Counterfactual Field Knockouts as the Verifier for Closed-Loop Laser Electron Acceleration

When an intense laser pushes through a low-density plasma, it can drive electrons to hundreds of MeV by "direct laser acceleration" (DLA): the laser carves out a positively charged channel, electrons oscillate across it, and when that oscillation stays in step with the laser field, energy pours into them. Whether a given shot works depends on a tangle of competing electric fields that the plasma generates itself. Some help. A short longitudinal field can reset an electron's phase so the laser keeps pushing it. Others hurt. A retarding "sheath" field at the channel wall drains energy when the focus is too wide. Today these fields are diagnosed after the fact and tuned around by hand, or the whole problem is handed to a black-box optimizer that sees only one number per shot. This note proposes a closed search loop that runs the plasma simulation, deletes individual field components in silico to measure what each one causally contributed to the fastest electrons, and uses that diagnosis to decide which laser and target parameters to try next.

The idea borrows from two places. The first is the electrode short-circuit experiment on a spherical tokamak, which proved an internal electric field was doing real work by switching it off. The second is retrieval-augmented question answering, where revising a query after critiquing a first draft answer beats rewriting it blind. The claim to be tested is specific. In DLA, attributing a shot's outcome by counterfactual field knockout gives a better search signal than either a scalar figure of merit or the standard accounting of work done by each field. The reason is that the most important helpful field does little direct work. It acts as a gate.

Core Claim

Robinson et al. showed that a brief longitudinal electric field met near a zero of the laser's vector potential lowers the dephasing rate $R = \gamma - p_x/(m_e c)$. Here $\gamma$ is the electron's Lorentz factor and $p_x$ its forward momentum. Lowering $R$ lets the laser's magnetic force push electrons past three times the free-electron limit. Their PIC simulations attribute over 30% of the forward momentum to that laser force rather than to the longitudinal field itself.

This motivates the proposal and constrains it. If the decisive field contributes little $\int q\,\mathbf{E}\cdot\mathbf{v}\,dt$ (the standard per-field work tally available from any particle-in-cell code), then work accounting will systematically undervalue it. A search loop guided by work accounting will happily trade it away. A counterfactual answers a different question: "how much energy would the top electrons have had if this field component in this region were absent?" That question can see the gate.

Tang et al. found the opposite kind of field. Sheath fields at the channel boundary decelerate electrons at a 16 µm focus, while a matched 8 µm focus reaches beyond 120 times the ponderomotive energy $U_p$. The research file's central tension is that blanket suppression of electrostatic fields is counterproductive. It is the same lesson Inomoto et al. found when suppressing the in-plane field by more than 50% delayed the plasma's structural transition. The target, therefore, is selective: suppress the retarding transverse sheath while preserving or enhancing the axial anti-dephasing field. Our loop searches explicitly for that.

Research Program

The physical search space contains only knobs that could be set on a real facility. These are focal spot size, focal plane position, peak density and longitudinal density ramp length, pulse duration at fixed energy, and an optional applied axial or transverse magnetic field. Applied fields at kilotesla scale are demonstrated in Weichman et al. to reshape surface fields and electron trapping. The density ramp is included because Bhakta et al. showed in a test-electron model that a slowly rising azimuthal magnetic field along the axis breaks the reversibility of DLA energy exchange. A density gradient is the practical way to produce such a rise. The field knockouts are never search knobs. They are purely a diagnostic, the in-silico analogue of Inomoto's switched electrodes. This sidesteps the fact that physical electrodes cannot survive relativistic intensities.

Each iteration has three steps, mirroring the Score–Judge–Revise structure of CoV-RAG. There, a draft answer is scored, judged reliable or not, and a targeted follow-up query is written. The authors report that this post-draft revision beats revising queries before retrieval.

In the Score step, a 2D PIC run (OSIRIS or EPOCH) is executed at the proposed parameters. Fields are recorded on a coarse spatiotemporal grid, and the top-energy electrons are tracked.

In the Judge step, those tracked electrons are re-pushed through the recorded fields several times. Each time, one field component in one region is zeroed or scaled. The regions are the transverse field inside the channel wall, the longitudinal field in the channel core, the quasi-static azimuthal magnetic field, and the boundary sheath. The result is a counterfactual attribution vector: the change in top-decile energy and in a dephasing summary when each field is removed. Re-pushing test particles is cheap compared with a full simulation, but it ignores the plasma's response to the knockout. A subset of iterations therefore also runs full self-consistent knockout simulations to measure how faithful the cheap version is.

In the Revise step, a surrogate model predicts the attribution vector, not just the final energy, as a function of the knobs. The next query maximizes predicted energy subject to a learned coupling, for example "sheath loss grows with spot size times density" or "the anti-dephasing contribution collapses below a ramp length". In practice this means multi-output Bayesian optimization whose acquisition step uses the predicted mechanism breakdown. It can, for instance, propose moves that reduce predicted sheath loss without reducing predicted longitudinal-field benefit.

A second deliverable falls out of the loop: a map of which field dominates the outcome across parameter space. That map includes where the sheath stops mattering and where the gate turns on, and it is a physics result in its own right.

Baselines

The naive reference is the configuration the literature already identifies: Tang et al.'s matched 8 µm focus in their helium gas-jet setup. They report energies exceeding $120\times U_p$ in OMEGA EP experiments and OSIRIS 2D simulations. We reproduce that simulation setup and also the Willingale et al. OMEGA EP setup, which reported $(505 \pm 75)$ MeV and $(140 \pm 30)$ nC at $0.028\,n_{crit}$. Every method starts from these points. Beating them means exceeding the reproduced simulated top-decile energy at equal laser energy.

The strongest existing method is scalar Gaussian-process Bayesian optimization with expected improvement, as in Sha et al.. They found it roughly 3× more efficient than uniform scanning over a 4D laser-target space, with 83.3% of uniform-scan simulations wasted. We run it over the same knobs, the same simulation code and resolution, and the same simulation budget, with the same hyperparameter-tuning effort. Its objective is top-decile energy, or energy times charge above a threshold. We also run a contextual variant in the spirit of DynaBO, which uses a fast learned dynamics prior plus a slow Gaussian process. No reported result on DLA was found on Lacuna for either method, so the comparison numbers will be our own.

Two ablations isolate the key ingredient. The first is identical to our loop but replaces counterfactual attribution with the per-field work tally from the same tracked particles. This tests the specific claim that causal attribution beats energy accounting. The second keeps our multi-output structure but feeds it shuffled attribution vectors. This checks whether gains come from the extra outputs rather than their content.

Beating the strongest baseline means one of two things. Either we reach the best scalar-BO energy in significantly fewer simulations, measured across at least five random seeds, or we reach a significantly higher final energy at equal budget. Beating the work-accounting ablation is the result that justifies the proposal.

Why not just…?

Why not just pick the right spot size, as Tang et al. did? Spot matching is the strongest passive answer, and in the nominal setup it may be enough. Its weakness is that it was established with a three-point scan (5, 8, 16 µm). It also leaves sheath deceleration in place whenever power, density, or pointing push the system off the matched point. We test this directly. One experimental arm forces the system off-optimum (fixed wide focus, or constrained density) and asks whether the loop finds other knobs that recover the lost energy. If it cannot, spot matching wins and we say so.

Why not just run scalar Bayesian optimization with more simulations? This is the real threat. In a low-dimensional, smooth space, scalar BO may match us, and extra diagnostics could add nothing. Our expected advantage appears only as dimension grows and as the landscape contains regime changes, such as the onset of the gate field. The budget-matched comparison across 4, 6, and 8 knobs settles this. A null result there is a publishable finding that mechanism-aware search is unnecessary for DLA.

Why not just use the work done by each field, which PIC codes already output? This is cheaper and requires no counterfactual runs. Robinson et al.'s finding is the reason to doubt it. A gating field can be causally essential while doing little direct work. The work-accounting ablation and the faithfulness check against full knockouts decide this. If the work tally ranks fields the same way as the knockouts across the explored space, we drop knockouts.

Why not just ask a language model to propose the next simulation? The query-refinement literature motivates the Score–Judge–Revise structure, not language as the medium. The quantities here are continuous fields and spectra, and a numeric surrogate handles them directly. We use no language model, and claim nothing about one.

Proposed Experiments

All experiments use 2D PIC. The 2D-versus-3D gap is a stated limitation, and a final best configuration is spot-checked in 3D at reduced length. We report four things: top-decile electron energy in units of $U_p$, charge above 100 MeV, divergence, and simulations-to-target.

The first decisive check is faithfulness. On about 30 configurations spanning the space, compare the cheap test-particle knockout attribution with full self-consistent knockout runs. Rank agreement of field importance must be high for the loop's premise to hold.

The second is the head-to-head. Our loop is compared against the naive baseline, scalar BO, contextual BO, and both ablations, at equal budget, across seeds, at increasing dimension, and in the forced off-optimum arm.

The third is selectivity. Among the best configurations found, check whether the attribution vectors show reduced sheath loss together with an intact or larger longitudinal-field contribution. That is the selective-suppression signature the tension in the literature calls for. Also check whether including the magnetic ramp knob reproduces Bhakta et al.'s multi-fold gain once space charge is self-consistent, which their test-electron model omitted.

The proposal is refuted if any of three things happens. Work accounting matches knockout attribution everywhere. Scalar BO matches our loop at equal budget in the 8-knob setting. Or no configuration achieves sheath reduction without losing the anti-dephasing contribution.