Microfluidic technology, which involves the manipulation of fluids in microchannels, faces challenges in channel design and performance optimization due to its complex, multi-parameter nature. Traditional design and optimization approaches usually rely on time-consuming numerical simulations, or on trial-and-error methods, which entail high costs associated with experimental evaluations. Additionally, commonly used optimization methods require many numerical simulations, and to avoid excessive computation time, they approximate simulation results with faster surrogate models. Alternatively, machine learning (ML) is becoming increasingly significant in microfluidics and technology in general, enabling advancements in data analysis, automation, and system optimization. Among ML methods, Bayesian optimization (BO) stands out by systematically exploring the design space, usually using Gaussian processes (GP) to model the objective function and guide the search for optimal designs. In this paper, we demonstrate the application of BO in the design optimization of the microfluidic systems, by enhancing the mixing performance of a micromixer with parallelogram barriers and a Tesla micromixer modified with parallelogram barriers. Micromixer models were made using Comsol Multiphysics software® and their geometric parameters were optimized using BO. The presented approach minimizes the number of required simulations to reach the optimal design, thus eliminating the need for developing a separate surrogate model for approximation of the simulation results. The results showed the effectiveness of using BO for design optimization, both in terms of the execution speed and reaching the optimum of the objective function. The optimal geometries for efficient mixing were achieved at least an order of magnitude faster compared to state-of-the-art optimization methods for microfluidic design. In addition, the presented approach can be widely applied to other microfluidic devices, such as droplet generators, particle separators, etc.

Thermal power plants are essential for maintaining a stable electricity supply but face high operational costs and potential economic losses due to unexpected equipment failures. Predictive maintenance (PM) aims to reduce these risks by detecting early anomalies, but existing anomaly detection methods lack statistical control over the false positive rate (FPR), leading to frequent false alarms and inefficient maintenance interventions. We hypothesize that a conformal anomaly detection (CAD)-based approach, incorporating calibration-conditional p-value adjustments, can effectively mitigate the issue of uncontrolled FPR by ensuring statistically valid anomaly detection for power plant PM, even when multiple dependent hypothesis tests are performed on operational data. We introduce a CAD framework for power plant PM, which integrates calibration-conditional validity adjustments to systematically control FPR while maintaining sensitivity to operational faults. The evaluation results of the proposed CAD framework for PM in thermal power plants show that it ensures statistically valid FPR control, significantly reducing false alarms while maintaining high sensitivity to faults and strong discriminative ability between normal and faulty operations. By providing automatic and statistically valid FPR control without manual threshold tuning, the proposed framework offers a robust and scalable solution for anomaly detection in thermal power plant PM, improving maintenance efficiency and reducing unnecessary interventions.

This paper introduces a novel two-stage active learning (AL) pipeline for automatic speech recognition (ASR), combining unsupervised and supervised AL methods. The first stage utilizes unsupervised AL by using x-vectors clustering for diverse sample selection from unlabeled speech data, thus establishing a robust initial dataset for the subsequent supervised AL. The second stage incorporates a supervised AL strategy, with a batch AL method specifically developed for ASR, aimed at selecting diverse and informative batches of samples. Here, sample diversity is also achieved using x-vectors clustering, while the most informative samples are identified using a Bayesian AL method tailored for ASR with an adaptation of Monte Carlo dropout to approximate Bayesian inference. This approach enables precise uncertainty estimation, thereby enhancing ASR model training with significantly reduced data requirements. Our method has shown superior performance compared to competing methods on homogeneous, heterogeneous, and OOD test sets, demonstrating that strategic sample selection and innovative Bayesian modeling can substantially optimize both labeling effort and data utilization in deep learning-based ASR applications.

The goal of the state estimation (SE) algorithm is to estimate complex bus voltages as state variables based on the available set of measurements in the power system. Because phasor measurement units (PMUs) are increasingly being used in transmission power systems, there is a need for a fast SE solver that can take advantage of high sampling rates of PMUs. This paper proposes training a graph neural network (GNN) to learn the estimates given the PMU voltage and current measurements as inputs, with the intent of obtaining fast and accurate predictions during the evaluation phase. GNN is trained using synthetic datasets, created by randomly sampling sets of measurements in the power system and labelling them with a solution obtained using a linear SE with PMUs solver. The presented results display the accuracy of GNN predictions in various test scenarios and tackle the sensitivity of the predictions to the missing input data.