Researchers Yibo Peng and Caixia Deng, affiliated with Ningbo University and the Ningbo Institute of Materials Technology and Engineering of the Chinese Academy of Sciences, set out to understand why a so-called senary high-entropy alloy containing platinum, ruthenium, nickel, cobalt, iron, and molybdenum shows tolerance to CO during the alkaline hydrogen oxidation reaction. This reaction is the anode-side half of an anion-exchange membrane fuel cell, a technology prized for its potential to use cheaper, non-precious components than conventional proton-exchange devices. The catch is that hydrogen oxidation proceeds sluggishly in alkaline conditions, and platinum-based anodes are acutely vulnerable to poisoning by even trace amounts of carbon monoxide that latch onto active sites and refuse to let go.

High-entropy alloys have emerged as a compelling answer. Unlike traditional alloys with one dominant metal and minor additives, these materials mix five or more elements in roughly equal proportions, producing a crystalline surface where the identity of every atom’s neighbors is essentially random. That randomness means the surface is not a single, uniform catalyst but a vast landscape of distinct local environments, each potentially binding hydrogen or CO differently. The trouble for theorists is obvious: there are astronomically many such environments, and calculating the adsorption energy of a molecule on each one with conventional density functional theory would be computationally prohibitive.

Peng and Deng tackled this challenge with what they call a compositional ensemble framework, a pipeline that fuses geometric machine-learning descriptors, intelligent sampling, high-throughput quantum calculations, and a state-of-the-art graph neural network. First, they used smooth overlap of atomic positions, or SOAP, descriptors to encode each surface site’s local chemical neighborhood in a form a machine can compare. Then, farthest point sampling allowed them to select a diverse, representative subset of configurations from that enormous space, ensuring the training data spanned the full variety of local environments rather than clustering around a few common motifs.

On that curated training set, the team ran high-throughput density functional theory calculations to obtain accurate adsorption energies, and used the results to train EquiformerV2, an equivariant transformer architecture designed to respect the rotational and translational symmetries of three-dimensional atomic systems. The payoff was striking: the trained model predicts CO adsorption energies with a mean absolute error of just 0.090 electron volts. With that level of accuracy in hand, the researchers could do something previously impractical, namely statistically evaluate CO adsorption across 120,000 distinct surface sites of the PtRuNiCoFeMo alloy, building a distribution rather than a handful of anecdotal data points.

The results reveal a subtle and somewhat counterintuitive picture. Compared with the flat platinum (111) surface, the canonical benchmark in this field, the high-entropy alloy does not uniformly weaken CO adsorption across its surface. Instead, the distribution of CO binding energies becomes continuously broadened, stretching from sites that bind CO far more weakly than platinum to sites that grip it even more tightly. The overall CO adsorption distribution remains dominated by intermediate-to-strong binding configurations, particularly at bridge and hollow geometries where the molecule can bond to multiple surface atoms simultaneously. In other words, the average alloy surface is, by and large, still a welcoming host for CO.

Yet buried within that distribution lies the alloy’s real advantage. When the researchers performed a comparative two-dimensional analysis of hydrogen and CO adsorption energies, screening each site against the platinum-referenced criterion of accessible hydrogen binding combined with relatively weakened CO binding, they identified 9,479 sites that met both conditions. That corresponds to roughly 7.9 percent of the examined ensemble. These sites, the authors emphasize, should be interpreted as a minority but statistically resolvable reservoir of local motifs where hydrogen chemistry and CO poisoning are effectively decoupled, rather than as evidence that the entire alloy surface outperforms platinum in CO tolerance. It is a reservoir effect: the catalyst as a whole retains enough clean, hydrogen-friendly real estate to keep working even as other regions succumb to adsorbed CO.

Perhaps the most actionable finding concerns what those favorable sites look like at the atomic scale. The team found that the H/CO-favorable motifs are mainly associated with platinum atoms sitting at top-site positions whose nearest-neighbor shells are enriched in molybdenum and platinum. This suggests that nearby molybdenum-platinum coordination is a prominent local environment for balancing hydrogen accessibility against reduced CO affinity. The result dovetails with decades of experimental observations that molybdenum-containing platinum catalysts, from early PtMo alloys to modern MoOx-Pt composites, exhibit exceptional CO tolerance, often attributed to molybdenum’s oxophilicity and its electronic modifying influence on adjacent platinum atoms. The new work reframes that intuition in statistical terms, pinpointing the specific coordination motif worth engineering.

The methodological significance of the study may prove as durable as its catalytic insights. By demonstrating that SOAP descriptors, farthest point sampling, DFT training data, and an equivariant transformer can be chained into a reliable surrogate model for adsorption energies on chemically disordered surfaces, the authors offer the electrocatalysis community a blueprint for interrogating other high-entropy systems, from oxygen reduction catalysts to CO2 conversion electrodes. The statistical framing itself is a corrective: rather than asking whether a high-entropy alloy binds a poison more weakly than a pure metal on average, designers should ask how large the subpopulation of protective local motifs is, and whether synthetic strategies can enlarge it.

For the fuel cell industry, the implications are tantalizing though still computational. Anion-exchange membrane fuel cells need anodes that combine fast alkaline hydrogen oxidation kinetics with robustness against fuel impurities, and a surface in which nearly eight percent of sites are naturally H/CO-decoupled represents a meaningful margin of tolerance. The study also suggests a concrete design lever: tuning synthesis and annealing conditions to promote molybdenum-enriched neighborhoods around surface platinum atoms could, in principle, expand the favorable reservoir further. As hydrogen energy infrastructure scales up globally, turning statistical portraits of disorder like this one into practical catalyst recipes may become one of the field’s central pursuits, bridging the gap between atomic-scale randomness and real-world device durability.

Subject of Research: Statistical machine-learning analysis of CO tolerance in a PtRuNiCoFeMo high-entropy alloy catalyst for alkaline hydrogen oxidation in fuel cells

Article Title: Statistical origin of CO tolerance during alkaline hydrogen oxidation on a PtRuNiCoFeMo high-entropy alloy

Article References: Peng, Y., & Deng, C. (2026). Statistical origin of CO tolerance during alkaline hydrogen oxidation on a PtRuNiCoFeMo high-entropy alloy. Ionics. https://doi.org/10.1007/s11581-026-07511-1

Image Credits: AI Generated

DOI: 10.1007/s11581-026-07511-1

Keywords: high-entropy alloy, CO tolerance, hydrogen oxidation reaction, anion-exchange membrane fuel cell, PtRuNiCoFeMo, adsorption energy, density functional theory, machine learning, graph neural network, EquiformerV2, electrocatalysis, molybdenum-platinum coordination

The core challenge the team tackled is one that haunts deep learning across domains: neural networks are insatiably hungry for large, diverse, well-labeled datasets. In software engineering, that hunger is hard to satisfy. Annotated code smell datasets are scarce, expensive to produce, and heavily skewed—some smell types dominate the labeled examples while others appear only in tiny numbers. Class imbalance of this kind pushes a neural network toward the majority categories, leaving it nearly blind to the minority ones, which are often the most damaging defects. Data augmentation, the technique of programmatically expanding a training set without collecting new real-world samples, has transformed fields like image recognition and natural language processing, but translating it to source code has proven far trickier, because code must remain syntactically valid and semantically faithful after any transformation.

The researchers’ solution leans on a practice every programmer already knows: refactoring, the discipline-preserving rearrangement of code that changes its internal structure without altering its external behavior. Because a refactor is guaranteed to keep a program’s functionality intact while reshaping its form, it provides a principled way to generate legitimate new training variants of the same code. The team compared two augmentation engines: a rule-based approach that applies classical refactoring operations directly, and a large language model-based approach that uses generative models to produce refactored variants of source code. Crucially, they paired these with a loss penalty, an algorithmic balancing scheme for neural network training that penalizes errors on underrepresented classes more heavily, aiming to counteract the dataset’s skew from the optimization side as well as the data side.

To represent code in a form a neural network can digest, the team turned to pre-trained transformer models from the BERT family: the original BERT, which treats code as plain text; CodeBERT, which was pre-trained jointly on programming and natural languages; and GraphCodeBERT, which additionally incorporates the data-flow structure of programs during pre-training. Each model converts snippets of source code into dense numerical embeddings—vector fingerprints of a program’s semantics—which then serve as inputs to the downstream classification network. This embedding stage matters because two pieces of code that differ only in formatting or naming should map to nearby points in the embedding space, and the more structurally aware the embedding model, the more stable that mapping should be under refactoring.

The experimental results delivered a clear verdict. Code refactoring-based augmentation powered by GraphCodeBERT embeddings and large language models measurably enhanced classifier performance, and its benefits were most pronounced precisely where deep learning models are weakest: the imbalanced, underrepresented smell classes. When the team analyzed how the volume of augmented data influenced outcomes, they discovered a pattern of diminishing returns with an important twist. The greatest improvement arrived with the first increments of augmentation applied to minority classes; beyond that point, adding more synthetic examples yielded progressively smaller gains. The impact of augmentation volume also varied depending on the size of the class being boosted, a finding the researchers argue should inform how practitioners budget their augmentation efforts rather than blanket-duplicating synthetic data across all categories.

Equally instructive were the negative results. The loss penalty mechanism, which was expected to amplify the benefit of augmentation by rebalancing training dynamics, had a negligible effect on overall performance—and in some configurations proved actively detrimental. Augmentation based on vanilla BERT and CodeBERT embeddings likewise failed to move the needle or even hurt accuracy in certain settings. The implication is that not all code-aware embeddings are created equal: models that ignore program structure, such as plain BERT, appear unable to fully exploit refactored variants, while GraphCodeBERT’s data-flow-informed pre-training gives it the structural sensitivity needed to recognize that a refactored snippet and its original are two faces of the same entity.

Beyond the headline results, the study makes a methodological contribution of its own. Evaluating augmentation strategies requires repeatedly testing how performance responds to different volumes of synthetic data across multiple classes, a combinatorially expensive process. To tame this, the team developed a novel weighting strategy that optimizes the evaluation of augmentation volume distribution, cutting the computational overhead of the analysis while preserving its analytical precision. This kind of efficiency matters in practice, because organizations auditing large industrial codebases cannot afford exhaustive, compute-hungry tuning cycles for every detection model they deploy. The weighting scheme, the authors suggest, could become a reusable tool for anyone designing augmentation pipelines for code intelligence systems.

The broader significance of the work lies in what it says about the mechanics of learning from code. The findings reinforce a growing consensus that source code should not be treated as ordinary text. Programs carry rich graph-structured semantics—call graphs, data dependencies, inheritance hierarchies—and augmentation strategies that respect and exploit those structures deliver gains that text-level tricks cannot match. The results align with a wave of recent research on code-specific augmentation, from clone-aware contrastive learning to mixup-based transformations, but they are among the first to systematically quantify the relationship between augmentation volume, class size, and model architecture in the code smell domain, and to do so while comparing rule-based and LLM-based generation head to head.

For software engineering teams, the practical takeaway is immediately actionable. Code smell detection models trained with refactoring-based augmentation and structure-aware embeddings should be more reliable at flagging underrepresented defect patterns—precisely the smells that human reviewers most often miss. Because refactoring transformations are behavior-preserving by construction, the synthetic data they generate avoids the hallucination risks that come with purely generative augmentation, offering a safer middle ground between hand-crafted rules and free-form LLM generation. The researchers have also made their work reproducible, releasing all related source code and data through public repositories on GitHub and Zenodo, lowering the barrier for other teams to adopt and extend the technique.

Looking ahead, the authors suggest that code refactoring-based augmentation could drive the development of more efficient data augmentation strategies not only for code smell classification but for deep learning applications across software engineering more broadly—vulnerability detection, code search, clone detection, and automated repair among them. The lesson generalizes: when real-world labeled data is scarce, the discipline of software engineering itself, with its guarantees of behavioral equivalence, can become the engine that manufactures trustworthy training data. In a field racing to replace human code reviewers with machines, teaching models to see the same code through many refactored lenses may prove one of the most cost-effective ways to give them the experience they need—without a single additional line of manually annotated code.

Subject of Research: Data augmentation via code refactoring to improve deep learning-based code smell classification.

Article Title: Enhancing code smell classification with code refactoring-based data augmentation

Article References: Nizam, A., Aydin, M., Islamoglu, E., Sahmoud, S., & Ozaydin, S. B. (2026). Enhancing code smell classification with code refactoring-based data augmentation. Knowledge and Information Systems, 68(1), Article 254. https://doi.org/10.1007/s10115-026-02859-2

Image Credits: AI Generated

DOI: 10.1007/s10115-026-02859-2

Keywords: code smell, data augmentation, code refactoring, code embedding, GraphCodeBERT, CodeBERT, BERT, large language models, class imbalance, deep learning, software quality, Enhancing

The core problem the researchers identify is a dual bottleneck. Modern large language models have grown from the roughly 12 to 60 million parameters of ResNet in 2016 to hundreds of billions or even trillions of parameters in systems like Llama-4, released in 2025. That hundredfold expansion in less than a decade means inference can demand tens to hundreds of gigabytes of memory, far beyond the 4 to 32 gigabytes typical of consumer devices. Compression techniques such as quantization and pruning help, but they trade away model capability. At the same time, communication networks strain under the load: split-inference schemes that ship high-dimensional activation features from devices to servers can generate tens to hundreds of megabytes per inference step, while multi-agent systems that synchronize reasoning traces amplify overhead further. Congestion, jitter, and wireless instability compound the challenge.

AI Flow responds with three interlocking pillars. The first is a device-edge-cloud architecture that treats the network itself as a computational hierarchy. End devices handle lightweight tasks and real-time interaction; edge servers at base stations and roadside units provide nearby, low-latency processing; and cloud clusters supply the scalable horsepower for training and compute-intensive inference. By orchestrating workloads across these tiers, the framework balances resource scalability against latency, offloading latency-critical inference to the edge while reserving the cloud for heavy operations.

Within that hierarchy, the team introduces two collaboration techniques designed to cut communication costs. The first, task-oriented feature compression, targets vision-language model inference. Rather than transmitting raw images, the device merges visual features produced by a CLIP-style encoder using density peaks clustering based on K nearest neighbors, then encodes the merged features with a hyperprior-based entropy model whose parameters are modeled on a Laplacian distribution. A router network selects the best entropy model for each feature. In experiments on the LLaVA-OneVision-7B model using an NVIDIA Jetson AGX Orin device and an RTX 4090 edge server, the method reduced transmitted data by 25 to 45 percent compared with WebP and 35 to 60 percent compared with JPEG at equal accuracy on the RealWorldQA benchmark, and cut inference latency to roughly a third of server-only inference on the MME benchmark.

The second technique, hierarchical collaborative decoding, accelerates large language model generation through speculative decoding spread across network tiers. A lightweight model on the device drafts tokens locally, while a larger edge model validates and corrects them using a soft acceptance strategy, shifting the big model’s role from full generation to error correction. The researchers extend this into a parallel pipeline in which the device keeps generating without blocking while the edge server refines tokens at intervals. On the MATH-500 benchmark, a two-tier configuration pairing a 1.5-billion-parameter device model with a 7-billion-parameter edge model achieved about 40 tokens per second, a 1.25-fold speedup over edge-only decoding at the same accuracy, with a three-tier setup adding a 14-billion-parameter cloud model for further gains.

The second pillar of AI Flow is the concept of familial models: families of different-sized models whose hidden features are aligned so that intermediate results from a small model can be directly reused by a larger one without any middleware. Two enabling techniques make this possible. Early exit allows inference to terminate at intermediate layers while preserving acceptable accuracy, with lightweight branch modules refining features before prediction. Weight decomposition splits the linear layers of transformer blocks into pairs of low-rank matrices whose combined parameter count is smaller than the original, with the hidden dimension tuned to hit nearly any target size. Initialization via singular value decomposition on whitened data keeps distortion low, and the team shows that compression loss is quantitatively determined by the squared singular values of discarded components, enabling per-layer compression decisions.

The researchers demonstrate two implementation strategies. Hierarchical principal component decomposition trains a series of low-rank components that progressively fit the residuals of earlier ones, producing TeleChat-based models from 2.38 billion to 6.30 billion parameters that, despite limited training tokens, perform comparably to established models such as LLaMA2-7B and ChatGLM2-6B on benchmarks including MMLU, CMMLU, C-Eval, GSM8K, MATH, and BBH. The second strategy, early exiting with scalable branches, inserts decomposed transformer blocks between exit points and a shared language model head. Applied to LLaVA-1.5-7B, it retained 98.2 percent of the backbone’s average performance on six visual question answering benchmarks using only 3.17 billion parameters, while a baseline without the branch design needed at least 4.63 billion parameters to reach 90 percent.

The third pillar is perhaps the most provocative: connectivity- and interaction-based intelligence emergence. Here, the network becomes a medium through which heterogeneous models, including large language models, vision-language models, and diffusion models, collaborate to achieve capabilities exceeding any single model. A device-server collaboration scheme lets specialized on-device models generate preliminary responses in parallel, which a central server model aggregates into a unified answer that is then returned to devices for revision. Evaluations on MT-Bench, AlpacaEval 2.0, and Arena-Hard showed consistent gains, with weaker models benefiting most, and performance on Arena-Hard rose nearly linearly with the number of participating agents, suggesting practical scalability.

Diffusion models receive their own collaboration paradigms. A serial scheme for multi-person motion generation chains an interleaved interaction synthesis module with a relative coordination refinement module, achieving state-of-the-art results on the InterHuman benchmark, including a 25.3 percent improvement in Top-1 R-Precision and a 50.6 percent reduction in Fréchet inception distance compared with prior methods. A parallel scheme for monocular depth estimation splits processing into near-field and far-field decoder branches fused around a sliding anchor, topping benchmarks on both indoor NYU-V2 and outdoor KITTI data. A networked scheme, OmniVDiff, unifies RGB, depth, segmentation, and edge modalities within a single video diffusion transformer, outperforming baselines on depth-conditioned video generation.

The authors ground the framework in application scenarios that include embodied AI, where drones and ground robots share aligned intermediate features to avoid redundant computation; wearable devices, where smart glasses offload heavy recognition tasks to edge and cloud tiers while keeping latency-sensitive processing local; and smart cities, where the low-altitude economy of delivery drones and aerial mobility systems demands ultra-low-latency coordination across thousands of heterogeneous devices. Future directions include federated learning adapted to large models, distributed edge inference resilient to device churn, and adaptive network orchestration for volatile wireless conditions. The team also articulates guiding principles, including a Law of Information Capacity that defines efficiency as the ratio of text compression gain to inference cost, and a Law of Multi-model Collaboration showing that ensembles of diverse models follow power-law scaling with a better loss floor than single-series collaboration. Together, the researchers argue, these ideas chart a path toward AI that is not confined to data centers but flows through the networks that already surround us.

Subject of Research: A multidisciplinary framework integrating AI and communication technologies for ubiquitous, low-latency intelligence across device-edge-cloud networks

Article Title: AI Flow: perspectives, scenarios, and approaches

Article References: An, H., Hu, W., Huang, S., Huang, S., Li, R., Liang, Y., Shao, J., Song, Y., Wang, Z., Yuan, C., Zhang, C., Zhang, H., Zhuang, W., & Li, X. (2026). AI Flow: perspectives, scenarios, and approaches. Vicinagearth, 3(1), Article 1. https://doi.org/10.1007/s44336-025-00031-y

Image Credits: AI Generated

DOI: 10.1007/s44336-025-00031-y

Keywords: AI Flow, edge AI, device-edge-cloud collaboration, familial models, large language models, speculative decoding, intelligence emergence, task-oriented feature compression, vision-language models, diffusion models, ubiquitous intelligence, 6G networks

Researchers led by Agung Bella Putra Utama, Aji Prasetya Wibawa, and Anik Nur Handayani of Universitas Negeri Malang, together with Andrew Nafalski of Adelaide University, have introduced a deterministic duel-based imputation framework that puts forecasting performance itself in charge of deciding how missing values should be filled. The central insight is deceptively simple: most imputation methods are judged by how statistically similar their reconstructions are to the original data, not by whether they actually help a forecasting model make better predictions. A filled-in series can look numerically plausible on paper and still wreck a downstream forecast by smoothing away the very peaks and rhythms the model needs to learn.

The framework works like a structured tournament. For every missing observation, the system generates a set of candidate values using familiar baseline techniques, including mean, median, and mode substitution, K-Nearest Neighbors, Multiple Imputation by Chained Equations, and Last-Observation-Carried-Forward. Each candidate is then scored with a composite loss function that combines three complementary forecasting metrics: Mean Absolute Percentage Error, which captures proportional error; Root Mean Square Error, which penalizes large absolute deviations; and the coefficient of determination, or R-squared, which measures how much of the variance in the real series the reconstruction explains. The weights are set at 0.4, 0.4, and 0.2 respectively, so error-based criteria dominate while explanatory power still plays a supporting role.

What separates this approach from conventional candidate ranking is the way winners are chosen. Instead of collapsing every candidate into a single aggregated score and picking the top one, the framework stages deterministic pairwise duels. Candidates face each other one-on-one, and a sensitivity threshold of 0.001 prevents negligible score differences from flipping outcomes. After all comparisons, candidates are ranked by cumulative wins, and the top three advance to a final aggregation stage. Four aggregation modes are available: averaging the survivors for low-variance stability, taking the median for outlier robustness, taking the maximum to emphasize peak responsiveness, or a winner-take-all selection of the candidate with the lowest forecasting loss. Because every step follows fixed rules, the same input always produces the same output, a property the authors argue is essential for auditing, validation, and accountability in operational systems.

The team evaluated the framework on the Beijing PM2.5 dataset, a widely used benchmark of hourly meteorological and air-pollution measurements collected between 2010 and 2014. After cleaning and temporal alignment, 41,776 valid records remained out of 43,800, with roughly 2,068 missing values concentrated in the PM2.5 variable itself. Crucially, the missingness is not random: gaps cluster in cold seasons and during periods of rapid atmospheric change, exactly the conditions where forecasting matters most and where naive imputation does the most damage. The dataset includes dew point, temperature, pressure, cumulative wind speed, snowfall, rainfall, and temporal indicators as explanatory variables, with PM2.5 concentration serving as the prediction target.

The imputed datasets were then fed into four recurrent forecasting architectures trained under identical conditions: a vanilla Recurrent Neural Network, a Long Short-Term Memory network, a Bidirectional LSTM, and a Gated Recurrent Unit. Hyperparameters for all models were tuned with Particle Swarm Optimization, though the researchers are careful to note that this stochastic tuning happens entirely outside the imputation stage and does not compromise its determinism. The optimized configurations converged on surprisingly lightweight designs: three hidden layers of 24 neurons each, sigmoid activations, the Adam optimizer with mean squared error loss, a batch size of 32, 46 training epochs, and a dropout rate of 0.2.

The results are striking. Compared with conventional imputation methods, the duel-based framework reduced MAPE by up to 70 percent and RMSE by 12 to 15 percent, while maintaining R-squared values above 0.95 across all four architectures. The Mean aggregation variant delivered the strongest overall performance, preserving both proportional variation and the amplitude structure of the signal. Dropping missing observations entirely, by contrast, produced the worst forecasts, confirming that simply discarding incomplete rows destroys the temporal coherence recurrent models rely on. Even BRITS, a sophisticated deep-learning imputation method, showed less consistent gains, suggesting that reconstruction quality alone does not guarantee forecasting quality.

Statistical testing reinforced the case. Paired t-tests and Wilcoxon signed-rank tests comparing the proposed Mean variant against KNN, the strongest conventional baseline, produced p-values below 0.01 across all architectures, with each experiment repeated ten times under fixed random seeds. Visual analysis of predicted versus observed PM2.5 trajectories showed close alignment across smooth and volatile segments alike, with the Bi-LSTM achieving the highest R-squared values and the GRU delivering the most efficient runtime while matching LSTM accuracy. The framework’s computational overhead scales quadratically with the number of candidates but linearly with data size, and because the candidate set is small and fixed, runtimes remain predictable even as data volume grows.

To test generalizability, the researchers extended the evaluation beyond air quality to three additional domains: the Heart Disease dataset, which has weak temporal structure; the KEDS e-journal dataset, which shows irregular behavioral patterns and moderate sparsity; and the Sunspot dataset, which exhibits strong periodic behavior. The framework achieved lower MAPE and RMSE than baselines including drop-missing, mean imputation, KNN, and BRITS across all of them, with the largest gains appearing in datasets with strong sequential patterns such as Sunspot. The authors acknowledge limitations, including the quadratic cost of pairwise comparison for large candidate pools and the fixed forecasting horizons tested, and they point to cluster-based candidate reduction and online learning as future directions. But the broader message stands: imputation should not be a passive preprocessing chore. Treated as a forecasting-driven decision process, it becomes a strategic lever for building AI systems that are not only accurate but consistent, transparent, and worthy of trust.

Subject of Research: A deterministic duel-based imputation framework that integrates forecasting performance metrics into missing-data selection for reliable AI-driven time-series forecasting.

Article Title: A deterministic duel-based imputation framework for reliable AI-driven time-series forecasting

Article References: Utama, A. B. P., Wibawa, A. P., Handayani, A. N., & Nafalski, A. (2026). A deterministic duel-based imputation framework for reliable AI-driven time-series forecasting. Discover Informatics, 1(1), Article 3. https://doi.org/10.1007/s44564-026-00001-6

Image Credits: AI Generated

DOI: 10.1007/s44564-026-00001-6

Keywords: time-series imputation, missing data, AI forecasting, deterministic imputation, deep learning, LSTM, GRU, RNN, PM2.5, forecasting reliability, machine learning, data preprocessing

The urgency underlying the analysis is stark. Climate change is intensifying biological disturbances, accelerating the frequency and geographic reach of pest outbreaks and plant invasions. Warmer winters raise insect survival rates, hotter springs speed up life cycles, and extreme events such as heatwaves, droughts, and unseasonal storms create sudden ecological shocks that traditional linear models struggle to capture. Meanwhile, conventional surveillance, built on labor-intensive manual inspection and subjective visual assessment, simply cannot keep pace with modern epidemics. The authors point out that the fusion of AI with IPM enables earlier detection, higher diagnostic precision, and targeted interventions, marking a paradigm shift from experience-based decision-making toward high-dimensional, data-driven inference across individual plants, whole fields, and entire continents.

At the technical core of the review is a detailed account of how machine perception has matured. Early disease recognition relied on handcrafted features, in which researchers manually defined the shapes and textures of leaf lesions and matched them against predefined libraries. These approaches proved fragile under shifting illumination and cluttered field backgrounds. The advent of convolutional neural networks changed the equation, allowing models to learn hierarchical features directly from pathological imagery, from low-level edge structures to high-level lesion textures. Equally important, the sensory frontier has expanded beyond the visible spectrum. Multispectral and hyperspectral imaging capture reflectance signatures that reveal internal biochemical changes before physical symptoms appear, while thermal and chlorophyll fluorescence imaging quantify physiological stress through indicators such as reduced fluorescence intensity.

Temporally, the review highlights how sequential models process continuous biological signals, from sap flow to transpiration rates, distinguishing natural metabolic variation from the early signatures of infestation. For long-horizon challenges, such as tracking the spread of invasive species across decades and forecasting climate-driven shifts, Transformer-based architectures use attention mechanisms to process entire time series at once, identifying critical historical climate anomalies that traditional models miss. Yet no single sensor is sufficient. Modern systems increasingly depend on heterogeneous data fusion, aligning observable physical traits from standard cameras with internal biochemical changes from spectral sensors, and bridging the spatial resolution gap between coarse satellite coverage and ground-level verification through air-ground synergy. These fusions underpin high-resolution risk networks that track biological invasions at global scale.

Data scarcity remains one of the field’s thorniest problems, particularly for newly emerged pests and rare invaders. The review documents a family of adaptive learning strategies designed to bridge this gap. Transfer learning exploits the insight that the basic logic of recognizing shapes and textures is universal, pre-training models on large general datasets and fine-tuning them on small numbers of agricultural samples. Meta-learning goes further, teaching systems a general rule for adapting to novel species traits rather than memorizing specific symptoms. When data are extremely limited, generative models such as Generative Adversarial Networks synthesize realistic training images of rare diseases or weather events, improving recognition of threats rarely encountered in the wild and enabling deployment in new scenarios at minimal data cost.

Deployment in the field imposes its own constraints, since IoT sensors and autonomous drones operate under strict power and hardware limits while advanced deep learning models typically demand server-grade computing. The review describes how model compression techniques, including pruning and quantization, evaluate millions of connection weights, discard redundant neural pathways, and reduce numerical precision to shrink memory footprints without sacrificing diagnostic performance. Natively lightweight architectures factorize computationally expensive operations into simplified sequential stages, and one-stage detection frameworks such as improved YOLO variants analyze images in a single read, simultaneously pinpointing a threat’s location and identifying its species. These efficient designs are essential for the real-time, closed-loop control loops at the heart of intelligent agriculture.

The review then maps intelligent diagnosis across three spatial scales. At the individual plant level, hyperspectral imaging can detect viral latency in ultra-early, pre-symptomatic phases, while deep learning excels at post-symptomatic identification once lesions, spots, or eggs become visible, and LiDAR captures three-dimensional structural data for larger plants and forest canopies. At field scale, drones and IoT sensors track canopy traits and microclimate conditions in real time; a far-view and close-look strategy scans whole fields for anomalies before descending for high-resolution imaging of specific targets. Acoustic sensors identify the vibration signatures of wood-boring insects, and electronic noses decode volatile organic compounds emitted by stressed plants, although atmospheric interference remains a hurdle. At regional and global scales, satellite remote sensing with red-edge spectral bands and high revisit frequencies supports continuous surveillance, with phenological correction algorithms distinguishing normal seasonal defoliation from disease-induced canopy change.

Crucially, the review emphasizes that detection is only the beginning; anticipation and action complete the loop. Short-term early-warning frameworks use wireless sensor networks and deep neural networks to model the precise microclimate windows that trigger spore germination and egg hatching, replacing rigid calendar-based spraying with data-driven intervention. Dynamic economic decision-support systems now weigh fluctuating market prices, expected yields, and crop growth stages, and even incorporate the migration patterns of beneficial insects to avoid disrupting natural pest control. For seasonal planning, recurrent neural networks such as long short-term memory estimate pest development timelines, while graph convolutional networks map how shifting winds and geography drive the migration routes of highly mobile threats like rice planthoppers, allowing growers to anticipate high-risk years and adjust crop varieties accordingly.

On the action side, AI is transforming machinery into intelligent execution systems. Modern agricultural UAVs integrate high-resolution sensing with precision spraying, executing optimized flight trajectories above the canopy without soil compaction or crop damage. Variable-rate systems adjust fluid delivery in near real time using pump modulation, matching dose to canopy volume and density, while AI-driven prescription maps ensure chemicals are applied only when and where needed. Flight parameters, nozzle technology, and platform design all interact: lower altitudes enhance droplet penetration, multi-rotor downdrafts push droplets into dense crops, and air-induction flat-fan nozzles balance deposition efficiency with drift control. Ground-based robots complement aerial platforms, using multi-modal sensor fusion for sub-centimeter target localization, vision-guided arms for pruning and canopy management, and soft robotic grippers with force feedback for damage-free harvesting. Multi-agent coordination increasingly links air and ground platforms into fully automated management chains.

Invasive species emerge as a special frontier. Unlike endemic pests managed through economic thresholds that balance crop loss against control costs, invasive populations grow exponentially without natural predators, triggering systemic ecological losses and demanding early eradication. Detection confronts a dual dilemma of sparse, scattered targets and severe annotation scarcity, met with tailored vegetation indices, centimeter-level drone hyperspectral imagery, LiDAR and radar fusion to penetrate canopy occlusion, and phenology-timed observations when invaders’ spectral signatures are most distinct. Data-efficient learning, including lightweight networks and active learning, reduces labeling burdens. Climate anomalies and global trade corridors drive non-linear, jump-dispersal spread that models now trace using diffusion equations and graph theory, while biomass estimation, still underexplored, is becoming a prerequisite for precise eradication budgets and sustainable biological control alternatives.

The review closes with an honest appraisal of the field’s limits. Predictive modeling still lags well behind identification in both volume and maturity, because forecasting is a non-stationary time-series problem vulnerable to error accumulation, whereas visual identification is a static, closed-set task. Complex models trade interpretability for accuracy, correlational approaches lack mechanistic causality, and representation biases in training data risk generalization failures across crops, pathogens, and climates. Field deployment is further constrained by sensor fragility, environmental heterogeneity, high costs, and unstable rural networks, challenges the authors propose to address through cloud-edge-device coordination, open-source cross-climatic benchmark datasets, standardized knowledge graphs, and stringent regulatory frameworks. Their ultimate vision is a globally interconnected, AI-enabled agroecological immune network: a closed-loop system in which machine intelligence does not merely interpret signals but actively directs smart machinery to safeguard productivity and ecological resilience as climate change and global trade intensify the biological threats facing entire biomes.

Subject of Research: Application of artificial intelligence, remote sensing, and robotics to integrated pest management and invasive plant control in agriculture

Article Title: From detection to action: artificial intelligence in integrated pest and invasive plant management

Article References: Li, Y., Zha, L., Liu, W., Luo, F., & Xu, C. (2026). From detection to action: artificial intelligence in integrated pest and invasive plant management. Advanced Biotechnology, 4(3), Article 25. https://doi.org/10.1007/s44307-026-00118-7

Image Credits: AI Generated

DOI: 10.1007/s44307-026-00118-7

Keywords: artificial intelligence, integrated pest management, invasive species, precision agriculture, remote sensing, hyperspectral imaging, drones, agricultural robotics, deep learning, plant disease detection, early warning systems, climate change

The research was led by Mohammad Matin Rouhani of Amirkabir University of Technology in Tehran, together with Mahdi Hasanipanah of Duy Tan University, Xin Yin of Wuhan University of Science and Technology, Hesam Dehghani of Hamedan University of Technology, and Mohammad Rezaei of the University of Kurdistan. Their goal was ambitious: to build a predictive system that could generalize reliably to new blasting events at a major copper mine, while giving engineers a transparent view of the physical factors that matter most. Traditional empirical equations for flyrock, the team notes, often fall short because they cannot capture the complex, nonlinear interplay between explosive energy, rock properties, and blast geometry.

To train their models, the researchers assembled a dataset of 252 blasting events recorded at the Sungun Copper Mine in northwestern Iran, one of the country’s largest open-pit operations. Each event was characterized by a set of controllable blast design parameters, including the powder factor, which describes the amount of explosive energy delivered per unit of rock; the burden, the distance between boreholes and the free face of the rock; the spacing between holes; the stemming length, the inert material packed atop the explosive charge; the bench height; and the diameter of the drill holes. The output variable was the measured flyrock distance, the maximum throw of rock fragments beyond the blast zone.

At the heart of the framework sits XGBoost, or extreme gradient boosting, a machine learning algorithm that builds a strong predictive model from an ensemble of decision trees, each new tree correcting the errors of its predecessors. XGBoost has become a workhorse of applied machine learning because of its speed and accuracy, but its performance depends heavily on the tuning of internal settings known as hyperparameters, such as tree depth, learning rate, and the number of boosting rounds. Poorly tuned hyperparameters can leave substantial predictive power on the table, or worse, cause a model to memorize training data rather than learn generalizable patterns.

The team’s key innovation was to pair XGBoost with five different optimization algorithms that automatically search for the best hyperparameter configuration. These included Bayesian optimization, a statistically guided search method that models the relationship between hyperparameters and performance; the geometric mean optimization algorithm, a newer population-based technique; the osprey optimization algorithm, inspired by the hunting behavior of the fish-eating raptor; reptile search optimization, a nature-inspired metaheuristic mimicking reptilian hunting strategies; and the Archimedes optimization algorithm, which draws on principles of buoyancy and physics. Each hybrid model was trained and tested on the Sungun dataset, and their performances were compared using multiple statistical indicators, radar plots, and Taylor diagrams, which visualize how closely each model’s predictions match observed values in terms of correlation and variability.

The results were striking. The GMO-XGBoost model, which combines XGBoost with geometric mean optimization, delivered the best generalization performance, achieving the highest prediction accuracy on unseen testing data. This matters because a model that excels only on data it has already seen is of little practical use; engineers need forecasts they can trust for future blasts. Interestingly, the AOA-XGBoost variant, built on the Archimedes optimization algorithm, showed superior performance during the training phase, illustrating a common tension in machine learning between fitting known data and generalizing to new situations. The fact that different optimizers excelled at different stages underscores why the team ran the full comparison rather than assuming a single best approach.

But the researchers went a step further than raw prediction. To open the black box of their best-performing models, they applied Shapley additive explanations, or SHAP analysis, a technique borrowed from cooperative game theory that quantifies each input variable’s contribution to every individual prediction. The analysis revealed a clear hierarchy of influence: the powder factor, burden, and stemming length emerged as the most powerful controls on flyrock behavior, with powder factor showing the strongest effect. This finding aligns with physical intuition, since the powder factor directly governs the explosive energy available to launch rock fragments, while burden and stemming determine how that energy is contained and directed.

The practical implications for mine operators are significant. With an interpretable model in hand, blast engineers can run what-if scenarios before a single hole is drilled, adjusting the powder factor or stemming design to keep predicted flyrock within safe exclusion zones. That translates into fewer evacuations, less equipment damage, reduced liability, and a smaller environmental footprint around the mine. The framework also supports sustainable mine planning more broadly, because blasting that throws rock unpredictably can disturb surrounding habitats, contaminate nearby land with debris, and erode community trust. By making flyrock a quantified, decision-oriented variable, the study moves hazard assessment from reactive investigation to proactive design.

The work also reflects a broader shift in the geosciences toward explainable artificial intelligence. For years, machine learning models in mining and rock engineering have been criticized for offering high accuracy without insight, leaving practitioners unable to justify safety-critical decisions. By embedding SHAP analysis directly into the modeling pipeline, the Sungun study demonstrates that accuracy and interpretability need not be competing goals. The authors describe this combination of optimized hybrid machine learning and explainable AI as the study’s central contribution, positioning it as a template for other geo-environmental hazards such as ground vibration, air blast, and back-break, where similar hybrid frameworks are already gaining traction.

Limitations remain, as they do in any data-driven study. The models were trained on data from a single mine, and while the GMO-XGBoost variant generalized well within that setting, transferring the framework to sites with different rock types, explosives, or drilling practices would require retraining and validation. The researchers indicate that the underlying data will be shared on reasonable request, which should help other groups test the approach. Still, the message of the study is clear: with 252 real blasts, five competing optimizers, and a transparent view of what drives the danger, flyrock prediction has entered a new era, one in which the rocks thrown by an explosion can be forecast before the fuse is ever lit.

Subject of Research: Intelligent prediction of blast-induced flyrock distance in surface mining using optimized hybrid XGBoost machine learning models and explainable AI

Article Title: Intelligent Assessment of Blast-Induced Flyrock as a Geo-Environmental Hazard Using Optimized Hybrid XGBoost Models

Article References: Rouhani, M. M., Hasanipanah, M., Yin, X., Dehghani, H., & Rezaei, M. (2026). Intelligent Assessment of Blast-Induced Flyrock as a Geo-Environmental Hazard Using Optimized Hybrid XGBoost Models. Natural Resources Research. https://doi.org/10.1007/s11053-026-10757-1

Image Credits: AI Generated

DOI: 10.1007/s11053-026-10757-1

Keywords: flyrock, surface mining, XGBoost, machine learning, blasting hazard, Sungun Copper Mine, hyperparameter optimization, explainable AI, SHAP analysis, geo-environmental hazard, sustainable blasting, powder factor

The new system, described in the journal Results in Engineering, is called SRGAN-ConvFormer, and its architecture is deliberately modular. Rather than asking a single neural network to do everything, the framework splits the work among three specialized components. A generative adversarial network known as SRGAN sharpens the spatial detail of the imagery. A hybrid convolution-transformer module called ConvFormer models the long-range dependencies that run along each pixel’s spectral signature. Finally, a dual-channel convolutional neural network, or DCCNN, fuses the two streams of information into a single classification decision. Each module addresses a distinct failure mode of earlier approaches, and the ablation experiments show that removing any one of them measurably degrades performance.

The spatial branch begins with a pragmatic preprocessing step. Because hyperspectral cubes contain hundreds of correlated bands, the researchers first apply principal component analysis to compress the data down to five components without discarding the bulk of its information content. Around each pixel, the team extracts a small patch, and it is here that SRGAN earns its place in the pipeline. Originally developed to super-resolve natural photographs, SRGAN uses a generator network built from convolutional layers, residual blocks with batch normalization, and a global skip connection that preserves low-level spatial features while upsampling the image. A discriminator network, trained in adversarial competition with the generator, pushes the reconstructed patches toward photorealistic texture. Applied patch-wise across the scene, this process yields spatially enhanced inputs in which fine structures such as field boundaries, road edges, and vegetation textures become legible to the downstream classifier.

In parallel, the spectral branch treats each pixel’s reflectance spectrum as a one-dimensional sequence, much as a language model treats a sentence. The ConvFormer module projects this sequence into a 256-dimensional embedding space using a one-dimensional convolutional patch-embedding layer, then passes the resulting spectral tokens through six encoder blocks. Each block combines depthwise convolutions, which excel at local pattern mixing, with multi-head self-attention using eight heads, which can relate any band to any other band regardless of distance along the spectrum. Learnable positional embeddings preserve the ordering of the spectral bands, feed-forward networks with GELU activations add nonlinear capacity, and residual connections with layer normalization keep training stable. The design borrows from ConvFormer architectures originally proposed for medical image segmentation, but the authors adapted it specifically to hyperspectral spectral sequences, a contribution they emphasize alongside the integration rather than a redesign of the underlying transformer.

The fusion stage is where the framework departs most clearly from single-branch designs. The DCCNN accepts the super-resolved spatial features from the SRGAN branch, processed through convolutional, max-pooling, and flattening layers, and concatenates them with the flattened spectral features from the ConvFormer branch. Dense layers with dropout regularization then map the fused vector to class probabilities through a softmax output. The authors argue that this division of labor matters because each module solves a different information-loss problem: super-resolution recovers spatial detail that coarse sensors destroy, the transformer captures spectral dependencies that convolutional networks tend to overlook, and the dual-channel fusion ensures neither modality dominates the final decision.

The experimental evaluation covers four benchmark scenes that span the practical diversity of hyperspectral remote sensing. Botswana, captured by NASA’s EO-1 satellite over the Okavango Delta, presents 14 land-cover classes across wetlands and woodlands with 11,275 labeled samples. Kennedy Space Center, acquired by the AVIRIS sensor over Florida wetlands, offers 13 classes with subtle spectral fluctuations among marsh and forest types. Pavia University, recorded by the ROSIS sensor over an Italian city at 1.3-meter resolution, contains 9 urban classes and more than 42,000 labeled pixels. Salinas, another AVIRIS acquisition over California farmland, packs 16 agricultural classes and 54,129 samples into a high-resolution scene. Training used roughly 10 percent of labeled pixels per class, with another 10 percent for validation, and results were averaged over five runs.

The numbers are remarkable. On Botswana, SRGAN-ConvFormer achieved an overall accuracy of 99.26 percent and a kappa coefficient of 99.19 percent, reaching perfect classification on ten of fourteen classes, including spectrally tangled categories that reduced competing methods to accuracies as low as 60 percent. On Kennedy Space Center, the framework posted 99.78 percent overall accuracy with seven classes classified perfectly. Pavia University yielded 99.86 percent, and Salinas 99.75 percent, with eleven of sixteen classes at 100 percent. By comparison, a 3D-CNN baseline ranged from about 91 to 93 percent across the datasets, standalone SRGAN and ConvFormer variants landed in the low-to-mid 90s, and established dual-branch methods such as DBMA and DBDA peaked around 97 to 98 percent. A diffusion-model-based competitor, DDPM, reached 94 to 97 percent but struggled on spectrally complex classes. Against the newest wave of Mamba state-space models and diffusion-based approaches, the proposed framework was competitive or superior on three of the four datasets, with only EnMambaHSI edging it out slightly on Botswana.

The ablation studies illuminate why the combination works. Feeding raw, unenhanced imagery into the same ConvFormer-DCCNN backbone consistently lowered accuracy across all four datasets, isolating the SRGAN module’s contribution to the performance gains. Standalone SRGAN, lacking spectral modeling, faltered on spectrally mixed classes; standalone ConvFormer, blind to spatial context, collapsed on spatially demanding scenes such as Kennedy Space Center, where it dropped to 85.91 percent. The dual-channel fusion alone reached the high 90s, and only the full three-module pipeline crossed the 99 percent threshold. Sensitivity analyses added practical guidance: moderate training sets of roughly 20 to 30 percent of labeled pixels proved optimal, batch sizes of 8 or 16 outperformed larger batches, and larger spatial patches, from 25 by 25 up to 41 by 41 pixels, consistently improved accuracy by widening the spatial context available to the classifier.

Computational costs remain reasonable for a research-grade system. The model carries roughly 16.7 to 17.6 million trainable parameters and occupies about 191 to 202 megabytes, with inference speeds between roughly 650 and 960 frames per second on an NVIDIA RTX A4000 GPU. Attention maps show the network concentrating on discriminative, spatially coherent regions rather than treating all locations equally, t-SNE visualizations reveal compact, well-separated class clusters in the learned feature space, and confusion matrices display strong diagonal dominance across all four scenes, indicating few residual confusions between spectrally similar classes.

The authors are candid about the limits of the current protocol. Because training and test pixels are sampled at the pixel level from spatially continuous scenes, neighboring patches may partially overlap, potentially inflating performance estimates relative to fully spatially disjoint partitions. Future work, they write, will adopt stricter train-test separation, cross-scene and cross-dataset evaluation with domain adaptation, explicit spectral variability modeling through atmospheric correction and spectral calibration, and privacy-preserving federated learning for distributed hyperspectral data. Even with those caveats, the message of the study is clear: by teaching a generative model to restore the spatial detail that hyperspectral sensors give up, and pairing it with a transformer that reads the full spectral story of every pixel, classification accuracy on some of remote sensing’s hardest benchmarks has been pushed to the edge of perfection. For applications from precision agriculture to mineral exploration and environmental monitoring, that margin could translate into maps that are not just detailed, but trustworthy.

Subject of Research: A dual-channel transformer framework combining SRGAN super-resolution and ConvFormer spectral modeling for hyperspectral image classification

Article Title: SRGAN-ConvFormer: A Dual-Channel transformer framework for hyperspectral image classification

Article References: Khan, M. M., Kumar, B., Saini, A., Khan, A. A., & Kryvinska, N. (2026). SRGAN-ConvFormer: A Dual-Channel transformer framework for hyperspectral image classification. Results in Engineering, 32, Article 112818. https://doi.org/10.1016/j.rineng.2026.112818

Image Credits: AI Generated

DOI: 10.1016/j.rineng.2026.112818

Keywords: hyperspectral imaging, SRGAN, transformer, image classification, super-resolution, remote sensing, deep learning, generative adversarial network, ConvFormer, spectral-spatial fusion, dual-channel CNN, land-cover mapping

The problem, according to the authors, stems from how current self-supervised learning methods are built. Most approaches fall into three broad families. Transformation-based methods ask models to predict the temporal order of shuffled frames or recognize motion patterns. Contrastive learning approaches maximize agreement between differently augmented views of the same video. Masked video modeling techniques, which currently lead the field, reconstruct heavily masked spatiotemporal tokens, with systems such as VideoMAE demonstrating that this strategy yields highly transferable representations. More recent methods like SMILE inject semantic supervision from pretrained vision-language models and motion-aware masking to sharpen temporal sensitivity.

Yet all of these methods share a common limitation: they learn correlations between observed frames without modeling the underlying mechanisms that generate temporal dynamics. Real-world video is produced by structured interactions between objects, agents, and environments, where actions lead to observable consequences over time. Correlation-driven objectives allow models to exploit shortcuts, such as static appearance cues or temporal redundancy, a phenomenon the authors link to well-documented shortcut learning in deep networks. The result is representations that may lack robustness under distribution shifts, fail to generalize to unseen dynamics, and struggle with tasks requiring reasoning about actions and their effects.

TRACE, short for Temporal Causal Representation Learning for Video Understanding, tackles this gap by borrowing an idea from causal inference: the intervention. Inspired by Judea Pearl’s do-operator, the framework approximates the effect of intervening on latent temporal factors by modifying motion dynamics in a learned latent space and observing how future representations change. Crucially, the authors emphasize that TRACE does not perform true causal discovery or identifiable causal inference. Instead, it offers a practical approximation of intervention-based learning that captures intervention-sensitive temporal dependencies without requiring explicit causal supervision.

Technically, the framework rests on three pillars. First, a transformer-based encoder maps video frames into latent tokens that are explicitly decomposed into two 384-dimensional components: a content branch that preserves stable scene semantics and a motion branch that captures temporal dynamics. Second, a temporal intervention module perturbs the motion component through three structured operations. Motion perturbation scales temporal changes and injects noise to simulate faster or slower action dynamics. Token-level intervention permutes latent tokens along trajectories to disrupt temporal correspondence. Structural intervention masks selected edges in a learned temporal dependency graph, simulating altered interactions between scene components. Third, a counterfactual prediction module is trained to forecast future latent representations conditioned on the intervened state, with a consistency loss aligning predictions with approximated counterfactual outcomes and a separation term preventing trivial identity mappings.

The authors are careful to distinguish these interventions from conventional data augmentation. Temporal shuffling, frame dropping, and playback speed variation treat perturbed samples as additional views of the same video, aiming for invariance. TRACE instead intervenes after decomposing the latent representation, acting exclusively on motion while an invariance constraint keeps content fixed. The intervened representation becomes the input to the prediction module, so the model must learn how changes in motion affect future evolution rather than simply ignoring perturbations. An additional intervention-aware contrastive objective uses hard negatives drawn from temporally inconsistent or intervention-mismatched trajectories, pushing the model to separate causally valid evolution from implausible alternatives.

The empirical results are striking. Under linear probing, TRACE outperformed the strongest baseline, SMILE, by 3.9 percent on Something-Something V2, a dataset demanding fine-grained temporal reasoning, and by 1.7 percent on EPIC-Kitchens, while also gaining 2.7 percent on appearance-dominated Kinetics-400 and 1.3 percent on UCF-101. Under full fine-tuning, it improved over SMILE by 2.2 percent on SSv2 and 1.5 percent on K400, reaching 74.3 and 84.6 percent Top-1 accuracy respectively. All comparisons were statistically significant in paired two-tailed t-tests over five independent runs, with p-values below 0.05, and standard deviations remained consistently low, indicating stability across random initializations.

Ablation studies confirmed that every component contributes, with the temporal intervention module proving most critical: removing it cost 3.8 percent on Kinetics-400 and 2.7 percent on SSv2. Cross-dataset transfer showed gains of 3.9 percent for Kinetics-400 to SSv2 and 2.6 percent in the reverse direction, and under controlled motion perturbations at test time TRACE’s accuracy drop was nearly halved, from minus 9.1 percent for SMILE to minus 4.8 percent. On a CLEVRER-style synthetic benchmark with known physical rules, TRACE reached 76.9 percent causal reasoning accuracy versus 71.4 for SMILE, and achieved the highest counterfactual prediction consistency at 0.81 cosine similarity. Notably, the intervention and counterfactual modules are used only during pretraining, so inference cost remains identical to standard transformer encoders, with total computational overhead of roughly 4 to 5 percent during training.

Qualitative visualizations reinforced the story. t-SNE and UMAP projections showed content representations forming compact clusters aligned with action categories, while motion representations grouped actions sharing similar dynamics regardless of semantics. Attention and motion sensitivity maps revealed that TRACE concentrates on hands, manipulated objects, and interaction points, highlighting the take-off phase of a basketball dunk, the release of a javelin, or the subtle hand-object interactions in egocentric kitchen videos, precisely the regions where altering motion would most change future outcomes.

The authors frame TRACE as a scalable pathway toward causally grounded video representation learning that integrates with modern architectures without manual annotation. They acknowledge open challenges, including principled identification of true causal factors in real-world video, extension to multimodal signals such as language and audio, incorporation of physical constraints and structured world models, and application to video question answering, planning, and embodied decision-making. If the approach generalizes, it could mark a meaningful step toward machines that do not merely watch the world unfold, but understand what would happen if it unfolded differently.

Subject of Research: Intervention-aware self-supervised temporal representation learning for causal video understanding

Article Title: TRACE: Intervention-aware temporal representation learning for video understanding

Article References: Chaudhry, H. N., Kulsoom, F., Mohsin, S. M., Aslam, S., & Ashraf, N. (2026). TRACE: Intervention-aware temporal representation learning for video understanding. Machine Learning with Applications, 25, Article 100976. https://doi.org/10.1016/j.mlwa.2026.100976

Image Credits: AI Generated

DOI: 10.1016/j.mlwa.2026.100976

Keywords: TRACE, video understanding, self-supervised learning, causal representation learning, temporal interventions, counterfactual learning, masked video modeling, action recognition, VideoMAE, machine learning, computer vision, structural causal models

The New Innovator Award occupies a distinctive niche in the NIH funding landscape. Rather than requiring preliminary data or a conventional track record, the program seeks out scientists whose ideas are considered transformative precisely because they depart from established approaches. Daniela Friedman, the Arnold School’s associate dean for research and leadership development, described the recognition as extremely well deserved, noting that since joining the school Shi has built an outstanding research program and distinguished herself as an innovative investigator whose work has the potential to reshape her field. For a researcher who has been on the faculty for only a short time, the award signals both the ambition of the science and the confidence the institute has placed in it.

Shi’s path to this project began far from South Carolina. She studied preventive medicine at Shanghai Jiao Tong University in China, where her involvement in tobacco control and other public health research projects led her to a realization that would define her career: improving health requires addressing not only diseases themselves but also the social and behavioral factors that shape how people live. That conviction deepened during an intervention project for people living with HIV in China. Although antiretroviral therapy has transformed HIV from a fatal diagnosis into a manageable chronic condition, Shi observed firsthand how many patients continued to struggle with stigma, fear of disclosure, and discrimination. Even individuals who were effectively controlling the virus medically tended to isolate themselves from family and friends, a pattern that convinced her that biomedical advances alone are not enough and that social and structural barriers must be confronted alongside them.

That experience brought her to the Arnold School, where she enrolled in the doctoral program in health promotion, education, and behavior and later completed a postdoctoral fellowship with the department and the South Carolina SmartState Center for Healthcare Quality. During that period she worked closely with faculty members Xiaoming Li and Xueying Yang, whom she credits as exceptional mentors who encouraged her to ask meaningful research questions, think creatively, and pursue innovative approaches that combine big data and artificial intelligence to advance HIV prevention and care. She has said their guidance was instrumental in her development as an independent researcher, fostering an environment of collaboration, curiosity, and innovation while emphasizing that research should ultimately improve people’s lives. It is also the community she built there, she explains, that drew her to remain at the Arnold School as a faculty member.

The new project brings together the threads of that training into a single, data-intensive research program. Shi and her team will draw on the NIH’s All of Us Research Program, one of the largest and most diverse health data resources ever assembled, which offers researchers access to longitudinal health information from hundreds of thousands of participants across the United States. Using advanced computational techniques, including machine-learning tools, the team will analyze how patterns of sexually transmitted infections relate to HIV risk and to HIV treatment outcomes over time. The analytical challenge is considerable: STI diagnoses arrive in clinical systems as scattered events, and connecting them to downstream HIV outcomes requires models that can capture multilevel social, structural, and clinical determinants of health simultaneously.

The scientific rationale for the project rests on a well-documented but underexploited relationship. Sexually transmitted infections are known to biologically increase the risk of acquiring HIV, and they may also signal lapses in engagement with HIV care among people already living with the virus. Yet, as Shi points out, STI data remains markedly underutilized by healthcare systems and researchers when it comes to predicting who may be at higher risk of HIV infection or adverse HIV-related health outcomes. In most clinical settings, an STI diagnosis is treated as a discrete event to be treated and closed, rather than as a signal that could trigger risk stratification, intensified prevention counseling, pre-exposure prophylaxis evaluation, or re-engagement efforts. Her project aims to close that gap by turning routinely collected clinical data into actionable predictive insight.

The stakes of better prediction are substantial. Of the roughly 1.1 million Americans living with HIV, an estimated 13 percent are unaware of their status, 25 percent do not receive care, and 35 percent are not virally suppressed. Each of those gaps represents a point at which the care continuum fails both the individual and public health, since unsuppressed viral load sustains transmission risk while untreated infection progresses. The extraordinary advances of the past three decades in HIV treatment, including reduced transmission achieved through careful management and modern prevention measures, mean that these numbers do not have to be what they are. What is missing, Shi argues, is a systematic way for clinicians to identify which patients need additional support and when. Her findings are intended to help clinicians do exactly that, and to give other scientists and healthcare providers a foundation for developing more effective strategies for HIV prevention and care.

Technically, the project sits at the intersection of infectious disease epidemiology, behavioral science, and data science. By training machine-learning models on the rich, longitudinal All of Us dataset, the team hopes to detect combinations of STI histories, demographic characteristics, social determinants, and clinical indicators that reliably precede HIV acquisition or poor treatment outcomes. Such models, if validated, could eventually be embedded in electronic health record systems as risk-stratification tools, flagging patients who warrant proactive outreach. The guidelines Shi’s team plans to develop for healthcare providers will translate these statistical findings into concrete clinical workflows, addressing the persistent divide between what population data can reveal and what individual practitioners can act upon during a routine visit.

Beyond its immediate clinical aims, the project reflects a broader vision of prevention that Shi has carried since her early days in preventive medicine: one in which social context, stigma, and structural barriers are treated as measurable, modifiable components of risk rather than as background noise. Her own trajectory, from studying preventive medicine in China to leading federally funded research at a major American research university, is one she hopes will encourage other early-career researchers and international scholars to pursue ambitious ideas with the potential for meaningful public health impact. The award, by its design, rewards exactly that kind of trajectory, betting on investigators whose careers are just taking shape and whose questions are still unconventional.

Looking ahead, Shi describes the Arnold School and the Center for Healthcare Quality as a unique environment where expertise in public health, clinical research, and data science converge. She intends to use the New Innovator Award to build an independent research program that leverages large-scale data and artificial intelligence to improve HIV prevention and care, while eventually extending her methods to other infectious diseases and health disparities. She also plans to invest in the next generation of the field, mentoring students and collaborating across disciplines to translate research findings into real-world impact. For a scientist whose formative insight was that data alone is never enough, the project she has now begun represents an attempt to prove the converse as well: that when large-scale data is joined to an understanding of human behavior and social barriers, prediction can become prevention.

Subject of Research: Use of large-scale health data and artificial intelligence to identify sexually transmitted infection patterns that predict HIV infection risk and treatment outcomes in at-risk populations

Article Title: New faculty member Fanghui Shi awarded NIH grant to protect at-risk populations from contracting HIV

Article References: New faculty member Fanghui Shi awarded NIH grant to protect at-risk populations from contracting HIV. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: HIV, NIH Director's New Innovator Award, Fanghui Shi, sexually transmitted infections, machine learning, All of Us Research Program, HIV prevention, public health, health disparities, University of South Carolina, Arnold School of Public Health, data science

The research, led by Alan Victor de Souza Pinho and Rosalvo Ferreira de Oliveira Neto of the Federal University of San Francisco Valley, together with Edilson Beserra de Alencar Filho, tackles a stubborn bottleneck in environmental risk assessment. Experimental toxicity testing is expensive, slow, and ethically fraught, and for pollinators the available data are especially sparse. Quantitative structure-activity relationship, or QSAR, modeling offers a computational alternative, predicting a compound’s biological effects directly from its molecular structure. But QSAR models are only as good as the molecular representations fed into them, and building those representations traditionally requires laborious feature engineering with specialized descriptor libraries.

The team drew its data from ApisTox, a recently released open-access benchmark containing curated toxicity information for 1,035 compounds, of which 296 are classified as toxic to honey bees and 739 as non-toxic. The classification follows the United States Environmental Protection Agency’s regulatory convention: a compound is deemed toxic if its acute LD50 value, whether by oral or contact exposure, is at or below 11 micrograms per bee. The dataset aggregates information from authoritative sources including the EPA’s ECOTOX knowledgebase and the Pesticide Properties DataBase, and for each chemical it retains the most toxic exposure route, a conservative strategy designed to capture the highest potential risk to pollinators.

Into this benchmark the researchers introduced three competing ways of describing molecules. The first used PaDEL, a widely adopted open-source software that calculates hundreds of handcrafted molecular descriptors. The second used Morgan fingerprints computed with RDKit, circular substructure patterns that encode the atomic neighborhoods of each molecule at fixed radii. The third and most novel approach extracted embeddings from MolFormer, a BERT-style chemical language model developed by IBM that treats SMILES strings, the text-based notation for molecular structures, as a language. MolFormer was pretrained on millions of molecules from the PubChem and ZINC databases using a masked language modeling objective, learning to predict hidden tokens from their context and thereby absorbing structural, semantic, and physicochemical information without any labeled toxicity data.

Crucially, the researchers did not fine-tune the massive model. Instead, they adopted a feature-based transfer learning strategy, using the pretrained network as a frozen feature extractor that converts each SMILES string into a dense 768-dimensional vector. These embeddings then served as input to three classical machine learning classifiers: Random Forest, Support Vector Machine, and a Multilayer Perceptron. All models were evaluated with five-fold cross-validation using the area under the receiver operating characteristic curve, or ROC-AUC, a threshold-independent metric well suited to the dataset’s class imbalance. Notably, the embeddings came from a publicly available reduced-scale MolFormer checkpoint trained on only about 100 million molecules, roughly ten percent of the combined ZINC and PubChem corpora, since the full-scale checkpoints trained on the reported 1.1 billion molecules are not publicly available.

The results delivered a clear verdict. Random Forest paired with Morgan fingerprints achieved the best overall discrimination, with a mean ROC-AUC of 0.866, confirming that substructure-based representations remain formidable baselines. Yet Support Vector Machine combined with MolFormer embeddings came astonishingly close, reaching a ROC-AUC of 0.859, a gap of just 0.007. Considering that the embeddings originated from a comparatively lightweight pretraining regime, the near-parity is striking, and it raises an tantalizing question the authors pose explicitly: had the full-scale model trained on the complete 1.1-billion-molecule corpus been accessible, the remaining gap might have narrowed further or even reversed.

Against the older PaDEL descriptors, the verdict was unambiguous. MolFormer embeddings outperformed PaDEL across all three classifiers, with ROC-AUC gains ranging from 0.005 to 0.021, the largest improvement appearing for the Support Vector Machine. The authors attribute this to the alignment between MolFormer’s learned representations and margin-based decision boundaries. More fundamentally, the advantage reflects a difference in representational philosophy. Morgan fingerprints encode the presence of local substructures without capturing the global molecular context in which a functional group sits, and their binary hashing is vulnerable to bit collisions. MolFormer’s bidirectional self-attention, by contrast, encodes every atomic position in relation to all other tokens in the SMILES sequence simultaneously, producing a holistic representation in which toxicophoric groups are described within the context of the complete molecular architecture.

The qualitative analysis of individual predictions illuminated this distinction vividly. Examining cases where the embedding-based model succeeded while the fingerprint-based model failed, the researchers found compounds rich in phosphate esters and carbamate groups, or polyhalogenated aromatic systems, structural motifs associated with acetylcholinesterase inhibition and modulation of neuronal ion channels, the classic mechanisms of insecticidal action. Conversely, when the fingerprint model won, the compounds tended to belong to well-defined agrochemical classes with established low bee toxicity, such as triazine herbicides and triazole fungicides, where simple class-defining substructures are the dominant signal and local encoding suffices. The two models, in other words, see different things in the same molecules.

That complementarity was quantified as well. Although the two best models achieved similar average performance, the Pearson correlation between their predicted probabilities was only 0.7483, indicating partially overlapping but non-identical predictive patterns. In some cross-validation folds the correlation dropped as low as 0.6417 without any degradation in predictive accuracy, evidence that each representation captures distinct aspects of chemical space. The authors argue that this makes the embedding-based model a strong candidate for ensemble frameworks, in which majority voting, probability averaging, or stacking with a meta-learner could combine the strengths of both representations, potentially alongside graph neural network architectures, to build unified predictive platforms accessible through online tools and mobile applications.

The practical implications extend well beyond the leaderboard. By eliminating the need to integrate multiple descriptor calculation libraries and perform complex variable selection, the transfer learning approach substantially simplifies the modeling pipeline, lowering the barrier for ecotoxicological screening under limited labeled data, precisely the conditions that prevail in pollinator protection. The authors caution that their models were developed and evaluated exclusively within the chemical space of the ApisTox dataset, and they recommend formal applicability domain analysis, using tools such as leverage statistics or nearest-neighbor distances in embedding space, before any regulatory deployment. Future work, they note, should explore full-scale pretrained models and incorporate chronic toxicity endpoints. As honey bee populations continue to decline worldwide under the combined pressures of pesticide exposure, habitat loss, and climate change, tools that can rapidly and reliably flag dangerous chemicals before they reach the field may prove not just convenient but essential.

Subject of Research: Transfer learning with chemical language models for predicting honey bee toxicity

Article Title: Transfer learning for honey bee toxicity prediction: MolFormer versus classical QSAR representations

Article References: de Souza Pinho, A. V., de Alencar Filho, E. B., & de Oliveira Neto, R. F. (2026). Transfer learning for honey bee toxicity prediction: MolFormer versus classical QSAR representations. Ecotoxicology, 35(7), Article 166. https://doi.org/10.1007/s10646-026-03149-x

Image Credits: AI Generated

DOI: 10.1007/s10646-026-03149-x

Keywords: transfer learning, MolFormer, QSAR, honey bee toxicity, Apis mellifera, ApisTox, chemical language models, Morgan fingerprints, PaDEL descriptors, ecotoxicology, pesticide risk assessment, machine learning