It gets the same simulated shots, noise, and forward operator, and a matched compute budget for sampling. To beat it, our model must give clearly better sign calibration and peak-field accuracy on the microtube family, where filament structure breaks the simple shape. It must also not lose on the twisted-beam family. If it ties on both, the parametric fit wins.

MSE-trained regressor. A network of comparable size, trained on the same data to minimize field-wise squared error. This baseline isolates the averaging pathology. We expect it to look competitive on field MSE and to fail on reversal-ambiguous shots by returning near-zero core fields.

Mode-seeking MAP. The VML-MAP solver, given our forward operator and the same trained diffusion prior. This isolates "sharp but overconfident." We expect good peak estimates and poor sign calibration. The file reports VML-MAP's advantage only qualitatively, as better LPIPS and FID on ImageNet, FFHQ and CelebA restoration. No reported result was found on Lacuna for any method on plasma field reconstruction, so all baseline numbers here will come from our own runs.

Ablations. We isolate the key ingredient by retraining our sampler with a forward model that ignores timing (all layers treated as instantaneous) and, separately, without the divergence-free parameterization. If the timing-unaware version is equally calibrated, the space-to-time mapping is not doing the work.

Why not just…?

Why not just minimize MSE, perhaps with a smoothing filter as in PAMSE? Filtering helps rank image distortions against human scores. It does not change the fact that the squared-error optimum is the posterior mean. On a 50/50 sign posterior that mean is near zero whatever filter comes first. The MSE-regressor baseline tests this directly on the reversal-ambiguous subset. If it somehow calls signs correctly there, our premise is wrong.

Why not just fit a parametric decay model? This is a real threat. For twisted-beam fields we expect it to match us, and we will say so if it does. Its weakness is model misspecification. The microtube reversal involves filaments penetrating the wall before stagnation, and a few-parameter decay cannot represent that structure, so its sign probabilities may be confidently wrong. The family-by-family comparison settles the question. If the parametric fit matches on microtubes too, we will recommend it and limit the learned model to designing the measurement.

Why not just build a faster detector? Electronic detectors hit avalanche jitter of 50–100 ps, per Fluorescence lifetime Hong-Ou-Mandel sensing. Interference tricks that escape this limit need seconds of photon accumulation. Both the 7.22 ± 0.04 ps result and the 3.5 s retrieval come from that work, and seconds of accumulation are useless for single, non-repeatable shots. Proton energy-to-time mapping is already the fastest practical channel, and its window is set by the proton source, not the electronics. Better hardware narrows $K(t)$ but leaves the inverse problem in place. Our sweep over window widths shows how much hardware improvement would make inversion unnecessary.

Why not just guide reconstruction with an off-the-shelf perceptual metric such as LPIPS or a CLIP-based score? These metrics measure image texture, not physics. Human Preference Score documents that distribution-level metrics reward structurally impossible images. BiRQA reports that undefended neural metrics fall to 0.30–0.57 SROCC under small adversarial perturbations on KADID-10k. A reconstruction optimized against such a metric can satisfy it while breaking physical laws. We include an LPIPS-guided variant only as a diagnostic and expect it to fail the prior-only control.

Decisive Checks

The project succeeds if three things hold:

  1. On microtube shots whose true sign is held out, our sampler's sign calibration error is clearly lower than that of the parametric fit and of MAP.
  2. The MSE regressor's core-field estimates cluster near zero exactly on the shots where the simulated sign is ambiguous.
  3. The mismatched-shot and prior-only controls drive sign probabilities to 0.5, wherever the Fisher analysis says the films carry no sign information.

A secondary check retrains on 2D runs and tests on held-out 3D runs. A large loss of calibration would mean the 2D-only evidence behind sign reversal cannot yet support claims about real experiments.

The last check is a design result. Using the calibrated model, we will compute which film-layer energy spacing makes sign reversal distinguishable within a single shot. This gives experimenters a concrete configuration to field.