The central challenge the authors confront is that pulsed laser deposition is a strongly nonlinear, fast-changing process. When an excimer or femtosecond laser pulse, typically lasting tens of nanoseconds, is absorbed by the target surface, the material does not evaporate gently. Instead, it is ejected in a plasma plume whose composition, density, temperature and velocity depend sensitively on fluence, spot size, target morphology and background gas pressure. Downstream, the plume interacts with oxygen or nitrogen atoms, thermalizes, and finally condenses on the substrate, where surface diffusion, nucleation and growth kinetics decide whether the film crystallizes in the desired phase. Small drifts anywhere along this chain, such as target roughening after thousands of shots or a subtle shift in plume stoichiometry, can push the film away from its intended properties without any obvious warning until post-deposition characterization reveals the damage.

Traditional control approaches have largely relied on static recipes: fix the laser fluence, the repetition rate, the background pressure and the substrate temperature, and hope the recipe is robust enough to survive normal process variability. Some facilities have added monitoring, such as optical emission spectroscopy of the plume or reflection high-energy electron diffraction at the substrate, but the information gathered is usually used for post hoc diagnostics rather than for closed-loop correction. The new work argues that this is precisely the gap where modern data-driven methods can make a decisive difference. By learning a model of the process-state dynamics directly from real-time measurements, the controller can predict where the process is heading and act before deviations accumulate into defects.

The heart of the framework is a predictive model of the process state, constructed from data acquired during deposition rather than from first-principles simulation alone. The process state is a compact mathematical description of the quantities that actually govern film formation, including indicators of plume behavior, ablation dynamics and growth conditions at the substrate. Instead of requiring an engineer to derive governing equations by hand, the approach trains a model to map the recent history of process measurements and actuator settings onto the future evolution of that state. This model predictive structure then solves, at every control step, an optimization problem: choose the adjustments to the manipulated variables, such as laser fluence or timing, that drive the predicted process trajectory toward the desired reference while respecting physical and hardware constraints.

Predictive feedback of this kind has a crucial advantage over purely reactive control. A reactive controller waits for the measured state to deviate and then corrects, which is often too slow in a process whose characteristic timescales are milliseconds to seconds and whose nonlinearities can amplify errors. A predictive controller, by contrast, anticipates the deviation several steps ahead. If the model indicates that the plume conditions are drifting toward a regime associated with off-stoichiometric growth, the controller can modulate the laser energy before that regime is reached. This anticipatory action smooths transients, suppresses oscillations and keeps the process state within the corridor where the intended film phase forms reliably.

The data-driven element is equally important for practical deployment. Physical models of laser ablation and plume transport exist, but they are computationally expensive, parameter-sensitive and awkward to embed in a controller that must make decisions in real time. Learned models, in contrast, can be trained on the actual deposition system in question, capturing chamber-specific quirks that no generic simulation would encode. The authors emphasize that the framework treats the learned model as a surrogate for the true process dynamics and wraps it in the well-established machinery of model predictive control, which brings guarantees of constraint handling and stability that pure black-box machine-learning controllers often lack. The result is a hybrid: machine learning supplies the speed and adaptability, while predictive control theory supplies the discipline.

Real-time implementation is the aspect that makes the work notable for manufacturing rather than merely for laboratory science. Thin-film deposition for quantum devices, power electronics, sensors and photonic components increasingly demands not just one good film but reproducible good films, batch after batch, tool after tool. Variability between depositions is a persistent tax on yields and costs. A controller that senses the process state continuously and corrects drift automatically reduces dependence on operator expertise, shortens the tuning time for new materials, and offers a pathway to qualification of deposition processes in industrial settings where every shot of the laser must count. In effect, the technique moves pulsed laser deposition closer to the kind of statistically controlled, closed-loop manufacturing that semiconductor fabs take for granted.

The implications extend beyond the specific chamber geometry or material system studied. Pulsed laser deposition is prized for complex oxides, where perovskites, superconductors and magnetic heterostructures must be grown with layer-by-layer control of cation stoichiometry. It is also central to emerging applications such as solid-state battery electrolytes, ferroelectric memories and wide-bandgap semiconductor films. Each of these areas suffers from the same underlying difficulty: the relationship between process parameters and film quality is mediated by a fast, chaotic intermediate stage, the ablation plume, which no one can observe directly at the substrate with full fidelity. A predictive state-space controller sidesteps part of this difficulty by inferring the unobservable state from accessible measurements, much as modern automotive control systems infer combustion dynamics from a handful of sensors.

The work also illustrates a broader trend in advanced manufacturing: the migration of machine learning from offline analysis, where models are trained on historical process data to predict yield or classify defects, into the tight feedback loop of the equipment itself. In offline mode, machine learning informs the engineer; in closed-loop mode, it acts. That transition raises the bar for reliability, interpretability and safety, and it is why the authors pair their learned dynamics model with the constrained optimization structure of predictive control. Constraint handling matters in a real deposition tool: the laser fluence cannot exceed what the optics and target tolerate, the repetition rate is bounded by plume clearance times, and abrupt setpoint changes can itself induce defects. A controller that respects these limits while optimizing performance is far more credible on a production floor than one that merely minimizes a loss function on paper.

Looking forward, the authors’ framework suggests a future in which deposition tools continuously learn from every film they grow. Each deposition adds data that can refine the model of process-state dynamics, making the controller better calibrated for the next run, the next material or the next chamber condition. Combined with increasingly rich in-situ diagnostics, such as plume imaging, spectroscopic monitoring and real-time film thickness gauges, predictive feedback control could reduce the development cycle for new functional materials from months of parameter sweeps to a handful of instrumented runs. For a technique that helped pioneer the growth of some of the most celebrated quantum materials of the past three decades, that would represent a quiet but profound modernization: the same flashes of laser light, now guided by an algorithm that watches, predicts and corrects thousands of times faster than any human ever could.

Subject of Research: Real-time data-driven predictive feedback control of process-state dynamics in pulsed laser deposition thin-film growth

Article Title: Real-time data-driven predictive feedback control of process-state dynamics for pulsed laser deposition

Article References: Horide, T., Kawabata, K., & Yoshida, Y. (2026). Real-time data-driven predictive feedback control of process-state dynamics for pulsed laser deposition. npj Advanced Manufacturing, 3(1), Article 28. https://doi.org/10.1038/s44334-026-00112-w

Image Credits: AI Generated

DOI: 10.1038/s44334-026-00112-w

Keywords: pulsed laser deposition, predictive control, model predictive control, machine learning, thin films, advanced manufacturing, laser ablation, real-time feedback, process control, complex oxides, npj Advanced Manufacturing, Real-time