Conceptualizing A Closed-Loop Digital Twin for Real-Time Plasma Control in Large-Area Magnetron Sputtering Systems

to convert line intensities into relative species densities, provides a route from raw spectra to physically meaningful concentrations [24]. OES is fast, robust, optically coupled through a window and therefore non-invasive, and can be spatially multiplexed with fibre arrays to observe several positions along a large source—properties that make it the primary real-time sensor for a sputtering twin. Its limitations are that emission is a convolution of density and excitation conditions, so quantitative inversion requires modelling, and that line selection and calibration must be tailored to each material system. 3.3 Mass spectrometry and in-situ metrology Energy-resolved mass spectrometry characterizes the flux of ions and neutrals arriving at the substrate plane, resolving their mass and energy distributions [17], [20]. This is the diagnostic most directly connected to the quantities HiPIMS control seeks to manipulate—ionized metal fraction, ion energy, and charge state—and it is therefore a powerful calibration and validation instrument for the flux-composition sub-model of a twin. It is, however, comparatively slow, expensive, and geometrically intrusive, and is best regarded as a reference sensor rather than a continuous control input. Finally, in-situ film metrology closes the gap between plasma state and film property. Real- time spectroscopic ellipsometry can monitor film thickness and optical constants during growth, and quartz-crystal microbalances measure deposition rate directly; optical monitoring is standard in precision optical-coating production. These metrology signals are the ground truth for the twin’s ultimate objective—film thickness, composition, and uniformity—but they are typically local, may lag the plasma state, and can be difficult to deploy over a full large-area substrate. 3.4 Sensing constraints for the twin Several cross-cutting constraints follow. Latency and update rate vary by orders of magnitude across these instruments—from sub-microsecond electrical waveforms to second-scale spectral or metrology updates—so the twin must fuse asynchronous, multi-rate data streams. Robustness is paramount: window fouling degrades optical diagnostics, probe coating corrupts electrical characteristics, and reactive environments are hostile to intrusive hardware. No single sensor observes the full distributed state, and each observes it indirectly. This motivates a sensing philosophy of heterogeneous redundancy feeding a model-based estimator: fast, robust, non- perturbing signals (electrical, OES) drive continuous control, while richer but slower or more intrusive diagnostics (probes, mass spectrometry, ellipsometry) calibrate and periodically correct the model. The estimator, not any individual instrument, delivers the synchronized state on which control acts—precisely the role of the state-estimation layer developed in Sections 4 and 5. 4. Digital Twin Foundations 4.1 Definitions and taxonomy Table 1. The digital model–shadow–twin taxonomy and its implication for real-time plasma control. Property Digital model Digital shadow Digital twin Physical → virtual data flow Manual Automatic Automatic Virtual → physical data flow Manual Manual Automatic Real-time synchronisation No Partial Yes Closes the control loop No No Yes Suitability for plasma control Design / analysis only Monitoring Real-time control The term “digital twin” originated in aerospace as the notion of an ultra-high-fidelity, continuously updated simulation of an as-built vehicle, mirroring its physical counterpart throughout its life [12], [27]. Subsequent work in manufacturing has refined the concept and, importantly, distinguished it from adjacent ideas. A widely adopted taxonomy classifies the physical–virtual coupling by its degree of automation: a digital model has no automatic data exchange; a digital shadow has automatic data flow from the physical object to the virtual one but not back; and a true digital twin has automatic, bidirectional data flow, so that changes in the physical asset update the model and decisions in the model act on the asset [14]. This distinction is central to the present work: real-time plasma control requires the fully closed, bidirectional coupling of a genuine twin, not merely a monitoring shadow [13], [15], [16]. The bidirectional, predictive twin has since been demonstrated across engineering domains well beyond aerospace and discrete manufacturing— including energy systems, where tiered digital twins deliver predictive diagnostics to distributed, resource-constrained assets [48], environmental-compliance monitoring in process industries [49], and edge-intelligence architectures for real-time control of resilient microgrids [50]— underscoring the generality of the closed-loop concept that this paper specialises to large-area plasma processing. Related efforts couple automation with digital twins to reduce process errors and turnaround time [51], and systematic analyses of hierarchical, decentralised, and hybrid control architectures for distributed energy resources illuminate the control-layer design choices that a plasma twin must also confront [52].Reviews of the field emphasize that a digital twin is characterized less by any single technology than by the integration of models, data, and services around a specific decision or control purpose [28], [29], [30], [31], [32]. Wright and Davidson stress that the defining additions beyond a conventional model are the live data stream that keeps the twin current and a clear, quantified purpose against which fidelity is judged [35]. Fuller and colleagues survey the enabling technologies—sensing, connectivity, modelling, and analytics— and the open challenges of latency, security, and standardization [34], while Rasheed and co- workers frame the twin from a modelling perspective and articulate the values and enablers of physics-based, data-driven, and hybrid approaches [33]. For a sputtering process, the “asset” is the discharge and the growing film, the “purpose” is regulation of deposition rate, stoichiometry, and uniformity within specification, and the twin’s value lies in reasoning about the unobserved distributed state in real time. 4.2 Physics-based, data-driven, and hybrid models The twin’s virtual model may be built along a spectrum from first-principles physics to pure data [33]. At the physics-based end, magnetron discharges can be described by particle-in-cell/Monte- Carlo collision simulations that resolve the kinetics of charged particles in the magnetized sheath and bulk, and by fluid or hybrid plasma models; these capture the mechanisms of confinement, ionization, and transport but are far too computationally expensive to run in the control loop [4], [25]. Finite-element simulation of energy-harvesting devices offers a further illustration of high- fidelity physics-based modelling whose computational cost motivates the reduced-order surrogates discussed below [81]. Reactive-process dynamics are captured by the Berg-type balance models, which are inexpensive and physically interpretable but coarse-grained [10], [11]. At the data- driven end, regression and machine-learning models—Gaussian-process regressors, neural networks, and virtual-metrology models—can map process settings and sensor signals to outcomes with high fidelity where training data are abundant, at the cost of extrapolation risk and limited physical guarantees [26], [39]. Hybrid and physics-informed strategies aim to combine the strengths of both. Physics-informed neural networks embed governing equations as soft constraints in the training loss, improving data efficiency and physical consistency [36]; sparse-regression methods can discover parsimonious governing equations directly from data [37]. For low- temperature and processing plasmas specifically, machine learning is increasingly applied to modelling, diagnostics inversion, and control, and is regarded as a strategic direction for the field [26]. The design choice for a sputtering twin is therefore not physics versus data but the appropriate hybridization: a physically structured skeleton, calibrated and corrected by data, that is both trustworthy and fast enough to evaluate in real time. 4.3 Model-order reduction and surrogates Real-time operation makes computational speed a hard constraint, which is the province of model- order reduction and surrogate modelling. Projection-based MOR methods approximate a high-dimensional parametric model by its behaviour on a low-dimensional subspace identified from representative solutions, yielding reduced models that are orders of magnitude cheaper to evaluate while retaining a controllable accuracy and, often, a physical interpretation [38]. Data- driven surrogates—including neural-network and Gaussian-process emulators trained on offline high-fidelity simulations or experiments—achieve similar acceleration and are naturally suited to the machine-learning toolchain [26], [39]. A predictive digital twin at scale can be organized as a library of such component reduced-order models, assembled and updated as the asset’s condition changes [38], [42]. For the sputtering twin, MOR and surrogates are what render an otherwise intractable plasma-and-film model executable within a control cycle, and the fidelity-versus-speed trade-off they mediate is a central design axis (Section 7). 4.4 Synchronization and state estimation Synchronization is the mechanism that makes a model a twin. Because the discharge state is only partially and indirectly observed, the twin must estimate its full state by optimally combining model predictions with sensor data—the classical problem of recursive state estimation [40]. The Kalman filter provides the optimal linear-quadratic solution for linear-Gaussian systems, propagating both a state estimate and its covariance and correcting it as measurements arrive [40]. Its non-linear extensions—the extended and unscented Kalman filters—and, for high-dimensional systems, the ensemble Kalman filter, extend this framework to the non-linear, distributed dynamics typical of plasma and flow problems and are the workhorses of data assimilation [41]. Model- based monitoring of digital twins is now commonly framed explicitly around Kalman filtering, in which the filter both estimates unmeasured states and detects model–data discrepancy [40], [41]. Probabilistic-graphical-model formulations generalize this to enable predictive, uncertainty-aware twins that assimilate data to update a belief over asset state and thereby drive decisions [42]. In the proposed architecture (Section 5), a Kalman-family estimator is the pivot that turns heterogeneous, multi-rate sputtering diagnostics into a coherent, uncertainty-quantified estimate of the discharge and film state on which control can act. The modelling ingredients for such a twin are well developed in adjacent literatures. Magnetron discharges have been simulated with 3D particle-in- cell/Monte-Carlo and two- and three-dimensional fluid models [82]–[85], reactive-process behaviour captured by extended Berg-type balance models [86], [87], and ionized HiPIMS discharges described by ionization-region and global models [89]–[91]; Langmuir-probe and energy-resolved mass-spectrometry diagnostics characterise such plasmas experimentally [88]. On the data-driven side, machine learning is increasingly applied to low-temperature-plasma modelling and diagnostics [92], [93], with physics-informed neural networks [97], [98], reduced- order modelling [99], and surrogate- and Bayesian-optimisation approaches—including autonomous sputter synthesis and neural-network control of plasma etching—yielding fast, control-oriented models [94]–[96]. Recursive state estimation via the unscented Kalman filter provides the assimilation machinery to keep such models synchronised with the asset [100], and digital-twin implementations in the process and manufacturing industries demonstrate the surrounding architecture [104], [105]. 5. Conceptual Architecture of the Closed-Loop Digital Twin We now assemble these foundations into a concrete conceptual architecture. The proposed closed- loop digital twin is organized as five interacting layers—physical asset, sensing and data pipeline, twin models, state estimation, and control policy with actuation—linked in a continuous cycle and governed by an explicit latency budget. Figure 1 is described in prose below. 5.1 The layered framework The proposed architecture is shown in Figure 1. Figure 1. Conceptual architecture of the closed-loop digital twin for real-time plasma control in large-area magnetron sputtering. Figure 1 (described). Envision the architecture as a vertical loop. At the bottom sits the physical asset: the magnetron source or sources, target, magnet pack, power supply, gas-delivery and pumping systems, and the moving large-area substrate within the vacuum chamber. Arrows rise from the asset into a sensing and data pipeline layer, which conditions and time-stamps heterogeneous signals—discharge voltage/current waveforms, OES spectra from a multiplexed fibre array along the source, mass-spectrometer and probe data where available, and in-situ ellipsometry or rate signals. These feed upward into two parallel blocks that constitute the twin models: a reduced-order physics block (a Berg-type reactive-dynamics model, an erosion-and- uniformity model, and a flux-composition surrogate) and a data-driven correction block (a learned residual/emulator). Their outputs and the sensor streams converge in the state-estimation layer— a Kalman-family estimator that produces an uncertainty-quantified estimate of the distributed discharge-and-film state and a synchronization/discrepancy signal. This estimated state passes to the control-policy layer (model predictive control and/or reinforcement learning), which computes actuation set-points. Arrows then descend back to the physical asset’s actuators—reactive-gas flow, applied power and, for HiPIMS, pulse waveform parameters, magnetic-field or magnet- motion settings where adjustable, and substrate speed—closing the loop. A slower outer path carries reference metrology and periodic diagnostics back into the twin models for recalibration. 5.2 Physical asset and actuation channels The controllable inputs available on a modern coater define the twin’s action space. The dominant fast actuator in reactive processes is reactive-gas mass flow (or, equivalently, a partial-pressure set-point), which moves the operating point along and across the hysteresis loop [8], [9], [10]. Applied power sets the overall sputtering rate; in HiPIMS, the pulse parameters—peak power, pulse length, frequency, and, in bipolar schemes, the reverse-voltage phase—constitute a rich additional channel that shapes the ionized flux and its energy [17], [18], [19], [20]. Magnetic configuration, whether through adjustable or moving magnet packs, influences the erosion profile and hence uniformity, though it is often a slow or commissioning-time actuator [6], [7]. In dynamic coaters, substrate transport speed directly trades throughput against integrated dose and uniformity [1]. Fast arc-management electronics constitute a specialized, near-instantaneous actuation loop of their own [9], [20]. The twin must respect the differing bandwidths and constraints of these channels. 5.3 Sensing and data pipeline The sensing layer implements the heterogeneous-redundancy philosophy of Section 3. Always-on, low-latency electrical and OES signals form the continuous backbone; probe, mass-spectrometry, and in-situ metrology signals enter at lower rates or during dedicated calibration windows [22], [23], [24]. The pipeline must synchronize and time-stamp multi-rate, asynchronous streams; perform online quality monitoring to detect window fouling or probe degradation; and pre-process raw data—for example, extracting characteristic line ratios from spectra or features from pulse waveforms—into the observation vectors the estimator expects [24], [26]. Because a large source is spatially extended, spatial multiplexing (a fibre array along the target) is essential to observe, not merely a point, but the profile that uniformity control requires. 5.4 Twin models and state estimation Within the twin, the reduced-order physics block supplies interpretable, fast predictions of the process’s dominant dynamics: reactive-gas partial pressure and target/collector coverage from a Berg-type model [10], [11]; a slowly evolving erosion state and its mapping to the deposition profile [6], [7]; and a flux-composition surrogate for the ion/neutral distribution, distilled via MOR or emulation from expensive particle-in-cell or hybrid simulations and calibrated against mass- spectrometry references [4], [25], [38]. The data-driven correction block learns the residual between this physics skeleton and reality, absorbing unmodelled effects—chamber conditioning, geometry, ageing—using physics-informed or purely empirical models trained on historical and streaming data [26], [36], [37], [39]. The state-estimation layer fuses these predictions with the live observations. A non-linear or ensemble Kalman filter propagates the estimated state and its covariance through the reduced model and corrects it against the electrical, optical, and metrology measurements as they arrive, yielding both the best estimate of unobserved quantities—electron density and temperature, coverage, ionized flux fraction, local rate—and a quantified uncertainty [40], [41]. Crucially, the filter’s innovation sequence (the model–measurement discrepancy) serves double duty as a fault- and drift-detector and as a trigger for recalibration or model updating, realizing the self-updating character that distinguishes a twin from a static model [15], [40], [42]. 5.5 Control policy: MPC and reinforcement learning The control layer consumes the estimated, uncertainty-tagged state and computes actuation. Two complementary paradigms are appropriate. Model predictive control uses the twin’s reduced model to optimize a finite-horizon sequence of actions subject to actuator and safety constraints, re-optimizing at each cycle as new state estimates arrive [43]. MPC is the natural fit for constrained, multivariable regulation—holding partial pressure on the unstable hysteresis branch while simultaneously balancing a uniformity profile—because it explicitly handles constraints, coupling, and prediction, and its theory and practice are mature [43]. Reinforcement learning (RL) offers a complementary route in which a control policy is learned, potentially against the twin itself as a training environment, and can capture non-linear, hard-to-model regimes; deep RL has demonstrably controlled magnetically confined fusion plasmas in real time and pre- emptively avoided instabilities [44], [45], [46], [47]. A pragmatic design uses the twin as a high- fidelity simulator in which RL policies are trained and validated offline before deployment, with MPC providing a constraint-respecting, interpretable baseline and safety fallback. Both paradigms depend on the twin: MPC on its predictive reduced model, RL on it as a training environment and state estimator. 5.6 Latency budget Real-time feasibility hinges on a coherent latency budget that matches each loop’s cycle time to the dynamics it must regulate. We propose a hierarchy of nested loops. An innermost arc- suppression loop, implemented in dedicated power-supply electronics, must react on microsecond time scales and is best left as a specialized hardware function informed, but not gated, by the twin [9], [20]. A fast plasma-regulation loop—reactive-gas and power control on the hysteresis branch—must close within milliseconds to tens of milliseconds, set by the reactive-process dynamics; this budget must accommodate OES/electrical acquisition, feature extraction, one estimator update, and one MPC re-optimization, which is why the model in this loop must be a reduced-order surrogate rather than a first-principles simulation [8], [9], [38], [43]. A uniformity/profile loop operates on the seconds-to-minutes scale of substrate passage and profile measurement. An outermost drift-compensation loop tracks target erosion and chamber conditioning over minutes to hours and updates the twin’s slow states and calibration [6]. Explicitly allocating the computational budget—sensing latency, estimator update, and control optimization—across these nested loops is what converts the digital-twin concept from an aspiration into an implementable, real-time controller. For the control layer specifically, established model-predictive-control practice [101] and a growing body of reinforcement-learning process control—together with systematic comparisons of the two paradigms—provide the decision machinery a closed-loop plasma twin would embed [102], [103]. 6. Control Strategies and Use Cases The value of the architecture is best seen through concrete control problems. Four canonical use cases exercise different layers and time scales of the twin. 6.1 Thickness and profile uniformity control Uniformity is a distributed regulation problem: the objective is a target thickness (and composition) profile across a large substrate, and the manipulated variables are those that shape the integrated flux—magnetic configuration and magnet motion, spatially resolved power or gas distribution in segmented sources, and substrate speed [1], [6], [7]. A twin adds two capabilities beyond static shimming and empirical shutter/masking. First, by estimating the erosion-dependent deposition profile from electrical and spatially multiplexed OES signals, it can predict how the profile will drift as the target ages and pre-emptively adjust the manipulated variables, rather than reacting after out-of-specification product appears [6]. Second, cast as a multivariable MPC problem, uniformity control can coordinate several actuators against a profile objective while respecting their constraints and their coupling to rate and stoichiometry [43]. In-situ metrology or downstream inspection provides the profile feedback that periodically corrects the twin’s uniformity model. 6.2 Deposition-rate and stoichiometry control in reactive sputtering This is the flagship use case. Operating a reactive process at the high-rate, fully compound point on the unstable transition branch is the classic motivation for fast feedback [8], [9], [10], [11]. Conventional PEM regulates a single emission line to a set-point; a twin generalizes this in three ways. It fuses OES with discharge voltage, partial-pressure, and rate signals into an estimate of the actual target and collector coverage—the physically meaningful state—rather than a single proxy [10], [11], [24]. It uses MPC to hold that state on the unstable branch while explicitly enforcing constraints and rejecting flow and pressure disturbances, with the reduced Berg-type model providing the short-horizon prediction that stabilizes the branch [10], [43]. And it anticipates the slow migration of the hysteresis loop as chamber conditioning and target state change, adjusting set-points to keep stoichiometry constant over a campaign. In reactive HiPIMS, the twin can additionally exploit pulse-parameter actuation to manage poisoning and ionized flux jointly, a coupling too intricate for single-loop control [17], [19]. 6.3 Arc suppression and instability management Arcs and self-organized instabilities are fast, localized disturbances that damage films and hardware [9], [16], [20], [21]. Their suppression today is a hardware function of the power supply, reacting to the electrical signature of an incipient arc within microseconds [9], [20]. The twin’s role here is not to replace this fast loop but to reduce the incidence of instabilities by steering the process away from conditions that provoke them: excessive poisoning, pressure excursions, or discharge regimes prone to spoke formation [16], [21]. By estimating proximity to known instability boundaries and feeding an arc-rate signal back as a soft constraint or cost term in the MPC objective, the twin manages instability at the process-planning time scale while the hardware loop handles individual events—a clean separation of concerns across the latency hierarchy of Section 5.6. 6.4 Target-erosion compensation and campaign management Target erosion is the archetypal slow, non-stationary drift [6], [7]. As the racetrack deepens, deposition rate at fixed power, discharge voltage, uniformity profile, and even the tendency to arc all evolve. A twin that maintains a slowly updated erosion state—estimated from cumulative charge, discharge-voltage trends, and periodic profile metrology—can compensate these effects continuously: trimming power and gas set-points to hold rate and stoichiometry, re-optimizing uniformity actuators as the profile shifts, and forecasting remaining target life to schedule changes and maximize utilization [6]. This outermost loop is where the twin’s economic value is most tangible, converting a source of yield loss and unplanned downtime into a predictable, optimized campaign. Across all four use cases, the same twin infrastructure—reduced models, state estimation, and constrained optimization—is reused at different time scales, which is the architectural economy the concept is designed to deliver. 7. Implementation Challenges and Validation Pathway Translating the concept into a deployable system raises substantial challenges, which we group into model fidelity, calibration, uncertainty, verification and validation, and scale-up, and for which we propose a staged validation pathway. 7.1 Model fidelity versus real-time execution The central tension is between fidelity and speed. First-principles particle-in-cell and hybrid plasma simulations, while mechanistically faithful, are far too slow for the millisecond fast loop [4], [25]. The resolution is disciplined model-order reduction and surrogate modelling: offline high-fidelity simulations and experiments generate the data from which fast reduced or emulated models are constructed for online use [38], [39]. The risk is that a reduced model, valid over its training envelope, degrades outside it—precisely where disturbances push the process. Mitigations include hierarchical modelling (fast surrogate in the loop, higher-fidelity model consulted less frequently or for validation), physics-informed structure that constrains extrapolation [36], [37], and continuous residual learning that adapts the model to the observed asset [26]. Determining, for each control loop, the minimal fidelity that still yields acceptable closed-loop performance is itself a core design activity. 7.2 Calibration, identifiability, and uncertainty quantification A twin must be calibrated to a specific machine and re-calibrated as it drifts. Many internal states— coverages, spatial erosion, flux composition—are not directly or fully observed, raising identifiability concerns: different parameter sets may explain the same measurements. Rich diagnostics used during commissioning (probes, mass spectrometry, ellipsometry) constrain these parameters, and recursive estimation maintains them online, but the analyst must verify that the available observations render the controlled states identifiable [22], [23], [40]. Uncertainty quantification is not optional: the twin must propagate uncertainty from sensor noise, model error, and parameter uncertainty into its state estimates and predictions so that control can be made robust or risk-aware [33], [42]. Probabilistic twin formulations and ensemble estimators provide the machinery for this, yielding not just a state but a calibrated belief over it [41], [42]. 7.3 Verification, validation, and the staged pathway Because the twin will drive actuators on expensive, safety-relevant equipment, it must be verified (the software correctly implements the intended models and algorithms) and validated (the twin adequately represents reality for its purpose) before and during deployment [33], [35]. We propose a four-stage validation pathway. Stage 1 — offline model validation: reduced and surrogate models are validated against high-fidelity simulations and reference diagnostics (probes, mass spectrometry, ellipsometry) across the intended operating envelope [22], [23], [24], [25]. Stage 2 — open-loop shadow (digital-shadow) operation: the twin runs alongside a production tool, ingesting live data and predicting outcomes without acting, so that its state estimates and forecasts can be scored against measured film properties—this is the digital-shadow stage of the taxonomy, and it de-risks the closed loop [14], [15]. Stage 3 — closed-loop pilot: the twin is given authority over selected actuators on a pilot coater, initially with conservative constraints and human oversight and a hardware safety fallback, exercising each use case of Section 6. Stage 4 — production hardening and continual validation: the deployed twin is monitored via its own innovation/discrepancy signals, with drift detection triggering recalibration, closing the loop on the twin’s own health [40], [42]. This staged progression from model to shadow to controlled twin directly instantiates the field’s conceptual distinctions [13], [14], [15], [35]. 7.4 Scale-up Finally, the large-area, multi-source, dynamic nature of industrial coaters compounds every challenge. A production coater may host many magnetrons in series along a moving substrate, so the twin becomes a distributed system of coupled sub-twins whose interactions (shared gas environment, cumulative dose) must be modelled, and whose sensing and computation must be orchestrated with bounded latency across the line [1], [3]. Roll-to-roll and in-line glass coaters add web dynamics and throughput coupling. These realities argue for modular, component-based twin construction—libraries of reduced-order source models composed into a line-level twin— consistent with the scalable digital-twin architectures proposed in the literature [38], [42], and for a computing and connectivity infrastructure engineered to the latency budget of Section 5.6 [34]. 8. Discussion and Research Roadmap The preceding sections argue that the ingredients for a closed-loop plasma-control digital twin in large-area sputtering now exist in adjacent, mature literatures—magnetron discharge physics and reactive-process modelling [4], [10], [11]; non-perturbing diagnostics suitable for real-time use [24]; digital-twin architecture and data assimilation [14], [40], [42]; model-order reduction and machine learning for fast, hybrid models [26], [38], [39]; and predictive and learning-based control demonstrated on genuinely hard plasmas [43], [46], [47]. What is largely missing is their integration around this specific process. Several research thrusts follow. The proposed twin also connects to a wider body of work on data-driven decision systems and cyber-physical automation across domains. Predictive-analytics and forecasting methods, geospatial analytics, real-time key-performance-indicator tracking, and AI-assisted decision frameworks illustrate the analytics layer on which any twin depends [53]–[56], as do data-driven analytics and digital-transformation programs for service and utility operations [57], [58]. Autonomous-systems research—multi-agent coordination and swarm intelligence for autonomous mobile robots [59]–[61]—exemplifies the tightly coupled sense–decide–act loop a control-grade twin must implement, while cloud- and edge-computing models for resource allocation, scalable architectures, and machine-learning-based scaling [62]–[67] speak to the low-latency compute substrate such twins require. Perception and assurance methods, including robustness against adversarial inputs and AI-based software testing [68], [79], bear on twin verification and validation, and a substantial literature on UAV-, LiDAR-, and GIS-based monitoring and inspection of power and utility infrastructure demonstrates digital-twin-adjacent cyber-physical monitoring at industrial scale [69]–[78]. Finally, IoT-based sensing architectures underscore the pervasive instrumentation that feeds all such systems [80]. These domains differ from plasma processing, but they share the twin’s essential architecture of synchronized sensing, modelling, and actuation. First, real-time reduced-order models of reactive and HiPIMS discharges are needed that are fast enough for the fast loop yet capture poisoning, erosion, and flux-composition dynamics with quantified error—likely hybrids of Berg-type balance models, MOR of hybrid plasma simulations, and learned residuals [10], [11], [25], [36], [38]. Second, estimator design for partially observed, multi-rate sputtering must establish which combinations of diagnostics render the controlled states identifiable, and how ensemble/non-linear Kalman methods perform under realistic fouling and noise [22], [40], [41]. Third, control-policy research should benchmark MPC against RL for the reactive-branch and uniformity problems, quantifying the safety, interpretability, and performance trade-offs and developing the twin-as-simulator training methodology that fusion-plasma control has validated [43], [44], [46], [47]. Fourth, spatially distributed control for multi-source, large-area lines remain largely unaddressed and is where the industrial payoff concentrates [1], [38]. Fifth, uncertainty quantification and verification-and- validation methodology tailored to actuating twins on production tools must mature from principle to practice [33], [35], [42]. Sixth, standardization and data infrastructure—sensor interfaces, model exchange, and latency-bounded computing—will determine whether such twins are transferable across tools and vendors rather than bespoke [34]. Two broader observations frame this roadmap. The demonstrated feasibility of deep-RL control of confined fusion plasmas [46], [47] is a strong existence proof that real-time, model-and- learning-based control of strongly non-linear plasmas is achievable, and processing plasmas— lower energy, better instrumented, and economically motivated—are arguably a more tractable target. At the same time, the digital-twin literature’s insistence that a twin is defined by a closed, purposeful, live coupling [14], [15], [35] is a useful discipline against over-claiming: much of what is marketed as a sputtering “digital twin” today is, in the taxonomy’s precise terms, a digital shadow. The contribution of a conceptual framework such as this one is to specify, layer by layer and loop by loop, what closing that coupling actually requires. 9. Conclusion Large-area magnetron sputtering is a technology of enormous industrial importance whose control has outgrown the single-loop feedback that has served it for decades. The process is a distributed, strongly non-linear, multi-time-scale, partially observed plasma in which reactive hysteresis, target-erosion drift, uniformity coupling, and fast instabilities interact to challenge conventional control [1], [8], [10], [16]. This paper has conceptualized a closed-loop digital twin as a principled response. We reviewed the discharge fundamentals and instability phenomena that define the control problem, the diagnostic instruments that can supply a twin with real-time observations under demanding industrial constraints, and the digital-twin, model-order-reduction, and data- assimilation foundations on which such a system rests. We then proposed a concrete, five-layer architecture—physical asset, sensing and data pipeline, hybrid twin models, Kalman-family state estimation, and MPC/RL control—governed by an explicit nested latency budget, and we mapped it onto four canonical use cases in uniformity, stoichiometry, arc/instability management, and target-erosion compensation. Finally, we set out the implementation challenges, a staged model- to-shadow-to-twin validation pathway, and a research roadmap. The central thesis is that the enabling components now exist in adjacent, mature fields, and that the outstanding work is their disciplined integration—fast hybrid models, identifiable estimation, safe predictive and learning- based control, and rigorous validation—around the specific physics of the sputtering discharge. Realizing this integration would transform large-area sputtering from a process that is stabilized reactively into one that is understood, predicted, and optimized in real time, with direct consequences for the yield, quality, and material efficiency of the coatings on which photovoltaics, displays, and architectural glass depend. References [1] P. J. Kelly and R. D. Arnell, “Magnetron sputtering: a review of recent developments and applications,” Vacuum, vol. 56, no. 3, pp. 159–172, 2000, doi: 10.1016/S0042-207X(99)00189-X. [2] G. Bräuer, B. Szyszka, M. Vergöhl, and R. Bandorf, “Magnetron sputtering – Milestones of 30 years,” Vacuum, vol. 84, no. 12, pp. 1354–1359, 2010, doi: 10.1016/j.vacuum.2009.12.014. [3] J. E. Greene, “Review Article: Tracing the recorded history of thin-film sputter deposition: From the 1800s to 2017,” Journal of Vacuum Science & Technology A, vol. 35, no. 5, art. 05C204, 2017, doi: 10.1116/1.4998940. [4] J. T. Gudmundsson, “Physics and technology of magnetron sputtering discharges,” Plasma Sources Science and Technology, vol. 29, no. 11, art. 113001, 2020, doi: 10.1088/1361- 6595/abb7bd. [5] P. Sigmund, “Theory of Sputtering. I. Sputtering Yield of Amorphous and Polycrystalline Targets,” Physical Review, vol. 184, no. 2, pp. 383–416, 1969, doi: 10.1103/PhysRev.184.383. [6] R. K. Waits, “Planar magnetron sputtering,” Journal of Vacuum Science and Technology, vol. 15, no. 2, pp. 179–187, 1978, doi: 10.1116/1.569451. [7] B. Window and N. Savvides, “Charged particle fluxes from planar magnetron sputtering sources,” Journal of Vacuum Science & Technology A, vol. 4, no. 2, pp. 196–202, 1986, doi: 10.1116/1.573470. [8] J. A. Thornton, “Influence of apparatus geometry and deposition conditions on the structure and topography of thick sputtered coatings,” Journal of Vacuum Science and Technology, vol. 11, no. 4, pp. 666–670, 1974, doi: 10.1116/1.1312732. [9] W. D. Sproul, D. J. Christie, and D. C. Carter, “Control of reactive sputtering processes,” Thin Solid Films, vol. 491, no. 1–2, pp. 1–17, 2005, doi: 10.1016/j.tsf.2005.05.022. [10] S. Berg and T. Nyberg, “Fundamental understanding and modeling of reactive sputtering processes,” Thin Solid Films, vol. 476, no. 2, pp. 215–230, 2005, doi: 10.1016/j.tsf.2004.10.051. [11] K. Strijckmans, R. Schelfhout, and D. Depla, “Tutorial: Hysteresis during the reactive magnetron sputtering process,” Journal of Applied Physics, vol. 124, no. 24, art. 241101, 2018, doi: 10.1063/1.5042084. [12] J. Musil, P. Baroch, J. Vlček, K. H. Nam, and J. G. Han, “Reactive magnetron sputtering of thin films: present status and trends,” Thin Solid Films, vol. 475, no. 1–2, pp. 208–218, 2005, doi: 10.1016/j.tsf.2004.07.041. [13] D. Depla and S. Mahieu, Eds., Reactive Sputter Deposition (Springer Series in Materials Science, vol. 109). Berlin, Germany: Springer, 2008. [14] V. Kouznetsov, K. Macák, J. M. Schneider, U. Helmersson, and I. Petrov, “A novel pulsed magnetron sputter technique utilizing very high target power densities,” Surface and Coatings Technology, vol. 122, no. 2–3, pp. 290–293, 1999, doi: 10.1016/S0257-8972(99)00292-3. [15] K. Sarakinos, J. Alami, and S. Konstantinidis, “High power pulsed magnetron sputtering: A review on scientific and engineering state of the art,” Surface and Coatings Technology, vol. 204, no. 11, pp. 1661–1684, 2010, doi: 10.1016/j.surfcoat.2009.11.013. [16] J. T. Gudmundsson, N. Brenning, D. Lundin, and U. Helmersson, “High power impulse magnetron sputtering discharge,” Journal of Vacuum Science & Technology A, vol. 30, no. 3, art. 030801, 2012, doi: 10.1116/1.3691832. [17] D. Lundin and K. Sarakinos, “An introduction to thin film processing using high-power impulse magnetron sputtering,” Journal of Materials Research, vol. 27, no. 5, pp. 780–792, 2012, doi: 10.1557/jmr.2012.8. [18] A. Anders, “Deposition rates of high power impulse magnetron sputtering: Physics and economics,” Journal of Vacuum Science & Technology A, vol. 28, no. 4, pp. 783–790, 2010, doi: 10.1116/1.3299267. [19] A. Anders, “Tutorial: Reactive high power impulse magnetron sputtering (R-HiPIMS),” Journal of Applied Physics, vol. 121, no. 17, art. 171101, 2017, doi: 10.1063/1.4978350. [20] A. Anders, “A review comparing cathodic arcs and high power impulse magnetron sputtering ,” Surface and Coatings Technology, vol. 257, pp. 308–325, 2014, doi: 10.1016/j.surfcoat.2014.08.043. [21] D. Lundin and J. T. Gudmundsson, Eds., High Power Impulse Magnetron Sputtering: Fundamentals, Technologies, Challenges and Applications. Amsterdam, The Netherlands: Elsevier, 2020. [22] M. A. Lieberman and A. J. Lichtenberg, Principles of Plasma Discharges and Materials Processing, 2nd ed. Hoboken, NJ, USA: Wiley, 2005. [23] I. H. Hutchinson, Principles of Plasma Diagnostics, 2nd ed. Cambridge, U.K.: Cambridge Univ. Press, 2002. [24] J. W. Coburn and M. Chen, “Optical emission spectroscopy of reactive plasmas: A method for correlating emission intensities to reactive particle density,” Journal of Applied Physics, vol. 51, no. 6, pp. 3134–3136, 1980, doi: 10.1063/1.328060. [25] I. Adamovich et al., “The 2022 Plasma Roadmap: low temperature plasma science and technology,” Journal of Physics D: Applied Physics, vol. 55, no. 37, art. 373001, 2022, doi: 10.1088/1361-6463/ac5e1c. [26] A. Mesbah and D. B. Graves, “Machine learning for modeling, diagnostics, and control of non-equilibrium plasmas,” Journal of Physics D: Applied Physics, vol. 52, no. 30, art. 30LT02, 2019, doi: 10.1088/1361-6463/ab1f3f. [27] E. H. Glaessgen and D. S. Stargel, “The Digital Twin Paradigm for Future NASA and U.S. Air Force Vehicles,” in Proc. 53rd AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics and Materials Conf., Honolulu, HI, USA, 2012, art. 2012-1818, doi: 10.2514/6.2012-1818. [28] F. Tao, H. Zhang, A. Liu, and A. Y. C. Nee, “Digital Twin in Industry: State-of-the-Art,” IEEE Transactions on Industrial Informatics, vol. 15, no. 4, pp. 2405–2415, 2019, doi: 10.1109/TII.2018.2873186. [29] E. Negri, L. Fumagalli, and M. Macchi, “A Review of the Roles of Digital Twin in CPS-based Production Systems,” Procedia Manufacturing, vol. 11, pp. 939–948, 2017, doi: 10.1016/j.promfg.2017.07.198. [30] W. Kritzinger, M. Karner, G. Traar, J. Henjes, and W. Sihn, “Digital Twin in manufacturing: A categorical literature review and classification,” IFAC-PapersOnLine, vol. 51, no. 11, pp. 1016– 1022, 2018, doi: 10.1016/j.ifacol.2018.08.474. [31] D. Jones, C. Snider, A. Nassehi, J. Yon, and B. Hicks, “Characterising the Digital Twin: A systematic literature review,” CIRP Journal of Manufacturing Science and Technology, vol. 29, pp. 36–52, 2020, doi: 10.1016/j.cirpj.2020.02.002. [32] C. Semeraro, M. Lezoche, H. Panetto, and M. Dassisti, “Digital twin paradigm: A systematic literature review,” Computers in Industry, vol. 130, art. 103469, 2021, doi: 10.1016/j.compind.2021.103469. [33] A. Rasheed, O. San, and T. Kvamsdal, “Digital Twin: Values, Challenges and Enablers From a Modeling Perspective,” IEEE Access, vol. 8, pp. 21980–22012, 2020, doi: 10.1109/ACCESS.2020.2970143. [34] A. Fuller, Z. Fan, C. Day, and C. Barlow, “Digital Twin: Enabling Technologies, Challenges and Open Research,” IEEE Access, vol. 8, pp. 108952–108971, 2020, doi: 10.1109/ACCESS.2020.2998358. [35] L. Wright and S. Davidson, “How to tell the difference between a model and a digital twin,” Advanced Modeling and Simulation in Engineering Sciences, vol. 7, art. 13, 2020, doi: 10.1186/s40323-020-00147-4. [36] M. Raissi, P. Perdikaris, and G. E. Karniadakis, “Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations,” Journal of Computational Physics, vol. 378, pp. 686–707, 2019, doi: 10.1016/j.jcp.2018.10.045. [37] S. L. Brunton, J. L. Proctor, and J. N. Kutz, “Discovering governing equations from data by sparse identification of nonlinear dynamical systems,” Proceedings of the National Academy of Sciences, vol. 113, no. 15, pp. 3932–3937, 2016, doi: 10.1073/pnas.1517384113. [38] P. Benner, S. Gugercin, and K. Willcox, “A Survey of Projection-Based Model Reduction Methods for Parametric Dynamical Systems,” SIAM Review, vol. 57, no. 4, pp. 483–531, 2015, doi: 10.1137/130932715. [39] C. E. Rasmussen and C. K. I. Williams, Gaussian Processes for Machine Learning. Cambridge, MA, USA: MIT Press, 2006. [40] R. E. Kalman, “A New Approach to Linear Filtering and Prediction Problems,” Transactions of the ASME—Journal of Basic Engineering, vol. 82, no. 1, pp. 35–45, 1960, doi: 10.1115/1.3662552. [41] G. Evensen, “The Ensemble Kalman Filter: theoretical formulation and practical implementation,” Ocean Dynamics, vol. 53, no. 4, pp. 343–367, 2003, doi: 10.1007/s10236-003- 0036-9. [42] M. G. Kapteyn, J. V. R. Pretorius, and K. E. Willcox, “A probabilistic graphical model foundation for enabling predictive digital twins at scale,” Nature Computational Science, vol. 1, no. 5, pp. 337–347, 2021, doi: 10.1038/s43588-021-00069-0. [43] D. Q. Mayne, “Model predictive control: Recent developments and future promise,” Automatica, vol. 50, no. 12, pp. 2967–2986, 2014, doi: 10.1016/j.automatica.2014.10.128. [44] R. S. Sutton and A. G. Barto, Reinforcement Learning: An Introduction, 2nd ed. Cambridge, MA, USA: MIT Press, 2018. [45] V. Mnih et al., “Human-level control through deep reinforcement learning,” Nature, vol. 518, no. 7540, pp. 529–533, 2015, doi: 10.1038/nature14236. [46] J. Degrave et al., “Magnetic control of tokamak plasmas through deep reinforcement learning,” Nature, vol. 602, no. 7897, pp. 414–419, 2022, doi: 10.1038/s41586-021-04301-9. [47] J. Seo et al., “Avoiding fusion plasma tearing instability with deep reinforcement learning,” Nature, vol. 626, no. 8000, pp. 746–751, 2024, doi: 10.1038/s41586-024-07024-9. [48] N. M. Komi and I. O. Adeniji, “A tiered digital twin that brings predictive diagnostics to resource-constrained renewable energy sites,” International Journal of Engineering and Modern Technology, vol. 9, no. 3, pp. 287–347, 2023, doi: 10.56201/ijemt.v9.no3.2023.pg287.347. [49] P. N. Ike, J. S. Okojie, S. B. Nnabueze, J. O. O. Idu, O. M. Filani, and S. I. Ihwughwavwe, “Digital twin-driven environmental compliance models for sustainable procurement in oil, gas, and utilities,” International Journal of Advanced Multidisciplinary Research and Studies, vol. 5, no. 5, pp. 562–576, 2025. [50] N. M. Komi and O. S. Ganiu, “Edge intelligence for resilient microgrid control: Advances, energy sovereignty, and open challenges,” Gyanshauryam, International Scientific Refereed Research Journal, vol. 6, no. 3, pp. 525–586, 2023, doi: 10.32628/GISRRJ236339. [51] O. F. Akinlade, O. M. Filani, and P. S. Nwachukwu, “Automation and digital twins framework reducing procurement errors and turnaround time,” International Journal of Scientific Research in Humanities and Social Sciences, vol. 1, no. 1, pp. 197–216, 2024. [52] N. M. Komi and A. Adamolekun, “Hierarchical, decentralized, or hybrid? A systematic review of control architectures for distributed energy resources,” International Journal of Engineering and Modern Technology, vol. 8, no. 5, pp. 141–195, 2022, doi: 10.56201/ijemt.v8.no5.2022.pg141.195. [53] A. Tonoyan, O. Dada, and S. S. Ayivi-Donkor, “Advances in demand forecasting: Machine learning algorithms for revenue projection and financial planning accuracy,” Gyanshauryam, International Scientific Refereed Research Journal, vol. 4, no. 1, pp. 282–317, 2021, doi: 10.32628/GISRRJ120361. [54] A. Tonoyan, O. Dada, and S. S. Ayivi-Donkor, “Advances in geospatial analysis: Predictive analytics methods for store location selection and market entry return on investment,” International Journal of Marketing and Communication Studies, vol. 6, no. 2, pp. 78–105, 2022, doi: 10.56201/ijmcs.v6.no2.2022.pg78.105. [55] A. Tonoyan, O. Dada, and S. S. Ayivi-Donkor, “Real-time KPI tracking systems: A review of automated performance monitoring and data-driven decision making,” World Journal of Innovation and Modern Technology, vol. 8, no. 6, pp. 247–281, 2024, doi: 10.56201/wjimt.v8.no6.2024.pg247.281. [56] R. Shah, S. T. Oziri, and O. B. Seyi-Lande, “A framework for leveraging artificial intelligence in strategic business decision-making,” Gulf Journal of Advance Business Research, vol. 3, no. 11, pp. 1517–1558, 2025. [57] S. B. Nnabueze, J. K. Sakyi, O. M. Filani, J. S. Okojie, O. M. Babatope, and A. Adanyin, “Transforming utility and service operations through automation, data-driven analytics, and customer-centric innovation,” International Journal of Multidisciplinary Evolutionary Research, vol. 4, no. 2, pp. 40–57, 2023, doi: 10.54660/IJMER.2023.4.2.40-57. [58] J. K. Sakyi, S. B. Nnabueze, O. M. Filani, J. S. Okojie, and O. M. Babatope, “Digital transformation in service delivery leveraging automation and risk reduction for long-term commercial efficiency,” International Journal of Advanced Multidisciplinary Research and Studies, vol. 4, no. 4, pp. 1446–1464, 2024. [59] O. Akanbi and E. A. Sunday, “An integrated path-planning and slotting optimization model for AMR-enabled high-density warehousing,” International Journal of Advanced Multidisciplinary Research and Studies, vol. 4, no. 6, pp. 3226–3243, 2024, doi: 10.62225/2583049X.2024.4.6.6181. [60] O. Akanbi, O. S. Ganiu, and E. A. Sunday, “A multi-agent AI framework for swarm intelligence in autonomous mobile robot fleet coordination,” International Journal of Scientific Research in Humanities and Social Sciences, vol. 2, no. 1, pp. 81–109, 2025. [61] O. Akanbi and E. A. Sunday, “A systematic review of AI-driven autonomous mobile robots in scalable micro-fulfillment centers,” International Journal of Advanced Multidisciplinary Research and Studies, vol. 5, no. 6, pp. 2363–2379, 2025, doi: 10.62225/2583049X.2025.5.6.6182. [62] K. S. Ahmed and O. D. Odejobi, “Resource allocation model for energy-efficient virtual machine placement in data centers,” IRE Journals, vol. 2, no. 3, pp. 1–10, 2018. [63] K. S. Ahmed and O. D. Odejobi, “Conceptual framework for scalable and secure cloud architectures for enterprise messaging,” IRE Journals, vol. 2, no. 1, pp. 1–15, 2018. [64] O. D. Odejobi and K. S. Ahmed, “Performance evaluation model for multi-tenant Microsoft 365 deployments under high concurrency,” IRE Journals, vol. 1, no. 11, pp. 92–107, 2018. [65] O. D. Odejobi, N. I. Hammed, and K. S. Ahmed, “Approximation complexity model for cloud-based database optimization problems,” IRE Journals, vol. 2, no. 9, pp. 1–10, 2019. [66] K. S. Ahmed, O. D. Odejobi, and T. O. Oshoba, “Algorithmic model for constraint satisfaction in cloud network resource allocation,” IRE Journals, vol. 2, no. 12, pp. 516–532, 2019. [67] K. S. Ahmed, O. D. Odejobi, and T. O. Oshoba, “Predictive model for cloud resource scaling using machine learning techniques,” Journal of Frontiers in Multidisciplinary Research, vol. 1, no. 1, pp. 173–183, 2020. [68] A. I. Adeniyi, O. Odejobi, and T. Taiwo, “Countermeasures against bias and spoofing in modern facial recognition systems,” World Journal of Advanced Research and Reviews, vol. 25, no. 1, pp. 1914–1930, 2025. [69] D. Dagodzo and M. C. Ahiaeke Patrick, “A framework for national-scale UAV deployment in power infrastructure: Lessons from developing economies,” International Journal of Scientific Research in Computer Science, Engineering and Information Technology, vol. 11, no. 4, pp. 779– 824, 2025, doi: 10.32628/CSEIT251116286. [70] D. Dagodzo and M. C. Ahiaeke Patrick, “A review of right-of-way encroachment detection methods using geospatial technologies,” International Journal of Scientific Research in Computer Science, Engineering and Information Technology, vol. 8, no. 1, pp. 638–667, 2022, doi: 10.32628/CSEIT2281226. [71] D. Dagodzo and M. C. Ahiaeke Patrick, “A framework for integrating drone operations into enterprise GIS systems for utility companies,” International Journal of Scientific Research in Computer Science, Engineering and Information Technology, vol. 9, no. 4, pp. 924–957, 2023, doi: 10.32628/CSEIT23564536. [72] D. Dagodzo and M. C. Ahiaeke Patrick, “A review of UAV regulatory frameworks in developing economies: Progress, gaps, and recommendations,” International Journal of Scientific Research in Computer Science, Engineering and Information Technology, vol. 9, no. 4, pp. 958– 994, 2023, doi: 10.32628/CSEIT23564537. [73] D. Dagodzo and M. C. Ahiaeke Patrick, “An integrated framework for UAV, LiDAR, and GIS in infrastructure corridor management,” International Journal of Scientific Research in Computer Science, Engineering and Information Technology, vol. 7, no. 5, pp. 497–524, 2021, doi: 10.32628/CSEIT217566. [74] D. Dagodzo, M. C. Ahiaeke Patrick, and C. Aliliele, “AI and deep learning for vegetation classification in power corridor management: A review,” International Journal of Scientific Research in Computer Science, Engineering and Information Technology, vol. 8, no. 1, pp. 668– 698, 2022, doi: 10.32628/CSEIT2281227. [75] D. Dagodzo, “A conceptual framework for UAV integration into national power grid inspection programs,” IRE Journals, vol. 2, no. 5, pp. 391–412, 2018, doi: 10.64388/IREV2I5- 1716082. [76] D. Dagodzo, “A review of UAV applications in electrical transmission line inspection: Methods, technologies, and challenges,” IRE Journals, vol. 2, no. 6, pp. 234–254, 2018, doi: 10.64388/IREV2I6-1716083. [77] D. Dagodzo and M. C. Ahiaeke Patrick, “UAV-based pipeline and corridor monitoring: A review of current practices and emerging technologies,” IRE Journals, vol. 3, no. 10, pp. 574–597, 2020, doi: 10.64388/IREV3I10-1716084. [78] D. Dagodzo and M. C. Ahiaeke Patrick, “A review of GIS applications in utility asset management and infrastructure planning,” IRE Journals, vol. 5, no. 3, pp. 468–492, 2021, doi: 10.64388/IREV5I3-1716085. [79] S. Borah, K. C. Aliliele, S. Rakshit, and N. R. Vajjhala, “Applications of artificial intelligence in software testing,” in Cognitive Informatics and Soft Computing, Lecture Notes in Networks and Systems, vol. 375. Singapore: Springer, 2022, pp. 727–736, doi: 10.1007/978-981-16-8763-1_60. [80] D. Ussher-Eke, O. F. Okoh, and O. M. Ijiga, “The role of biometric and IoT-based attendance systems in streamlining HR administrative functions, enhancing workforce accountability, and reducing labor inefficiencies,” International Journal for Multidisciplinary Research, vol. 7, no. 4, 2025. [81] O. Ganiu, “Design and simulation of an ocean wave energy contact separation mode triboelectric nanogenerator using finite element methods,” M.S. thesis, North Carolina Agricultural and Technical State Univ., Greensboro, NC, USA, 2025. [82] R. Tonneau, A. Pflug, and S. Lucas, “Magnetron sputtering: Determining scaling relations towards real power discharges using 3D particle-in-cell Monte Carlo models,” Plasma Sources Science and Technology, vol. 29, no. 11, art. 115007, 2020, doi: 10.1088/1361-6595/abb3a0. [83] S. B. Q. Tran, F. Y. Leong, R. Hariharaputran, and D. V. Le, “A fluid model of pulsed direct current planar magnetron discharge,” Scientific Reports, vol. 13, art. 9017, 2023, doi: 10.1038/s41598-023-36231-z. [84] C. Costin, L. Marques, G. Popa, and G. Gousset, “Two-dimensional fluid approach to the dc magnetron discharge,” Plasma Sources Science and Technology, vol. 14, no. 1, pp. 168–176, 2005, doi: 10.1088/0963-0252/14/1/018. [85] A. Pflug, M. Siemers, T. Melzig, L. Schäfer, and G. Bräuer, “Simulation of linear magnetron discharges in 2D and 3D,” Surface and Coatings Technology, vol. 260, pp. 411–416, 2014, doi: 10.1016/j.surfcoat.2014.09.042. [86] T. Nyberg, H. Högberg, G. Greczynski, and S. Berg, “A simple model for non-saturated reactive sputtering processes,” Thin Solid Films, vol. 688, art. 137413, 2019, doi: 10.1016/j.tsf.2019.137413. [87] S. Berg, E. Särhammar, and T. Nyberg, “Upgrading the ‘Berg-model’ for reactive sputtering processes,” Thin Solid Films, vol. 565, pp. 186–192, 2014, doi: 10.1016/j.tsf.2014.02.063. [88] A. Palmero, E. D. van Hattum, H. Rudolph, and F. H. P. M. Habraken, “Characterization of an Ar/O2 magnetron sputtering plasma using a Langmuir probe and an energy resolved mass spectrometer,” Thin Solid Films, vol. 494, no. 1–2, pp. 18–22, 2006, doi: 10.1016/j.tsf.2005.07.154. [89] M. A. Raadu, I. Axnäs, J. T. Gudmundsson, C. Huo, and N. Brenning, “An ionization region model for high-power impulse magnetron sputtering discharges,” Plasma Sources Science and Technology, vol. 20, no. 6, art. 065007, 2011, doi: 10.1088/0963-0252/20/6/065007. [90] H. Eliasson, M. Rudolph, N. Brenning, H. Hajihoseini, M. Zanáška, M. J. Adriaans, M. A. Raadu, T. M. Minea, J. T. Gudmundsson, and D. Lundin, “Modeling of high power impulse magnetron sputtering discharges with graphite target,” Plasma Sources Science and Technology, vol. 30, no. 11, art. 115017, 2021, doi: 10.1088/1361-6595/ac352c. [91] J. Zgheib, P.-Y. Jouan, and A. Rhallabi, “A high-power impulse magnetron sputtering global model for argon plasma–chromium target interactions,” Journal of Vacuum Science & Technology A, vol. 39, no. 4, art. 043004, 2021, doi: 10.1116/6.0000865. [92] A. D. Bonzanini, K. Shao, D. B. Graves, S. Hamaguchi, and A. Mesbah, “Foundations of machine learning for low-temperature plasmas: Methods and case studies,” Plasma Sources Science and Technology, vol. 32, no. 2, art. 024003, 2023, doi: 10.1088/1361-6595/acb28c. [93] J. Trieschmann, L. Vialetto, and T. Gergs, “Review: Machine learning for advancing low- temperature plasma modeling and simulation,” Journal of Micro/Nanopatterning, Materials, and Metrology, vol. 22, no. 4, art. 041504, 2023, doi: 10.1117/1.JMM.22.4.041504. [94] D. M. Fébba, K. R. Talley, K. Johnson, S. Schaefer, S. R. Bauers, J. S. Mangum, R. W. Smaha, and A. Zakutayev, “Autonomous sputter synthesis of thin film nitrides with composition controlled by Bayesian optimization of optical plasma emission,” APL Materials, vol. 11, no. 7, art. 071119, 2023, doi: 10.1063/5.0159406. [95] A. Yanguas-Gil and J. W. Elam, “Surrogate models to optimize plasma-assisted atomic layer deposition in high aspect ratio features,” Physics of Plasmas, vol. 32, no. 7, art. 073507, 2025, doi: 10.1063/5.0275766. [96] T. Xiao and D. Ni, “Multiscale modeling and neural network model based control of a plasma etch process,” Chemical Engineering Research and Design, vol. 164, pp. 113–124, 2020, doi: 10.1016/j.cherd.2020.09.013. [97] G. E. Karniadakis, I. G. Kevrekidis, L. Lu, P. Perdikaris, S. Wang, and L. Yang, “Physics- informed machine learning,” Nature Reviews Physics, vol. 3, pp. 422–440, 2021, doi: 10.1038/s42254-021-00314-5. [98] S. Cuomo, V. Schiano Di Cola, F. Giampaolo, G. Rozza, M. Raissi, and F. Piccialli, “Scientific machine learning through physics-informed neural networks: Where we are and what’s next,” Journal of Scientific Computing, vol. 92, no. 3, art. 88, 2022, doi: 10.1007/s10915-022- 01939-z. [99] C. W. Rowley and S. T. M. Dawson, “Model reduction for flow analysis and control,” Annual Review of Fluid Mechanics, vol. 49, pp. 387–417, 2017, doi: 10.1146/annurev-fluid-010816- 060042. [100] E. A. Wan and R. van der Merwe, “The unscented Kalman filter for nonlinear estimation,” in Proc. IEEE Adaptive Systems for Signal Processing, Communications, and Control Symp., 2000, pp. 153–158, doi: 10.1109/ASSPCC.2000.882463. [101] S. J. Qin and T. A. Badgwell, “A survey of industrial model predictive control technology,” Control Engineering Practice, vol. 11, no. 7, pp. 733–764, 2003, doi: 10.1016/S0967- 0661(02)00186-7. [102] V. S. Devarakonda, W. Sun, X. Tang, and Y. Tian, “Recent advances in reinforcement learning for chemical process control,” Processes, vol. 13, no. 6, art. 1791, 2025, doi: 10.3390/pr13061791. [103] T. H. Oh, “Quantitative comparison of reinforcement learning and data-driven model predictive control for chemical and biological processes,” Computers & Chemical Engineering, vol. 181, art. 108558, 2024, doi: 10.1016/j.compchemeng.2023.108558. [104] M. Perno, L. Hvam, and A. Haug, “Implementation of digital twins in the process industry: A systematic literature review of enablers and barriers,” Computers in Industry, vol. 134, art. 103558, 2022, doi: 10.1016/j.compind.2021.103558. [105] F. Tao and M. Zhang, “Digital twin shop-floor: A new shop-floor paradigm towards smart manufacturing,” IEEE Access, vol. 5, pp. 20418–20427, 2017, doi: 10.1109/ACCESS.2017.2756069.