Multi-objective and multi-fidelity Bayesian optimization of laser-plasma acceleration

• Laser-plasma acceleration provides a setting where growing simulation and experimental datasets motivate machine-learning methods for multidimensional optimization.Source: §I · p. 1
• Optimizing charge alone can reduce beam energy because divergence and energy information disappear during integration.Source: §I · p. 1
• Weighted combinations of charge, energy, and spread depend on empirically chosen weights, often requiring repeated trial-and-error optimization runs.Source: §I · pp. 1–2
• Single-objective functions compress complex beam distributions into one value, inevitably losing information about features such as central tendency, dispersion, skewness, and amplitude.Source: §I · pp. 1–2
• Multi-objective optimization instead exposes trade-offs among objectives through a Pareto front, allowing objective combinations to be selected after optimization.Source: §I · p. 2

• The simulated accelerator uses shock injection, where a density downramp traps electrons and the plasma density and laser conditions shape the final energy spectrum.Source: §II · p. 2
• The optimization varies plateau density, focus position, upramp length, and downramp length.Source: §II · p. 2, Figure 1 · p. 2, Table I · p. 3
• GPU execution, cylindrical geometry, boosted-frame windows, and spatially varying macro-particle density reduce the cost of particle-in-cell simulations.Source: §II · p. 2
• Transverse and longitudinal simulation-box sizes are scaled with laser waist and estimated wakefield size to avoid wasting computational resources.Source: §II · pp. 2–3
• The optimized simulation scan is constrained by numerical instabilities, artifacts, and the inability to use boosted-frame geometry near critical density.Source: §II · p. 3

• A Gaussian-process surrogate and acquisition function select new evaluations by modeling expected improvement rather than directly probing the costly system.Source: §III · p. 3
• Expected hypervolume improvement extends Bayesian optimization to multiple objectives and their trade-offs.Source: §III · pp. 3–4
• Approximately ∼102 evaluations are typically needed for the four-input accelerator, and hour-scale simulations would make a full run take several days.Source: §III · p. 4
• Multi-fidelity optimization adds fidelity χ to the input space, using low-resolution simulations that capture essential injection and acceleration physics at much lower cost.Source: §III · p. 4
• Computing time scales approximately as cost(χ) ∝χ3.5, while accounting for fidelity and cost yields an average order-of-magnitude speed-up and typical ten-hour runs.Source: §III · p. 4
• Single- and multi-objective results use the same MOMF algorithm, constraints, computational budget, and repeated-run protocol for comparison.Source: §III · p. 4

• The optimization seeks quasi-monoenergetic electron beams with a high-quality energy spectrum.Source: §IV · p. 4

Objective definitions

• Objective functions combine charge, target-energy distance, and spectral dispersion using alternative statistical metrics such as mean, median, standard deviation, and median absolute deviation.These choices encode different biases toward the desired beam properties.Source: Optimization parameters · pp. 4–5, Table III · p. 5
• The mean emphasizes points farther from the target, whereas median-based descriptors can be more suitable in the presence of noise.The median-based formulation uses median energy, absolute target distance, and median absolute deviation, with a square root reducing charge emphasis.Source: Optimization parameters · pp. 4–5
• Alternative charge weights are empirical: increasing the charge exponent raises total charge, but the best choice cannot be known beforehand and requires multiple optimization runs.The compared variants use Q^1/2, Q^2, and Q^3 weightings.Source: Optimization parameters · pp. 5–6
• Division-based objectives can diverge as their denominator decreases, motivating an alternative that optimizes charge within a specified energy window around the target.The windowed formulation implicitly optimizes target energy and energy spread without division; the relevant charge is integrated over the interval around E0.Source: Optimization parameters · p. 5

Results and discussion

• Different scalarizations produce distinct spectra: mean-based objectives suppress tails, median-based objectives permit high-energy tails, and implicit target-window optimization produces a higher peak charge.The third objective also explicitly penalizes charge outside the 250–350 MeV window and appears to outperform explicit target-energy optimization.Source: Optimization parameters · p. 5, Figure 2 · p. 6
• Peak energy can differ substantially from mean or median energy because beam loading skews highly charged spectra, requiring explicit peak-energy optimization when the peak matters.The peak is identified as the energy with the highest spectral charge density and may lie far from the target despite other central-tendency metrics.Source: Optimization parameters · pp. 5–6

Method

• The method directly optimizes the trade-off among mean energy distance, energy spread, and total beam charge using expected hypervolume improvement.The resulting Pareto front contains non-dominated output-space solutions, while the hypervolume measures the diversity of covered solutions.Source: §V · p. 6

Results and discussion

• Comparable performance to scalarized objectives is achieved by a single multi-objective run while exposing the full Pareto surface of non-dominated solutions.The Pareto surface is visualized in three dimensions, with pairwise fronts for charge, energy distance, and energy spread.Source: §V · p. 6, Figure 3 · p. 7

Trade-offs

• Above 500 pC, increasing charge increases distance from the target energy because beam loading reduces wakefield strength and mean energy.The physical mechanism links higher bunch charge to weaker wakefields and lower mean energy.Source: §V · p. 6
• Higher charge also trades off against mono-energetic beams, producing wider electron-beam spectra and distinct input settings for high charge versus narrow spread.High-energy, low-charge beams are largely excluded because the objectives penalize distance from the target and favor higher charge.Source: §V · p. 7

Benefits

• The multi-objective formulation characterizes system capability across objectives without strongly biasing the solution set toward one objective combination.Hypervolume optimization also avoids the denominator-offset problem that can distort single-objective objectives.Source: §V · p. 7

A posteriori objectives

• Existing multi-objective scans can be re-evaluated for new objectives by probing Gaussian-process predictions of means and variances and selecting conservative candidates with a lower confidence bound.This enables a posteriori objective evaluation without rerunning the original optimization; inferred solutions improved O1 by approximately 40% and reached 90% of the corresponding single-objective O2 result.Source: §V · pp. 7–8

Input-space analysis

• The Gaussian-process analysis maps pairwise and individual effects of four inputs on charge, energy distance, and energy spread by averaging over the remaining parameters.Pair plots cover six input pairs, while one-dimensional views average over the other three inputs.Source: §VI · p. 8, Figure 5 · p. 9

Density (ne)

• Charge rises and then falls with density because higher-density injections increasingly lose electrons through dephasing and defocusing near the accelerator end.The optimal charge and energy occur at the same density, indicating an appropriate charge for beam loading to reach 300 MeV.Source: §VI · p. 8

Upramp length (lup)

• Longer upramps increase injected charge approximately linearly and appear to support reaching the target energy, likely through stronger laser self-focusing.A longer upramp allows higher laser intensity at the injection point.Source: §VI · p. 8

Downramp length (ldown)

• Shorter downramps increase injection through faster wakefield expansion, while downramp length shows no influence on mean energy within the scanned range.Within these ranges, downramp length changes the injection point rather than the acceleration length.Source: §VI · p. 8

Focus position (zfoc)

• The focus position that maximizes charge tends to produce the worst energy and energy spread, most likely because increased charge causes beam loading.The charge-optimal focus position shifts forward at higher density and longer upramp length.Source: §VI · pp. 8–9

Conclusion and outlook

• Multi-objective optimization achieves similar or superior results to several single-objective optimizers without iterative fine-tuning of objective parameters.It produces a more general result that supports evaluating objectives without repeatedly redesigning scalarized objective functions.Source: §VII · p. 9
• Combining GPU-based simulation with multi-fidelity optimization enables extensive multidimensional optimization beyond prior particle-in-cell studies of similarly expensive physical systems.The authors describe this as a milestone toward digital twins for optimizing real-life experiments and infrastructure.Source: §VII · p. 9
• The multi-objective techniques can transfer directly to experiments, where surrogate models may help reveal input-output couplings and inform Bayesian-optimization priors.The multi-fidelity component is most immediately useful for simulations with different numerical resolutions.Source: §VII · p. 9