The team, led by Maya Houacine of the Laboratory of Research in Artificial Intelligence at the University of Science and Technology Houari Boumediene in Algiers, together with Khadidja Belattar and Naila Aziza Houacine, set out to correct what they see as persistent blind spots in the violence-detection literature. Many existing approaches lean on computationally heavy models that cannot plausibly run on the hardware guarding a real camera network. Others sidestep the timing question entirely, reporting accuracy without demonstrating that their systems can process video as it arrives. And a surprising number evaluate their methods on staged footage or cinematic fight scenes rather than the grainy, awkwardly angled, often distant video that real cameras actually produce. The Algerian group built its pipeline around real surveillance recordings and made real-time throughput a first-class design requirement rather than an afterthought.

The first stage of the system addresses a deceptively simple problem: video contains enormous redundancy. Consecutive frames in a surveillance feed are nearly identical, and training a neural network on every single frame wastes computation without adding information. The researchers compared two keyframe extraction techniques, methods that select the most informative frames from a clip while discarding duplicates, to determine which one best stripped redundancy while preserving the visual evidence of violent motion. Keyframe selection has a long history in video indexing, stretching back to early interfaces for browsing video archives in the 1990s, and modern variants rely on cues such as color histograms and adaptive clustering. Choosing the right extraction strategy matters because the quality of the frames fed into a detector directly shapes what the network can learn about fast, chaotic events like fights.

With preprocessing settled, the team adapted and evaluated three distinct deep learning architectures for the detection task. The first was an enhanced version of MobileNet, a convolutional neural network family originally engineered by Google researchers for mobile and embedded vision applications, where every parameter and every multiplication must be justified. The second combined MobileNetV2, which introduces inverted residual blocks and linear bottlenecks to squeeze more accuracy from fewer computations, with a long short-term memory network, a recurrent architecture designed to capture how visual features evolve across time, which is essential because violence is fundamentally a temporal phenomenon rather than a single frozen pose. The third contender was a three-dimensional convolutional neural network, or 3D CNN, which applies convolutions across both space and time simultaneously and is a popular choice for action recognition, though typically at a steep computational price.

Evaluation took place on two of the most challenging public benchmarks in the field. The first, RWF-2000, is a large-scale video database built specifically for violence detection using footage from real surveillance cameras, and it is openly available through the Zenodo repository. The second, UCF-Crime, comes from the University of Central Florida’s Center for Research in Computer Vision and contains real-world anomaly detection videos captured in uncontrolled settings, a subset of which the team used. The researchers supplemented these with material drawn from the Open Video Project archive. Testing on such data is deliberately unforgiving: lighting shifts, camera shake, occlusion, and the sheer variety of human behavior in public spaces all conspire to confuse models that were tuned on cleaner inputs.

A rigorous validation scheme separated this work from many predecessors. Rather than splitting the data once and reporting a single number, the team applied k-fold cross-validation to all three architectures, rotating through multiple partitions of the data so that performance estimates reflect the model’s behavior across different subsets rather than one lucky or unlucky split. This process exposed an unexpected complication in the enhanced MobileNet model: the folds were imbalanced, meaning some partitions contained uneven distributions of the classes the network needed to distinguish, which distorts both training and evaluation. Imbalanced folds are a subtle but consequential problem, because a model can appear to perform well simply because the data it sees is skewed.

The solution the researchers devised is one of the more inventive elements of the study. They turned to a genetic algorithm, a metaheuristic optimization method inspired by biological evolution, in which candidate solutions compete, mutate, and recombine over successive generations until a strong solution emerges. Their initial genetic algorithm reduced the fold imbalance cost, measured in misallocated images, from nearly 100 images down to 15. They then refined the approach into a dynamic genetic algorithm that adapted its behavior during the search, cutting the imbalance cost further to just 4 images while reaching the improved result in less time. The improvement demonstrates how classical optimization techniques can still play a decisive supporting role in modern deep learning pipelines, quietly fixing data plumbing problems that the headline models cannot solve on their own.

When the numbers settled, the MobileNet-based architecture emerged as the clear winner. It reached 90 percent accuracy, 91 percent precision, and 90 percent recall on the benchmark tasks, meaning it correctly identified violent incidents most of the time while keeping both false alarms and missed detections in check. Precision and recall matter in different ways for security applications: a system with poor precision floods operators with false alerts and quickly earns their distrust, while a system with poor recall lets genuine incidents slip through unnoticed. The MobileNet model balanced both while sustaining a processing speed of up to 130 frames per second, comfortably above the 25 to 30 frames per second that live video typically demands, leaving headroom for the other computational tasks a surveillance platform must perform.

The speed figure is the detail most likely to resonate with practitioners, because it addresses the gap between laboratory results and deployable systems. A detector that processes video slower than the cameras produce it accumulates an ever-growing backlog, and its alerts arrive after the moment to respond has passed. By building the system around efficient architectures and disciplined preprocessing, the Algerian team showed that state-of-the-art detection does not require data-center hardware. The choice of MobileNet as the backbone reflects a broader trend in computer vision toward models designed for edge devices, cameras, and embedded processors, where power budgets and memory constraints rule out the giant networks that dominate headline benchmarks.

The researchers were explicit about the ethical dimensions of their work, noting that surveillance data raises genuine privacy concerns. They mitigated these risks by relying exclusively on publicly available datasets, ensuring that no attempt was made to identify individuals, and confirming that no personally identifiable information was accessed or processed. All data were used solely for research purposes and handled in accordance with relevant privacy and data protection regulations, in line with principles of responsible data usage, data minimization, and respect for individual privacy. The code itself is not publicly available, but the datasets underpinning the study remain open, allowing other groups to reproduce the evaluation. As cities continue to multiply their camera networks and human attention stays fixed in supply, systems of this kind, fast enough to watch in real time and accurate enough to be trusted, are likely to move from research papers into the operational core of public safety infrastructure.

Subject of Research: Real-time violence detection in surveillance videos using lightweight deep learning models

Article Title: An intelligent system for real-time violence detection in surveillance videos

Article References: Houacine, M., Belattar, K., & Houacine, N. A. (2026). An intelligent system for real-time violence detection in surveillance videos. Multimedia Tools and Applications, 85(11), Article 815. https://doi.org/10.1007/s11042-026-21958-4

Image Credits: AI Generated

DOI: 10.1007/s11042-026-21958-4

Keywords: violence detection, surveillance video, deep learning, MobileNet, LSTM, 3D CNN, keyframe extraction, genetic algorithm, cross-validation, computer vision, real-time systems, RWF-2000

The study, published in the journal Results in Engineering, was led by Wei Leng, Zihan Zhao, Jiaxiang Chen, and Lei Wang of Sichuan University of Science and Engineering, working with cuttings supplied by the Oilfield Engineering Services Branch of Sinopec in Zigong, China. Rather than treating the cuttings as mere refuse, the team exploited their hidden chemistry. X-ray fluorescence analysis revealed that the dried waste is rich in silicon dioxide, at roughly 40 percent by weight, and aluminum oxide, at about 11 percent, precisely the two ingredients needed to synthesize zeolites, the porous aluminosilicate crystals that underpin applications from water softening to catalysis and pollutant adsorption.

The researchers adopted an alkali fusion-hydrothermal route to activate the otherwise inert mineral phases in the cuttings. But the process is governed by a tangle of interacting variables: the solid-to-liquid ratio, the silicon-to-aluminum ratio, the alkali-fusion temperature, and the slag-to-alkali ratio. Exhaustively testing every combination at seven levels of each factor would have required 2,401 experiments, a prohibitively expensive undertaking. Instead, the team first ran single-factor screening experiments to bracket plausible ranges, then applied a uniform design, an experimental strategy that spreads a handful of representative runs evenly across the entire factor space, cutting the workload to just seven trials.

Those seven experiments produced a modest dataset, far too small to train a neural network reliably. The researchers therefore turned to data augmentation, generating synthetic training samples by adding zero-mean Gaussian noise around the original observations. To guard against information leakage and overfitting, they wrapped the whole procedure in a nested leave-one-out cross-validation scheme: in each round, one experimental observation was held out for testing, one served as an unaugmented validation point, and the remaining five were expanded into a synthetic training set. This disciplined structure allowed the team to extract honest estimates of how their models would perform on data they had never seen.

At the heart of the work lies a well-known weakness of the back-propagation neural network, the workhorse algorithm the team chose for modeling the synthesis process. Standard back-propagation is notoriously sensitive to initialization and training parameters, and it can stall in local optima or converge erratically on noisy inputs. The classic remedies are hybrid schemes in which evolutionary algorithms tune the network’s weights and hyperparameters. Genetic algorithms, inspired by natural selection, search the solution space through selection, crossover, and mutation, while particle swarm optimization, proposed by Kennedy and Eberhart in 1995, mimics the collective movement of bird flocks to explore parameter landscapes efficiently.

The innovation in this study is that both hybrid schemes were made adaptive. In the conventional versions, crossover and mutation probabilities, or the swarm’s inertia weight and learning coefficients, are held fixed, which often causes premature convergence and the loss of high-quality solutions. The researchers’ adaptive genetic algorithm instead raises these probabilities early in evolution to maximize exploration, then lowers them later to refine the best solutions, while elite retention and tournament selection preserve the strongest candidates. The adaptive particle swarm variant similarly decreases its inertia weight and cognitive coefficient over time while increasing the social coefficient, steering the swarm cooperatively toward promising regions. Early stopping during network training further stabilized both frameworks.

The results were evaluated with a battery of statistical tools, including Bland-Altman plots, parity plots, error distributions, and violin plots of relative error. Both adaptive models outperformed their non-adaptive counterparts, showing tighter error distributions and better agreement with the ideal one-to-one prediction line. Between the two, the AGA-BP model edged ahead, achieving a coefficient of determination of 0.9088 on the training set and 0.8264 on the test set, with relative errors confined within plus or minus 0.2 across all data partitions. Stability analysis confirmed that predictions from both models clustered tightly around the reference values across the entire range of synthesis parameters, with no systematic drift.

The decisive test came when the models’ recommended synthesis conditions were put into practice. The AGA-BP network prescribed a solid-liquid ratio of 1:3.38, a silicon-to-aluminum ratio of 1.80:1, an alkali-fusion temperature of 336 degrees Celsius, and a slag-alkali ratio of 1:1.240, while the APSO-BP model suggested a somewhat different recipe. Zeolites synthesized under all three optimized conditions, including those from the conventional regression approach, showed the characteristic X-ray diffraction peaks of Na-X zeolite, confirming that the target crystalline phase had formed in every case. But the adsorption experiments told a sharper story: material made under AGA-BP conditions achieved an ammonia-nitrogen adsorption capacity of 4.8852 milligrams per gram, significantly higher than both the APSO-BP result of 4.6782 and the regression-based 4.5800, according to Tukey’s honestly significant difference testing.

Scanning electron microscopy revealed just how dramatic the transformation was. The raw cuttings appeared as a disordered jumble of bulk, granular, and needle-like particles, while the AGA-BP-optimized product was dominated by densely intergrown, interlaced lamellar crystallites riddled with interparticle voids and crevices, exactly the kind of open architecture that gives zeolites their adsorptive power. The images capture, in effect, hazardous waste being remade into a functional environmental material capable of stripping ammonia nitrogen from contaminated water.

Beyond the immediate result, the study points to a broader shift in how complex materials synthesis is conducted. By pairing frugal experimental designs with adaptive machine learning, the researchers extracted near-optimal process conditions from a dataset that would have been hopeless for conventional modeling, and their framework could in principle be applied to other waste-derived precursors, scaled to industrial production, or tested against real wastewater matrices. For an industry searching for ways to close the loop on its own waste streams, the message is compelling: the cuttings left behind by drilling for clean-burning gas may themselves become part of the solution, one carefully optimized crystal at a time.

Subject of Research: Machine learning optimization of Na-X zeolite synthesis from oil-based drilling cuttings

Article Title: Optimization of synthesis parameters for Na-X zeolite using AGA-BP and APSO-BP neural network

Article References: Leng, W., Zhao, Z., Chen, J., & Wang, L. (2026). Optimization of synthesis parameters for Na-X zeolite using AGA-BP and APSO-BP neural network. Results in Engineering, 32, Article 113386. https://doi.org/10.1016/j.rineng.2026.113386

Image Credits: AI Generated

DOI: 10.1016/j.rineng.2026.113386

Keywords: Na-X zeolite, oil-based drilling cuttings, neural network, genetic algorithm, particle swarm optimization, data augmentation, ammonia adsorption, waste valorization, hydrothermal synthesis, machine learning, X-ray diffraction, shale gas waste

A new study published in Frontiers in Water tackles this bottleneck head-on. The research team, led by scientists working across multiple regional settings, developed an enhanced Genetic Algorithm–optimized SEIR-V model, known as GA-SEIR-V, that couples the classic compartmental epidemic framework with evolutionary optimization and temperature-dependent viral decay. The work addresses the two central weaknesses of traditional approaches: their susceptibility to poor local optima during least-squares fitting, and their inability to capture the seasonal fluctuations that distort wastewater signals.

The SEIR-V framework explicitly links viral concentration in wastewater to the epidemic state. Susceptible, exposed, infectious, and recovered populations evolve according to mechanistic equations, while a viral load variable accumulates the shedding from infectious individuals. The day-to-day change in this variable represents the model’s predicted viral concentration, which can be compared directly with measured values. The challenge lies in estimating the transmission rate, shedding parameters, and initial conditions from noisy data—a high-dimensional, non-convex optimization problem where conventional gradient-based methods frequently stall at unstable solutions.

The genetic algorithm approaches this problem differently. Rather than following a single trajectory through parameter space, it maintains a population of candidate solutions that evolves through selection, crossover, and mutation. A carefully tuned configuration—population sizes, selection pressures, and mutation rates—preserves diversity while driving the search toward configurations that fit the data. The team validated this approach across three contrasting regions: Greater Boston, California, and Switzerland, together with a fourth long multimodal record from Luxembourg used for robustness checks.

The results reveal a clear pattern. On the complex California record, which spans multiple years and includes highly variable epidemic waves, the genetic algorithm reached a calibration-window R² of 0.7979, whereas conventional least-squares fitting of the SIR-V baseline managed only 0.4373. The unified evaluation across 30 independent seeds confirmed this advantage, with the algorithm achieving 0.78 against 0.20 for a fairly tuned least-squares baseline. On the simpler Greater Boston record, however, a well-tuned least-squares fit was essentially competitive, and on the short Swiss calibration window neither method produced a meaningful fit.

Out-of-sample performance tells a more sobering story. Long multi-wave records remained hard for all methods beyond the calibration window, with extrapolation R² values dropping to 0.05 for the algorithm and 0.06 for least squares on California. The advantage of the genetic algorithm therefore lies primarily in producing more stable parameter estimates on difficult data, rather than in delivering dramatically better long-range forecasts. This finding aligns with the growing consensus that parameter estimation, rather than model structure, is the primary bottleneck in wastewater-based epidemic modeling.

The study also compared the genetic algorithm against alternative optimizers within the same mechanistic framework. Particle Swarm Optimization, Differential Evolution, and Bayesian estimation all reached broadly similar calibration performance, suggesting that the benefit stems from using a global or Bayesian estimator rather than from the genetic algorithm per se. Data-driven approaches that bypass the mechanistic model entirely—fitting statistical surrogates directly to the viral signal—achieved strong in-window fits but produced wildly unstable extrapolation, with R² values collapsing to negative values on multi-wave records.

Temperature-dependent viral decay emerged as a second key innovation. The degradation rate of SARS-CoV-2 RNA in wastewater varies substantially with temperature, with warmer conditions accelerating breakdown and distorting the relationship between shedding and measured concentration. By replacing a fixed decay rate with a temperature-driven function, the model captured seasonal fluctuations that a static parameter cannot. This proved particularly important for records spanning multiple seasons, where a fixed rate would introduce systematic bias into the viral signal.

The team also tested whether memory effects—deviations from the standard exponential dynamics—improved performance. A fractional-order SEIR formulation, which allows the epidemic state to depend on its entire history rather than just the present, detected memory across all regions, with selected orders between 0.79 and 0.88. Yet the fractional formulation did not systematically improve calibration or extrapolation performance, and the team retained the simpler integer-order model as adequate and more parsimonious for these datasets.

The study’s most important contribution may be its honest delineation of where advanced estimation helps. The genetic algorithm delivers its clearest advantage on complex, multi-wave records where conventional fitting is most fragile, while remaining competitive rather than superior on simpler data. The integration of temperature-driven decay captures seasonal distortions that a static parameter cannot. The unified evaluation across contrasting regions, alternative optimizers, and data-driven surrogates provides a template for honest benchmarking in wastewater-based epidemic modeling. As surveillance networks expand worldwide, this work suggests that the path to better forecasts runs through more robust estimation on difficult data, rather than through ever-more-elaborate model structures.

Subject of Research: Computational Biology

Article Title: Genetic algorithm–optimized SEIR-V: robust parameter estimation for wastewater-based infectious disease modeling across multiple regions

Article References: Li, B., Chachar, S., Zhang, C., Rui, Z., Lu, N., Liu, J., Chachar, B., Chachar, Z., Huang, F., & Cai, Y. (2026). Genetic algorithm–optimized SEIR-V: robust parameter estimation for wastewater-based infectious disease modeling across multiple regions. Frontiers in Water, 8, Article 1874972. https://doi.org/10.3389/frwa.2026.1874972

Image Credits: AI Generated

DOI: 10.3389/frwa.2026.1874972

Keywords: wastewater epidemiology, genetic algorithm, SEIR-V model, epidemic forecasting, parameter estimation, SARS-CoV-2, temperature-dependent decay, Frontiers in Water, Genetic, algorithm, optimized, SEIR-V

Against this backdrop, a new study published in Theoretical and Applied Climatology introduces a hybrid machine learning framework that promises to sharpen the predictive lens through which we view climate hazards. The research, conducted by Bahjat Fakieh of the Information Systems Department at King Abdulaziz University in Jeddah, Saudi Arabia, combines two powerful but fundamentally different algorithms: Long Short-Term Memory networks, known as LSTM, and XGBoost, an extreme gradient boosting method. The innovation lies not merely in pairing these techniques but in wrapping the entire pipeline within an optimization layer driven by a genetic algorithm, a search procedure inspired by the mechanics of biological evolution. The result is a system that can both classify climate risk into discrete categories and predict it as a continuous value, offering flexibility for the varied ways in which decision-makers consume climate information.

The architecture of the model reflects a deliberate division of labor between its two core components. The LSTM network, a form of recurrent neural network equipped with memory cells that can retain information over long sequences, is tasked with extracting temporal representations from sequential climate observations. Climate data are inherently time-dependent: today’s temperature, humidity, and pressure readings carry information that shapes tomorrow’s conditions. By processing these sequences, the LSTM learns patterns that unfold across time, such as seasonal cycles, gradual warming trends, and the buildup of conditions that precede extreme events. These learned temporal features are then combined with static climate attributes, such as geographic or categorical variables that do not change from one observation to the next, and the merged representation is passed to XGBoost for the final prediction.

XGBoost, which builds an ensemble of decision trees in a stage-wise fashion, each new tree correcting the errors of its predecessors, excels at handling tabular data with mixed feature types and complex nonlinear interactions. Its gradient boosting framework is computationally efficient and includes regularization mechanisms that guard against overfitting, making it a robust choice for the classification and regression stages of the pipeline. The hybrid design therefore leverages the strengths of deep learning for sequence modeling while relying on gradient-boosted trees for the final decision boundary, a combination that has proven effective in other domains but had not been systematically optimized for climate risk prediction in this way.

What elevates the framework beyond a simple model pairing is the role of the genetic algorithm, which operates at two distinct levels. First, GA-based feature selection identifies the most relevant input features from the available climate variables, discarding those that add noise rather than signal. In high-dimensional climate datasets, where dozens of correlated variables may be measured simultaneously, feature selection is critical for both model performance and interpretability. Second, GA-based hyperparameter optimization determines suitable parameters for both the LSTM and XGBoost components. Hyperparameters, such as the number of memory units in the LSTM, the learning rate, the depth of the trees in XGBoost, and the regularization coefficients, are settings that are not learned from data but must be specified before training. Choosing them poorly can cripple even the most sophisticated architecture. The genetic algorithm addresses this by evolving a population of candidate configurations, evaluating their fitness through model performance, and iteratively breeding the best performers until a near-optimal set of settings emerges.

Quantitatively, the model delivered impressive results across both prediction tasks. For classification, where the goal is to assign climate observations to risk categories, the framework achieved an accuracy of 91 percent, a precision of 0.88, a recall of 0.90, an F1-score of 0.89, and a ROC-AUC of 0.93. These metrics indicate that the model not only classifies correctly in the majority of cases but also maintains a strong balance between false positives and false negatives, a crucial property when the cost of missing a genuine climate risk can be measured in lives and livelihoods. The ROC-AUC of 0.93, in particular, suggests that the model’s ability to discriminate between risk levels across all possible thresholds is excellent, well above the 0.5 that would indicate random guessing.

For regression tasks, where the model predicts climate risk as a continuous numerical value rather than a category, the performance was equally encouraging. The framework recorded a root mean square error of 1.74, a mean absolute error of 1.31, and a coefficient of determination, or R-squared, of 0.87. An R-squared value of 0.87 means that the model explains 87 percent of the variance in the target variable, leaving only 13 percent unexplained. The relatively small gap between RMSE and MAE suggests that the model’s errors are fairly consistent and not dominated by a few extreme outliers, which speaks to the stability of the predictions across the range of climate conditions represented in the data.

Compared with the other models evaluated under the same experimental conditions, the proposed GA-optimized LSTM-XGBoost framework demonstrated favorable predictive performance, confirming that the combination of deep temporal feature extraction, gradient boosting, and evolutionary optimization yields tangible benefits over simpler alternatives. The study contributes meaningfully to the field of climate informatics by providing a framework that is not only accurate but also scalable and interpretable. Interpretability, often a casualty of complex machine learning systems, is preserved here partly through the feature selection stage, which makes explicit which climate variables drive the predictions, and partly through the tree-based structure of XGBoost, which allows researchers to trace how individual features influence the final output.

The practical implications of this work extend across several domains. For real-time environmental monitoring, the framework could process streams of climate data as they arrive, flagging emerging risks before they fully materialize. For policy planning, the ability to predict risk levels with quantified confidence supports more informed decisions about infrastructure investment, resource allocation, and land use. For disaster preparedness and emergency response systems, early and accurate warnings of climate-related hazards could save lives by giving communities the time needed to evacuate, reinforce structures, or stockpile supplies. The framework’s scalability means it can handle the massive volumes of climate data generated by modern observation networks, from satellite remote sensing to ground-based weather stations, without requiring prohibitive computational resources.

However, the study is candid about its limitations. The promising results were obtained on the evaluated climate dataset, and the author acknowledges that additional validation across diverse geographic regions and climatic conditions is necessary before considering large-scale deployment. Climate systems vary enormously between tropical, temperate, arid, and polar zones, and a model trained on data from one region or one set of conditions may not generalize seamlessly to another. The author notes that no datasets were generated or analysed during the current study in the data availability statement, while the source code associated with the research is publicly available on Zenodo, allowing other researchers to reproduce the experiments, test the framework on their own data, and build upon the methodology. This transparency is a welcome feature in a field where reproducibility is often hampered by proprietary data and closed-source implementations.

The broader significance of this research lies in its demonstration that the marriage of deep learning and evolutionary optimization can bridge the gap between raw data-driven knowledge and actionable climate information. As global climate change continues to intensify, the demand for precise, timely, and interpretable risk predictions will only grow. Frameworks like the one presented here, which combine the pattern-recognition power of LSTM networks, the tabular-data mastery of XGBoost, and the adaptive search capability of genetic algorithms, offer a template for how machine learning can support more responsible decision-making in sustainable development, resource management, and emergency response. While the path toward proactive climate resilience planning remains long, tools of this kind represent a meaningful step forward, turning the torrent of climate data that surrounds us into a stream of insight that can guide humanity toward a safer and more sustainable future.

Subject of Research: Climate risk prediction

Article Title: Harnessing massive climate data for enhanced risk prediction: a GA-optimized LSTM and XGBoost hybrid approach

Article References: Fakieh, B. (2026). Harnessing massive climate data for enhanced risk prediction: a GA-optimized LSTM and XGBoost hybrid approach. Theoretical and Applied Climatology, 157(11), Article 729. https://doi.org/10.1007/s00704-026-06649-2

Image Credits: AI Generated

DOI: 10.1007/s00704-026-06649-2

Keywords: climate risk, machine learning, LSTM, XGBoost, genetic algorithm, climate informatics, extreme weather, disaster preparedness, deep learning, predictive modeling, Harnessing, massive

Muhammad Zubair and Syed Khaliq Shah of Southeast University in Nanjing tackled what they call the integrated rebalancing and faulty bike recycling problem. Instead of sending trucks out on separate missions—one to redistribute bikes, another to collect broken ones—their framework folds both tasks into a single vehicle routing model. Each service truck carries a partitioned cargo hold: a main compartment for rebalancing bikes and a dedicated section for faulty ones. This partitioned capacity design means a vehicle can deliver bikes to deficit stations, collect surplus bikes from oversupplied ones, and scoop up damaged bicycles along the way, all in one tour.

The mathematics behind the model is demanding. The researchers formulated it as a variant of the pickup and delivery problem, with an objective function that balances three competing costs: user dissatisfaction when riders cannot find a bike, operational expenses including travel and loading time, and the handling cost of faulty bicycles. Loading or unloading an operational bike takes about one minute, while a damaged bike requires roughly two minutes of handling time—values drawn from real observations of Nanjing operators. Vehicles must complete their tours within a six-hour shift, and travel times are adjusted by traffic congestion factors derived from Baidu Maps data, ranging from free-flow conditions to heavy congestion.

Exact optimization tools like IBM’s CPLEX solver can crack such models, but only at small scale. The team first validated their formulation on 38 real stations near Southeast University, solving it to near-perfect optimality with percentage gaps as low as 0.02 percent. In 35 of the 38 stations, rebalanced inventory aligned almost exactly with forecast demand, and all faulty bikes were successfully routed back to the depot for repair or recycling. But real-world networks dwarf this testbed: Shanghai operators manage more than 50,000 dockless bikes, and New York’s CitiBike runs over 1,000 stations. Exact mixed-integer programming becomes computationally intractable beyond roughly 50 stations, which is where the researchers’ metaheuristic enters.

Their solution combines a genetic algorithm with adaptive large neighborhood search, wrapped in a cluster-first, route-second decomposition. The genetic algorithm evolves populations of candidate routes using tournament selection, order crossover, and mutation operators, while preserving the best solutions through elitism. The ALNS component then intensifies the search by destroying parts of a solution—removing stations at random, targeting the worst-performing segments, or dismantling entire clusters—and repairing them with greedy, regret, or cheapest insertion strategies. An adaptive weighting mechanism learns which destroy-and-repair combinations perform best, and a simulated annealing acceptance criterion decides whether to embrace worse solutions temporarily to escape local optima.

The clustering stage is where the approach departs most sharply from prior work. Rather than dividing a city into fixed geographic zones, the dynamic clustering procedure jointly optimizes the number of clusters, their centroids, and station assignments based on expected inter-cluster traversal costs. This captures a fundamental trade-off: more clusters shrink the distance vehicles travel within each zone but inflate the cost of moving between zones. In experiments on a 200-station instance, increasing clusters from 3 to 20 cut intra-cluster travel by 65 percent while inflating inter-cluster travel by more than 700 percent. The sweet spot landed at 11 clusters, where total routing distance fell to 362 kilometers—a 26.1 percent improvement over three clusters and a 21.8 percent improvement over twenty.

The performance numbers are striking. On 200-station synthetic instances, the hybrid algorithm produced solutions within 3.2 percent of best-known values in about 15 minutes, a fourfold scale improvement over the most comparable integrated studies, which stall at 50 to 100 stations. Against 50-station benchmarks, the method matched exact solvers with gaps below 0.5 percent while running 3.75 times faster. Dynamic clustering itself shaved 6.5 percent off routing distance compared with static, geography-only approaches. Meanwhile, folding faulty bike collection into rebalancing routes added just 8 to 12 percent computational overhead—yet eliminated the separate collection trips entirely, cutting total vehicle distance by 15 to 25 percent compared with running the two operations sequentially.

The practical implications are substantial. For a 200-station city, the researchers estimate daily savings of 50 to 75 kilometers of vehicle travel, translating to roughly 10 to 15 liters of fuel, two to three hours of driver time, and annual cost reductions of 15,000 to 25,000 US dollars per vehicle. The recycling loop itself follows a four-stage pipeline: IoT sensors and GPS monitoring flag faulty bikes based on battery status, signal strength, failed QR scans, and idle periods exceeding seven days; trucks with partitioned holds collect them during routine rebalancing; capacity constraints keep the two cargo types physically separated throughout the route; and the depot sorts arrivals into immediate repair, scheduled maintenance, or parts recovery.

Empirical research from major Chinese cities underscores why this integration matters. Faulty bike rates can exceed 2 percent during peak usage, and delayed collection measurably damages user satisfaction and safety. Earlier integrated models relied on exact optimization that buckled beyond 50 stations, and most treated vehicle capacity as a single undifferentiated constraint, implicitly assuming sequential loading and unloading. The new framework’s partitioned capacity model and jointly optimized clustering address both shortcomings, and the authors argue the approach could extend beyond bike sharing to fleet management, ride-sharing, and last-mile delivery, where decomposition is often treated as a necessary evil rather than an optimization opportunity.

Limitations remain. The model assumes deterministic demand within a single operational period and a homogeneous fleet, simplifications that real cities may strain against. The researchers point to several frontiers: incorporating stochastic demand through robust optimization, extending to multi-depot networks, adapting the algorithm for real-time rebalancing as conditions change mid-shift, integrating dynamic pricing to encourage users to self-balance the system, and pairing the optimizer with machine learning demand forecasts. For now, the study offers something operators have lacked: a computationally feasible way to keep dockless fleets balanced, functional, and profitable at the scale modern cities actually demand.

Subject of Research: Integrated rebalancing and faulty bike recycling optimization in dockless bike-sharing systems using a hybrid metaheuristic with dynamic clustering

Article Title: Integrated rebalancing and faulty bike recycling in dockless bike-sharing systems: a hybrid metaheuristic approach with dynamic clustering

Article References: Zubair, M., & Shah, S. K. (2026). Integrated rebalancing and faulty bike recycling in dockless bike-sharing systems: a hybrid metaheuristic approach with dynamic clustering. Frontiers in Future Transportation, 7, Article 1919819. https://doi.org/10.3389/ffutr.2026.1919819

Image Credits: AI Generated

DOI: 10.3389/ffutr.2026.1919819

Keywords: bike-sharing, dockless bikes, vehicle routing, metaheuristics, genetic algorithm, adaptive large neighborhood search, dynamic clustering, faulty bike recycling, urban mobility, optimization, Nanjing, transportation systems

In a study published in Frontiers in Batteries and Electrochemistry, Yasmin Shabeer, Seyed Saeed Madani, Satyam Panchal, and Michael Fowler systematically investigated the corrosion behavior of a commercial Al-Mg-Si alloy in sodium hydroxide (NaOH) and potassium hydroxide (KOH) electrolytes spanning concentrations from 1 to 10 molar. The team combined three complementary electrochemical techniques—open-circuit voltage monitoring, potentiodynamic polarization, and electrochemical impedance spectroscopy—with X-ray diffraction to assess structural stability. They then trained an artificial neural network on the resulting dataset and optimized it using two genetic algorithm variants, achieving coefficients of determination exceeding 0.999 for both corrosion potential and corrosion current density.

The experimental campaign revealed two distinct corrosion regimes. At low-to-moderate alkalinity, anodic activation dominates: the corrosion potential shifts progressively negative and the corrosion current density climbs as hydroxide concentration increases, reflecting accelerated chemical dissolution of the native Al2O3/AlOOH surface film. At high concentrations, the behavior transitions to transport-limited kinetics, where the accumulation of soluble aluminate species, increased electrolyte viscosity, and hydrogen bubble coverage of 20 to 50 percent of the surface physically throttle the reaction rate. This shift from charge-transfer control to mass-transport control was confirmed by the weak coupling between corrosion potential and current density observed in the high-molarity polarization data.

Perhaps the most striking finding is that the two electrolytes behave very differently despite supplying identical hydroxide activity at the same molarity. KOH consistently produced more negative corrosion potentials and higher corrosion current densities than NaOH at equivalent concentrations. The researchers attribute this to cation-specific interfacial effects: potassium ions have a larger ionic radius, lower hydration energy, and higher mobility than sodium ions, facilitating faster charge transport, stronger double-layer perturbation, and more aggressive hydrogen evolution. In practical terms, this means that electrolyte identity—not just hydroxide concentration—governs corrosion kinetics in alkaline aluminum systems.

Impedance spectroscopy provided mechanistic confirmation of these trends. The charge-transfer resistance decreased markedly with increasing hydroxide concentration in both electrolytes, consistent with progressive activation of anodic dissolution. At equivalent molarities, NaOH consistently exhibited higher charge-transfer resistance than KOH, corroborating the milder corrosion environment inferred from polarization measurements. The constant phase element exponent remained below unity across all conditions, confirming a heterogeneous and dynamically evolving interface shaped by continuous oxide dissolution, aluminate formation, and gas evolution.

X-ray diffraction added a crucial structural dimension to the analysis. The diffraction patterns were dominated by the characteristic reflections of face-centered cubic aluminum, with invariant peak positions across all concentrations and both electrolytes. No crystalline aluminum hydroxides, oxyhydroxides, or aluminate phases were detected under any condition, indicating that corrosion products are predominantly amorphous, nanocrystalline, or highly hydrated. This finding validates the electrochemical evidence for continuous surface dissolution with transient, non-protective films rather than bulk phase transformation.

On the modeling side, the team conducted a comprehensive grid search of 351 artificial neural network configurations, varying network depth, neuron allocation, learning rate, and Adam optimizer parameters. The best-performing model used a three-layer architecture with 20, 25, and 25 neurons, a learning rate of 0.001, and a beta-1 value of 0.92. This configuration achieved R-squared values of 0.999888 for corrosion potential and 0.999945 for corrosion current density, representing an almost perfect mapping between experimental variables and corrosion indicators. Notably, 76 percent of all tested configurations achieved mean validation R-squared values exceeding 0.97, indicating that the underlying electrochemical relationships are highly learnable yet sensitive to architectural choices.

The researchers then applied two fundamentally different optimization strategies to the trained network. A scalar genetic algorithm, which combined the two corrosion objectives into a single weighted fitness function, and a non-dominated sorting genetic algorithm (NSGA-II), which preserved the independence of both objectives and resolved the complete Pareto trade-off front. Remarkably, both approaches converged on the same optimal condition: approximately 1.35 molar KOH, with a predicted corrosion potential of −1.6199 volts and a corrosion current density of 0.0876 A cm−2. This knee-point solution represents the least-compromise operating condition, where further reductions in corrosion current would require a disproportionate loss in corrosion potential stability.

The agreement between the two optimization formulations provides strong validation of the framework. The Pareto front revealed a smooth, continuous trade-off between corrosion thermodynamics and kinetics, with pronounced curvature at low current densities where small increases in corrosion rate yield diminishing gains in potential stability. This nonlinear structure reflects the fundamental electrochemical coupling between anodic dissolution kinetics and corrosion thermodynamics in alkaline aluminum systems, and it underscores why empirical trial-and-error electrolyte selection often fails to identify truly optimal conditions.

The implications extend beyond aluminum-air batteries. The framework demonstrates that integrating systematic electrochemical experimentation with neural network-assisted evolutionary optimization enables rapid, quantitative evaluation of corrosion behavior without exhaustive testing. The researchers suggest that this approach could support the development of digital-twin platforms for the design, optimization, and control of next-generation aluminum-based electrochemical energy systems, where parasitic corrosion losses must be minimized without sacrificing electrochemical stability. As the global transition toward cleaner energy accelerates and demand for alternative battery chemistries grows, such corrosion-aware electrolyte optimization may prove essential for unlocking the full potential of aluminum-based energy storage.

Subject of Research: Machine learning prediction of aluminum anode corrosion kinetics in alkaline electrolytes for aluminum-air batteries

Article Title: Artificial neural network-genetic algorithm prediction of aluminum anode corrosion kinetics in high-concentration NaOH and KOH

Article References: Shabeer, Y., Madani, S. S., Panchal, S., & Fowler, M. (2026). Artificial neural network-genetic algorithm prediction of aluminum anode corrosion kinetics in high-concentration NaOH and KOH. Frontiers in Batteries and Electrochemistry, 5, Article 1815822. https://doi.org/10.3389/fbael.2026.1815822

Image Credits: AI Generated

DOI: 10.3389/fbael.2026.1815822

Keywords: aluminum-air batteries, corrosion, artificial neural networks, genetic algorithm, NSGA-II, alkaline electrolytes, KOH, NaOH, Al-Mg-Si alloy, machine learning, electrochemistry, energy storage

The core problem the researchers set out to solve is one of data scarcity. Machine learning thrives on large, consistent datasets, but additively manufactured austenitic stainless steels, the family that includes workhorses such as 304L and 316L, have only been studied intensively for roughly a decade. Every published experiment reports different process parameters, different powder chemistries, different build orientations and different testing conditions, which makes the mapping from composition to mechanical properties noisy and fragmented. Casting, by contrast, has more than a century of carefully documented composition-property relationships. The insight behind the new work is that although casting and additive manufacturing produce very different microstructures, the fundamental thermodynamics that govern whether an alloy solidifies as stable austenite, forms ferrite, or sits near a phase boundary are shared across processes. If the model could learn those underlying phase-stability relationships from casting data, the relative ordering of compositions might transfer to printing even when the absolute strength numbers do not.

To make that transfer possible, the team needed descriptors that capture metallurgical meaning rather than raw elemental percentages. They turned to two of the oldest and most trusted quantities in stainless steel metallurgy: the chromium equivalent and the nickel equivalent. These equivalents, familiar to generations of engineers through the Schaeffler diagram used to predict weld microstructures, condense the effects of ferrite-stabilizing elements such as chromium, molybdenum and silicon, and austenite-stabilizing elements such as nickel, manganese, carbon and nitrogen, into two intuitive axes. By expressing each alloy’s composition in terms of these empirical descriptors, the researchers gave the machine learning algorithm a physically informed coordinate system in which alloys with similar phase behavior cluster together, regardless of whether their properties were measured in a foundry or beside a laser powder bed.

On top of this descriptor choice, the authors introduced a divide-and-conquer strategy for model construction. Rather than forcing a single algorithm to describe the entire composition space, including the treacherous regions near phase boundaries where small chemical changes cause dramatic microstructural shifts, the divide-and-conquer model partitions the data into sub-regions defined by phase stability and trains separate predictors within each partition. The study evaluated this framework against three targets: ultimate tensile strength, elongation, and a combined tensile property defined as the product of strength and elongation, a single figure of merit that rewards alloys that are both strong and ductile rather than merely one or the other. The results were impressive. The divide-and-conquer model achieved coefficients of determination of 0.944 for ultimate tensile strength, 0.812 for elongation, and 0.899 for the combined property, and its mean absolute percentage errors were lower than those of ten commonly used machine learning models tested for comparison.

But a model that only predicts well on the data it was trained on is of limited use for designing new alloys. The real test came when the researchers coupled their model to a genetic algorithm, an optimization technique inspired by biological evolution that iteratively breeds candidate compositions toward better predicted performance. This inverse design loop allows the researchers to start from a desired property profile and work backwards to a chemical recipe, rather than laboriously testing one composition after another. The genetic algorithm searches the composition space through the lens of the divide-and-conquer model, proposing candidate austenitic stainless steels whose predicted strength-ductility combinations exceed those of conventional grades. The team then took the most promising candidates out of the computer and into the laboratory, fabricating them by additive manufacturing and measuring their actual tensile behavior.

Here the study delivers its most important and nuanced finding. When the printed alloys were tested, their absolute strength and ductility values did not match the model’s numerical predictions exactly, a systematic deviation the authors attribute to the profound process differences between slow, equilibrium-like casting solidification and the rapid, cyclic thermal history of layer-by-layer printing. Casting produces coarse, slowly formed grains, while additive manufacturing generates fine, rapidly solidified structures with unique sub-grain features, dislocation densities and segregation patterns that can dramatically alter mechanical response. Yet when the researchers compared the relative rankings of different compositions, the agreement was essentially perfect. The order in which the model predicted the compositions would perform, from best to worst, matched the order observed in the printed specimens. For the predicted improvement in ultimate tensile strength between two specific compositions, the relative error was as low as 0.81 percent, a level of agreement that would have been unthinkable without the phase-stability-based partitioning scheme.

This distinction between absolute and relative prediction is more than a statistical curiosity; it is the conceptual heart of the paper. In alloy development, engineers often care less about predicting an exact tensile strength to the nearest megapascal than about knowing which of a hundred candidate chemistries deserves the expense of a printing trial. The new work demonstrates that the relative composition-property relationships encoded in chromium and nickel equivalents possess genuine cross-process transferability. The physics of phase stability, it turns out, is a kind of universal language spoken by both the foundry crucible and the laser powder bed, and a model that learns that language from abundant casting data can say meaningful things about alloys it has never seen printed. This effectively converts a century of casting knowledge into a screening tool for a technology that is barely out of its adolescence.

The practical implications are considerable. Austenitic stainless steels are among the most widely printed structural metals, appearing in aerospace brackets, marine components, chemical processing equipment, nuclear applications and biomedical implants. Each new application pushes toward compositions tuned for specific balances of strength, ductility, corrosion resistance, printability and cost, and researchers have explored everything from in-situ alloying and nanoparticle additions to grain-boundary engineering to squeeze more performance from these alloys. Traditionally, exploring that composition space meant printing and testing dozens or hundreds of variants, an expensive and slow process given the cost of spherical metal powders and machine time. The new approach enables efficient preliminary screening using nothing more than compositional information and prior casting data, allowing researchers to discard poor candidates computationally and reserve laboratory resources for the handful of compositions most likely to succeed. Trend prediction, in which the model indicates how strength or ductility will shift as an element is added or removed, further guides the direction of experimental campaigns.

The study also speaks to a broader movement in materials science toward hybrid approaches that blend data-driven methods with physical understanding. Purely black-box machine learning models can interpolate well within their training data but often fail catastrophically when extrapolating to new processing routes or composition regimes, and their predictions can be difficult for metallurgists to trust or interpret. By anchoring the model in empirical descriptors with clear physical meaning and by structuring the learning problem around phase stability, the researchers preserved interpretability while retaining the efficiency of modern algorithms. The strategy echoes and extends recent successes in which machine learning has accelerated the discovery of high-entropy alloys, titanium alloys and aluminum alloys, but the cross-process element is distinctive: it explicitly addresses the situation every emerging manufacturing technology faces, where the data needed for good models do not exist yet and must be borrowed, intelligently, from adjacent fields.

There remain, of course, limits to what the approach can do. The systematic deviations in absolute property predictions mean that final qualification of any new alloy will still require real additive manufacturing experiments, and the model’s transferability rests on the assumption that phase stability remains the dominant organizing principle across processes, an assumption that may weaken for compositions near critical boundaries or for properties dominated by process-specific defects such as porosity and residual stress. The authors are candid that their method provides a pathway for preliminary screening and trend prediction rather than a replacement for experimental validation. Even so, the demonstration that a relative error of 0.81 percent is achievable for cross-process strength predictions marks a genuine advance. For a field racing to design alloys tailored to the extreme thermal conditions of 3D printing, the message is liberating: the data you need may already exist, sitting in decades of casting handbooks, waiting for the right physics-informed algorithm to translate them into the language of the laser.

Subject of Research: Machine learning-driven composition design of austenitic stainless steels for additive manufacturing using cross-process casting data and empirical phase-stability descriptors

Article Title: Composition design of austenitic stainless steels for additive manufacturing: a cross-process data and empirical knowledge-driven approach

Article References: Chen, C., Li, B., Tang, M., & Chen, H. (2026). Composition design of austenitic stainless steels for additive manufacturing: a cross-process data and empirical knowledge-driven approach. Journal of Materials Science. https://doi.org/10.1007/s10853-026-13666-z

Image Credits: AI Generated

DOI: 10.1007/s10853-026-13666-z

Keywords: additive manufacturing, austenitic stainless steel, machine learning, alloy design, chromium equivalent, nickel equivalent, divide-and-conquer model, genetic algorithm, tensile properties, phase stability, casting data, materials science

The researchers built a machine learning framework they call the Enhanced Fractional Probabilistic Self-Organizing Map, or EF-PRSOM. The name sounds forbidding, but the underlying idea is elegant. Self-organizing maps, first invented by Teuvo Kohonen in 1990, are neural networks that learn to arrange high-dimensional data, such as a household’s daily electricity curve, onto a two-dimensional grid, so that similar consumption patterns end up clustered together. A probabilistic variant, developed in the late 1990s, adds statistical rigor by modeling each cluster with radial basis functions, giving a measure of uncertainty rather than a hard, brittle assignment. These tools have long been workhorses for load profiling, but they share a well-known weakness: their optimization is memoryless. At each training step, the algorithm updates its clusters based only on the current state, discarding the rich history of how the solution has evolved. That amnesia makes the methods sensitive to initialization and prone to getting trapped in local minima, the mediocre solutions that plague iterative optimization.

Safouan and El Moutaouakil attacked this memory problem with an unexpected weapon from mathematics: a fractional derivative. Classical calculus deals with derivatives of integer order, first, second, third. Fractional calculus generalizes the concept to non-integer orders, and its operators are inherently non-local, meaning the value of a fractional derivative at a given point depends on the entire history of the function, not just its immediate neighborhood. By embedding the Atangana–Baleanu–Caputo fractional derivative, a formulation introduced in 2016 with a non-singular kernel, into the learning rule of the probabilistic self-organizing map, the researchers gave their algorithm a form of long-term memory. In effect, the network remembers where it has been during optimization and uses that accumulated experience to steer its updates, smoothing out erratic trajectories and escaping the shallow traps that snare conventional methods.

There was one more twist. The fractional order itself, the parameter that controls how much historical information the derivative retains, is not obvious in advance. Setting it by hand would be guesswork. So the team turned to a genetic algorithm, an evolutionary search method inspired by natural selection, to tune the fractional-order parameter automatically. The result is a hybrid system in which fractional calculus supplies memory and evolutionary computation supplies calibration, a combination the authors had previously explored in earlier work on fuzzy clustering and on memory-enhanced probabilistic maps.

To prove the framework was more than mathematical exotica, the researchers first tested it on a well-known benchmark: a French smart-meter dataset of individual household electricity consumption, publicly available through the UCI machine learning repository. This is where smart-meter data is abundant, so the ground truth of household behavior can be examined in fine detail. The EF-PRSOM delivered the best clustering quality of any method compared, achieving a Silhouette score of 0.6709, a measure of how tightly grouped and well separated the clusters are, and a Davies–Bouldin Index of 0.8769, a measure where lower values indicate better separation. Against the plain probabilistic self-organizing map, the enhanced version improved the Silhouette score by 124.8 percent and reduced the Davies–Bouldin Index by 84.2 percent. It also outperformed K-means, the standard self-organizing map, and Gaussian mixture models, the classic statistical clustering approach. Those are not marginal gains; they represent a qualitative leap in how cleanly the algorithm distinguishes different kinds of energy consumers.

With the method validated on rich European data, the team turned to the real target: Morocco. They applied the framework to the Moroccan Household Consumption and Living Conditions Survey, a nationally representative dataset collected by the country’s statistical authority, the Haut-Commissariat au Plan. Unlike smart-meter streams, survey data captures household characteristics, expenditures, and living conditions, precisely the kind of information available in developing economies where metering infrastructure lags. The clustering analysis revealed three distinct household energy consumption profiles: low, medium, and high consumers. Crucially, these profiles were not random. They aligned strongly with income, poverty status, and whether a household lived in an urban or rural area, associations that held with statistical significance at p < 0.001. In other words, the algorithm independently rediscovered the social geography of Moroccan energy use, but in a systematic, quantifiable form that policymakers can act on.

The spatial analysis may prove the most consequential finding of all. When the researchers mapped consumption profiles across Morocco’s regions, they uncovered a structural mismatch: the northern coastal regions, where demand is highest, are not the places endowed with the greatest renewable energy resources. That wealth sits in the southeast, home to the solar potential that has made Morocco a renewable poster child, from the Noor complex onward. High demand in the north, abundant supply potential in the south and east, and a transmission network that must bridge the two: this is exactly the kind of insight that transmission infrastructure planning requires, and exactly the kind that has been missing from household-level analysis in the region. Grid expansion decisions worth billions can now be grounded in a data-driven portrait of who uses energy, where, and how much.

Beyond infrastructure, the findings speak directly to energy poverty. Morocco has made notable progress on electrification, but access to a connection is not the same as access to affordable, adequate energy services. Households in the low-consumption cluster, disproportionately poor and rural, may be underconsuming not by choice but by constraint, a pattern long documented in the energy ladder literature across Africa, where families stack biomass and electricity rather than progressing cleanly from one fuel to the next. By attaching poverty status to consumption profiles with rigorous statistical confidence, the framework gives governments a targeting tool: subsidies, efficiency programs, and clean cooking initiatives can be aimed at the households that need them, identified by their position in the consumption landscape rather than by crude geographic averages.

The broader significance of the study extends well beyond Morocco’s borders. Across Africa and much of the Global South, rapid urbanization and economic development are driving steady growth in energy demand, while reliable household-level consumption data remains scarce. The lesson of EF-PRSOM is that this gap is not a dead end. With a clustering framework that extracts maximal structure from survey data, validated against smart-meter benchmarks where they exist, data-scarce regions can produce energy profiles of genuine analytical quality. The technique also offers something to wealthy countries: the memory-enhanced optimization could improve load profiling, demand response design, and customer segmentation wherever consumption data flows. As Morocco pursues its target of roughly half renewable electricity by 2030, the quiet mathematics of fractional derivatives and self-organizing maps may turn out to be among the most practical tools in its transition toolkit, proving that the path to an equitable energy future sometimes begins with teaching an algorithm to remember.

Subject of Research: Machine learning clustering of household energy consumption patterns in Morocco

Article Title: Household energy consumption patterns in Morocco’s energy transition: a clustering framework using enhanced fractional probabilistic self-organizing maps

Article References: Safouan, S., & El Moutaouakil, K. (2026). Household energy consumption patterns in Morocco’s energy transition: a clustering framework using enhanced fractional probabilistic self-organizing maps. Neural Computing and Applications, 38(19), Article 790. https://doi.org/10.1007/s00521-026-12528-8

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12528-8

Keywords: Morocco, household energy consumption, self-organizing maps, fractional calculus, clustering, energy poverty, smart meter data, renewable energy planning, genetic algorithm, energy transition, machine learning, survey data

To understand why the study matters, it helps to grasp what fractional crystallization actually does. Refined, bleached, and deodorized palm oil, known in the industry as RBDPO, is a natural mixture of triglycerides with different melting behaviors. When the oil is cooled in carefully controlled stages, the more saturated molecules begin to solidify first, forming crystals that can be separated from the still-liquid fraction. The liquid portion, palm olein, is prized as a cooking oil that stays clear at room temperature. The solid portion, palm stearin, finds uses in shortenings, margarines, and specialty fats. The economics of the entire operation hinge on how much olein can be squeezed out of each batch without degrading its quality.

The trouble is that the process is exquisitely sensitive to the raw material. Palm oil is a natural product, and its composition shifts with the season, the plantation, and the fruit’s ripeness at harvest. One of the most important indicators of this variability is the iodine value, or IV, a measure of the degree of unsaturation of the oil’s fatty acids. A batch labeled IV51 behaves differently in the crystallizer than a batch labeled IV53, because the degree of unsaturation influences how crystals nucleate, how they grow, and how efficiently they can ultimately be separated from the mother liquor. A cooling profile that delivers a bumper olein yield on Monday may underperform on Tuesday when the feedstock shifts.

Traditionally, fractionation plants have coped with this variability through operator experience and relatively rigid process recipes. Some researchers have applied response surface methodology, a statistical technique that models how a handful of process variables affect an outcome, to map the crystallization behavior. Response surface models have a real virtue: they are interpretable, allowing engineers to see exactly how cooling rate or temperature influences yield. But they struggle when the underlying relationships are strongly nonlinear, and they become unwieldy when a process unfolds across multiple stages, each with its own dynamics and its own interactions between variables.

That is where the new study, led by Noor Hidayu Othman of SD Guthrie Technology Centre and Universiti Sains Malaysia, with colleagues including Musfirah Zulkurnain as corresponding author, breaks new ground. The team designed a structured three-stage experimental program mirroring the physical reality of crystallization: a supercooling stage, a nucleation stage, and a crystal growth stage. By systematically varying process conditions across these stages for RBDPO with iodine values of 51, 52, and 53, they generated 145 datasets, a substantial foundation for data-driven modeling. Each dataset captured not just the headline number, olein yield, but a suite of quality measures including the iodine value of the olein, the cloud point at which it turns hazy when chilled, the slip melting point of the stearin, and the stearin’s own iodine value.

On top of this experimental foundation, the researchers built two complementary modeling frameworks. The first was a set of stage-specific response surface methodology models, which quantified the effects of individual process variables and served as an interpretable benchmark against which the machine learning approach could be judged. The second was a multilayer feedforward artificial neural network with a 7-13-5 architecture, meaning it took seven inputs, processed them through a hidden layer of thirteen neurons, and produced five outputs corresponding to the five yield and quality measures. Neural networks of this kind excel at capturing nonlinear interactions that would defeat simpler statistical models, learning the hidden structure of the process directly from data.

The reported performance of the neural network was strong, with coefficients of determination, or R-squared values, ranging from 0.79 to 0.96 across the five outputs. In practical terms, that means the network explained up to 96 percent of the variance in the measured outcomes, though the spread indicates that some quality attributes were harder to predict than others. But a good predictive model is only half the battle. The researchers wanted a tool that could actively find the best operating conditions, not merely describe what had already happened.

To achieve that, they turned to a genetic algorithm, an optimization technique inspired by biological evolution. Genetic algorithms work by maintaining a population of candidate solutions, in this case proposed cooling profiles, and iteratively selecting, combining, and mutating them so that better solutions survive and reproduce. Because the neural network can evaluate a candidate profile almost instantly, the algorithm can search thousands of possibilities without running a single physical experiment. When the hybrid ANN-GA framework was unleashed on the data, it identified feed-specific optimum cooling conditions for each iodine value, predicting olein yields of 83.80 percent for IV51, 83.10 percent for IV52, and 87.98 percent for IV53. The notably higher yield for the IV53 feed reflects the underlying chemistry: more unsaturated oil leaves more liquid fraction to harvest.

The crucial test came when the team took those computer-generated optima back to the laboratory. Experimental validation of the predicted conditions produced overall prediction accuracies of 91.94 percent, 96.97 percent, and 97.75 percent for the three feedstocks, with corresponding mean absolute percentage errors of 8.06 percent, 3.03 percent, and 2.25 percent. Those error figures are remarkably low for a process as variable as palm oil fractionation, and they demonstrate that the hybrid framework is not just a statistical curiosity but a genuinely reliable guide to real-world operation. The authors note that response surface methodology provided robust interpretability, whereas the ANN-GA combination offered greater flexibility for the nonlinear, multistage interactions that define the process.

The implications stretch well beyond a single refinery. Palm oil fractionation is a multibillion-dollar industry, and even a few percentage points of additional olein yield translate into enormous gains in efficiency, profitability, and sustainability, since more usable product from the same feedstock means less waste and less pressure on plantations. More broadly, the study offers a template for feed-responsive optimization in any food or chemical process where raw material variability confounds fixed recipes. The researchers frame their work as establishing a basis for adaptive crystallization control and, ultimately, the digitalization of palm oil fractionation, in which sensors feed live data to models that continuously retune cooling profiles as each new batch arrives. The humble crystallizer, it seems, is finally getting a brain.

Subject of Research: Hybrid artificial neural network and genetic algorithm modeling and optimization of multistage palm oil fractional crystallization

Article Title: Hybrid Artificial Neural Network-Genetic Algorithm modelling and optimization of multistage palm oil fractional crystallization

Article References: Othman, N. H., Mohd Kasihmuddin, M. S., Bhagya Raj, G. V. S., Dash, K. K., Saparin, N., & Zulkurnain, M. (2026). Hybrid Artificial Neural Network-Genetic Algorithm modelling and optimization of multistage palm oil fractional crystallization. npj Science of Food. https://doi.org/10.1038/s41538-026-01133-7

Image Credits: AI Generated

DOI: 10.1038/s41538-026-01133-7

Keywords: palm oil, fractional crystallization, artificial neural network, genetic algorithm, response surface methodology, olein yield, iodine value, food processing, process optimization, machine learning, palm stearin, digitalization

The stakes are higher than most people realize. Traditional geotechnical design has long relied on empirical rules, simplified analytical models, and generous safety factors to cope with the enormous uncertainty inherent in soil behavior. That conservatism keeps structures safe, but it also produces overdesigned foundations, wasted concrete and steel, inflated construction costs, and unnecessary embodied carbon. Optimization algorithms promise a way out: by mathematically searching the design space for solutions that satisfy stability, settlement, and bearing capacity requirements while minimizing cost and environmental impact, they allow engineers to explore far more of the possible design space than trial-and-error ever could. The trouble, Dehghanian argues, is that most existing reviews of these methods are purely descriptive, cataloguing applications without ever comparing them under consistent, criteria-driven rules.

To fix that, the study classifies optimization techniques into four families: classical deterministic methods, metaheuristic algorithms, surrogate-based approaches, and artificial intelligence-driven techniques. Deterministic methods, including gradient-based techniques, linear and quadratic programming, and sequential quadratic programming, converge reliably when objective functions are smooth and convex, but they stumble on the nonlinear, discontinuous, and multimodal landscapes that soil behavior typically produces. Metaheuristics, inspired by evolution, swarming birds, annealing metals, and foraging ants, dominate the field precisely because they can navigate such messy terrain without needing gradients. Genetic algorithms evolve populations of candidate solutions through selection, crossover, and mutation; particle swarm optimization moves a flock of particles through the search space guided by individual and collective memory; and a growing menagerie of newer algorithms, from the Grey Wolf Optimizer to the Whale Optimization Algorithm, has been applied to foundations, slopes, and parameter identification problems.

Yet the review delivers a sobering assessment of the metaheuristic boom. Each optimization iteration may demand a full finite-element simulation, and a typical study can require thousands of such evaluations before converging, making computational cost a serious bottleneck. Performance also depends heavily on algorithm-specific tuning parameters such as population size, mutation rate, and inertia weight. Perhaps most provocatively, the study notes that recent benchmarking research has shown performance differences among many newly proposed metaheuristics often become statistically insignificant once parameters are properly calibrated. The No-Free-Lunch theorem looms over the entire field: averaged across all possible problems, every optimization algorithm performs identically, so algorithm selection must rest on problem characteristics rather than novelty.

To move beyond qualitative judgments, Dehghanian constructed a unified evaluation framework with explicitly quantified criteria: convergence efficiency, measured by how quickly an algorithm reaches 95 percent of its final optimum; data efficiency, measured by the number of objective-function evaluations required; interpretability; uncertainty quantification capability; scalability; constraint-handling robustness, expressed as the fraction of feasible solutions generated; and practical maturity. This framework was then stress-tested with a benchmark problem: optimizing the width of a square shallow footing on homogeneous medium-dense sand, minimizing the foundation footprint while satisfying a minimum bearing capacity factor of safety of 3.0 and a settlement limit of 25 millimeters. Three algorithms, a genetic algorithm, particle swarm optimization, and Gaussian Process Regression-based Bayesian Optimization, tackled the identical constrained problem.

The results are striking. All three methods converged to nearly the same optimal design, a footing width of roughly 0.80 meters, with a computed settlement of about 2.55 millimeters, far below the allowable limit, and a bearing capacity factor of safety exceeding 5.5. But the paths to that answer differed dramatically. The metaheuristic methods required many objective-function evaluations and occasionally violated settlement constraints during early iterations. Bayesian Optimization, by contrast, terminated after just 30 evaluations in a total elapsed time of 13.38 seconds, with the Gaussian process model’s best-predicted feasible point exactly matching the best-observed solution. The surrogate-based approach achieved equivalent design quality at a fraction of the computational cost, a decisive advantage when each evaluation means running an expensive numerical simulation.

Even more important than raw speed is what Bayesian Optimization knows about its own ignorance. Because Gaussian Process Regression produces both a prediction and a predictive variance for every candidate design, the algorithm can translate statistical uncertainty directly into engineering reliability. The review illustrates this with a compelling example: two footing designs, one predicting 19 millimeters of settlement with a standard deviation of 8, the other predicting 21 millimeters with a standard deviation of only 2. A deterministic optimizer would prefer the first design, but its large uncertainty carries a real risk of exceeding the 25-millimeter threshold. The second design, though nominally worse on average, is far more defensible. By feeding uncertainty into the acquisition function, Bayesian Optimization balances exploiting promising regions against exploring poorly understood ones, effectively steering itself away from unreliable designs.

The study also tackles the black-box problem that has long plagued machine learning in safety-critical engineering. Using SHAP, or SHapley Additive exPlanations, values to interpret a surrogate-assisted footing design model, the analysis revealed that soil friction angle exerted the strongest influence on optimization outcomes, a finding that maps directly onto classical bearing capacity theory, in which the bearing capacity factors depend strongly on the friction angle. This kind of physically meaningful explanation, the review argues, is essential for engineers who must accept or reject algorithmic recommendations, and it distinguishes genuine geomechanical mechanisms from mere statistical correlations. The study similarly highlights physics-informed neural networks, which embed governing equations into the learning process and require less training data than purely data-driven models.

The review’s scope extends well beyond foundations. For slope stability, optimization algorithms automate the search for the critical slip surface within limit equilibrium methods, where the search space is riddled with local minima that defeat traditional grid searches. Ant colony optimization, genetic algorithms, and particle swarm optimization have all proven effective here, and reliability-based design optimization now allows engineers to minimize excavation volume subject to explicit limits on the probability of failure. In site characterization, optimization coupled with geostatistics can determine optimal borehole layouts that minimize estimation uncertainty, while inverse analysis calibrates constitutive model parameters against laboratory and field data. Machine learning models, tuned by optimization algorithms themselves, increasingly predict soil properties from cone penetration and standard penetration test data, and transfer learning is allowing data-scarce regions to leverage models trained on large regional databases.

Looking forward, Dehghanian identifies digital twins, reinforcement learning for adaptive excavation, real-time sensor-driven optimization, and integration with Building Information Modeling as the field’s next frontiers, alongside persistent challenges in computational scalability, software interoperability, and the ethics of autonomous decision-making. The practical guidance distilled from the benchmark is refreshingly concrete: for inexpensive problems, genetic algorithms and particle swarm optimization remain robust workhorses; for expensive finite-element analyses, Bayesian and surrogate-assisted methods are generally preferable; for reliability-critical projects, uncertainty-aware frameworks should be prioritized; and explainability should be treated as a requirement, not an optional extra. The overarching message is one of disciplined pragmatism. Optimization algorithms are not replacements for engineering judgment but powerful decision-support tools, and choosing among them should be an evidence-based engineering decision, guided by computational budget, problem dimensionality, data availability, and the non-negotiable demands of safety.

Subject of Research: Comparative assessment of optimization algorithms for geotechnical engineering design

Article Title: A systematic review and comparative assessment of optimization algorithms in geotechnical engineering with a benchmark foundation design case study

Article References: Dehghanian, K. (2026). A systematic review and comparative assessment of optimization algorithms in geotechnical engineering with a benchmark foundation design case study. Discover Geoscience, 4(1), Article 304. https://doi.org/10.1007/s44288-026-00642-y

Image Credits: AI Generated

DOI: 10.1007/s44288-026-00642-y

Keywords: geotechnical engineering, optimization algorithms, Bayesian optimization, genetic algorithm, particle swarm optimization, foundation design, metaheuristics, uncertainty quantification, surrogate modeling, explainable AI, slope stability, machine learning