Abstract Tuning a laser--wakefield accelerator to a desired electron energy spectrum is normally done by scanning the machine one setting at a time, which consumes scarce beam time and expert attention. We address the inverse problem directly, recovering the experimental settings that produce a target spectrum. To do so we train a conditional diffusion model, built on the Elucidating the Design Space of Diffusion-Based Models (EDM) framework, on measured electron spectra from the ALFA beamline, conditioned on laser energy, gas-jet backing pressure, and gas-valve opening time. Because EDM samples by integrating a probability-flow ordinary differential equation, the map from settings to spectrum is differentiable end to end, and the settings themselves can be treated as free variables and optimized by backpropagation through the sampler. We propose and evaluate a three-stage strategy, \mbox{BO $\to$ L-BFGS $\to$ RAdam}, in which Bayesian optimization provides a global warm start across flat regions, L-BFGS exploits local curvature, and a final stochastic phase relaxes the solution away from sharp, bin-aligned minima. Against seven alternative single- and multi-stage strategies, this ordering recovers the target settings with the lowest mean squared error. We position this as a tool for differentiable inverse design in experiment planning, complementary to particle-in-cell simulation rather than a substitute for it.
Abstract Recent advances in quantum computing experiments have enabled new pathways for feedback control. Here, we investigate the role of information in interactive quantum many-body dynamics to engineer out-of-equilibrium phases of matter. To this end, we train reinforcement learning agents to prevent entanglement spreading in $(1{+}1)$-dimensional stabilizer circuits with up to $128$ qubits. We find that, above an information threshold, agents learn stochastic near-optimal strategies that prevent the system from reaching volume-law entangled states and instead stabilize area-law scaling. These strategies rely on bottlenecks that divide the system and cap the entanglement in the steady state. Crucially, learned strategies are inherently out of equilibrium, require real-time active feedback, and could not be reproduced by simple human-designed control rules. This work establishes the foundations for information-driven individual control of many interacting quantum degrees of freedom, demonstrating the capabilities of reinforcement learning to stabilize and uncover novel critical properties of many-body nonequilibrium steady states.
Abstract Gaussian processes are widely used probabilistic models for regression and classification tasks. Their flexibility and ability to quantify uncertainties make them well-suited for many applications. However, their applicability to large datasets is limited by the cubic computational time complexity. While reduced-rank methods such as Hilbert space approximations reduce this complexity, scaling to very large datasets remains challenging. Quantum algorithms are promising candidates to further accelerate such computations. Quantum principal component analysis has previously been applied to Hilbert space Gaussian process regression to address this scaling limitation. However, its applicability is limited by the requirement for fault-tolerant quantum hardware. In this work, we present a hybrid quantum-classical model for reduced-rank Gaussian process regression, where we apply a quantum principal component analysis algorithm based on parameterized quantum circuits to Hilbert space Gaussian process regression. In noiseless simulations on a one-dimensional test function, the resulting model reaches a regression accuracy comparable to that of the exact Gaussian process and the classical Hilbert space Gaussian process. The framework uses shallow parameterized circuits and does not require fault-tolerant quantum hardware, which makes it a candidate for noisy intermediate-scale quantum devices.
Hyperparameter optimisation is one example of a broad class of multivariate minimisation problems with a complex landscape and often many local minima. As a result, gradient descent typically performs poorly, and easily parallelisable sampling methods are most commonly used, which often require the set of trial points to be chosen in advance. In most practical tuning problems, the loss depends strongly only on a few hyperparameters or their combinations, yet which ones matter is rarely known before the search begins. A sampling strategy must therefore be efficient regardless of which directions turn out to be important. Although a grid is the most intuitive way to sample the full space, it has been shown that in practice, selecting trials randomly outperforms a grid search in the typical case in which the variables are of widely differing importance, since a grid contains the same point multiple times when projected to lower dimension. However, random search is also suboptimal, as selected points can be either nearly identical or leave large gaps. Here, a new method is proposed based on disordered stealthy hyperuniform point distributions, which have smaller density fluctuations than random distributions while simultaneously performing well in projection. On a set of benchmark hyperparameter optimisation and simple minimisation problems, it is demonstrated that this strategy provides an improvement over grid and random search. While low-discrepancy sequences such as Sobol and Halton are competitive when the most ‘natural’ axes coincide with the coordinate axes, we show that applying random rotations to the benchmark functions removes their advantage, while leaving the stealthy hyperuniform patterns virtually unaffected. Since the relevant directions are rarely known in advance and are generally not axis-aligned, this rotation-invariance is a key practical advantage of the proposed method.
Abstract Multi-qubit gates are crucial to realizing quantum algorithms. Direct implementation of these gates will facilitate a reduction in circuit depth. However, existing model-based methods rely on precise characterization of the system Hamiltonian as well as execution of the control fields, which present challenges in obtaining high fidelity multi-qubit gates. In this paper, we exploit the bang-bang control protocol and model-free reinforcement learning to realize several kinds of three-qubit gates for superconducting quantum computers with tunable/fixed frequency qubits, including the iToffoli gate, the parity check gate and the CCZ gate. The simulation results show that the fidelity of each gate can achieve 99.2% with the Adam Gradient Descent Optimizer. Moreover, the proposed method is robust against random noise in the control pulse. Gate fidelity can be maintained over 95% even in the case of Gaussian noise with variance of 0.3. Besides, the impacts of parameter iterative optimization strategies are also analyzed.
