The tool, called SALS-BioC, is described in the journal SoftwareX as a soft-sensor based adaptation learning simulator for predicting bacterial concentrations in membrane bioreactors. A soft sensor, in the engineering sense, is a machine learning model that estimates hard-to-measure quality variables from variables that are easy to measure continuously. In this case, the system takes routine physicochemical water quality readings such as pH, electrical conductivity, total suspended solids, biochemical oxygen demand, chemical oxygen demand, nitrate nitrogen, turbidity, and chlorine levels, and uses them to predict how many bacterial cells are present in the treated effluent. The idea is to replace days of laboratory work with an instant, data-driven estimate that plant operators can act on in real time.

The research team, composed of H. Bagci, I. N’Doye, F. Almulhim, and P.-Y. Hong, confronted a familiar obstacle in environmental machine learning: there is simply not enough data. Wastewater treatment plants face cost and confidentiality constraints that limit access to large datasets from water-based epidemiology studies, and microbial measurements are inherently scarce because they depend on weekly sampling campaigns. Machine learning models trained on small, static datasets tend to perform well in the laboratory but fail when confronted with new conditions, a problem known as poor generalization. When the researchers tested a conventional long short-term memory (LSTM) neural network on data from a biologically independent sampling period, the model achieved an impressive coefficient of determination of 0.96 on its own training and testing data, but collapsed to a negative value of minus 0.33 on the unseen replicate, meaning its predictions were worse than a naive average.

To solve the data scarcity problem, the team developed a novel synthetic data generation algorithm called EMCM-PS, short for an ensemble Markov chain model with a joint state representation and probabilistic sampling rule. The approach is a creative departure from mainstream generative techniques. Instead of using generative adversarial networks, which pit two neural networks against each other in a training game and often suffer from a failure mode known as mode collapse, EMCM-PS builds on Markov chains and Gaussian copulas. Each multivariate observation in the wastewater dataset is encoded as a single joint state, and new synthetic samples are drawn directly from the empirical probability distribution of those observed states rather than from a sequential transition matrix. The researchers argue this design choice matters because wastewater samples are collected at irregular intervals and do not form a strict time series, so forcing a random-walk structure onto the data would introduce spurious dependencies and trap the generator in a limited set of states.

The generation workflow begins by augmenting the original dataset with correlation-aware multiplicative noise, then discretizing the samples into quantile-based joint states. Synthetic states are sampled independently from the global frequency distribution of observed states and decoded back into continuous measurements using either a local mode, which samples from the observations associated with each state, or a uniform mode, which draws values within the bounds of the corresponding quantile bucket. The software includes a suite of visual diagnostics to verify that the synthetic data faithfully reproduce the statistical structure of the real measurements, including marginal distribution comparisons, Pearson correlation heatmaps, and dimensionality-reduction plots based on principal component analysis and t-distributed stochastic neighbor embedding. A fidelity scorecard summarizes these metrics so users can judge at a glance whether the generated data preserve the variability and dependency patterns of the original samples.

On top of this synthetic data engine sits the second key innovation: a lifelong learning framework built around the LSTM predictor. Unlike traditional machine learning models that are trained once on a static historical dataset and then frozen, the lifelong learning approach continuously updates the model as new batches of data arrive from the target environment, while preserving previously acquired knowledge through a dictionary learning mechanism. In the SALS-BioC workflow, a source-domain model is first trained on historical data, then adapted to a target-domain dataset processed in sequential batches of thirty samples. For each batch, the model predicts first and is updated with the true values only afterward, mimicking the realistic operational scenario in which laboratory confirmation lags behind the need for a prediction. The software even provides an animated visualization of how the root mean square error evolves batch by batch during adaptation, showing how quickly the model recovers its accuracy as it encounters the new domain.

The validation experiment drew on real data from the KAUST aerobic membrane bioreactor wastewater treatment plant, which treats a mix of municipal wastewater. Two biologically independent replicates were used, one collected weekly from July to October 2023 for model development and another from February to March 2024 to assess generalization. Fourteen physicochemical parameters were measured alongside bacterial abundances quantified by flow cytometry. When the lifelong learning model, trained on EMCM-PS-generated synthetic data, was adapted to the unseen replicate, it achieved a cross-validation coefficient of determination of 0.8723. When the same experiment was run using synthetic data from a Wasserstein generative adversarial network, the best-performing member of the GAN family, the score reached only 0.6450. That difference of 0.2273 translates into a 35.2 percent improvement for the new probabilistic approach, a striking margin in a field where incremental gains are the norm.

The software itself is designed to be accessible rather than the exclusive province of machine learning specialists. It is implemented as a modular Python application with a browser-based interface built on Flask, HTML, CSS, and JavaScript, using standard scientific libraries including NumPy, pandas, SciPy, scikit-learn, and PyTorch. Users upload CSV datasets, select target variables, configure parameters, and run simulations through three dashboard sections covering synthetic data generation, LSTM-based prediction, and lifelong learning adaptation. Trained models can be downloaded as serialized PyTorch files and reloaded later for validation without retraining, and Bayesian hyperparameter optimization is built in with configurable search iterations. The code is released under the MIT license on GitHub, and the authors note that the architecture deliberately separates the interface, the server orchestration, and the algorithms, so that new generators, predictive models, or adaptation strategies can be added through dedicated API routes without redesigning the application. Extensions to viral contaminant prediction are described as a natural next step.

The implications reach beyond one treatment plant in Saudi Arabia. Water-based epidemiology has surged in prominence since the COVID-19 pandemic demonstrated that sewage can serve as an early-warning system for disease outbreaks in entire communities, but the field remains bottlenecked by the pace of laboratory analysis. A reliable soft sensor that predicts bacterial concentrations from measurements plants already collect could enable continuous biological monitoring, early diagnosis of operational faults, and rapid response to contamination events, all without waiting for culture results. The authors also emphasize the educational value of the tool, arguing that its guided interface lowers the expertise required to design data-driven soft sensor systems and makes the technology accessible to interdisciplinary researchers and students.

The researchers are candid about limitations. EMCM-PS currently uses a fixed number of quantile-based buckets for all features, but different water quality variables have different distributions and may require different discretization resolutions, particularly under operational variations such as fluctuations in flow rate or for highly skewed measurements like total cell counts. Future work, they suggest, could develop an adaptive discretization strategy that tunes the bucket count for each feature dynamically. They also plan to validate the simulator across international wastewater datasets to establish its transferability. For now, SALS-BioC stands as a concrete demonstration that thoughtful statistical modeling, rather than ever-larger neural networks, can sometimes deliver the biggest wins when data are scarce, and that open, reproducible software may be the fastest route to putting adaptive artificial intelligence into the pipes and pumps of the world’s water infrastructure.

Subject of Research: Machine learning-based soft sensor prediction of bacterial concentrations in membrane bioreactor wastewater treatment

Article Title: SALS-BioC: A soft-sensor based adaptation learning simulator for predicting bacterial concentrations in membrane bioreactors

Article References: Bagci, H., N’Doye, I., Almulhim, F., & Hong, P.-Y. (2026). SALS-BioC: A soft-sensor based adaptation learning simulator for predicting bacterial concentrations in membrane bioreactors. SoftwareX, 36, Article 103084. https://doi.org/10.1016/j.softx.2026.103084

Image Credits: AI Generated

DOI: 10.1016/j.softx.2026.103084

Keywords: soft sensors, machine learning, wastewater treatment, membrane bioreactors, synthetic data generation, Markov chains, LSTM networks, lifelong learning, water-based epidemiology, bacterial concentration prediction, generative adversarial networks, environmental monitoring

Researchers led by Chengkai Yu, Fengling Zhang, and Francesco Marinello have built a framework that embeds data trustworthiness directly into the life cycle assessment of what they call the Digitization Footprint, the total resources consumed as data are acquired, transmitted, processed, stored, and reused. Their central insight is deceptively simple: when data quality degrades, through sensor drift, calibration failure, or environmental interference, the system does not merely produce bad numbers. It fights back. Verification routines, retransmissions, and rollback recomputations kick in, and these defensive behaviors generate no additional agricultural output while consuming enormous amounts of energy. Under the worst conditions simulated in the study, this hidden overhead consumed up to 48.13 percent of total information and communication technology energy on a modeled mid-scale farm.

The scale of the problem emerges from the architecture of modern agricultural IoT. Data from a field sensor typically traverse five to ten hops, moving through gateways and edge computing units before reaching cloud platforms where irrigation schedules and pest warnings are computed. Along the way, data are frequently processed, retransmitted, and replicated. Conventional life cycle assessment models attribute energy fluctuations to workload variation and hardware efficiency, assuming data remain trustworthy throughout their journey. The authors call the prevailing regime a blind forwarding, delayed correction paradigm: a data flow that passes through a faulty or compromised node continues to propagate contaminated values downstream, and only when terminal validation finally triggers does the system perform a full-path rollback and recomputation, with penalty costs that can reach ten times the baseline per hop.

To break that cascade, the team decomposed the entire farm infrastructure into what they term Data Trust Units, or DTUs. Each node, whether a soil probe, a gateway, an edge server, or a cloud platform, is abstracted as a unit whose trust state can be independently measured and dynamically updated across four dimensions: reliability, security, recoverability, and integrity. Reliability is quantified through a context-aware Markov chain in which nodes move among operational, degraded, failed, and repair states, with transition probabilities that shift as system load rises, capturing phenomena such as increased sensor failure during extreme weather. A complementary non-homogeneous Poisson process models how failure rates spike under event-driven triggers, environmental complexity, and peak loads. Security is scored through behavioral entropy, which flags dispersed and uncertain access patterns, and through Bayesian attack graphs that estimate how risk propagates across multi-step intrusion paths. Recoverability is measured by the shape of the recovery trajectory itself, not just the time to repair, while integrity relies on structural hash comparisons across multiple data versions and replicas.

What makes the framework distinctive is that trust is treated not as a passive score but as an active control variable. Before each forwarding operation, the system evaluates the instantaneous trust of candidate nodes. If trust falls below a threshold, the data flow is terminated through path pruning, a pre-emptive stop-loss that incurs a small evaluation cost but eliminates the far larger cascading penalty of letting contaminated data complete its journey. The energy model captures this elasticity explicitly: at high trust, flows proceed along their original paths, while at low trust, pruning blocks high-risk propagation and allows energy consumption to converge topologically. Crucially, the framework is built on classical probabilistic tools, including Markov chains, Poisson processes, Bayesian networks, and structural hashing, rather than opaque machine learning, keeping the trust evaluation interpretable.

The experimental results reveal a striking threshold-governed transition. Simulating a 500-node directed acyclic graph calibrated against published agricultural IoT parameters, with 20 data flows per step traversing 5 to 10 hops, the researchers generated four trust environments ranging from roughly 0.85 down to 0.40. Baseline energy stayed within a narrow band of 7.9 to 8.2 kilowatt-hours, yet traditional-mode energy climbed nonlinearly from 11.10 to 15.61 kilowatt-hours as trust fell, an added increment of 7.63 kilowatt-hours, nearly equivalent to one full baseline workload cycle. The DTU mechanism, by contrast, delivered savings that grew from a negligible 1.07 percent at high trust to 6.83 percent, then 43.93 percent, and finally 48.13 percent under crisis conditions, executing more than 17,000 path prunings in the lowest-trust scenarios to intercept contaminated flows before they could trigger downstream recomputation.

A full-factorial analysis of the four trust dimensions uncovered effects that linear models would miss entirely. Reliability degradation proved the steepest single driver, raising traditional-mode energy by roughly 1,820 watt-hours, followed by security at about 860 watt-hours, with integrity and recoverability contributing milder increments. But the real surprise lay in the interactions: when reliability and recoverability degraded simultaneously, the combined energy increment far exceeded the sum of their individual contributions. The mechanism is intuitive once stated, simultaneous sensor failures across a monitoring zone combined with delayed field repairs, a common reality on geographically dispersed farms, push the system into a high-trigger-probability region where expected end-point recomputations multiply. Under the most severe multi-dimensional degradation, traditional energy climbed to 17.4 kilowatt-hours while the DTU framework held it to approximately 12.0.

Threshold tuning emerged as the decisive operational question. Scanning the trust cutoff from 0.50 to 0.75 revealed three governance zones: below roughly 0.57, stop-loss interventions rarely fire and the governance overhead goes unrecovered; between 0.58 and 0.70 lies a critical benefit zone yielding stable 5 to 7 percent net savings; and above 0.73, the system over-prunes, cutting nearly all flows so that nominal energy reductions come from shrinking the business itself rather than eliminating abnormal consumption. Sensitivity tests showed the structure is robust: with cascading penalties set at five times baseline, savings moderated to about 35 percent, and at twenty times they rose to roughly 58 percent, while the critical zone shifted only marginally. The per-hop cost of the trust evaluation itself is two orders of magnitude smaller than baseline forwarding energy, meaning the framework operates on a principle of small fixed cost and large variable gain.

To test the trust-scoring logic against real measurements, the team applied the reliability and integrity dimensions to the Intel Berkeley Research Lab sensor dataset, a public deployment of more than 50 wireless motes. Composite trust scores across the 20 most data-rich motes averaged 0.490 with a standard deviation of 0.077, spanning from 0.672 at the top to 0.365 at the bottom, and the lowest-trust mote exhibited visibly more erratic temperature readings, more frequent dropouts, and larger deviations from peer medians. Ninety percent of motes fell below the 0.60 mark, suggesting that even in a stable indoor environment a substantial fraction of nodes would sit inside the critical benefit zone. The authors are careful to note, however, that this validates the internal logic of the scoring method, not performance under real agricultural conditions, where soil moisture drift under variable rainfall, UAV sensing noise, and livestock tag irregularities introduce degradation patterns an indoor dataset cannot capture.

The implications extend well beyond accounting. As the researchers emphasize, trustworthiness is not merely a descriptive quality metric but an endogenous variable that shapes execution paths and energy consumption across agricultural IoT. Optimal thresholds and dimension weights will differ by production context: open-field cropping may tolerate lower thresholds because spatial averaging corrects occasional bad readings, greenhouse climate control demands stricter integrity safeguards, and livestock monitoring elevates security and recoverability given privacy concerns and frequent device disconnections. Before field deployment, practical barriers remain, from protocol heterogeneity across LoRaWAN, NB-IoT, and Zigbee systems to the need for adaptive thresholds that track pest outbreaks, harvest peaks, and extreme weather. But the core message stands: the energy cost of distrusting your own data may be the largest hidden line item in the digital farm’s ledger, and now, for the first time, it can be measured, located, and cut off at the source.

Subject of Research: Embedding data trustworthiness into digitization footprint life cycle assessment for smart agriculture IoT systems

Article Title: Integrating data trustworthiness into digitization footprint life cycle assessment for smart agriculture: Quantifying data credibility through data trust units

Article References: Yu, C., Zhang, F., Dan, Y., Chen, Q., Huang, Q., & Marinello, F. (2026). Integrating data trustworthiness into digitization footprint life cycle assessment for smart agriculture: Quantifying data credibility through data trust units. Smart Agricultural Technology, 15, Article 102520. https://doi.org/10.1016/j.atech.2026.102520

Image Credits: AI Generated

DOI: 10.1016/j.atech.2026.102520

Keywords: smart agriculture, data trustworthiness, digitization footprint, life cycle assessment, agricultural IoT, energy consumption, Data Trust Units, sensor drift, edge computing, path pruning, Markov chains, sustainability

The challenge has become more complicated as artificial-intelligence coding assistants have entered programming education. A student may now submit correct code after receiving suggestions from ChatGPT, GitHub Copilot, or a similar tool, while the process that produced that code remains hidden. A flawless final program does not necessarily show whether the learner understood the algorithm, struggled for half an hour, or accepted a generated solution without grasping its logic. The researchers therefore focused on the process rather than only the product. Programming environments record a continuous stream of events, including edits, compilations, executions, pauses, browser navigation, documentation searches, and interactions with course platforms. These events can reveal patterns that are invisible in the final source code. But the signals are inherently ambiguous: a pause can reflect careful planning or confusion, and frequent edits can indicate either systematic debugging or increasingly random attempts. The proposed system addresses that ambiguity by examining several dimensions of behavior at once.

The first layer is a Markov Chain, a probabilistic model that estimates how likely one observable action is to follow another. It can recognize workflow structures such as fluent editing followed by a validation compile, as well as less productive loops involving hesitant editing, repeated compilation, and long pauses. In mathematical terms, the model assigns probabilities to transitions between behavioral states, using smoothing so that rare or unseen transitions do not produce extreme conclusions. The second layer is a Hidden Markov Model, or HMM. Rather than treating cognitive condition as directly measurable, the HMM infers latent states from the observed sequence. The operational categories used in evaluation were Progressing, Hesitating, Blocked, and Confused. These labels are not diagnoses of a student’s mind; they are probabilistic summaries of behavior that can guide instructional decisions. A student classified as Hesitating might benefit from a targeted hint, while one classified as Confused may need a question that clarifies the strategy being attempted. A student identified as Blocked may require direct help with a persistent error.

The third layer is a recurrent neural network with attention. The study describes a bidirectional gated recurrent unit architecture that processes activity in both temporal directions and represents each time window using features such as typing speed, deletion ratio, pause duration, navigation density, compilation frequency, repeated errors, and code progress. Attention assigns greater weight to moments that are especially informative for the current prediction. This allows the system to connect a present difficulty with events that occurred several minutes earlier, overcoming the short memory of a basic Markov model. The final prediction combines the outputs of all three components using confidence-adaptive weights. If the transition probabilities are uncertain, the Markov contribution is reduced. If the inferred HMM state changes erratically, its influence falls. If attention is diffuse rather than concentrated on particular moments, the neural component contributes less. The result is intended to be not just a blockage score, but a record of which behavioral transitions, latent state patterns, and time points shaped the alert.

To evaluate the framework, the researchers analyzed 287,236 timestamped actions gathered from 70 first-year students enrolled in an introductory C++ course. The students had no prior programming experience and completed six exercises of increasing complexity in a standardized software environment. The analysis concentrated on 220 annotated sequences from two representative exercises. Events were converted into overlapping 30-second windows advancing in five-second steps, allowing the models to track changes during a session rather than relying only on totals such as the number of compilations. Two experienced programming instructors independently labeled a subset of the windows, reaching a Cohen’s kappa of 0.81, a measure of strong agreement. The dataset was divided using student-level five-fold cross-validation, so all sequences from a student remained in either the training or testing portion. This design reduces the risk that a model simply learns an individual student’s habits and then appears to generalize.

The results contain a notable twist. The hybrid model achieved a Macro-F1 score of approximately 90.7 percent across the four cognitive-state categories, but so did the simpler comparison models, including a Random Forest, a Markov Chain alone, an HMM alone, and a recurrent neural network with attention. A Friedman test found no statistically significant differences among the eight evaluated configurations, with a reported p-value of 0.83. The authors interpret this equivalence as evidence that the behavioral taxonomy itself is highly discriminating: once the observable categories are defined precisely, several machine-learning approaches can learn to recognize them. The hybrid architecture should therefore not be presented as a more accurate classifier. Its distinctive contribution lies elsewhere. The HMM supplies pedagogically meaningful state labels, the Markov layer exposes workflow transitions, and attention highlights relevant moments in the sequence. Together, these outputs can provide more context than a single risk label, even when the final classification accuracy is nearly identical.

Signals associated with impending blockage included progressive typing deceleration, a rising proportion of deleted characters, and lengthening pauses. In the study’s corpus, these patterns often appeared three to five minutes before a blockage was fully visible. A transition from neutral activity cycles to destructive cycles was another strong warning sign: when hesitation increased across consecutive observation windows and repetitive error attempts continued, blockage followed in 78 percent of the sequences examined. The model’s attention mechanism could emphasize earlier failed compilations or pauses, while the HMM summarized the broader trajectory from Progressing to Hesitating to Blocked. In a pilot deployment involving 12 instructors and 180 students across three institutions, 82 percent of alerts were judged accurate and actionable by instructors. The report also describes 18 percent more completed exercises, a 12 percent reduction in completion time, and final programming examination scores 6.3 percentage points higher than in control classrooms. These pilot outcomes are promising, but they should be interpreted alongside the study’s limitations and the authors’ description of the system as real-time-capable rather than fully validated in live classroom operation.

The research team emphasizes that behavioral tracking cannot reveal cognition with certainty. A student may pause because they are thinking deeply, because they are distracted, or because they have lost their strategy. The rare Confused category, representing 5.9 percent of windows, had the lowest F1 score at 79.0 percent and was frequently confused with Hesitating. Short sessions also produced more missed blockages because there was not enough time for precursor signals to accumulate. The dataset came from one institution, one introductory C++ course, and a relatively small group of students, so the thresholds may not transfer directly to other languages, teaching styles, or learners. The study also warns that attention weights show where the model focused, not necessarily what caused its decision. Any educational deployment would need strong privacy protections, informed consent, and safeguards preventing formative monitoring from becoming a grading mechanism. The authors propose testing the framework across institutions and programming languages, incorporating additional signals such as self-reports, and developing an instructor dashboard. For now, the work suggests that the most useful educational AI may not be the system that claims to know exactly why a student is struggling, but one that notices a changing pattern early, explains the evidence cautiously, and leaves the final judgment to a human teacher.

An important methodological distinction is between recognizing a labeled behavioral category and establishing that a learner is cognitively blocked. The study’s four-class taxonomy—progression, hesitation, blockage, and confusion—provides an operational language for analyzing traces, but its categories remain model-based interpretations of observable activity. This matters because the reported similarity in Macro-F1 across the tested approaches suggests that performance depends substantially on how the behavioral states are defined and represented, not only on architectural complexity. The absence of significant differences among models also cautions against treating a more elaborate system as automatically more accurate.

The hybrid design is therefore most valuable as a decision-support framework. Markov transition scores can describe local workflow changes, while the HMM offers a probabilistic account of how activity may correspond to a changing latent state. The recurrent component adds a way to connect events separated in time, and confidence-adaptive fusion can reduce the influence of a component when its evidence is unreliable. These signals could help an instructor distinguish a single unusual pause from a sustained deterioration across successive activity windows. Such distinctions are particularly relevant in programming, where debugging often involves temporary failure and repeated experimentation that should not be mistaken for learning collapse.

The reported pilot findings provide an initial indication that interpretable alerts can be linked to instructional outcomes, but they do not by themselves establish effectiveness across settings. The evaluation involved a limited number of students and instructors, and the source describes the deployment as a pilot. Future testing would need to examine whether alerts remain calibrated when students use different programming languages, development environments, or assistance tools, and whether interventions prompted by the system produce benefits beyond those attributable to increased instructor attention. It will also be important to assess how students perceive monitoring and whether uncertainty information is presented clearly enough to prevent probabilistic alerts from being treated as definitive judgments.

Subject of Research: Machine-learning detection of novice programming difficulties from fine-grained activity traces

Article Title: A multi-dimensional hybrid stochastic model for early and interpretable blockage detection in programming education

Article References: Abdelkader, G., Mohammed, E., Patrick, E., & Thierry, N. (2026). A multi-dimensional hybrid stochastic model for early and interpretable blockage detection in programming education. Discover Informatics, 1(1), Article 9. https://doi.org/10.1007/s44564-026-00007-0

Image Credits: AI Generated

DOI: 10.1007/s44564-026-00007-0

Keywords: programming education, learning analytics, educational data mining, student blockage detection, Hidden Markov models, Markov Chains, recurrent neural networks, explainable AI, multi-dimensional, hybrid, stochastic, model