The research team, led by dermatologists and AI researchers in Turkey, fine-tuned ResNet-50, EfficientNet-B3, and Vision Transformer B/16 on 10,015 dermatologist-labelled dermoscopic images spanning seven diagnostic classes, from melanocytic nevi to melanoma and basal cell carcinoma. All three networks were initialised with ImageNet-pretrained weights, trained end-to-end in TensorFlow/Keras on a single NVIDIA Tesla P100 GPU, and evaluated three times each with identical hyperparameters, augmentation policies, and early-stopping criteria. The dataset was split at the lesion level rather than the image level, so that repeat photographs of the same lesion never straddled the training and test partitions. This detail matters more than it might seem: the authors report that the same architectures score three to six percentage points higher under the more common image-level splitting, which allows a model to recognise lesions it has already seen rather than generalise to new ones.

On the leakage-free lesion-level test set of 1456 images from 1068 lesions, accuracy was 0.807 for ResNet-50, 0.817 for EfficientNet-B3, and 0.808 for Vision Transformer B/16, with no statistically significant difference between any pair. Macro-averaged AUROC values clustered between 0.950 and 0.959. The one comparison that crossed the Bonferroni-corrected significance threshold was the transformer’s AUROC advantage over ResNet-50, a difference of 0.0097. The authors are careful about what these intervals do and do not license: the study was not designed as a non-inferiority comparison, and the confidence interval for the accuracy difference between the transformer and EfficientNet-B3 spanned minus 2.47 to plus 0.64 percentage points, a range that does not exclude differences some clinicians would consider material. The finding is that an advantage was not demonstrated, not that the models are equivalent.

Energy separated the architectures far more sharply than performance did. Measured gross training energy per run was 0.0143 kWh for ResNet-50, 0.0205 kWh for EfficientNet-B3, and 0.0630 kWh for Vision Transformer B/16, corresponding to 5.5, 7.9, and 24.4 grams of CO2-equivalent under the Turkish electricity grid scenario. The transformer consumed 3.07 times the training energy of EfficientNet-B3 and 4.41 times that of ResNet-50. Inference told a similar story: measured energy per 1000 single-image inferences was 0.115 Wh for ResNet-50, 0.071 Wh for EfficientNet-B3, and 0.307 Wh for the transformer, a 4.32-fold gap between the two extremes. Mean board power under load ranged from 111.5 W for EfficientNet-B3 to 214.5 W for the transformer, whose peak draw of 250.1 W was the only one to approach the card’s rated ceiling.

One result runs against a common intuition and is worth stating plainly: parameter count did not predict energy consumption. EfficientNet-B3 has 10.8 million trainable parameters against ResNet-50’s 23.6 million, yet consumed 43% more training energy. The reason is that EfficientNet-B3 was slower per epoch, 49 seconds against 37, and converged in more epochs, so a lower instantaneous power draw was more than offset by a longer time under load. A study that had inferred energy from model size, as much of the literature implicitly does, would have reached the opposite conclusion. Nor did training cost predict inference cost: EfficientNet-B3 was the most expensive convolutional network to train but the cheapest to run, consuming 38% less inference energy than ResNet-50 despite being 50% slower per image. The additional training energy is recovered after roughly 143,600 inferences, a threshold a screening programme would cross in its first year.

The melanoma results reveal why aggregate metrics are insufficient for clinical governance. Melanoma sensitivity was 0.446 for ResNet-50, 0.548 for EfficientNet-B3, and 0.593 for Vision Transformer B/16. Both differences from ResNet-50 crossed the corrected significance threshold, with EfficientNet-B3 gaining 0.102 and the transformer gaining 0.146. The architectures that detected more melanomas detected marginally fewer nevi, and because nevi constitute 63% of the test set, the two effects nearly cancel in the accuracy figure. An evaluation reporting only overall accuracy would have suggested little separation among the models while obscuring a 15-percentage-point difference in melanoma detection, the class with the most direct bearing on patient harm.

The same measurement framework therefore supported opposite verdicts in two adjacent comparisons, and this is the finding the authors ask readers to take from the study. The additional energy separating EfficientNet-B3 from ResNet-50 corresponded to a statistically significant and potentially clinically important absolute gain in melanoma sensitivity, roughly ten additional melanomas detected per hundred. The much larger increment separating Vision Transformer B/16 from EfficientNet-B3, 3.07 times the training energy and 4.32 times the inference energy, bought nothing detectable on accuracy, AUROC, or melanoma sensitivity. When a certain and substantial cost is set against an uncertain and undemonstrated benefit, the burden of justification falls on the party proposing the costlier option. Energy cannot decide architecture selection on its own, but it can make the exchange rate explicit, and neither verdict is available without measurement.

The measurement itself proved harder than expected, and the obstacles are cautionary in their own right. The team initially used CodeCarbon, a widely adopted software estimator, which on this hardware reported a GPU energy of exactly zero for every run, attributing the whole of the measured consumption to CPU and memory and omitting the component in which the architectures actually differ. The failure is silent, the tool reports a plausible non-zero total, and had it gone unchecked it would have inverted the conclusions. The researchers therefore sampled board-level GPU power directly through the NVIDIA management interface at 100-millisecond intervals and integrated it over time. Estimating inference energy from the accelerator’s thermal design power proved similarly unreliable: such estimates were 2.7 to 13.8 times the measured consumption, and the multiple differed by architecture, so the estimate is not a consistent upper bound and cannot be corrected by a constant factor. Rated maximum draw describes a card, not a workload.

Calibration, the property whose degradation post-market monitoring of clinical AI is most often designed to catch, produced a null result. Expected calibration error ranged from 0.052 to 0.064 across the three architectures, with run-to-run standard deviations of the same magnitude as the between-architecture differences. The reliability diagrams show the pattern typical of fine-tuned image classifiers, with confidence concentrated in the highest bin and modest overconfidence in the mid-range. The authors report this null result as a baseline against which post-deployment drift could be detected, rather than omitting a property that did not discriminate between the models.

The study’s limitations are acknowledged with unusual candour. HAM10000 is drawn from a limited number of mainly European centres, over-represents lighter skin phototypes, and contains images acquired under controlled dermoscopic conditions, so the performance figures may not transfer to larger and more heterogeneous datasets. The fixed 224 by 224 input resolution matches the canonical configuration of ResNet-50 and the transformer but not EfficientNet-B3, whose reference input is 300 by 300, so its performance here should be read as a lower bound. The energy figures are board-level measurements that exclude facility overhead such as cooling and power-supply losses, and they characterise one GPU card rather than the accelerator generations now in common use. The authors did not evaluate robustness to acquisition-device shift, a first-order concern for deployment, and note that if attention-based architectures prove more robust to changes in dermatoscope, that advantage would constitute exactly the kind of concrete clinical benefit capable of justifying the higher energy cost they document.

The wider argument is one of routine reporting and governance. The authors propose that pre-deployment evaluations and technical documentation for medical AI should report measured training energy, training carbon emissions under a specified grid, the hardware used, inference latency, measured energy per 1000 inferences, at least one calibration statistic, sensitivity on the highest-priority diagnostic class, and one or two normalised metrics such as accuracy per kWh. Their grid analysis shows that switching to lower-carbon electricity reduces absolute emissions for every architecture but does not close the gap between them, so efficient model selection and grid decarbonisation are complementary strategies rather than substitutes. Clinical AI models are not static: they are retrained when new images arrive, recalibrated when performance drifts, re-validated when they move sites, and updated when imaging hardware changes, and each of those steps consumes electricity. Reporting the energy used during initial training gives downstream stewards a baseline against which retraining cycles, model updates, and procurement decisions can be benchmarked over time. Two of the study’s findings are cautionary in this respect: parameter count did not predict energy consumption, and a widely used software estimation tool reported no GPU energy at all on the hardware used. Both would have produced a wrong answer had they been trusted without verification. Adding energy to the decision changed the conclusion here, and the authors expect it would change others.

Subject of Research: Artificial Intelligence

Article Title: Diagnostic performance and energy consumption of deep learning models for dermoscopic image classification

Article References: Ecer, N., Kankılıç, A. T., Karakoyun, Ö., Aslancan, B., & Kaldan, B. (2026). Diagnostic performance and energy consumption of deep learning models for dermoscopic image classification. Discover Artificial Intelligence, 6(1), Article 1530. https://doi.org/10.1007/s44163-026-02558-5

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02558-5

Keywords: deep learning, dermoscopy, skin cancer, melanoma, energy consumption, carbon footprint, clinical AI, sustainable AI, HAM10000, algorithmovigilance, Diagnostic, performance

The research, published in Frontiers in Environmental Engineering, benchmarked four widely used machine learning and deep learning architectures on a publicly available operational dataset from a municipal wastewater treatment plant in Melbourne, Australia. The team compared a feedforward neural network, a recurrent neural network, a one-dimensional convolutional neural network and the gradient-boosted tree algorithm XGBoost, all trained on identical data splits and feature sets. Their verdict is striking: a comparatively simple tree-based model matched or beat the deep learning alternatives while consuming between 98 and 99 percent less energy and emitting roughly 65 to 100 times less carbon dioxide equivalent than the recurrent network.

The motivation reaches far beyond one Australian plant. Globally, around 380 billion cubic metres of wastewater were generated annually as of 2015, and more recent estimates suggest 332 billion cubic metres of household wastewater alone in 2024, a figure expected to double by 2050 as cities expand. Only about 56 percent of household wastewater is currently treated for safe discharge, leaving a vast management gap. Treatment plants are increasingly being reframed as resource recovery facilities: anaerobic digestion of the organic carbon in wastewater can substitute for 40 to 76 percent of a plant’s electrical energy, while nutrient recovery technologies could capture up to 15 to 20 percent of the world’s phosphorus and nitrogen requirements from wastewater streams.

The stakes of getting these forecasts right are both economic and environmental. Reliable COD predictions can guide feed allocation to anaerobic digestion and fermentation processes that yield volatile fatty acids, the critical intermediates for biochemical and biofuel production, and polyhydroxyalkanoates, sustainable alternatives to petrochemical plastics. TN forecasts, meanwhile, can inform nutrient recovery, aeration control and compliance management. Untreated nitrogen released into water bodies drives eutrophication, the algal overgrowth that depletes oxygen, and nitrogen-removal processes themselves can emit nitrous oxide, a greenhouse gas with a global warming potential roughly 298 times that of carbon dioxide.

To build their comparison, the researchers started with 1,382 operational records containing sixteen variables. Data cleaning, variance screening and standardisation were fitted only on the training subset and then applied unchanged to validation and test data, a discipline designed to mimic how a model would behave on unseen plant data rather than under idealised retrospective conditions. Weather-related variables with near-zero variance, such as atmospheric pressure, total rainfall and average visibility, were discarded. A two-step feature screening combining Pearson, Spearman and distance correlations with ANOVA-based relevance ranking distilled the predictor space down to seven input features, used consistently across all four architectures to guarantee a fair fight.

The dataset was divided into 70 percent training, 15 percent validation and 15 percent test subsets, with hyperparameter selection guided by validation performance and regularisation plus early stopping applied to the neural models. The feedforward network used two hidden layers of 64 and 32 neurons; the recurrent network a single 32-unit layer; the convolutional network a one-dimensional layer with 32 filters and a kernel size of 3; and XGBoost 100 boosting trees with a maximum depth of 6 and a learning rate of 0.3. All neural models trained for 400 epochs with a batch size of 50, and the identical configuration was applied to both targets so that performance differences reflected the data rather than the setup.

For COD, the recurrent neural network captured temporal behaviour most effectively, achieving an R-squared of 0.71 and a root mean square error of 75.55 milligrams per litre, with points hugging the diagonal of the actual-versus-predicted plot and residuals more symmetrically clustered around zero than the feedforward baseline, which scored 0.65. XGBoost matched that test performance at 0.71 with the strongest generalisation, posting validation and test R-squared values of 0.72 and 0.71, though its near-perfect training fit of 0.99 signalled mild overfitting, a familiar trait of high-capacity tree ensembles on structured data. For Total Nitrogen, XGBoost was the clear front-runner with validation and test R-squared values of 0.59 and 0.63, outperforming the recurrent network at 0.56 and 0.53 and the other neural architectures, suggesting that nitrogen variability is governed more by nonlinear feature interactions than by sequence memory.

The convolutional network added little value over the feedforward baseline for either target, confirming that local convolutional filtering does not benefit low-dimensional tabular wastewater data. The moderate R-squared range of roughly 0.54 to 0.71 observed across all models is realistic for operational plant datasets, which typically lack dissolved oxygen, sludge retention time and explicit kinetic parameters, and the authors note that state-of-the-art wastewater modelling studies on real plant data usually land between 0.6 and 0.8. The ceiling here reflects missing process-state variables and inherent sensor noise rather than model inadequacy.

The study’s most consequential contribution is its sustainability audit. Using the CodeCarbon library, the team recorded electricity use, water consumption for data-centre cooling and carbon dioxide equivalent emissions for every model-target combination. For COD, XGBoost required only 0.000018 kilowatt-hours and emitted 0.0000129 kilograms of CO2-equivalent, compared with 0.001490 kilowatt-hours and 0.001063 kilograms for the recurrent network; for Total Nitrogen, XGBoost consumed 0.000028 kilowatt-hours against 0.002563 for the recurrent model. From these measurements the researchers derived an Environmental Cost Index, a Green Performance Index and a Triple Green Sustainability Score, with XGBoost topping the green rankings at 0.63 to 0.70 and 0.64 to 0.72 respectively while the recurrent network carried the heaviest environmental burden.

The broader context makes these small numbers matter. AI systems could be responsible for 32.6 to 79.7 million tons of CO2 emissions in 2025, data centres consumed about 4.4 percent of United States electricity in 2023, and under current trends they could approach 9 percent of global electricity demand by 2030. Because machine learning models are trained and redeployed repeatedly across facilities, the cumulative footprint of an energy-hungry architecture compounds. The authors’ conclusion reframes model selection for environmental engineering as a multi-criteria decision problem in which algorithmic greenness is a core technical consideration rather than an afterthought: when a lean tree ensemble predicts as well as a deep network, the conventional instinct toward architectural complexity is unjustified on both performance and sustainability grounds, and the future of wastewater analytics may belong to the models that do more with far less.

Subject of Research: Green machine learning model selection for predicting COD and Total Nitrogen in wastewater valorisation

Article Title: Green machine learning assisted analytics for wastewater valorisation: a case study of Melbourne, Australia

Article References: Debnath, B., Pal, R. S. C., & Chattopadhyay, A. K. (2026). Green machine learning assisted analytics for wastewater valorisation: a case study of Melbourne, Australia. Frontiers in Environmental Engineering, 5, Article 1839957. https://doi.org/10.3389/fenve.2026.1839957

Image Credits: AI Generated

DOI: 10.3389/fenve.2026.1839957

Keywords: green machine learning, wastewater valorisation, XGBoost, recurrent neural network, chemical oxygen demand, total nitrogen, CodeCarbon, carbon footprint, resource recovery, circular economy, Melbourne wastewater dataset, sustainable AI

The framework, called Adaptive-CGAN, is a conditional generative adversarial network, a class of AI architecture in which two neural networks compete against each other. A generator produces synthetic images from random noise, while a discriminator tries to distinguish those fakes from real samples. As training proceeds, the generator becomes progressively better at fooling the discriminator, and the discriminator becomes sharper at catching deception. What makes Adaptive-CGAN distinctive is how it handles class information—the labels that tell the model whether an image depicts COVID-19, viral pneumonia, lung opacity, or a healthy chest. Rather than injecting the class label only once at the generator’s input, as conventional conditional GANs do, Adaptive-CGAN re-projects a learned class embedding at every layer of the generator, so that class information shapes both the global anatomical structure and the fine pathological texture of the output.

The second key innovation is a dual-output discriminator. Instead of merely judging whether an image is real or fake, the discriminator simultaneously classifies the image into its disease category. This means the generator receives two complementary streams of feedback during training: a realism signal that pressures it to suppress artifacts and produce anatomically plausible textures, and a class-alignment signal that forces generated images to carry the diagnostically relevant features of their target condition. The authors describe this as a dual-feedback mechanism, with both gradients flowing through a shared discriminator trunk back into the generator, aligning fine-scale, class-diagnostic cues without sacrificing visual plausibility.

To train and evaluate the model, the team compiled a multi-institutional chest X-ray dataset of roughly 8,000 unique images spanning four clinical classes, expanded through training-only augmentation to approximately 80,000 instances per epoch. The researchers took unusual care to prevent data leakage: duplicate and near-duplicate studies were removed before splitting, and partitioning into training, validation, and test sets was performed at the anonymized patient level, so no patient appeared in more than one subset. Augmentation was applied only after splitting and exclusively to the training data. Images were resized to 256 by 256 pixels, normalized, enhanced with histogram equalization, and denoised with median filtering before being paired with class labels encoded as 100-dimensional embedding vectors.

The results are striking. Against the closest baseline, AC-GAN, Adaptive-CGAN raised downstream classification accuracy from 93.75 percent to 99.60 percent, improved the Inception Score—a measure of generated image quality and diversity—by 17.65 percent, and reduced the Fréchet Inception Distance and Kernel Inception Distance, both standard measures of how closely synthetic images mimic real data distributions, by roughly 8 percent each. Training stability metrics reinforced the picture: conditional entropy fell from about 3.8 to 0.15 nats, the Jensen–Shannon divergence between real and generated distributions dropped to 0.041 compared with 0.119 for AC-GAN, and mode coverage reached 94.2 percent versus 63.2 percent for the baseline, indicating the model avoided the classic GAN failure mode of producing a narrow, repetitive set of outputs. An ablation study confirmed that both the layer-wise embedding and the dual-feedback discriminator contribute independently, and that combining them yields the largest gains.

Skeptics of generative medical imaging often worry about circularity: if you evaluate synthetic images with the same classifier used during training, the results may flatter the model. The authors addressed this with two independent validation strategies. A ResNet-50 network trained exclusively on real X-rays achieved 94.20 percent accuracy on unseen synthetic images, a modest 3.4-point degradation that suggests the synthetic data preserve diagnostically meaningful information. Separately, the team computed quality metrics in the feature space of a DenseNet-121 encoder pre-trained on the external CheXpert dataset of more than 224,000 chest X-rays, and Adaptive-CGAN outperformed AC-GAN in that independent, domain-specific representation as well.

The study also tackles a dimension that most generative modeling papers ignore: the environmental cost of training. Using the CodeCarbon tool, the researchers measured energy consumption and estimated CO₂ emissions with region-specific carbon-intensity factors. On identical local hardware, Adaptive-CGAN trained 15.8 percent faster and emitted 9.5 percent less CO₂ than AC-GAN, gains attributable to the architecture itself. Deployed on Google Cloud Platform, the model emitted just 0.0023 kilograms of CO₂, roughly half the 0.0047 kilograms of a local setup, thanks to optimized resource allocation and a TFRecords-based data pipeline that accelerates storage and loading. The cloud deployment also doubled throughput to 1 gigabit per second and cut latency from 100 milliseconds to 10, while the CWGAN baseline, by contrast, produced the highest emissions of any model tested at 0.0336 kilograms.

The cloud infrastructure is more than a convenience. The TFRecord pipeline serializes images into an efficient format for TensorFlow’s data-loading API, reducing input-output overhead and supporting scalable training on GPU-equipped virtual machines. The authors frame this as a distinct engineering contribution, arguing that as medical datasets grow toward multi-hospital and multi-national scale, optimized sharding, larger GPU memory, and distributed training will become essential. They also note that Google Cloud’s renewable energy commitments amplify the sustainability benefits, aligning the work with the United Nations Sustainable Development Goals on climate action and industrial innovation.

The researchers are careful to draw boundaries around what the technology can and cannot do. Generated images are synthetic and must never be treated as real patient cases or standalone diagnostic evidence; the intended use is research-level data augmentation and model development, not clinical decision-making. The team acknowledges remaining limitations, including potential artifacts, class ambiguity between visually similar thoracic conditions, the risk that biases in training data could be reproduced or amplified, and the possibility of memorization, which they probed with nearest-neighbor inspection. They call for prospective clinical validation, blinded specialist review, privacy safeguards such as differential privacy and membership-inference testing, and fairness audits across patient subgroups before any clinical-facing deployment.

Even within those limits, the work signals where medical AI is heading. Generative models that can produce faithful, class-controlled synthetic images could relieve the data bottleneck that has long constrained diagnostic deep learning, particularly for rare diseases and fast-moving epidemics. By demonstrating that architectural innovation, rigorous validation, cloud-scale engineering, and carbon accounting can be pursued together, the Port Said team offers a template for what sustainable, clinically responsible generative AI might look like as the field matures.

Subject of Research: A conditional generative adversarial network for synthesizing medical chest X-ray images on cloud infrastructure to improve disease diagnosis and reduce carbon emissions

Article Title: Adaptive-CGAN: a comprehensive framework for generative modeling on google cloud platform

Article References: Saber, W., El-Baz, A., & Rizk, R. (2026). Adaptive-CGAN: a comprehensive framework for generative modeling on google cloud platform. Cluster Computing, 29(12), Article 710. https://doi.org/10.1007/s10586-026-06496-w

Image Credits: AI Generated

DOI: 10.1007/s10586-026-06496-w

Keywords: generative adversarial networks, medical imaging, synthetic data, COVID-19, chest X-ray, Google Cloud Platform, deep learning, diagnostic accuracy, sustainable AI, carbon emissions, data augmentation, TFRecords

The challenge the team set out to solve is one familiar to anyone who manages a vast heritage site. Pompeii covers roughly 44 excavated hectares and contains approximately 1,500 buildings, and every one of them, along with its frescoes, statues, and structural elements, is slowly deteriorating under the open sky. Continuous monitoring is essential, but human inspection of every roof is slow, expensive, and sometimes dangerous. Drones offer a solution: they can fly over inaccessible areas and capture high-resolution, georeferenced photographs. Yet the raw images are only the beginning. Someone, or something, must interpret thousands of aerial photographs and decide which roofs need urgent attention.

HydraML’s answer is a two-stage pipeline that blends the strengths of deep learning with the efficiency of classical machine learning. In the first stage, a deep neural network, selected automatically by a deep AutoML process built on Amazon’s AutoGluon framework, examines each drone image and extracts a rich numerical representation, or embedding, of what it sees. Rather than using this network to make the final decision, HydraML passes those embeddings through a mathematical compression step based on the Fuzzy-Transform, a technique rooted in fuzzy set theory that reduces the dimensionality of the feature space while preserving the information that matters for classification. The compressed features then feed into a second AutoML system, which automatically selects and trains the best traditional classifier for the task.

This hybrid design is what sets HydraML apart from earlier approaches. Previous combinations of deep and classical methods required researchers to manually choose and tune both the feature extractor and the final classifier, a laborious process involving extensive hyperparameter optimization. By bringing both stages into the AutoML paradigm, HydraML becomes something close to plug-and-play: users specify only a training time budget for the deep component and a few parameters for the fuzzy compression, and the system handles the rest. The Fuzzy-Transform compression also pays dividends before training even begins, because smaller images mean less energy spent transmitting data from drone to cloud.

The mathematical heart of the compression step is elegant in its simplicity. The Fuzzy-Transform partitions an image or feature matrix using overlapping basic functions, which can be sinusoidal, Gaussian, or pi-shaped, each covering a neighborhood of the data. By projecting the original matrix onto these functions, the method produces a much smaller compressed representation. Crucially, prior theoretical work has shown that under reasonable conditions, a learning system trained on Fuzzy-Transform-compressed data can achieve error rates of the same order as one trained on the original data. In HydraML’s experiments, compressing 768-dimensional deep embeddings down to a 12-by-16 grid roughly halved both classifier training time and CPU energy consumption on the Pompeii datasets, with almost no loss in predictive performance.

To test the system, the researchers built two novel datasets from orthophotos of Pompeii captured between June 2023 and September 2024 by a DJI Matrice 300 RTK drone flying at 80 meters, equipped with a Zenmuse L1 camera. Nine orthophotos, each roughly 10 gigabytes, were georeferenced so that any pixel could be traced back to its exact location on the ground. With the help of domain experts, the team curated PompeiiTile, containing 1,327 images of tiled roofs labeled as normal, broken, or covered in weedy vegetation, and PompeiiFlat, containing 922 images of flat roofs labeled as normal, broken conduits, stained, or weedy. Careful data augmentation, applied separately to each cross-validation fold to avoid leakage, balanced the training and validation sets.

The results were striking. On PompeiiTile, HydraML using a cosine fuzzy partition achieved an F1-score of 0.9406, with an accuracy of 0.9404 and the lowest variance among all tested methods, signaling unusual stability. On PompeiiFlat it reached an F1-score of 0.9117. To probe generalization beyond cultural heritage, the team also evaluated HydraML on RescueNet, a publicly available dataset of 8,354 aerial images collected after Hurricane Michael struck Florida in 2018, categorized by FEMA-inspired damage grades from no damage to total destruction. There, HydraML achieved an F1-score of 0.8686, outperforming a DenseNet201-based hybrid approach and a VGG16-based transfer learning method, and beating the standalone deep AutoML system by 1.30 percentage points.

The energy savings may prove even more consequential than the accuracy gains. On the urban dataset, HydraML cut GPU energy consumption by about 62.5 percent compared with a DenseNet201-SVM hybrid, dropping from roughly 139,000 joules to 52,000 joules. On PompeiiTile, reducing training from 480 to 300 seconds saved around 40.92 percent of GPU energy, and on PompeiiFlat the system saved 43.64 percent of GPU energy and 12.28 percent of CPU energy relative to the standalone deep AutoML baseline. Notably, HydraML beat that baseline even though it added just 20 seconds for compression and classical classifier training. The team measured power directly, sampling CPU and GPU draw every 100 milliseconds during training and integrating over time, giving the comparison unusual rigor.

Beyond simply detecting damage, HydraML can grade its severity. Using the same deep embeddings without retraining the expensive neural stage, the system trains additional classifiers to sort each detected problem into small, medium, or high severity. It distinguished broken-tile severity with an F1-score of 0.97, and performed well on vegetation and conduit issues, though it struggled with damp patches on flat roofs, where F1 barely exceeded 0.75. Most classification errors occurred between adjacent severity levels, a difficulty the researchers note also plagues human experts defining the ground truth. This severity information matters because heritage managers use vulnerability indices to prioritize interventions, and accurate severity grading feeds directly into maintenance planning.

The system is not infallible, and the researchers are candid about its failure modes. Dirt and mold on tiles can mimic missing tiles, shadows on flat roofs can be mistaken for damp stains, and dry, reddish vegetation is harder to spot than lush green growth. In one telling example, a clearly broken roof section at the edge of an image was classified as normal because the model focused on the center. The team also acknowledges that the Pompeii datasets contain relatively few original samples and that the approach has not yet been tested at other archaeological sites. Still, the researchers point out that false positives are the safer error: a wrongly flagged roof gets checked by an expert, whereas a missed problem can quietly destroy irreplaceable frescoes and structures. With the code released openly and the Pompeii datasets available on request, HydraML offers heritage sites worldwide a template for monitoring that is simultaneously smarter, faster, and greener.

Subject of Research: A hybrid automated machine learning approach using deep learning, Fuzzy-Transform compression, and UAV imagery to classify damaged roofs at the archaeological site of Pompeii

Article Title: HydraML: a hybrid damaged roofs automated machine learning classifier for cultural heritage

Article References: Colace, F., De Santo, M., Gaeta, R., & Zuchtriegel, G. (2026). HydraML: a hybrid damaged roofs automated machine learning classifier for cultural heritage. Neural Computing and Applications, 38(19), Article 798. https://doi.org/10.1007/s00521-026-12499-w

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12499-w

Keywords: HydraML, automated machine learning, cultural heritage, Pompeii, deep learning, Fuzzy-Transform, roof damage classification, UAV imagery, sustainable AI, AutoGluon, energy efficiency, archaeological site monitoring

The bet paid off. Individually, each of these units possesses limited intelligence and only fragmentary information about the world. Together, however, by exchanging local signals and coordinating their behaviour, they can build a shared representation of their existence, their environment and their goals. The researchers call this collaborative awareness, and it became the conceptual backbone of the entire project. The consortium brought together an unusually broad set of disciplines: artificial intelligence and robotics researchers from the University of Pisa, philosophers and cognitive scientists from Ludwig Maximilian University of Munich, swarm robotics specialists at the University of Bristol, soft robotics experts at Delft University of Technology, and the French research hub Da Vinci Labs. Over four years, this group produced not just papers, but a philosophical, mathematical and technological framework for awareness in artificial systems—along with robotic prototypes, more sustainable AI architectures, startup ventures and follow-up European projects.

One of the project’s most important contributions is a careful conceptual distinction that separates awareness from consciousness. Consciousness, in the philosophical sense, is generally associated with subjective or phenomenal experience: having feelings, or experiencing the world from a particular first-person perspective. Awareness, as the EMERGE team defines it, is something far more tractable. “In EMERGE, we understand awareness more operationally, as the capacity of an agent to process and integrate information in a way that is relevant to its actions,” explains Ophelia Deroy, philosopher and cognitive scientist at LMU Munich. Crucially, the project’s experiments showed that people can understand an artificial system as aware without ever assuming that it has subjective experience. That separation matters enormously for public perception and for the ethics of human–machine interaction, because it allows engineers to build and describe aware machines without drifting into claims about machine sentience.

To make the concept scientifically useful rather than merely evocative, the researchers decomposed awareness into distinct dimensions that can be applied across individuals and collective systems: spatial awareness, temporal awareness, self-awareness, agentive awareness and metacognitive awareness. Rather than treating awareness as an all-or-nothing property—a binary switch that a system either possesses or lacks—the framework ties each dimension to specific capabilities and concrete tasks. This move transforms a philosophical puzzle into an experimental programme. It becomes possible, for the first time, to test empirically whether increasing a system’s awareness along a particular dimension actually improves its performance on a defined task, and to measure by how much. The result is a research methodology in which awareness is not asserted but demonstrated, dimension by dimension, capability by capability.

The mathematical and computational core of the project addresses a deceptively simple question: how can awareness emerge across a physically distributed system in which no single agent sees the whole picture? The EMERGE framework describes how simple local information, exchanged with neighbouring agents, can be integrated so that the collective coordinates its actions toward solving a task. “The beauty of the approach is that the system operates in a distributed way,” says Sabine Hauert, professor at the University of Bristol. “This way systems that can scale to large numbers of agents and remain operational when individual robots fail or conditions change. It also allows humans to interact with the systems as a coordinated collective rather than controlling each robot separately.” In practical terms, that means a swarm does not collapse when one of its members breaks down, and a human supervisor can issue intentions to the group as a whole instead of micromanaging every unit.

At Delft University of Technology, the framework was pushed toward physical robotic bodies. “We want robots that can understand how their actions affect the world, anticipate what people and other robots are doing, and adapt accordingly,” explains Cosimo Della Santina, associate professor at TU Delft. “That is a key step toward robotic systems that can operate more autonomously and work naturally alongside humans.” This anticipation of consequences—understanding how one’s own movements reshape the environment, and predicting the trajectories of both human partners and fellow machines—is precisely the kind of integrated, action-relevant information processing that the project’s operational definition of awareness was built to capture. It is also what separates a genuinely collaborative robot from a pre-programmed arm repeating fixed motions.

The consortium’s mathematical framework also delivered an unexpected dividend for machine learning itself: parsimony. Systems built on the EMERGE principles proved capable of learning from substantially smaller quantities of data than conventional architectures require. “Many AI systems today depend on large, centralized computing infrastructures,” says Claudio Gallicchio, associate professor at the University of Pisa. “Our results open the way to a more decentralized and environmentally sustainable models, in which data can be processed locally using small and energy-efficient devices, such as neuromorphic hardware, reducing energy consumption, network traffic and latency while improving privacy.” In an era when training frontier models consumes enormous energy budgets and concentrates computational power in a handful of data centres, the prospect of aware, adaptive systems running on small, low-power edge hardware carries implications far beyond robotics—touching sustainability, data sovereignty and the geopolitics of computing.

The project also confronted the human side of the equation, and the findings are sobering. Bahador Bahrami, director of the Crowd Cognition group at LMU Munich, notes that “our work also examined the ethical side of human interaction with collaboratively aware systems, with important implications for future environments in which groups of humans and artificial agents will have to interact, negotiate and cooperate.” Among the studies was research into how people behave toward automated systems such as self-driving cars. The result: users may be more willing to take advantage of artificial agents than of human counterparts—a phenomenon the researchers describe as algorithmic exploitation. People who would never cut in front of another driver may happily exploit a machine, precisely because it cannot be offended, retaliate, or hold them socially accountable. Understanding this asymmetry, the team argues, is essential for designing and governing artificial systems that interact directly with people.

EMERGE was deliberately structured as a pathway from foundational research to technological innovation, and its results are already spinning out into ventures. Embodied AI develops efficient AI models for future generations of humanoid robots. ContinualIST builds AI models efficient enough to run on smaller, lower-power hardware. Collective Robotics focuses on swarms of robots that coordinate and operate collectively, and Phoebe is dedicated to systems in which humans and artificial agents solve problems together. The concepts will also anchor new research projects: EIC Pathfinder SWIFT-BUILD will apply collaboration among multiple robots and humans to automated construction, while HE MINGLE and ADA-COLLAB investigate how general-purpose, agentic AI systems can form societies, collaborate with one another, and interact with humans. Even public engagement got its own artefact: nOdes, an artistic installation in which an interactive swarm exchanges information and changes its behaviour in response to both its agents and its human visitors, letting anyone experience collaborative awareness firsthand.

“EMERGE started from a fundamental question about awareness, but it has ended with very concrete results: new robotic systems, more sustainable AI architectures, new startup initiatives and follow-up projects that are already carrying forward the idea of building artificial systems that are more efficient, sustainable, collaborative, and imbued with human values,” concludes Davide Bacciu, professor at the University of Pisa and coordinator of the project. The project was funded by the European Union under Grant Agreement 101070918, with UK participants supported by UKRI grant number 10038942. Its legacy is a reframing of one of technology’s oldest ambitions: intelligence may not need to be concentrated, enormous and opaque. It can be distributed, modest, energy-lean and—perhaps most importantly—legible to the humans who live alongside it. The swarms are small, but the idea they carry is anything but.

Subject of Research: Collaborative awareness in distributed artificial agents and swarm robotics

Article Title: European project EMERGE concludes advancing collective awareness in artificial systems

Article References: European project EMERGE concludes advancing collective awareness in artificial systems. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: EMERGE project, collaborative awareness, swarm robotics, distributed AI, EIC Pathfinder, machine learning, neuromorphic hardware, human-robot interaction, algorithmic exploitation, sustainable AI, philosophy of mind, cognitive science