In the case study, smart display units originated at electronics manufacturing centers in Oragadam and Sriperumbudur and were assigned to freight depots in Kanchipuram and Vellore. Deliveries could use either an automated electric truck or an automated electric van over one of two designated routes for each source–destination pairing. The model selected shipment quantities from the available combinations; it did not generate new road routes, determine sequences of stops or simulate driving behavior.

Its scope was therefore a mathematical allocation model for a specified network, not a field trial of automated vehicles. Two competing objectives were minimized simultaneously. The first combined direct transportation expenses, automation costs and energy-related costs.

The second sought to reduce total transportation time. The researchers represented the tradeoff by treating the cost and time objectives as parties in a cooperative bargaining calculation, with a deviation term favoring proximity to ideal outcomes and a satisfaction term favoring plans that perform well across both objectives. In the payoff calculation, the cost-only optimum was ₹147,480 with a transportation-time value of 13,076.1 hours, while the time-only optimum was ₹169,615.76 with 10,159.20 hours.

The distinct optima showed that minimizing one measure did not produce the same allocation as minimizing the other. Uncertain inputs were represented with triangular Fermatean fuzzy numbers, a framework designed to accommodate both membership and non-membership information under the constraint that the cubes of the two degrees do not exceed one. This permits assessments that would be excluded by intuitionistic or Pythagorean fuzzy frameworks in some cases.

The study modeled transportation costs, unit time and energy-consumption costs as fuzzy because market fluctuations, congestion, road conditions and weather can make such values difficult to specify precisely. Route distances, battery capacities, vehicle payloads, supply quantities and destination demands were treated as exact values. The case used two vehicle categories, each with defined capacity and battery limits, and a small set of designated routes rather than a comprehensive road network.

Because directly solving a bi-objective model with fuzzy coefficients would be computationally complex, the researchers first applied a score function to transform each triangular Fermatean fuzzy number into a deterministic equivalent. They then solved the cost and time objectives separately to identify their ideal and anti-ideal values. The hybrid Fermatean fuzzy game-theoretic method normalized each objective against its corresponding ideal and anti-ideal values, making deviations dimensionless despite the objectives having different units.

It then combined these normalized disagreements with Fermatean fuzzy satisfaction and dissatisfaction measures. The resulting single-objective model was implemented in LINGO with its global solver, converting the original bi-objective problem into a computationally tractable optimization task. The selected plan sent 2.0542 quintals from Oragadam to Kanchipuram by automated electric truck on route 1, 57.9457 quintals from Oragadam to Vellore by truck on route 2, 42.9457 quintals from Sriperumbudur to Kanchipuram by truck on route 1, and 7.0542 quintals from Sriperumbudur to Vellore by automated electric van on route 1.

The quantities satisfied the stated supply, demand, vehicle-capacity, battery-range and non-negativity constraints. For this allocation, the estimated transportation cost of ₹148,490 comprised ₹108,901.64 in vehicle movement and distribution expenses, ₹34,575.84 in automation costs and ₹5,012.52 in energy-consumption costs. A separate optimality test found no feasible allocation that could improve both cost and time simultaneously, confirming that the result was Pareto optimal for the two stated objectives.

In a comparison using the same deterministic model and normalization values, the Fermatean fuzzy programming method returned a cost of ₹169,607.90 and a time of 13,076.1 hours. A weighted sum method with equal weights of 0.5 for the normalized objectives returned ₹148,490 and 13,076.1 hours. The hybrid game-theoretic result matched the weighted sum method’s cost but reduced its time estimate by 133.16 hours, or about 1.02%.

Satisfaction levels for the two objectives were 99.32% and 72.60% under the hybrid method, compared with 99.32% and 71.29% under the weighted sum approach. Its degree of closeness was 0.1109, lower than the weighted sum method’s 0.1149 and the fuzzy programming method’s 0.1768; the paper treats a smaller degree of closeness as preferable. These comparisons concern the model’s calculated allocations and objective values, not measured operational performance in a live logistics system.

The computational tests indicated that the method could handle larger arrays of sources, destinations, vehicles and routes. Reported solution times rose from 0.01 seconds for a 2 × 2 × 2 × 2 instance with 16 decision variables and 12 structural constraints to 0.05 seconds for a 3 × 3 × 3 × 3 instance, 0.18 seconds for a 4 × 4 × 4 × 4 instance and 0.76 seconds for a 5 × 5 × 5 × 5 instance with 625 decision variables and 60 structural constraints. Sensitivity tests preserved feasibility when supply increased by 1% to 10% and when demand decreased by 1% to 10%.

With a 10% supply increase, the cost estimate fell by approximately 1.07% while time rose by approximately 0.69%. With a 10% demand reduction, time fell from 12,942.94 hours to 9,606.514 hours, while cost changed from ₹148,490 in the main results to ₹148,032.70 in the sensitivity table. These results describe the mathematical model’s response to parameter changes, not measured changes in an operating logistics network.

The study’s main limitation is that its sustainability and deployability claims extend beyond the variables included in the optimization. The model omits charging-station placement, communication-network requirements, data-transfer security, cybersecurity threats and unauthorized vehicle control. It also treats several uncertainty sources through fuzzy parameters rather than stochastic distributions, handles a limited product case and evaluates routes that are already predefined.

The authors note that lowering energy-related costs and time may indirectly reduce battery consumption and environmental impacts associated with electricity generation, but the case study did not measure emissions, fuel displacement, noise reduction or air-quality changes. They identify charging constraints, intrusion detection, multiple products, multilevel warehouse planning and stochastic uncertainty as possible extensions. Subject to those limitations, the case study shows how a Fermatean fuzzy game-theoretic method can convert uncertain cost and time inputs into one computationally tractable allocation for automated electric vehicle logistics.

Subject of Research: Fuzzy multi-objective optimization for automated electric vehicle transportation planning

Article Title: A sustainable Fermatean fuzzy bi-objective four-dimensional transportation model for automated electric vehicle logistics planning using a Fermatean fuzzy game-theoretic method

Article References: Kumar, E., & Dhanapal, A. (2026). A sustainable Fermatean fuzzy bi-objective four-dimensional transportation model for automated electric vehicle logistics planning using a Fermatean fuzzy game-theoretic method. Frontiers in Sustainability, 7, Article 1944956. https://doi.org/10.3389/frsus.2026.1944956

Image Credits: AI Generated

DOI: 10.3389/frsus.2026.1944956

Keywords: automated electric vehicles, logistics optimization, Fermatean fuzzy sets, game theory, four-dimensional transportation, supply chain management, Pareto optimization, uncertain inputs, smart display units, sustainable logistics, sustainable, Fermatean

The core difficulty is that cloud task allocation is not one optimization problem but several at once. Squeezing completion time, energy, SLA compliance, and balance into a single scalar reward, as many existing deep reinforcement learning (DRL) schedulers do, forces an implicit and fixed trade-off among objectives that operators cannot easily change. Population-based multi-objective evolutionary optimizers such as NSGA-II and NSGA-III can explore trade-offs explicitly, but repeatedly running them for online scheduling decisions incurs substantial computational overhead. MO-UDRL-H is designed to sidestep both pitfalls by learning a distributional value model conditioned on operator preferences and then shielding candidate actions with Pareto-aware safety checks.

Technically, the framework integrates several components. A preference-based global fusion embedding module encodes the observable scheduling state together with a preference vector that expresses how much weight the operator places on each objective. An implicit quantile network (IQN) value head estimates the full distribution of returns rather than only their mean, which matters because cloud workloads are stochastic and tail risk is as important as average performance. Preference-conditioned hypernetwork modulation then adapts the network’s behavior to the chosen preference, while hypervolume-aware training pushes the learned policy toward solutions that dominate as much of the objective space as possible.

The Hybrid Pareto Safety Shield is the component that gives the method its name. At each scheduling step, the shield constructs a set of candidate actions using only the currently observable state: neural candidates from the learned policy, latency-aware, energy-aware, SLA-aware, balance-aware, and risk-aware heuristic candidates, and a Pareto-greedy candidate. Feasibility- and preference-aware selection then picks among these candidates without any access to future workload information. This construction means the scheduler can never commit to an action that blatantly violates a hard constraint, and the Pareto-greedy candidate keeps the multi-objective character of the problem visible even when the neural policy is still learning.

To make the approach practical for online deployment, the authors introduce MO-UDRL-D, a distilled NumPy deployment policy that imitates the behavior of MO-UDRL-H. Policy distillation compresses the teacher’s decision function into a lightweight student that reduces inference latency to the same order of magnitude as fast rule-based heuristic schedulers. In experiments, MO-UDRL-D achieves approximately 0.015 milliseconds of inference latency while retaining substantial teacher behavior, providing a practical quality-latency trade-off for production clusters where the scheduler itself must not become a bottleneck.

The experimental evaluation uses Alibaba cluster trace-driven task allocation with five random seeds, comparing MO-UDRL-H against heuristic, evolutionary, scalar DRL, multi-objective RL, and distributional RL baselines, including DQN, PPO, NSGA-II-style, and NSGA-III-style methods. MO-UDRL-H achieves the highest mean Pareto hypervolume, the standard measure of how much of the objective space a solution set covers, and the lowest generational distance and Spacing among the applicable multi-solution methods. Under selected preference settings, it also achieves best or tied-best performance on latency, energy, SLA compliance, load balance, and CVaR@10%, a conditional value-at-risk metric that quantifies expected performance in the worst ten percent of outcomes.

The use of CVaR@10% is particularly noteworthy for risk-sensitive operators. Conditional value-at-risk, originally developed in financial risk management by Rockafellar and Uryasev, measures the average of the worst-case tail of a distribution rather than a single extreme point. By optimizing this metric, the scheduler explicitly controls tail latency and worst-case SLA violations, which are precisely the failure modes that erode user trust in cloud platforms. The distributional value head makes this possible because estimating the full return distribution is a prerequisite for reasoning about its tail.

The significance of this work lies in its synthesis of previously separate research threads. Distributional reinforcement learning, introduced by Bellemare and colleagues and refined through quantile regression and implicit quantile networks, provides the statistical machinery for reasoning about return distributions. Multi-objective reinforcement learning provides the preference-conditioned framework for representing entire families of policies. Hypernetworks, which generate network weights conditioned on an input vector, provide the mechanism for modulating a single network across preferences. The Pareto safety shield ties these threads together into an end-to-end scheduler that is both expressive and safe.

For cloud providers, the practical implications are immediate. A scheduler that can be re-tuned by changing a preference vector, rather than retrained from scratch, allows operators to shift priorities between energy saving and latency reduction as business needs change. The distilled deployment policy means the sophisticated teacher can run offline while the lightweight student handles online decisions at heuristic speed. The evaluation on public Alibaba and Microsoft Azure packing traces, both openly available for academic analysis, means other researchers can reproduce and extend the comparison.

The authors acknowledge contributions from other members of their research group to data preprocessing and model design. The work was conducted at the State Key Laboratory of Ocean Sensing and Ocean College, Zhejiang University, in Zhoushan, China, and the authors declare no competing interests. As cloud platforms grow and energy costs rise, schedulers that can balance multiple objectives explicitly, control tail risk, and still decide in microseconds may become as essential to datacenter operations as the hypervisors that manage the machines themselves. MO-UDRL-H offers a concrete template for how distributional, preference-conditioned, and Pareto-safe learning can be combined to meet that standard.

Subject of Research: Multi-objective distributional reinforcement learning for Pareto-safe cloud task allocation

Article Title: MO-UDRL-H: hybrid Pareto-safe distributional reinforcement learning for multi-objective cloud task allocation

Article References: Xu, C., & Hu, C. (2026). MO-UDRL-H: hybrid Pareto-safe distributional reinforcement learning for multi-objective cloud task allocation. Applied Intelligence, 56(16), Article 506. https://doi.org/10.1007/s10489-026-07534-1

Image Credits: AI Generated

DOI: 10.1007/s10489-026-07534-1

Keywords: reinforcement learning, cloud computing, multi-objective optimization, Pareto optimization, distributional reinforcement learning, task scheduling, policy distillation, hypervolume, MO-UDRL-H, hybrid, Pareto-safe, distributional

Researchers from Abylkas Saginov Karaganda Technical University in Kazakhstan and Peter the Great St. Petersburg Polytechnic University in Russia have developed a hierarchical, multi-criteria optimization model that simultaneously balances the workload of machine tools, transport vehicles, and storage systems across five nested time horizons: the calendar year, the quarter, the month, the working day, and the shift. When they tested the model on a real production unit running two shifts, 83 product groups, and 450 manufacturing operations, annual equipment efficiency rose by 15.11 percent and total system downtime fell by 19.31 percent.

The core insight is deceptively simple. Traditional planning approaches optimize one resource at a time: first the machines, then the vehicles, then the warehouse. Each step may look locally sensible, but the sequence creates information gaps and interface downtime that no single department can see, let alone fix. The new model instead treats the entire asset portfolio as one interconnected system, with the output of each higher planning level acting as a rigid constraint on the level below it.

Mathematically, the framework is built on vector optimization criteria evaluated through Pareto analysis. At each management level, the researchers generated sets of non-dominant alternatives, meaning solutions in which no single objective can be improved without degrading another. For the annual allocation of machining operations across 17 machines, including CNC lathes, turning-milling centers, boring machines, and cylindrical and surface grinders, the team evaluated eleven Pareto-optimal variants. For the fleet of three electric pallet trucks, they generated 23 non-dominated transport allocation scenarios. Choosing among these trade-off fronts, rather than collapsing everything into a single weighted score, is what allows the model to respect the distinct physics of each asset class.

The hierarchy itself is unusual. Where classical advanced planning systems decompose by time horizon, strategic, tactical, operational, the Kazakh-Russian team decomposes by both time and resource type. Sublevel 1.1 assigns machining operations to specific machines for the year; its output matrix then constrains sublevel 1.2, which assigns inter-operation storage locations, because each machine is physically tied to a particular warehouse. The storage solution in turn constrains sublevel 1.3, which allocates transport operations, since the type of transport move, inter-shop or intra-shop, determines route duration and the loading equipment required. The processing time of a part on a machine directly dictates the delivery schedule of the vehicle that serves it, and the physical capacity of the warehouse limits the production rhythm of the machines.

The practical results are striking. Redistributing production batches according to Pareto-optimal solutions cut load variability on milling machines by 6.78 percent and on grinding machines by 8.65 percent. Vehicle capacity utilization jumped by 37.06 percent while vehicle downtime dropped by 28.86 percent. At the quarterly level, equipment load variability fell by 34.68 percent and downtime unevenness by 54.24 percent. At the shift level, load variability declined by 36.59 percent and vehicle downtime unevenness by 14.28 percent. Across the whole system, load variability decreased by 15.9 percent.

One honest limitation stands out. At the annual planning level, the researchers deliberately excluded failure-probability parameters because the enterprise lacked reliable statistical data on breakdowns of specific machines and vehicles. Rather than inject unreliable reliability figures that could distort the Pareto front, they omitted those components from the global-level equations, a pragmatic decision that highlights how data quality, not just algorithm design, constrains real-world optimization.

The model’s modular architecture is arguably its most forward-looking feature. Each new class of production asset, whether a casting press, an automated guided vehicle fleet, or an automated storage and retrieval system, can be added as a separate management sublevel without redesigning the entire planning structure. This scalability matters as manufacturers digitize and reconfigure lines dynamically, and the authors suggest the framework could eventually integrate with ERP and MES systems to enable predictive analytics and automatic schedule generation.

The downstream consequences reach beyond the shop floor. By eliminating the gaps between production, logistics, and warehousing, the approach reduces production cycle times, shrinks work-in-process inventory, and cuts order delays. Coordinated transport planning means fewer idle runs and lower fuel consumption. Balanced schedules mean dispatchers spend less time firefighting and more time on strategy. For an industry under pressure to deliver more with less, the message is clear: the biggest efficiency gains may lie not in any single machine, but in the mathematics that binds them all together.

Subject of Research: Hierarchical multi-criteria optimization of production, transport, and storage asset utilization in machine-building systems

Article Title: Integrated multi-criteria optimization of production, transport, and storage assets in machine-building systems

Article References: Nurzhanova, O., Khrustaleva, I., Shkodyrev, V., Zharkevich, O., Bakenov, A., Zakirov, K., & Safaryan, A. (2026). Integrated multi-criteria optimization of production, transport, and storage assets in machine-building systems. Frontiers in Mechanical Engineering, 12, Article 1885264. https://doi.org/10.3389/fmech.2026.1885264

Image Credits: AI Generated

DOI: 10.3389/fmech.2026.1885264

Keywords: multi-criteria optimization, Pareto optimization, production planning, machine-building, hierarchical scheduling, equipment utilization, downtime reduction, industrial logistics, vector optimization, manufacturing systems, warehouse planning, vehicle fleet scheduling

The research, led by T. Thomas Leonid of KCG College of Technology in Chennai, together with Gaurav Vishnu Londhe of Symbiosis Institute of Technology in Nagpur and J. Robert Theivadas of Vel Tech Rangarajan Dr Sagunthala R & D Institute of Science and Technology in Avadi, presents a framework that fuses module-specific risk scores into what the authors call a monitoring-priority index, or MPI. Rather than pretending that a fall risk and a cardiac risk are the same kind of probability, the system treats them as evidence of different natures and stratifies them accordingly. The goal is not to diagnose disease but to decide, at any given moment, which older adult most needs human attention.

At the heart of the method sits an Adaptive Neuro-Fuzzy Inference System, or ANFIS, a hybrid architecture that combines the human-readable structure of fuzzy if-then rules with the learning capacity of neural networks. Fuzzy rules are attractive in medical settings because they can be read and audited: a rule might state that high cardiac risk combined with moderate mobility decline and poor signal quality should raise the monitoring priority. The trouble is combinatorics. With seven input modules each described by three linguistic levels, the full antecedent grid would contain three-to-the-seventh, or 2,187, possible rules, far too many to interpret or maintain. The team therefore reduced this grid to a training-derived library of 64 rules using an explicitly specified ranking criterion, and then let an optimizer decide which of those rules actually earn a place in the deployed policy.

That optimizer is the Non-dominated Sorting Genetic Algorithm-II, NSGA-II, a well-established evolutionary method for multi-objective optimization. Instead of collapsing every desideratum into a single score, NSGA-II evolves a population of candidate rule-selection policies and returns a Pareto front of solutions that trade off competing objectives, such as discrimination against parsimony. In the final frozen policy, 20 of the 64 candidate rules remained active. This two-stage design, pruning first and optimizing second, is what keeps the resulting model small enough for a human expert to inspect rule by rule, which is the entire point of choosing a fuzzy architecture in the first place.

Because no synchronized seven-modality cohort of older adults exists in the public domain, the authors evaluated the complete fusion mechanism on a controlled, subject-indexed synthetic benchmark of 240 subjects and 2,880 observation windows, while real datasets supplied module-level evidence and targets for future prospective validation. On 48 held-out subjects, the frozen 20-rule policy achieved an area under the receiver operating characteristic curve of 0.914, with a 95 percent participant-cluster bootstrap confidence interval of 0.871 to 0.948. The area under the precision-recall curve was 0.888, sensitivity 0.868 and specificity 0.751. These are respectable numbers, but the study is notable less for the numbers themselves than for the unusual honesty with which they are contextualized.

That honesty emerges most clearly in the comparisons. When the proposed model was tested against logistic regression, a generalized additive model, a random forest and a simple quality-weighted mean, paired participant-cluster statistical tests found that all of these comparators reached discrimination that was statistically indistinguishable from the ANFIS-NSGA-II policy after correction for multiple comparisons. The proposed model was significantly better than only one alternative: a naive max-risk rule that simply escalates on the worst single module. In other words, the interpretable machinery does not buy predictive accuracy; it buys auditability at essentially no cost in performance, which is itself a meaningful result for anyone designing clinical decision-support systems.

The ablation experiments sharpen this picture further. The authors identified 17 mandatory anchor rules, and a policy restricted to those anchors alone performed significantly worse, with an AUROC of 0.861. Intriguingly, when the anchor requirement was relaxed entirely, the optimizer found policies with only 8 rules that achieved significantly higher discrimination, suggesting that the anchor constraints trade accuracy for coverage of clinically important rule combinations. The team also examined how missing modalities should be handled, finding that quality-weighted marginalization is preferable to simply zero-filling absent signals, yet becomes significantly inferior to simple imputation when signal quality is uniformly high, with the alternatives converging only when quality genuinely varies across recordings.

Perhaps the most methodologically interesting section is a scenario analysis that varies interaction strength, interaction identity, label noise, the functional family of the target and the share of escalation propensity driven by unobserved latent states. This analysis quantifies how much of the reported performance reflects recovery of the researchers’ own data-generating mechanism rather than a generalizable capability. Without label noise, the target is deterministic given the seven scores and the generator’s subject-level offset, and the model attains between 72 and 98 percent of the oracle discrimination computed from the true propensity. Across 14 scenarios, the proposed model exceeded logistic regression in only one. The authors are explicit that every quantitative result is a proof-of-concept methodological finding on synthetic data, and that the contribution is an auditable, reproducible fusion and policy-selection architecture rather than a gain in predictive accuracy.

Reproducibility has been treated as a first-class deliverable. The published package includes the executable study-reproduction pipeline, figure-generation scripts, a fixed master seed and environment requirements, the synthetic dataset itself, the 64-rule numeric library, the selected active rules with their support and contribution values, initial and learned membership parameters, the 30-run final Pareto fronts, held-out confidence intervals, comparator outputs with paired test statistics, ablation and scenario-analysis outputs, runtime metadata and a provenance file recording the numeric values behind every figure. Restricted human data are not redistributed, and no new human participants were recruited for the fusion experiment. This level of transparency means any laboratory can rebuild the entire benchmark from scratch and check each claim independently.

The practical implications are twofold. For developers of elder-care wearables, the study offers a template for combining heterogeneous risk modules without collapsing them into a single opaque probability, and for documenting exactly why a system raises or lowers an alert. For clinicians and regulators, it demonstrates that interpretability need not come at the price of performance, even if it does not automatically exceed simpler baselines. The authors themselves stress that prospective, synchronized multimodal validation with clinician evaluation remains necessary before any clinical use. Until such trials exist, the framework stands as a carefully documented proof of concept: an argument that the path to trustworthy health monitoring may run through fewer, more legible rules rather than larger black boxes.

Subject of Research: An interpretable neuro-fuzzy and evolutionary optimization framework for fusing heterogeneous wearable health risk signals in older-adult monitoring

Article Title: An interpretable ANFIS-NSGA-II framework for evidence-stratified modular wearable risk fusion in elder-health monitoring

Article References: Leonid, T. T., Londhe, G. V., & Theivadas, J. R. (2026). An interpretable ANFIS-NSGA-II framework for evidence-stratified modular wearable risk fusion in elder-health monitoring. Scientific Reports. https://doi.org/10.1038/s41598-026-70620-4

Image Credits: AI Generated

DOI: 10.1038/s41598-026-70620-4

Keywords: ANFIS, NSGA-II, wearable sensing, interpretable artificial intelligence, elder-health monitoring, monitoring-priority index, fuzzy inference system, Pareto optimization, missing-modality handling, synthetic benchmark, risk fusion, reproducibility

The research, led by Raj Kumar Batchu of Amrita Vishwa Vidyapeetham together with colleagues at the National Kaohsiung University of Science and Technology, tackles a blind spot that has persisted for two decades. Since the early signature-based detectors of the 2000s, intrusion detection research has migrated toward deep learning architectures that achieve spectacular accuracy on benchmark datasets. Yet the authors point out that nearly all comparative studies report a single number, usually accuracy or F1-score, while ignoring the questions that actually decide whether a system survives in production: Can it run on a constrained edge device? Does it catch every category of attack or just the common ones? And what will it truly cost to operate over five years?

To answer those questions, the team built a multi-dimensional evaluation framework resting on three pillars. The first, Pareto Frontier Optimization, maps each candidate detector in a three-dimensional space of detection performance, computational efficiency and deployment feasibility, then identifies non-dominated solutions where no objective can be improved without sacrificing another. The second, a newly introduced metric called Attack Surface Coverage, compresses true positive rate, false positive minimization, adaptability to novel attacks and inference speed into a single weighted score. The third, a Total Cost of Ownership model, breaks down five years of deployment across 1,000 IIoT devices into five components: deployment, operations, incident response, scalability and compliance. A synthesis engine then fuses all dimensions into one composite ranking, with weights that organizations can tune to their own priorities.

The framework was stress-tested against three state-of-the-art intrusion detection models, all evaluated on the Edge-IIoTset 2022 dataset, a realistic benchmark containing 61 network traffic features and 15 distinct attack types spanning reconnaissance, exploitation and lateral movement. Model A was a hybrid convolutional neural network combined with gated recurrent units and long short-term memory layers, trained through asynchronous federated learning so that edge nodes keep their data local while a global model is aggregated centrally. Model B paired an autoencoder, which compresses network flows into low-dimensional representations, with a decision tree classifier, an unusually lightweight combination validated on an NVIDIA Jetson Nano. Model C ran parallel CNN and bidirectional LSTM branches topped with a self-attention mechanism, trained with SMOTE oversampling to handle class imbalance across rare attack categories.

The headline results upend the intuition that the most accurate model wins. On raw detection, the federated CNN-GRU-LSTM model was flawless, achieving 100 percent accuracy, precision, recall and F1-score with a measured zero false positive rate. But in the Pareto analysis, the humble decision tree with autoencoder dominated, earning a composite score of 0.999 against 0.333 for the federated deep model and 0.344 for the attention architecture. The reason is brutal arithmetic: the deep models carry footprints of 80 to 100 megabytes and training cycles measured in thousands of seconds, while the decision tree pipeline infers in roughly 0.185 milliseconds per sample with a compact model that edge hardware can actually host.

The Attack Surface Coverage scoring told a different story. Here the LSTM-CNN-Attention model led with 98.16 percent, edging out the federated model at 97.00 and the decision tree at 95.48. The attention mechanism, which re-weights the most salient temporal and spatial features of traffic sequences, gave the architecture an edge in generalizing to attack behaviors it had never seen, a property the authors encode as an architecture-informed adaptability score. The framework is candid that this sub-metric reflects a defensible relative ordering across architecture families rather than a directly measured zero-day detection rate, a transparency the authors extend to every formula, weight and cost assumption in the paper.

The economic dimension produced perhaps the most striking finding of all. Over a five-year horizon across 1,000 devices processing 10,000 flows each per day, the CNN-GRU-LSTM model achieved the lowest total cost of ownership at 1.385 million dollars, compared with 5.5 million for LSTM-CNN-Attention and a staggering 28.16 million for the decision tree with autoencoder. The paradox dissolves when the cost model is examined closely. Because the federated model’s false positive rate is exactly zero, its incident response cost, which multiplies every false alert by analyst investigation time and labor rates, is structurally zero. The lightweight decision tree, despite cheap hardware, generates enough false alarms at industrial traffic volumes to accumulate enormous analyst workload. A sensitivity analysis varying traffic volume, investigation time and labor rates confirmed that the federated model’s cost advantage is tied to its zero false positives rather than to any particular assumption.

When the synthesis engine combined all dimensions with its default weights, 30 percent detection, 25 percent security coverage, 20 percent cost, 15 percent deployment and 10 percent efficiency, the decision tree with autoencoder claimed the highest composite score at 78.85 percent, ahead of the federated model at 74.25 and the attention model at 71.98. Yet the authors emphasize that no single model dominates every criterion, and their sensitivity analysis shows the ranking shifts predictably as organizational priorities change. Weighting detection and cost favors the federated model; weighting edge deployment and efficiency favors the decision tree; weighting adaptability favors the attention architecture. The framework is deliberately parametric, letting a security team adjust the weight vector to match its own threat landscape and budget.

The study is honest about its limits. The authors did not independently retrain the models, drawing instead on previously published results, so the comparative conclusions should be treated as indicative rather than definitive, and rankings may vary with different datasets, weights and cost parameters. Still, the implications reach well beyond this single benchmark. For an industry where regulatory frameworks such as IEC 62443 demand auditable justifications for security decisions, and where black-box deep models can inflate compliance costs through required third-party validation, a transparent, reproducible selection framework that prices interpretability, alert fatigue and scalability into the decision could prove transformative. The full calculation logic has been released in a public repository, and the authors plan to extend the approach to broader families of lightweight algorithms and additional datasets.

What emerges is a deceptively simple message with far-reaching consequences: the best intrusion detector for a factory floor is not the one with the highest accuracy on a leaderboard. It is the one whose accuracy, coverage, hardware demands and five-year bill align with the specific constraints of the deployment. By turning an intuition into equations, weights and a tunable decision matrix, this work gives industrial defenders something they have rarely had, a defensible, quantified way to choose the shield that fits their world.

Subject of Research: Context-aware selection of intrusion detection systems for edge-based Industrial Internet of Things networks using multi-dimensional evaluation

Article Title: Multi-dimensional framework for context-aware IDS selection in edge-IIoT

Article References: Batchu, R. K., Padarthi, S. K., Nandanapu, R., Pandey, N., Malineni, R., & Shieh, C.-S. (2026). Multi-dimensional framework for context-aware IDS selection in edge-IIoT. Discover Artificial Intelligence, 6(1), Article 1473. https://doi.org/10.1007/s44163-026-02433-3

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02433-3

Keywords: intrusion detection systems, Industrial Internet of Things, edge computing, deep learning, federated learning, autoencoder, attention mechanism, total cost of ownership, Pareto optimization, cybersecurity, Edge-IIoTset dataset, multi-criteria decision making

The framework, called DMCBO, short for Diffusion-driven Multi-objective Collaborative Optimization for Building Clusters, was developed by Yalan Zheng of Zhengzhou Technical College and Taixiang Yin of North China University of Water Resources and Electric Power. Their central insight is that the joint scheduling of equipment settings and load distribution across an entire cluster of buildings can be treated as a high-dimensional data sample, much like an image. Diffusion models, the same family of generative techniques behind modern text-to-image systems, work by progressively corrupting data with Gaussian noise and then learning to reverse that corruption. The researchers reformulated multi-objective building optimization as exactly this kind of conditional denoising process: starting from random noise, the model strips away noise step by step until a complete, Pareto-optimal collaborative schedule for every piece of equipment in every building emerges.

Technically, the forward diffusion process is a fixed Markov chain that adds tiny amounts of Gaussian noise to real scheduling data over one hundred steps, with the noise coefficient increasing linearly from 0.0001 to 0.02 until the original data becomes indistinguishable from standard Gaussian noise. A neural network is then trained to predict the noise that was added at each step, given the noisy schedule and a conditional encoding vector describing the current situation. Once trained, the model can generate fresh solutions at inference time by sampling pure noise and running the reverse process, refining it over one hundred denoising steps into a schedule tailored to the prevailing weather, occupancy, and grid conditions. Because this generation is a gradient-based learning process rather than an iterative search, it avoids the local optima traps that plague traditional multi-objective evolutionary algorithms and the reward oscillations that destabilize deep reinforcement learning.

What makes the system genuinely cluster-aware is its use of graph neural networks. The researchers model the building cluster as a graph mapped onto a radial distribution microgrid topology, where nodes represent buildings and edges encode physical constraints such as maximum line thermal capacities and transformer limits. A two-layer graph convolutional network passes messages between neighboring buildings, fusing spatial-topological dependencies and network constraints into a single graph embedding. This embedding is concatenated with an adjustable preference weight vector to form the condition that steers the denoising trajectory. In effect, when the model generates a control strategy for one building, it simultaneously perceives the state of its neighbors and the limits of the shared electrical infrastructure, suppressing systemic coordination errors from 0.18 down to 0.06 in the team’s experiments.

The multi-objective aspect is handled through an elegant scalarization trick. Each objective, such as energy consumption, operating cost, or thermal comfort, is normalized to a dimensionless value between zero and one using Min-Max normalization based on historical operational bounds. The normalized objectives are then blended into a single scalar condition via a weight vector whose components sum to one. During training, these weights are sampled from a uniform Dirichlet distribution, exposing the model to the entire space of possible preferences so it learns the distribution of the full Pareto front. At deployment, a decision maker can slide the weight vector like an adjustment knob, instantly generating schedules that emphasize energy saving, comfort, or a balanced compromise, all within the same trained model.

The denoising backbone itself is a U-Net adapted to process one- and two-dimensional building scheduling tensors. Diffusion time steps are embedded through sinusoidal positional encodings, while the graph-derived condition is injected into every residual block via Adaptive Layer Normalization. A bottleneck layer with multi-head self-attention captures long-range dependencies across the entire daily scheduling horizon, and symmetric skip connections preserve high-resolution information during reconstruction. The authors trained the model for 300 epochs on dual NVIDIA Tesla V100 GPUs, and the loss curves dropped from 0.75 to below 0.03 with training and validation curves closely coinciding, indicating no overfitting. As the reverse diffusion step count decreased from 90 to 10 during analysis, the noise-prediction error distribution became sharply concentrated at zero, demonstrating extremely high denoising precision at the end of generation.

Validation was carried out on the open-access Building Data Genome Project 2 dataset, which supplied real hourly weather records, operational schedules, and annual electricity, cooling, and heating load profiles across diverse climate zones. Because that dataset provides metered energy data rather than physical building models, the team used United States Department of Energy Commercial Reference Buildings as templates and calibrated them in EnergyPlus, a high-fidelity thermodynamic simulation platform, to create a digital twin consistent with the real data. All results were benchmarked against a rule-based control strategy with static thermostat setpoints of 25 degrees Celsius in summer and 22 degrees in winter, and every experiment was repeated ten independent times to confirm statistical significance.

The headline numbers are impressive. In high-demand summer conditions, DMCBO achieved an average energy-saving rate of 28.6 percent and cut peak load by 38.5 percent, shaving the afternoon peak from roughly 560 kilowatts down to about 440 kilowatts and saving approximately 142.80 US dollars per day. Cumulative daily power savings exceeded 1,000 kilowatt-hours, compared with 400 kilowatt-hours for a traditional collaborative strategy and just 120 kilowatt-hours without coordination. In winter, peak shaving and valley filling reached about 39 and 42 percent respectively, far above the 22 and 25 percent of conventional methods. When the solution sets were projected into objective space, DMCBO dominated the baselines: it reached a hypervolume indicator of 0.95 within 12,000 function evaluations, versus 0.85 for NSGA-II and 0.70 for MOPSO, while keeping operating costs between 140 and 160 dollars against more than 210 dollars for deep reinforcement learning and 260 to 300 dollars for a GAN-based generative approach.

Robustness and scalability proved equally compelling. Under injected stochastic disturbances mimicking sudden weather changes and occupancy swings, the framework confined power fluctuations to a narrow band of plus or minus 5 kilowatts, whereas the deep reinforcement learning baseline swung between minus 15 and plus 20 kilowatts. Scaling from 5 to 50 buildings, the global energy-saving rate actually rose from 28.5 to 31.2 percent, because larger clusters unlock more load-complementary opportunities that the graph encoder can safely exploit within encoded grid boundaries. Computation time grew to 9.8 seconds at the 50-building scale, but that remains comfortably within the 15-minute control horizon used by real building management systems, and the hypervolume index slipped only slightly to 0.88. Ablation experiments confirmed that both the graph neural network module and the choice of 100 diffusion steps, which yields a 50-millisecond inference latency, were essential to this balance of speed, accuracy, and stability.

The authors are candid about limitations. The framework has so far been validated only in simulation, and EnergyPlus simplifies sub-minute mechanical actuation latencies that real HVAC hardware would introduce; they propose cascading PID loops and predictive safety margins to absorb those delays in practice. Data quality, generalization to residential buildings and extreme climates, the computational cost of scaling to city-sized networks, and the black-box nature of the generative process all remain open challenges, with explainable AI and transfer learning flagged as future work. Still, the architecture is designed to slot into standard building management systems as a supervisory layer, transmitting setpoints over protocols such as BACnet or Modbus. If the simulation gains survive contact with physical buildings, the idea of generating optimal energy schedules from noise, guided by a graph of the grid itself, could become a powerful new tool in the push toward decarbonized cities.

Subject of Research: Conditional diffusion models with graph neural networks for multi-objective energy optimization in building clusters

Article Title: Diffusion model-driven multi-objective collaborative optimization for building energy management using graph neural networks

Article References: Zheng, Y., & Yin, T. (2026). Diffusion model-driven multi-objective collaborative optimization for building energy management using graph neural networks. Discover Artificial Intelligence, 6(1), Article 1383. https://doi.org/10.1007/s44163-026-02140-z

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02140-z

Keywords: diffusion models, graph neural networks, building energy management, multi-objective optimization, EnergyPlus, smart buildings, Pareto optimization, deep learning, energy efficiency, HVAC control, microgrids, generative AI

The core challenge the researchers set out to solve is uncertainty. In an idealized computer model, bearing stiffness and damping are tidy numbers pulled from a manufacturer’s catalog. In the real world, they are anything but. Manufacturing tolerances, assembly errors, thermal expansion mismatches between the steel shaft and its housing, and speed-dependent centrifugal forces all conspire to shift these parameters by roughly 5 to 12 percent from their nominal values. Traditional deterministic designs may look excellent on paper but can fluctuate wildly or fail outright when reality deviates from the assumed values. Worse, most prior studies have relied on theoretical catalog data rather than measurements, ignoring what engineers call epistemic uncertainty, the gap between the model and the physical machine.

Wu and Ma’s first move was to measure rather than assume. They built a high-speed motorized spindle test rig spanning an operational range from 1,000 to 24,000 rpm, discretized into 50 speed points. At each point, the spindle ran under steady-state thermal conditions while high-precision eddy current sensors sampled vibration at 10 kilohertz. Repeating the identification procedure 20 times per speed produced a dataset of 1,000 identified parameter vectors. To extract bearing stiffness and damping coefficients from these measured frequency responses, the team employed the Multi-Innovation Stochastic Gradient, or MISG, algorithm, a recursive estimation technique that accelerates convergence by stacking innovation information from several consecutive time steps rather than using a single sample at a time.

The numerical verification of this algorithm is striking in its own right. In synthetic tests where the true parameter values were known, MISG reduced the number of convergence iterations from 606 to just 75, an 8.1-fold speedup and an 87.6 percent reduction compared with the standard stochastic gradient method, while also lowering the final normalized root-mean-square error by 15 percent. Although each iteration costs about 18 percent more computation, the faster convergence cuts total computation time by 28 percent. A sensitivity study further confirmed that algorithmic identification error contributes less than 3 percent of the total output variance, meaning the scatter in the identified parameters reflects genuine physical variability, not artifacts of the estimation procedure.

With the empirical distributions in hand, the team validated them statistically before feeding them into the optimizer. The identified stiffness parameters showed coefficients of variation between roughly 6.0 and 6.4 percent, and Shapiro-Wilk tests confirmed that a multivariate normal distribution was a physically permissible model, with quantile-quantile plots yielding coefficients of determination above 0.99. Bearing damping was not treated as an independent random variable; instead, it was coupled proportionally to the identified stiffness through a constant calibrated by modal hammer testing, keeping the probabilistic model lean while preserving fidelity. The identified stiffness values deviated only about 1.5 to 3.0 percent from the numerical baseline, with the small residual attributed to installation-specific effects such as thermal expansion mismatch and bearing-seat tolerances, exactly the factors catalog values ignore.

The optimization phase then framed spindle design as a bi-objective problem: minimize both the mean peak vibration amplitude at the tool tip and its standard deviation, subject to a probabilistic constraint that the probability of vibration exceeding a critical threshold of 195 micrometers stays below a target level. That threshold was calibrated from surface roughness requirements in precision milling, where larger displacements degrade finish beyond aerospace-grade limits. Rather than relying on first-order or second-order reliability methods, which linearize the problem and can incur errors of 20 to 30 percent in nonlinear bearing-rotor dynamics, the researchers propagated uncertainty through direct Monte Carlo simulation with 1,000 samples per candidate design. The Non-dominated Sorting Genetic Algorithm II then evolved populations of designs, with each run producing a Pareto front of optimal trade-offs.

Crucially, the team ran the optimization three times, for target failure probabilities of 1, 5, and 10 percent, generating a family of Pareto fronts that reveals the marginal cost of safety. The results expose sharp diminishing returns. Tightening the constraint from 10 to 5 percent delivered a high marginal benefit ratio of 0.55 micrometers of variance reduction per percentage point of reliability, but squeezing from 5 to 1 percent dropped that ratio to 0.39 and demanded a 7.0 percent increase in bearing preload, pushing it to its physical limit of 3,000 newtons, for only a marginal robustness gain. Such aggressive preload accelerates rolling-element wear and risks friction-induced thermal problems. Weighing these trade-offs against ISO 20816-3 vibration standards, the researchers identified the 5 percent configuration as the optimal engineering equilibrium.

The optimized design itself is physically revealing. The framework selected a front-bearing preload of about 2,804 newtons and shortened the bearing support span from roughly 750 to 537 millimeters, a geometric change that raises the first natural frequency and pushes the third critical speed from about 22,000 rpm to 25,800 rpm, safely beyond the 24,000 rpm maximum operating speed. Benchmarking against conventional approaches underscored the value of the joint formulation: a pure robust design optimization achieved low variability but left the failure probability at an unacceptable 21.07 percent, while a pure reliability-based design hit a low failure probability but suffered an inflated response standard deviation of 11.80 micrometers. The integrated framework achieved failure probability of 0.35 percent while keeping variability at 9.69 micrometers and mean amplitude at 16.44 micrometers.

Physical validation sealed the case. Twenty independent replication runs at the rated speed of 24,000 rpm showed the optimized spindle reducing mean peak vibration by 16.84 percent, from 19.77 to 16.44 micrometers, and the response standard deviation by 29.78 percent, from 13.80 to 9.69 micrometers, with all improvements statistically significant in paired-sample t-tests. Across the full 1,000 to 24,000 rpm spectrum, the optimized spindle maintained stable, constrained responses, eliminating the severe resonant amplification that plagued the baseline design near its third critical speed. The variance reduction of 50.70 percent in experimental terms also surpasses the roughly 22 percent typically reported in simulation-based robust optimization studies.

Beyond the immediate numbers, the study signals a broader shift in how rotating machinery can be engineered. By grounding probabilistic models in measured reality rather than catalog assumptions, and by mapping the full trade-off surface between performance, robustness, and reliability instead of fixing a single safety target, the framework gives engineers a quantitative language for deciding how much safety margin a given application truly needs. The authors note that the approach currently assumes linear bearing behavior within the Hertzian elastic regime and depends on sufficient experimental data, pointing toward amplitude-dependent identification schemes and non-probabilistic interval methods as future extensions for heavily loaded or multi-shaft systems. For now, the message is clear: the path to quieter, more dependable high-speed spindles runs directly through the test rig, with every ounce of uncertainty measured, quantified, and designed against.

Subject of Research: Reliability-based robust optimization of high-speed motorized spindle-bearing systems using experimentally identified bearing parameter uncertainties

Article Title: Experimental identification and robust optimization of spindle–bearing systems with reliability constraints

Article References: Wu, H., & Ma, J. (2026). Experimental identification and robust optimization of spindle–bearing systems with reliability constraints. Mechanical Sciences, 17(2), 839-855. https://doi.org/10.5194/ms-17-839-2026

Image Credits: AI Generated

DOI: 10.5194/ms-17-839-2026

Keywords: motorized spindle, bearing stiffness, vibration control, robust optimization, reliability-based design, uncertainty quantification, Monte Carlo simulation, NSGA-II, multi-innovation stochastic gradient, CNC machining, Pareto optimization, rotor dynamics

A team at the University of Jinan in China, working with a collaborator at the Second Qilu Hospital of Shandong University, now reports a systematic way out of that impasse. In a study published in the Journal of Materials Science, Qiyan Zhang, Shuhui Han, Junna Zhou, Dongfang Gao, and Yang Qiao describe a machine learning framework that treats alloy design as an inverse problem: instead of asking what properties a given composition produces, the algorithm starts from the desired properties and works backward to the recipe. The result is a recommended Zn–Mg–Li alloy, processed by hot extrusion, with predicted values of 325.9 megapascals yield strength, 442.8 megapascals ultimate tensile strength, and 14.8 percent elongation, a combination that would place it among the strongest and most deformable biodegradable zinc alloys on record.

The starting point for the study was a curated experimental database of 97 original records, each documenting an alloy composition, its processing route, and the resulting mechanical properties. Small datasets like this are the bane of machine learning in materials science, where algorithms trained on sparse, noisy data often generalize poorly. The researchers countered the problem with physics-informed data augmentation, a technique that expands the training set by generating synthetic data points consistent with known metallurgical behavior rather than arbitrary statistical noise. This approach, which has gained traction in studies of high-entropy alloys and other data-poor domains, allowed the team to build predictive models that map composition and processing conditions onto three target quantities: ultimate tensile strength, yield strength, and elongation.

The accuracy figures are striking for a dataset of this size. The optimized models achieved coefficients of determination, or R-squared values, of 0.924 for ultimate tensile strength, 0.771 for yield strength, and 0.930 for elongation. In practical terms, an R-squared above 0.9 means the model explains more than nine-tenths of the variance in the measured property, a level of fidelity that makes virtual screening credible. The somewhat lower score for yield strength reflects the well-known sensitivity of that property to subtle microstructural features that are difficult to capture in tabular data, but the team judged the performance sufficient to drive the next stage of the design pipeline.

What distinguishes this work from many machine learning studies in metallurgy is its insistence on explainability. Rather than treating the model as a black box, the researchers applied SHAP analysis, a technique borrowed from the machine learning community that attributes each prediction to the contributions of individual input variables. The analysis revealed something genuinely useful: within the composition and processing domain covered by the dataset, the strengthening contribution of magnesium appeared to plateau at roughly 1.5 weight percent. Below that threshold, adding magnesium buys substantial strength, likely through solid-solution hardening and the formation of fine secondary phases; above it, the model saw diminishing returns. That kind of quantitative, data-driven saturation point is exactly the information experimentalists need to avoid wasting effort on compositions that cannot pay off.

Lithium told a complementary story. The SHAP analysis associated lithium additions with the retention of ductility through synergistic alloying effects with magnesium, a finding consistent with prior experimental work showing that lithium can modify the deformation behavior of zinc-rich alloys. The models also flagged hot extrusion as beneficial, and the authors connect this to well-established metallurgy: the severe deformation and elevated temperature of extrusion promote dynamic recrystallization, a process in which new, strain-free grains nucleate during deformation, refining the microstructure and simultaneously improving strength and formability. In other words, the algorithm did not just find correlations; it pointed toward mechanisms that metallurgists can independently verify and exploit.

Armed with these mechanistic insights, the team turned to large-scale virtual screening. More than 100,000 candidate alloys were generated in silico and evaluated with the trained models under multi-objective Pareto optimization, a mathematical framework that searches for compositions that cannot be improved in one property without sacrificing another. Crucially, the optimization was constrained by biomechanical requirements, the strength and ductility thresholds that an implant must actually meet in service, rather than by abstract measures of model confidence. The Pareto front that emerged represented the best achievable compromises between strength and elongation, and the recommended composition of Zn-0.60Mg-0.54Li, processed by hot extrusion, was drawn from that frontier.

The predicted property set of the recommended alloy is worth unpacking. A yield strength of 325.9 megapascals and an ultimate tensile strength of 442.8 megapascals comfortably exceed the mechanical demands of many cardiovascular applications, where pure zinc’s roughly 100 to 200 megapascals falls short. At the same time, 14.8 percent elongation is far above what most high-strength zinc alloys achieve, since conventional strengthening strategies such as heavy secondary-phase precipitation typically embrittle the material. The study’s framing within the broader literature is notable: recent work has pushed biodegradable zinc toward 380 and even 400 megapascal grades, including a 2025 Nature paper reporting stronger, coarser-grained zinc alloys, and the new machine learning route offers a faster, cheaper way to explore that design space than trial-and-error casting and testing.

The implications extend beyond the specific Zn–Mg–Li system. All three alloying elements are metabolically familiar, magnesium is an essential nutrient and lithium has a long clinical history, which eases biocompatibility concerns that plague alloys containing rarer or potentially toxic additions. The workflow itself, a curated database augmented with physics-informed synthetic data, gradient-boosted predictive models, SHAP-based mechanistic interpretation, and Pareto-constrained inverse design, is a template that could be applied to magnesium alloys, iron-based biodegradable systems, or structural materials entirely outside medicine. The authors have made their 97 original records available in supplementary data, with models and source code available on reasonable request, a transparency measure that should accelerate adoption.

Caveats remain, as they always do at this stage. The predicted properties await experimental confirmation, and corrosion behavior, the property that makes zinc attractive in the first place, was not the optimization target here, so the recommended alloy must still prove that its degradation rate and its byproducts are acceptable in biological environments. The SHAP-derived magnesium threshold is explicitly valid only within the domain of the dataset used to train the models, not as a universal law of zinc metallurgy. Still, the study marks a convincing demonstration that inverse design, guided by explainable artificial intelligence, can navigate the century-old strength-ductility trade-off in a medically critical material, and it suggests that the next generation of implants that vanish harmlessly inside the body may be discovered first by algorithms, then confirmed in the laboratory.

Subject of Research: Machine learning-assisted inverse design of biodegradable Zn–Mg–Li alloys for medical implants

Article Title: Data-driven inverse design of biodegradable Zn–Mg–Li alloys using explainable machine learning and multi-objective optimization

Article References: Zhang, Q., Han, S., Zhou, J., Gao, D., & Qiao, Y. (2026). Data-driven inverse design of biodegradable Zn–Mg–Li alloys using explainable machine learning and multi-objective optimization. Journal of Materials Science. https://doi.org/10.1007/s10853-026-13880-9

Image Credits: AI Generated

DOI: 10.1007/s10853-026-13880-9

Keywords: biodegradable zinc alloys, Zn–Mg–Li, machine learning, SHAP analysis, multi-objective optimization, hot extrusion, dynamic recrystallization, biomedical implants, strength–ductility trade-off, materials informatics, Pareto optimization, yield strength

The core problem Fu set out to solve is one that has long frustrated educational researchers. Teaching quality is not a single number. It is a tangle of partially conflicting goals: keeping students engaged, raising academic achievement, maximizing instructional effectiveness and using institutional resources efficiently. Improving one of these dimensions can easily come at the expense of another. Traditional evaluation pipelines, which typically collapse everything into one weighted score, hide these trade-offs. Machine learning models described in earlier literature, including deep neural networks, convolutional classifiers and genetic-algorithm-based optimizers, generally optimized a single objective or relied on static indicator sets, leaving the multi-dimensional nature of the problem unaddressed.

The proposed system, called MODE–DCN, attacks the problem in two complementary stages. The first stage is a Deep Convolutional Network, a class of neural network best known for image recognition but here applied to tabular educational data. The network consists of two one-dimensional convolutional layers with 64 and 128 filters and a kernel size of three, followed by max-pooling and two fully connected layers ending in a Softmax output. The convolutional filters slide across the reduced feature sequence, learning nonlinear interactions between educational indicators that simpler models such as support vector machines, random forests or gradient-boosted tree ensembles cannot capture, because those models process features independently at each split rather than jointly.

Before any learning begins, the raw data passes through a carefully layered preprocessing pipeline. The study used a publicly available Teaching Quality Evaluation Dataset from Kaggle containing 1,000 evaluation instances, each describing a specific teacher, course and semester, with ten input attributes and a four-class target variable labeled Excellent, Good, Average or Poor. Records were checked for missing values and duplicates, categorical variables were encoded numerically, and the features were normalized with Min–Max scaling and standardized with Z-scores. Principal component analysis then compressed the feature space from eleven dimensions down to eight, and recursive feature elimination pruned the set further to the six most informative attributes. The authors stress that these two techniques complement rather than duplicate each other: PCA removes linear redundancy based on variance, while RFE performs supervised selection of task-relevant features, leaving the convolutional network free to learn nonlinear structure.

The second stage is where the framework earns its name. A Multi-Objective Decision Evolution algorithm, inspired by Differential Evolution, searches for the best configuration of the network’s hyperparameters, including the number of filters, kernel size, dropout rate and learning rate, together with a mask selecting which features to use. The optimizer works with a population of candidate solutions, each encoded as a decision vector. In each generation it applies mutation, in which a scaled difference between two randomly chosen individuals is added to a third, and crossover, which mixes parental and mutant vectors into trial solutions. Crucially, the algorithm does not rank candidates by a single scalar score. Instead it uses fast non-dominated sorting to arrange solutions into Pareto fronts and crowding distance to preserve diversity, so that the search converges toward a set of optimal compromises rather than one narrow answer.

Four objective functions guide that search. The first minimizes classification error on validation data, serving as a proxy for academic performance. The second minimizes an engagement-weighted cross-entropy loss, in which each student’s prediction error is scaled by a weight derived from their normalized engagement score, so the model is penalized more heavily for failing students who are highly engaged. The third measures the standard deviation of validation accuracy across folds, capturing teaching effectiveness as prediction stability. The fourth balances model complexity, measured by the number of trainable parameters, against training time, reflecting operational efficiency. The optimizer explored filters between 16 and 128, kernel sizes from 2 to 5, dropout rates from 0.1 to 0.5 and learning rates from 0.0001 to 0.01, ultimately settling on a population size of 30, a mutation factor of 0.5 and a crossover rate of 0.9.

The results are striking. On a held-out test set of 200 instances, the MODE–DCN model achieved 98.5 percent accuracy, 96.3 percent precision, 95.8 percent recall and a 96.6 percent F1-score, while the mean squared error fell to 14.67, compared with 56.41 for a gradient descent method and 22.63 for a backpropagation neural network with a random matrix model. The confusion matrix showed near-perfect separation across the four quality classes, with only a handful of errors, such as four Excellent cases misclassified as Good. Receiver operating characteristic analysis produced area-under-curve values above 0.99 for every class, and five-fold cross-validation with a paired t-test confirmed that the improvements over the strongest baseline were statistically significant at p below 0.001.

The framework also proved its mettle against re-implemented baselines under identical conditions, including the same data split, preprocessing pipeline and evaluation metrics. Models such as LSTM combined with CNN, KNN and a multi-source data CNN reached accuracies between 89.4 and 94.2 percent, all well short of the hybrid model. An ablation study reinforced the point: removing the MODE optimization stage caused the largest performance degradation, demonstrating that the evolutionary search over hyperparameters and features, rather than the network architecture alone, drives much of the gain. When the model was re-implemented on three external datasets, accuracy dropped to a range of 90.8 to 93.1 percent, a candid acknowledgment by the author that generalization across different educational domains remains a challenge requiring domain-specific calibration.

Interpretability, often the Achilles heel of deep learning in high-stakes settings, receives explicit attention. The study employs SHAP, or SHapley Additive exPlanations, as a post-training analysis tool to quantify how much each feature contributes to individual predictions. The beeswarm analysis identified teacher evaluation scores, student averages and technology integration scores as the three most influential drivers of the model’s judgments, giving administrators a transparent rationale rather than an opaque verdict. The Pareto-optimal solutions are further aggregated into a Teaching Quality Index, computed as a weighted sum of the objective functions with equal weights by default, and presented through an institutional decision dashboard that ranks instructor performance and surfaces engagement metrics. Because this index is applied after optimization, institutions can adjust the weights to reflect their own policy priorities without invalidating the underlying Pareto set.

The implications reach beyond vocational colleges. As education systems worldwide digitize their records, the combination of deep feature learning and Pareto-based optimization offers a template for any assessment problem where multiple legitimate goals pull in different directions, from curriculum design to faculty development. Fu notes that practical deployment will demand adequate hardware, staff training and ongoing model maintenance, and that future work should incorporate live classroom data from multiple institutions to validate and refine the model over time. For now, the study stands as a persuasive demonstration that the messy, contested question of what makes teaching good can be rendered measurable, optimizable and, perhaps most importantly, explainable.

Subject of Research: Deep learning and multi-objective optimization for assessing vocational education teaching quality

Article Title: Vocational education teaching quality assessment based on deep learning and multi-objective optimization decision algorithms

Article References: Fu, K. (2026). Vocational education teaching quality assessment based on deep learning and multi-objective optimization decision algorithms. Discover Artificial Intelligence, 6(1), Article 1342. https://doi.org/10.1007/s44163-026-02336-3

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02336-3

Keywords: deep learning, multi-objective optimization, vocational education, teaching quality assessment, convolutional neural networks, Pareto optimization, feature selection, PCA, recursive feature elimination, SHAP interpretability, educational data mining, Differential Evolution

The researchers describe a modular Safety-by-Design framework built around two complementary analytical branches. The first branch is dedicated to candidate prioritization: given a large space of possible RNA-lipid nanoparticle designs, it identifies which formulations best satisfy competing objectives of efficacy and safety. The second branch addresses a subtler question, namely whether the biological picture produced by the framework remains consistent with what is known about human immune organization. By separating these two tasks, the authors argue, the framework avoids the common pitfall of blending mechanistic interpretation, formulation ranking, and biological validation into a single opaque pipeline.

At the heart of the prioritization branch sits a deliberately constructed synthetic design space of 500 RNA-LNP formulations. Each formulation is evaluated across a heterogeneous cohort of virtual patients, computational stand-ins whose simulated immune responses capture variability among real people. From these simulations the framework derives five synthetic endpoints related to efficacy and safety, which are then condensed into a composite Safety-by-Design score. The score is not a single magic number imposed from above; rather, it is a transparent aggregation that can be interrogated, reweighted, and stress-tested, which is precisely what the authors do in a series of sensitivity analyses.

To make the exploration of 100,000 possible patient-formulation pairs tractable, the team trained a Random Forest surrogate model to reproduce the composite score. The model performed well within the synthetic design space: in formulation-grouped five-fold cross-validation, it achieved a pooled out-of-fold coefficient of determination of 0.887, with a root mean square error of 0.055 and a mean absolute error of 0.044. Grouping the cross-validation by formulation is an important methodological choice, because it tests whether the model generalizes to formulations it has not seen, rather than merely memorizing patient-level variation within a familiar formulation.

With the surrogate in place, the researchers turned to Pareto-based multi-objective analysis, a technique borrowed from optimization theory in which a candidate is considered non-dominated if no alternative is better on every objective simultaneously. The primary analysis identified 21 such non-dominated formulations, representing the efficient frontier of the efficacy-safety trade-off. Among these, one candidate, designated LNP_0462, ranked first globally. Crucially, it also ranked first when the analysis was repeated separately within strata of low, intermediate, and high inflammatory risk, suggesting that its leading position is not an artifact of averaging across patient subgroups with very different immune profiles.

Robustness was examined from several angles. The authors tested 3,876 alternative weight combinations for the Safety-by-Design score and found that the ranking was generally stable across them, indicating that the conclusions do not hinge on one particular way of trading efficacy against safety. They also performed a complete balanced analysis of all 100,000 patient-formulation pairs, rather than relying on sampled subsets, and the resulting ranking closely matched the primary analysis, with a Spearman correlation of 0.998 and an overlap of 18 of the top 20 candidates. For a framework intended to guide experimental prioritization, this kind of internal consistency matters as much as raw predictive accuracy.

The second branch of the framework tackles biological plausibility in a way that is deliberately kept apart from the ranking machinery. The authors drew on trajectories generated by the Universal Immune System Simulator, a mechanistic model published in previous work, and condensed descriptors from those trajectories into five Mechanistically Informed Immune Signatures, or MIIS. These signatures were used to constrain synthetic multi-omics profiles, which were then compared at the level of immune programs with an independent human single-cell RNA-sequencing dataset, GSE171964. Importantly, the MIIS variables and the synthetic multi-omics profiles were not used as predictors, endpoint inputs, or Pareto objectives, so the consistency assessment cannot inflate the ranking itself.

The comparison was quantified with a metric the authors call BPCI, computed across seven post-vaccination days. The mean BPCI was 0.760, and it exceeded an exact shuffled-label null with a p-value of 0.025. The authors are careful, however, not to oversell this result. Day-specific evidence was heterogeneous, and donor-level uncertainty was substantial, so the finding supports the general idea that the framework’s synthetic outputs organize immune programs in a manner consistent with real human data, rather than validating any individual formulation’s clinical behavior. The MIIS branch, as the authors put it, provides an assessment of immune-program organization, not a validation of specific candidates.

This restraint is characteristic of the paper’s overall posture. The work is explicitly labeled a proof of concept, and the authors state that experimental testing and evaluation in additional human datasets are required before any conclusions can be drawn about biological or clinical performance. The study involved no new human participants or animals; it relied on computational simulations, synthetic data, and secondary analysis of a publicly available, de-identified transcriptomic dataset, so no additional ethics approval was required. The research received no specific grant funding, and the authors declare no competing interests. The article was received on 28 July 2026, accepted on 27 August 2026, and published open access on 7 September 2026.

Even as a proof of concept, the framework points toward a shift in how RNA therapeutics might be developed. Instead of treating formulation design, safety assessment, and biological plausibility as separate silos handled by different teams at different stages, the Catania group’s approach integrates them into one transparent, modular pipeline in which every assumption can be varied and every ranking can be stress-tested. The surrogate model makes exhaustive exploration affordable, the Pareto analysis makes trade-offs explicit rather than hidden in a weighted average, and the MIIS branch provides an independent sanity check against real human immunology. If subsequent experimental work confirms that computationally prioritized candidates such as LNP_0462 translate into safe and effective formulations, frameworks of this kind could become a standard first filter in the design of RNA-LNP therapeutics, narrowing the laboratory search space and, ultimately, shortening the path from molecular design to clinical candidate.

Subject of Research: A computational Safety-by-Design framework for prioritizing RNA-lipid nanoparticle formulations and assessing biological consistency

Article Title: A modular Safety-by-Design framework for RNA-LNP prioritization and mechanistically informed biological consistency assessment

Article References: Di Salvatore, V., Russo, G., & Pappalardo, F. (2026). A modular Safety-by-Design framework for RNA-LNP prioritization and mechanistically informed biological consistency assessment. BMC Bioinformatics. https://doi.org/10.1186/s12859-026-06630-w

Image Credits: AI Generated

DOI: 10.1186/s12859-026-06630-w

Keywords: RNA therapeutics, lipid nanoparticles, Safety-by-Design, virtual patients, Random Forest, Pareto optimization, mechanistic modeling, systems immunology, synthetic multi-omics, single-cell RNA sequencing, explainable artificial intelligence, drug safety