The control of ventilation systems is often performed by automatic algorithms which often do not consider the future evolution of the system in its control politics. Digital twins allow system forecasting for a more sophisticated control. This paper explores a novel methodology to create a Machine Learning (ML) model for the predictive control of a ventilation system combining Computational Fluid Dynamics (CFD) with Artificial Intelligence (AI). This predictive model was created to forecast the temperature and humidity evolution of a ventilated room to be implemented in a digital twin for better unsupervised control strategies. To replicate the full range of annual conditions, a series of CFD simulations were configured and executed based on seasonal data collected by sensors positioned inside and outside the room. These simulations generated a dataset used to develop the predictive model, which was based on a Deep Neural Network (DNN) with fully connected layers. The model’s performance was evaluated, yielding final average absolute errors of 0.34 degrees Kelvin for temperature and 2.2 percentage points for relative humidity. The presented results highlight the potential of this methodology to create AI-driven digital twins for the control of room ventilation.

Forecasting influent flow in Wastewater Treatment Plants (WWTPs) is critical for managing operational risks during flash floods, especially in Spain’s Mediterranean coastal regions. These facilities, essential for public health and environmental protection, are vulnerable to abrupt inflow surges caused by heavy rainfall. This study proposes a data-driven approach combining historical flow and rainfall data to predict short-term inflow dynamics. Several models were evaluated, including Random Forest, XGBoost, CatBoost, and LSTM, using metrics such as Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and R-squared (R²). XGBoost outperformed the others, particularly under severe class imbalance, with only 1% of the data representing rainfall events. Hyperparameter tuning and input window size analysis revealed that accurate predictions are achievable with just 14 days of training data from a 10-year (2012–2022) dataset sourced from a single WWTP and on-site weather station. The proposed framework supports proactive WWTP management during extreme weather events.

Airborne dust emissions from bulk cargo handling in port terminals can degrade local air quality, but traditional dispersion models are often too slow or coarse to support rapid operational decisions. There is thus a pressing need for efficient tools that retain the spatial detail of CFD while enabling near-real-time scenario evaluation. In this work, we develop and test a hybrid framework that couples an RANS-based CFD model of dust dispersion with a neural network surrogate to rapidly predict exposure patterns for a bulk terminal under variable wind and operational conditions. The ML surrogate model, based on a decoder-style Multilayer Perceptron (MLP) architecture, processes two-dimensional slices of dispersion fields across particle diameter classes, enabling predictions in milliseconds with an acceleration factor of approximately $8\times {10}^6$ over traditional CFD while preserving high fidelity, as validated by performance metrics such as the F₁ score and precision values exceeding $0.8$ and $0.76$, respectively. This approach not only addresses computational inefficiencies but also lays the groundwork for real-time air-quality monitoring and sustainable urban planning, potentially integrating with digital twins fed by live weather data.

Accurate binarization of phase-detection probe signals (gas vs. liquid) is necessary for the estimation of local void fraction, interfacial velocity, and bubble statistics in gas–liquid flows. Classical threshold based methods—single or double level— performs well on clean laboratory signals but degrades under realistic industrial conditions where noise, baseline drift, and clustered (slug-like) events challenge fixed rules. This work investigates whether deep learning (DL) models trained exclusively on synthetic data can deliver robust, generalizable binarization on real probe measurements. We (i) build a parametric generator of realistic time series from bubbly pulse templates, extended to clusters/slug patterns and perturbed with controlled noise, drift, and oscillatory baselines; (ii) train four lightweight DL architectures—one-dimensional U-Net (UNET-1D), Temporal Convolutional Network (TCN), a minimal one-dimensional Convolutional Neural Network (CNN-1D), and a Bidirectional Long-Short Memory network (BiLSTM)—only on synthetic signals; and (iii) evaluate them against classical threshold methods using event-level and sample-level metrics. On synthetic signal evaluation, UNET-1D and TCN achieve near-perfect event detection and sub-millisecond onset errors. On real bubbly and slug flow sensor data, classical threshold based methods remain highly competitive on clean sensor signals, while DL models retain advantages under non-stationary baselines and clustered events, yielding accurate void and timing with no hand-tuned assumptions. Results support DL as a practical, data-driven complement to fixed algorithms, particularly in noisy or drift-dominated measuring conditions.

Airborne particulate emissions originating from bulk-material handling operations constitute an increasingly critical environmental and public health issue in port-industrial areas located near residential areas. This study introduces a novel hybrid framework integrating high-fidelity Computational Fluid Dynamics (CFD) with surrogate Machine Learning (ML) techniques for the rapid assessment of particle dispersion in port-industrial environments. Focusing on the Port of El Grao in Castellón de la Plana (Spain), the study employs detailed geometric reconstructions derived from LiDAR data and cadastral maps to build an accurate three-dimensional digital model of the area. The turbulent atmospheric boundary layer and particle dispersion dynamics were simulated using two different OpenFOAM solvers within a circular computational domain designed to reproduce realistic wind conditions. The ML surrogate model, based on a decoder-style Multilayer Perceptron (MLP) architecture, processes two-dimensional slices of dispersion fields across particle diameter classes, enabling predictions in milliseconds with an acceleration factor of approximately 8×106 over traditional CFD while preserving high fidelity, as validated by error metrics such as the F1 score and Precision values exceeding 0.8 and 0.76 respectively. This approach not only addresses computational inefficiencies but also lays the groundwork for real-time air quality monitoring and sustainable urban planning, potentially integrating with digital twins fed by live weather data.