Shahin Alipour Bonab

High-temperature superconducting bulk staggered-array undulators can generate record-breaking magnetic fields for compact light sources, but their performance is limited by non-uniform magnetic field profiles. This spatial inhomogeneity is directly related to variations in the critical current density among individual rare-earth barium copper oxide (ReBCO) bulk superconductors, which arise from microstructural inconsistencies formed during the melt-growth process. To mitigate these field errors, current state-of-the-art assembly methods rely on resource-intensive experimental pre-screening and iterative sorting of individual bulks based on their trapped field capabilities. However, predicting the exact critical current scaling factors required for an optimal spatial arrangement remains computationally demanding, a problem compounded by the limited availability and high generation cost of finite element method (FEM) training data. This paper presents a novel computational framework that integrates advanced generative data augmentation to learn underlying complex trends for producing complete artificial samples and machine learning (ML) to rapidly predict both the forward magnetic field integrals and the inverse critical current scaling factors of ReBCO bulks. To identify the optimal synthetic data generation strategy under severe data constraints, 8 deep generative architectures, encompassing adversarial, variational, and diffusion-based models, were systematically benchmarked alongside conventional statistical approaches. The utilised data synthesisers demonstrated superior synthesis of the physical manifold, achieving predictive accuracies of 99% and 88%, respectively. Ultimately, the fully augmented ML surrogate models accelerated the multidimensional FEM analyses by over 4 orders of magnitude compared to traditional numerical methods, completely bypassing the need for dedicated high-performance computing resources. The proposed framework establishes a highly data-efficient surrogate modelling approach for complex electromagnetic systems. By bridging the gap between limited simulation data and high-accuracy predictions, it provides critical computational support for the high-speed sorting and field optimisation of bulk superconductors during the experimental construction of next-generation synchrotron and free-electron laser facilities.

Hydrogen (H₂) is expected to play a critical role in modern industry, particularly in ammonia synthesis, petroleum refining, and low-carbon transportation. The safe storage of H₂ remains a major challenge due to its low volumetric density under ambient conditions. Metal–Organic Frameworks (MOFs), highly porous crystalline materials, have emerged as promising H₂ storage candidates owing to their high surface areas and tuneable pore structures. Molecular simulations such as grand canonical Monte Carlo or density functional theory are costly and limited in exploring large material spaces, motivating efficient predictive tools to accelerate discovery. Here, Machine Learning (ML) techniques are compared to an explainable artificial intelligence (XAI) approach using symbolic regression (SR), trained on 10,123 experimentally measured H₂ adsorption datapoints from real-world MOFs. The best performing model achieved a goodness of fit of 0.9986 with lower computational demand, but reduced interpretability, addressed using XAI analysis and clustering. SR achieves a lower goodness of fit of 0.914 but produces a physically meaningful equation highlighting structural features driving high gravimetric efficiencies. These results demonstrate strong ML capability for predicting how MOF properties and environmental conditions affect H₂ uptake. This offers engineers and researchers a practical means of screening potential MOFs for H₂ storage applications, with the XAI analyses providing additional confidence in the predictions. They allow researchers to understand the physical reasoning behind each output, assess the reliability of individual predictions, and make fully informed decisions, enabling predictive models to be acted upon with confidence in real-world contexts.

Direct methanol fuel cells (DMFCs) represent a promising energy conversion technology that offers compactness, high efficiency, and low emissions that are suitable for portable and clean energy applications. However, high market costs and resource limitations of platinum-group metal (PGM)-based catalysts remain a significant obstacle to their commercialized adoption. Despite PGM-free catalysts attracting recent attention, their design and development remain challenging, largely due to the complex non-linear correlation between control and performance parameters. This paper presents a machine-learning-based surrogate framework to predict the polarization curve and subsequently calculate power densities based on a set of model inputs: (i) temperature, (ii) Nafion concentration, (iii) methanol concentration, (iv) catalyst loading, and (v) current densities. Four separate gradient-boosting machine learning (ML) models, i.e. eXtreme gradient boosting (XGBoost), categorical boosting (CatBoost), histogram gradient-boosting regressor (HistGBR) and a light-gradient-boosting machine (LightGBM), were trained on experimental data corresponding to the set of predefined control parameters. The performance of each model also depends on how their hyperparameters were optimized using Bayesian optimization (BO), a tree-structured Parzen estimator (TPE), and Grey-Wolf Optimization (GWO). Interpolative analysis shows that the Grey-Wolf optimized CatBoost (CatGWO)-model was the best-performing model-optimizer framework, achieving an average R ² of 99.72% and an average mean absolute relative error of 2.16% with fairly limited standard deviation. To demonstrate its performance for possible out-of-the-range DMFC design applications, the model was subjected to extrapolation study with completely unseen hold-out datasets. Compared with prior studies, our CatGWO framework delivers promising cell-voltage predictions, and demonstrated only a modest gap between hold-in and hold-out performance. The proposed model offers a novel approach to improving the accuracy of machine learning predictions at very low computational latency. Potential applications include rapid pre-screening of operating conditions prior to multi-physics simulations, deployment as a surrogate model for optimization-inverse-AI modelling, and real-time use within model predictive control systems, accelerating the design and operational viability of practical Fe–N–C-based DMFCs.

Accurately predicting AC losses in high-temperature superconductors is critical for the efficient design of superconducting systems due to their contribution to the required cooling power at cryogenic temperatures. Experimental measurements are often expensive, time-consuming, and risky (in the case of measurements under liquid hydrogen cooling), so they are limited to a few measurement tests (datapoints), and analytical models lose fidelity as system geometries and boundaries become more complicated or operating conditions deviate from their underlying assumptions. On the other hand, although traditional finite element (FE) modelling techniques, such as H -formulation FE analysis, are accurate, they tend to have higher errors and get too slow in cases of complicated coil or system-level modelling. Despite the advantages of using purely data-driven artificial intelligence (AI) based models for AC loss prediction, this paper presents a physics-constrained deep learning (PCDL) framework for the first time that not only predicts AC losses accurately but also can extrapolate them from sparse experimental data. This novel PCDL model offers trustworthy predictions of AC losses even beyond the limited measured data range. By embedding the analytical formulation of the Norris model directly into the training process, the model can constrain the AI regressor to the physics-defined direction, which can significantly improve the prediction accuracy. Trained on a limited subset of experimental data, the proposed PCDL framework provides: (i) high-accuracy interpolation within the sparse experimental envelope—matching the accuracy and performance of purely data-driven AI models—and (ii) reliable extrapolation to current amplitudes and frequencies outside the measured range, where simple AI and machine learning techniques, including conventional neural networks and mathematical fittings fail. Benchmarked against a data-driven neural network trained on the same restricted set, PCDL consistently delivers the highest accuracy with a minimum goodness of fit of 91.05% at the worst-case scenario and 98.45% at the highest performance. Therefore, PCDL was found to be a potential intelligent solution for accurate AC loss estimations beyond experimental measurements.

Hydrogen sourced from renewable resources has been increasingly regarded as a promising alternative chemical energy carrier to fossil fuels, with support being given to efforts for climate change mitigation and the achievement of net-zero emission targets. Solid oxide electrolyzer cells (SOECs) operating in higher temperatures have shown greater system-level efficiency in comparison to other types of electrolyzers, making them a promising option for mass production of hydrogen. However, diverse SOEC architectures—such as varying anode and cathode materials—and wide operating ranges make it difficult to develop a single, multi-output, multi-parameter, broadly applicable performance model. In this paper, artificial intelligence (AI) techniques have been applied to SOEC data to develop an advanced model based on a cascade-forward neural network (CFNN) that holistically captures these design and operating variables. The resulting CFNN model reliably predicts ohmic resistance, current density, and hydrogen production rate with coefficients of determination ( R ² ) of 0.9981, 0.9752, and 0.9878, respectively, in comparison with experimentally sourced testing data and across a wide range of SOEC configurations and conditions. Essentially, this paper delivers a highly accurate and computationally efficient digital surrogate model based on an AI technique, enabling smarter design, operation, and scaling of SOECs, and contributing to broader decarbonization goals through cost-effective and efficient sustainable hydrogen production.

This paper presents the first intelligent estimator model of the critical current of high-temperature superconductor (HTS) tapes exposed to gamma or neutron radiation using machine learning (ML) techniques. A comprehensive benchmarking analysis of ten ML methods has been conducted to determine the best ML models for each type of radiation. To ensure the generalisability of the models, databases of experimental measurements were collected by an extensive review of 90 published papers in the literature, covering four and nine different rare-earth barium copper oxide (ReBCO) tapes for gamma and neutron irradiation tests, respectively. The results demonstrated that the cascade-forward neural network (CFNN) excels for both gamma and neutron irradiation prediction models. For the gamma irradiation model, the CFNN model’s performance in terms of goodness of fit and relative error was 99.979% and 0.2675%, respectively. For the neutron irradiation model, these metrics have shown a performance of 99.972% and 4.68%. The findings of this paper will advance the modelling of superconductors in terms of understanding their behaviour after irradiation for fusion applications.

Thermal stresses in superconducting tapes, resulting from quench, over-current conditions, or high-current faults, can rapidly elevate temperatures, potentially causing hot spots, performance degradation, localized damage, or catastrophic burnout failure if heat dissipation is inadequately managed. Traditional numerical approaches, such as finite element analysis (FEA), accurately simulate these transient thermal behaviors but suffer from significant computational burdens, complex meshing processes, and sensitivity to input and meshing parameters. Moreover, FEA results often lack generalization capability, necessitating new simulations for each variation in problem parameters. Recently, artificial intelligence (AI) and machine learning (ML) methods have emerged as alternatives for addressing superconductivity problems. However, existing AI/ML models typically operate as black boxes, providing little insight into the physics-based relationships between inputs and outputs. To address this limitation, this study introduces a physics-informed neural network (PINN) framework specifically designed to model transient thermal behaviors in superconducting tapes. By embedding governing physics laws directly into the neural network’s training process, PINNs offer increased transparency, accuracy, and interpretability compared to conventional AI/ML models. This paper, for the first time, aims to demonstrate PINNs as an innovative modeling approach for superconductors, focusing explicitly on transient thermal responses following a heat pulse in superconducting tapes. Validation against analytical solutions and benchmark FEA simulations confirms the PINN model’s effectiveness in accurately capturing sharp thermal gradients and transient heat transfer. Furthermore, results highlight the model’s excellent generalization and extrapolation capabilities, illustrating that a trained PINN can reliably predict scenarios beyond its original training dataset, thereby offering significant potential for tackling complex challenges related to superconductors.

Sustainability in space is now the centre of attention for different space research organisations given the scale of current investment in planetary search activities, ambitious plans for habitation in future, and focus on electric space propulsion systems. One potential propulsive means for future spacecraft is the applied-field magnetoplasmadynamic thruster (AF-MPDT). This type of thruster uses the principle of the Lorentz force, where ionized gas is propelled through the interaction of a current and a magnetic field. These thrusters are characterized by nonlinear and complex interaction between controllable parameters, such as current and magnetic field, and structural attributes like part dimensions including anode and cathode radii. Consequently, traditional empirical modelling approaches have encountered challenges in predicting certain outputs, such as thrust, with sufficient precision across different operational regimes. As an alternative to analytical/empirical formulas that approximate the true physics only partially, this paper demonstrates the potential of artificial intelligence (AI) techniques to predict thrust in AF-MPDTs. Through training and meticulous hyperparameter tuning, this study compares 7 different AI models fed with experimental data from 21 thrusters and their different configurations, reaching a total of 58 thruster designs, spanning decades of thruster research and development work. Results indicate that the supervised ensemble algorithm, eXtreme Gradient Boosting (XGBoost), outperforms all other utilized techniques such as random forest, Gradient Boosting Regressor, support vector regression, kernel ridge regression, K-nearest neighbors, and Gaussian process regression. With a Goodness of Fit (R ²) of 98.55%, root mean square error of 1.421 N, and mean absolute error of 0.453 N, XGBoost specifically, and AI in general, has demonstrated its superiority, by significantly improving on the accuracy of previously published empirical models for AF-MPDT thrust prediction. Additionally, the fast response associated with these techniques further expands their applicability to real-time data operation or to being used as a subroutine in thruster design procedure. This can potentially become a fundamental component of AF-MPDT designing software, whereby it may be used to check a configuration’s functionality, applicability and feasibility all in a few milliseconds. This data-driven approach can be helpful in upscaling or specially down-scaling AF-MPDTs to make them better suited for many lower power space applications.