Batch Bayesian optimization of attosecond betatron pulses from laser wakefield acceleration

Laser wakefield acceleration can generate a femtosecond-scale broadband X-ray betatron radiation pulse from electrons accelerated by an intense laser pulse in a plasma. The micrometer-scale of the source makes wakefield betatron radiation well-suited for advanced imaging techniques, including diffraction and phase-contrast imaging. Recent progress in laser technology can expand these capabilities into the attosecond regime, where the practical applications would significantly benefit from the increased energy contained within the pulse. Here we use numerical simulations combined with batch Bayesian optimization to enhance the radiation produced by an attosecond betatron source. The method enables an efficient exploration of a multi-parameter space and identifies a regime in which a plasma density spike triggers the generation of a high-charge electron beam. This results in an improvement of more than one order of magnitude in the on-axis time-averaged power within the central time containing half of the radiated energy, compared to the reference case without the density spike.

💡 Research Summary

This paper presents a comprehensive strategy to boost the performance of attosecond betatron X‑ray pulses generated by laser‑wakefield acceleration (LWFA) through the combined use of three‑dimensional particle‑in‑cell (PIC) simulations and batch Bayesian optimization (BBO). In LWFA a high‑intensity femtosecond laser pulse drives a plasma wake (the “bubble”) that both accelerates electrons to relativistic energies and provides a transverse focusing force that makes the electrons execute betatron oscillations, thereby emitting broadband X‑ray radiation. While femtosecond‑scale betatron pulses are well‑established, pushing the duration into the attosecond regime while simultaneously increasing the pulse energy remains a major challenge.

The authors introduce a localized plasma density spike downstream of the electron injection region. Three parameters describe the spike: its distance from the injection gradient (d u), its longitudinal length (d s), and its peak electron density (n p). The spike is expected to trigger a “phase‑reset” of the accelerated electrons, moving them toward the rear of the bubble where the accelerating fields are strongest, and to induce a second, intense electron injection. This secondary injection dramatically raises the charge and energy of the electron bunch, which in turn amplifies the betatron radiation.

To find the optimal combination of (d u, d s, n p) the authors employ Bayesian optimization, which builds a Gaussian‑process surrogate model of the expensive simulation output. Rather than the conventional sequential approach that evaluates a single candidate per iteration, they use batch Bayesian optimization: each iteration evaluates a batch of N candidate points in parallel, updates the surrogate model only after the whole batch finishes, and then proposes the next batch. This parallelism matches the high computational cost of 3‑D PIC runs (each requiring several thousand core‑hours) and reduces wall‑clock time dramatically.

The performance metric is a cost function C = −W₅₀/τ₅₀, where W₅₀ is the on‑axis radiated energy contained within the central 50 % of the cumulative intensity curve, and τ₅₀ is the temporal interval over which that energy is emitted. By minimizing C the optimizer simultaneously maximizes energy concentration and penalizes pulse broadening, ensuring the final solution remains in the attosecond regime.

Simulation details: a Ti:sapphire‑like laser (λ = 800 nm, a₀ = 3.44, peak intensity 2.53 × 10¹⁹ W cm⁻², pulse energy 37 mJ, 8.3 fs FWHM) is focused to a 3.8 µm spot at the plasma entrance. The baseline plasma density is n₀ = 2 × 10¹⁹ cm⁻³. An initial down‑ramp of total length 40 µm (linear rise to 1.2 n₀ then fall back to n₀) provides electron injection. The total acceleration length is set to the dephasing length (~256 µm). The density spike is superimposed after the injection region; its up‑ and down‑ramps are symmetric.

Optimization proceeds as follows: eight initial points are generated with a Sobol sequence, then eight iterations are performed with batch sizes N = 1, 4, 8 for comparison. The N = 4 configuration yields the fastest convergence, reaching a cost value 60 times better than the reference after only five iterations. The optimal parameters are approximately d u ≈ a few micrometers, d s ≈ 120 µm, and n p ≈ 4 n₀. Under these conditions the on‑axis peak radiated power increases by a factor of >25, while the energy contained in the central 50 % of the pulse grows by a factor of >6, and the pulse duration remains around 400 as.

The study demonstrates two key points: (1) a localized density spike can effectively reset the phase of the accelerated electrons, producing a second high‑charge injection that dramatically enhances attosecond betatron emission; (2) batch Bayesian optimization is a practical tool for navigating high‑dimensional, expensive‑to‑evaluate parameter spaces in laser‑plasma physics. Limitations include the predefined bounds of the search space (e.g., n p limited to ≤ 4 n₀, d s ≥ 5 µm) which may exclude even more optimal configurations such as multiple spikes or higher densities. Future work could explore broader parameter ranges, incorporate additional experimental constraints, and extend the cost function to include imaging‑quality metrics, thereby moving toward experimentally realizable, application‑driven designs of compact attosecond X‑ray sources.