The research team, led by Muhammad Shahid Mastoi and Delin Wang, targeted permanent magnet synchronous generators, or PMSGs, which dominate direct-drive wind turbines because of their high efficiency, wide speed range, and excellent low-speed performance. These machines require sophisticated converter interfaces to meet modern grid codes, including low-voltage and high-voltage ride-through requirements, dynamic reactive power injection, and strict harmonic limits. Conventional two-level back-to-back converters struggle here, suffering from high harmonic distortion, elevated switching losses, greater voltage stress on semiconductor devices, and larger electromagnetic interference.

The researchers selected a four-level nested neutral-point-clamped topology, known as 4L-NNPC, as their hardware platform. In this arrangement, each semiconductor device only needs to block about one-third of the DC-link voltage, compared with the full DC-link voltage in a two-level design. That reduction allows lower-rated, faster-switching devices, cuts the rate of voltage change that stresses insulation and bearings, and paves the way toward smaller LCL filters. The trade-off is complexity: each phase leg of the converter offers 64 theoretical switching combinations, and the three-phase system must evaluate 216 admissible states at every sampling instant to keep the flying capacitors balanced and the output waveform clean.

That combinatorial explosion is precisely where classical finite control set model predictive control, or FCS-MPC, hits a wall. In conventional FCS-MPC, the controller predicts the future behavior of currents and voltages for every possible switching state and picks the one that minimizes a cost function. It is elegant and fast-responding, but the computational burden grows steeply with the number of converter levels, and performance depends heavily on manually tuned weighting factors. For a four-level converter running at 10 kilohertz, exhaustive evaluation simply does not fit comfortably inside the 100-microsecond control window of a typical digital signal processor.

The team’s solution, called Neural-Augmented MPC, replaces the brute-force search with two lightweight neural surrogates. A multilayer perceptron with two hidden layers of 128 and 64 neurons predicts the one-step-ahead state of the system, while a compact convolutional neural network estimates and ranks the cost of candidate switching vectors. Only the top 10 to 15 candidates, out of 216, are then evaluated by the full predictive cost function. The networks were trained offline on roughly 1.5 million samples generated from high-fidelity simulations spanning torque ramps, grid faults, and reactive power support, with domain randomization over machine parameters and wind speeds to ensure robustness.

The results are striking. Neural inference takes just 8 to 10 microseconds, and the total control cycle drops from 350 to 420 microseconds under classical FCS-MPC to below 100 microseconds, enabling genuine real-time operation at 10 kilohertz on a TI TMS320F28335 DSP. Switching-state evaluations fall by 60 to 70 percent, and the memory footprint shrinks nearly fourfold after 8-bit quantization and 40 percent model pruning. Crucially, the controller does not blindly trust its neural components: a confidence-based mechanism measures the margin between the best and second-best ranked candidates, and if confidence falls below a threshold, the system falls back to conventional FCS-MPC for that instant. A safety projection layer further guarantees that only physically admissible switching vectors are ever applied.

Performance gains extend well beyond computation. In simulations of a 1.5-megawatt turbine with a 2.5-kilovolt DC link, the neural-augmented controller achieved a torque settling time of 21 milliseconds, nearly half the 40 milliseconds of conventional FCS-MPC, with overshoot below 1 percent. Grid current total harmonic distortion fell to 1.9 percent, comfortably compliant with IEEE 519 and IEC 61000-3-6 standards, while average conversion efficiency reached 96 percent, beating both three-level and two-level benchmarks. Under simulated low-voltage and high-voltage ride-through events, DC-link voltage fluctuations stayed within 2.9 percent and capacitor voltage imbalance remained below 1 percent, compared with drift exceeding 12 percent when balancing control was disabled.

The team validated the framework on a hardware-in-the-loop platform pairing a Speedgoat real-time simulator with the DSP controller. At rated power, the measured current distortion was 1.86 percent, versus 4.79 percent for the conventional controller under identical conditions. Under turbulent, gusty wind profiles with abrupt speed swings, the neural controller kept stator currents and power flows visibly smoother, and sensitivity tests with plus-or-minus 20 percent variations in machine parameters showed consistently better robustness than the classical approach, thanks to training data that deliberately included parameter mismatch.

The authors are candid about limitations. The neural models depend on the representativeness of their training data and may degrade under genuinely unseen conditions, and the study’s validation, while rigorous, remains simulation and hardware-in-the-loop rather than a full-scale physical prototype. Stability is framed as practical boundedness under bounded prediction error rather than a formal global guarantee. Still, the work demonstrates something power electronics has long sought: a way to marry the intelligence of model predictive control with the speed of embedded neural inference, making four-level converters viable for multi-megawatt wind turbines. As offshore farms push into higher power ratings and tighter grid codes, controllers that think faster while switching smarter may prove essential to keeping the cleanest electrons flowing.

Subject of Research: Neural-augmented finite-set predictive control of a four-level back-to-back converter in grid-connected PMSG wind energy conversion systems

Article Title: Neural-augmented finite-set predictive control for a four-level back-to-back converter in grid-connected PMSG-based wind energy conversion systems

Article References: Mastoi, M. S., Wang, D., Habib, S., Gulzar, M. M., Hassan, M., Khan, B., & MA, N. (2026). Neural-augmented finite-set predictive control for a four-level back-to-back converter in grid-connected PMSG-based wind energy conversion systems. Results in Engineering, 32, Article 113349. https://doi.org/10.1016/j.rineng.2026.113349

Image Credits: AI Generated

DOI: Not provided

Keywords: wind energy, PMSG, model predictive control, multilevel converter, neural networks, power electronics, grid integration, hardware-in-the-loop, capacitor voltage balancing, harmonic distortion, finite control set MPC, renewable energy

The researchers built a single hardware-aware neural architecture search framework and used it as a level playing field for three competing multi-objective optimizers: NSGA-II, a classic evolutionary algorithm; Multi-Objective Simulated Annealing, or MOSA, which borrows its logic from the physics of cooling metals; and Multi-Objective Bayesian Optimization, or MOBO, which builds a statistical model of the search space to guess where the best designs lie. Each optimizer was tasked with tuning three popular neural architectures for keyword spotting: a plain convolutional neural network, a convolutional recurrent hybrid known as CRNN, and a depthwise separable convolutional network called DS-CNN, which is prized for delivering strong accuracy at a fraction of the parameter count.

What makes the search genuinely difficult is that the two objectives pull in opposite directions. Maximizing accuracy generally means adding layers, channels, and parameters, which inflates the serialized model size. Shrinking the model saves memory but usually costs recognition performance. Rather than collapsing these goals into a single weighted score, the team kept the search bi-objective, allowing the optimizers to map out a Pareto front: the set of designs where no model can be made more accurate without becoming larger, and no model can be made smaller without losing accuracy. This Pareto framing gives designers a menu of defensible choices rather than a single answer that silently encodes someone’s assumptions about what matters most.

To keep the comparison statistically honest, the authors ran every optimization five times and subjected the results to non-parametric significance testing, including the Kruskal-Wallis and Mann-Whitney procedures with Holm correction for multiple comparisons. This is a methodological discipline that many architecture search papers skip, and it matters, because multi-objective optimizers are stochastic: a single lucky run can make a weak algorithm look brilliant. The headline result of the study is that MOSA delivered the most reliable convergence across all three architectures, achieving the lowest mean generational distance, a metric that measures how close the found solutions are to the true optimal front. Differences in hypervolume and spread, two other standard quality indicators, were not statistically significant at this sample size, a finding the authors report with appropriate caution.

The best model produced by the framework reached 97.41 percent accuracy, a figure that would have seemed unattainable for on-device keyword spotting only a few years ago. But accuracy alone does not guarantee a deployable model, and this is where the study’s second major contribution comes in. After the search completed, the team evaluated the selected candidates on real hardware, measuring peak SRAM usage, Flash storage requirements, and on-device inference latency. These measurements fed into a Hardware-in-the-Loop Deployability Index, a screening score that separates models that merely look good on paper from models that actually fit and run fast on a target microcontroller. Crucially, the HIL evaluation screens candidates after the search rather than steering it, which keeps the expensive physical measurements from slowing down the optimization loop itself.

The hardware results carry practical weight for anyone shipping embedded voice products. On the STM32F401, one of the tightest boards in the study, the DS-CNN selected by MOBO retained the most resource headroom among the DS-CNN candidates, leaving valuable SRAM and Flash margin for the rest of an application’s firmware. In the embedded world, that headroom is not a luxury; it determines whether a product team can add a second sensor pipeline, a firmware update mechanism, or a more sophisticated wake-word vocabulary later. A model that consumes 95 percent of available memory may benchmark beautifully and still be unshippable, which is precisely the failure mode the Deployability Index is designed to catch.

The final stage of the pipeline addresses a subtler question: given a Pareto front full of trade-off candidates, which one should a designer actually pick? The authors applied a Tchebycheff scalarization step, a technique from multi-objective decision theory that ranks the models according to a stated preference between accuracy and size. Instead of pretending there is one objectively best model, the method makes the designer’s priorities explicit and then identifies the candidate that best honors them. This separation of concerns, generating the trade-off frontier first and applying preferences second, reflects a maturing view of how automated design tools should serve human engineers rather than replace their judgment.

Underlying the whole effort is a quiet revolution in how neural networks are quantized and compressed. The study notes that moving from 32-bit floating point to 8-bit integer weights reduces model size roughly fourfold, which is what makes microcontroller deployment feasible at all. The search framework operates with these deployment realities in view, evaluating serialized model size rather than raw parameter counts, so the numbers the optimizers chase correspond to what will actually be written into Flash memory. The training data came from publicly available corpora, including Google Speech Commands v2 and the Multilingual Spoken Words Corpus, which supports reproducibility and opens the door for other groups to replicate the comparison.

The broader significance of the work lies in its insistence on fair comparison and real-hardware validation, two things the neural architecture search literature has often lacked. Many published NAS pipelines are evaluated on proxy tasks or simulated latency models, and the resulting architectures sometimes disappoint when they meet silicon. By grounding the evaluation in measured SRAM, Flash, and latency figures on actual boards, and by repeating every experiment with proper statistical controls, Garai and Samui have produced something rarer than a new record: a reproducible methodology for asking which optimizer serves TinyML designers best. Their answer, that MOSA converges most reliably while MOBO’s selections leave the most hardware headroom, is nuanced rather than triumphant, and that nuance is exactly what practitioners need.

As always-on voice interfaces spread into battery-powered sensors, hearing aids, industrial monitors, and smart home devices, the demand for tiny models that are simultaneously accurate, small, and fast will only intensify. This study offers a template for meeting that demand systematically: define the objectives honestly, compare optimizers under identical conditions with statistical rigor, verify the winners on real hardware, and only then apply human preferences to choose among the finalists. For the growing TinyML community, the message is that the path from a research prototype to a product that ships inside a two-dollar microcontroller is no longer a matter of trial and error. It can be searched, measured, and ranked, one carefully optimized neural network at a time.

Subject of Research: Multi-objective neural architecture search for hardware-efficient TinyML keyword spotting on microcontrollers

Article Title: Multi-objective neural architecture search for TinyML keyword spotting: a comparative study of metaheuristic and bayesian optimization with hardware-in-the-loop evaluation

Article References: Garai, S., & Samui, S. (2026). Multi-objective neural architecture search for TinyML keyword spotting: a comparative study of metaheuristic and bayesian optimization with hardware-in-the-loop evaluation. Neural Computing and Applications, 38(19), Article 784. https://doi.org/10.1007/s00521-026-12484-3

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12484-3

Keywords: TinyML, keyword spotting, neural architecture search, multi-objective optimization, Bayesian optimization, NSGA-II, simulated annealing, microcontrollers, hardware-in-the-loop, Pareto optimality, DS-CNN, embedded systems

The study, published in the journal Automotive and Engine Technology, addresses a long-standing trade-off in fuel cell simulation. Highly detailed stack models, which resolve the electrochemistry, heat transfer and fluid dynamics inside individual cells, are computationally expensive, making them impractical for full-vehicle or system-level investigations where thousands of operating seconds must be simulated. On the other end of the spectrum, reduced-order system models run quickly but often strip away the physical fidelity needed to capture how electrochemical processes respond to rapid changes in load, air supply and temperature. The new framework bridges this gap by running a physics-based stack model and a dynamic balance-of-plant model in parallel, exchanging data at every time step so that electrochemical, thermal and auxiliary losses are all accounted for within a single simulation environment.

Technically, the framework is implemented in AVL CRUISE™ M, a system simulation platform widely used in powertrain engineering, and MATLAB®, with the two tools synchronized during the simulation. The researchers evaluated their approach under representative transient operating conditions, the kind of rapidly fluctuating duty cycles that mirror real-world driving, including scenarios aligned with standard vehicle test cycles such as the Worldwide Harmonized Light Vehicles Test Cycle. A central contribution of the work is its systematic assessment of solver synchronization and coupling strategies, examining how the choice of data exchange scheme affects numerical stability, simulation accuracy and computational performance. This matters because co-simulation introduces a fundamental risk: if two solvers exchange information too infrequently or inconsistently, the coupled system can become unstable or produce physically meaningless results, particularly during sharp transients such as sudden acceleration demands.

The headline result is striking: the coupled framework achieved a real-time factor of 1.74, meaning the simulation ran faster than the physical process it represents. For automotive engineers, this figure is significant beyond mere convenience. Real-time capability is the entry ticket to hardware-in-the-loop testing, where a physical controller, such as an actual fuel cell management electronic control unit, is connected to the virtual plant and must respond to simulated signals within strict timing limits. It also opens the door to digital twin implementations, in which a running fuel cell system in a vehicle or test bench is mirrored by a live simulation model that can predict behavior, flag anomalies and support predictive maintenance. A framework that is both physically grounded and fast enough for real-time operation therefore has immediate practical value.

Equally important is what the framework reveals about where the energy actually goes. By integrating the stack model with the balance-of-plant model, the researchers performed an integrated loss analysis that links electrochemical stack losses with system-level inefficiencies. The results show that overall system efficiency declines from 49.6 percent at low load to 38.8 percent at high load, and that the relative contribution of individual loss mechanisms shifts dramatically with operating conditions. At low loads, activation losses associated with the sluggish kinetics of the oxygen reduction reaction at the cathode dominate, while at higher current densities ohmic losses from membrane resistance and concentration losses caused by mass transport limitations grow in importance. Simultaneously, the auxiliary components consume an increasing share of the produced power, since the air compressor must work harder to deliver the oxygen the stack demands at high current.

This load-dependent loss structure has direct implications for how fuel cell vehicles are designed and operated. A fuel cell powertrain that is efficient during gentle city driving may shed several percentage points of efficiency during highway cruising or hill climbing, not because the electrochemistry itself becomes dramatically worse, but because the balance of plant consumes more and the stack crosses into loss-dominated regimes. With a validated co-simulation environment, engineers can quantify these trade-offs before a single physical prototype is built. They can explore, for example, how a more efficient compressor, a revised humidification strategy or an altered operating temperature might shift the efficiency map, and how those choices interact with transient behavior during real driving profiles rather than idealized steady-state points.

The modular architecture of the framework is another distinguishing feature. Unlike existing co-simulation approaches that focus primarily on stack-internal behavior, the new design treats the stack and the balance of plant as interchangeable modules with well-defined interfaces. This means individual components can be swapped, refined or replaced without rebuilding the entire simulation. A researcher could substitute a more detailed compressor model, test a different membrane-electrode-assembly characterization, or plug in an alternative cooling concept while keeping the rest of the environment intact. Such flexibility is essential in industrial practice, where simulation tools and model detail levels evolve continuously across development programs and where different teams may be responsible for different subsystems.

The broader context makes this work timely. As automakers ramp up hydrogen programs for heavy-duty trucks, buses and passenger vehicles, the pressure to compress development cycles while meeting ever-stricter efficiency and durability targets has intensified. Fuel cell durability is highly sensitive to operating conditions: water management, membrane humidity, temperature gradients and load transients all influence degradation mechanisms. A system-level simulation environment that faithfully represents the interplay between the stack and its auxiliaries under realistic transient loads provides a virtual proving ground for control strategies that balance performance against long-term reliability. The authors specifically point to control strategy development, hardware-in-the-loop testing and digital twin implementations as the natural next applications of their framework.

In an era when the hydrogen economy is alternately hyped and dismissed, studies like this one do the unglamorous but essential work of turning chemistry into drivable engineering. By demonstrating that a physics-based stack model and a dynamic balance-of-plant model can be coupled stably, accurately and fast enough for real-time use, the Graz and Darmstadt researchers have given the fuel cell community a practical tool for answering one of its most consequential questions: not just how efficiently a membrane can convert hydrogen into electricity in isolation, but how an entire system behaves when the road demands everything at once. With system efficiency spanning nearly eleven percentage points between low and high load, and with loss mechanisms shifting beneath every change of the accelerator pedal, that system-level view may prove to be exactly what the next generation of hydrogen vehicles needs.

Beyond the headline performance figures, the publication carries several details that will interest specialists in simulation methodology. The framework’s modular design is consistent with the Functional Mock-Up Unit (FMU) standard, an industry convention that allows models developed in different tools to be packaged with standardized interfaces and exchanged between simulation environments. Adherence to such conventions is what makes a co-simulation genuinely cross-tool rather than a bespoke pairing of two specific software packages, and it explains how a physics-based stack model developed in MATLAB can coexist with a commercial powertrain platform like AVL CRUISE M without forcing either side to abandon its native solver.

The abbreviation list of the paper also hints at the physical depth retained inside the stack module. Terms such as area-specific resistance, gas diffusion layer, bipolar plate, and membrane-electrode-assembly indicate that the model resolves the layered structure of an individual cell rather than treating the stack as a black box. Likewise, the explicit distinction between the hydrogen oxidation reaction at the anode and the oxygen reduction reaction at the cathode, alongside side reactions such as hydrogen evolution, suggests that the electrochemical submodel distinguishes loss pathways that simplified models typically lump together.

Another noteworthy aspect is the choice of reference for efficiency. The reported values are given on the basis of the higher heating value, the more conservative convention that accounts for the latent heat of water vaporization. Efficiencies quoted on a higher heating value basis are systematically lower than those quoted on the lower heating value basis often used in combustion engineering, which is worth remembering when comparing the reported 49.6 percent peak against figures from other fuel cell studies.

The institutional collaboration itself is instructive. The work emerged from the Institute of Thermodynamics and Sustainable Propulsion Systems in Graz and the Institute for Internal Combustion Engines and Powertrain Systems in Darmstadt, both groups with deep roots in conventional powertrain simulation now applying that heritage to hydrogen systems. Published open access with funding from Graz University of Technology, and having passed through peer review between May and August 2026, the study is freely available to the development community it aims to serve.

Subject of Research: Integrated co-simulation of PEM fuel cell systems coupling electrochemical stack modeling with balance-of-plant dynamics for automotive applications

Article Title: Integrated co-simulation of PEM fuel cell systems for automotive applications: modeling, coupling, and system behavior

Article References: Strametz, F., & Endres, J. (2026). Integrated co-simulation of PEM fuel cell systems for automotive applications: modeling, coupling, and system behavior. Automotive and Engine Technology. https://doi.org/10.1007/s41104-026-00177-x

Image Credits: AI Generated

DOI: 10.1007/s41104-026-00177-x

Keywords: PEM fuel cell, co-simulation, balance of plant, fuel cell stack modeling, system-level simulation, real-time simulation, loss analysis, AVL CRUISE M, MATLAB, automotive powertrain, digital twin, hardware-in-the-loop