VNU University of Science

Polycyclic aromatic hydrocarbons (PAHs) are a typical group of organic pollutants and are commonly found in the soil environment. Analytical procedures for PAHs in soil samples typically consist of three main steps: sample extraction, extract clean up, and enrichment concentration prior to quantitative analysis. Ultrasonic extraction is a technique with several advantages such as simple, fast, effective, and chemical saving. In this study, the efficiency of PAHs extraction from soil samples was investigated using direct extraction (with a focused ultrasonic processor) and indirect extraction (with an ultrasonic bath) techniques with 3 different solvent systems such as acetone, dichloromethane, and acetone/hexane (1:1, v/v). The extract was purified on a glass column containing activated silica gel and anhydrous sodium sulfate with dichloromethane/hexane (1:3, v/v) as elution solvent. Concentrations of 18 PAHs in soil samples were determined by gas chromatography/mass spectrometry (GC/MS). Method detection limits of PAHs ranged from 0.20 to 3.0 ng/g. Total concentrations of 18 PAHs (Σ18PAHs) in the studied soil samples ranged from 30.8 to 55.7 ng/g and the benzo[a]pyrene equivalent (BaP-EQ) values ranged from 6.3 to 9.4 ng BaP-EQ/g. Based on Σ18PAHs and BaP-EQ values, soil extraction using ultrasonic bath with acetone provided the highest extraction efficiency. This sample extraction condition can be applied for further studies to assess the pollution levels and impacts on environmental and human health of PAHs in soil.

Reverse Osmosis (RO) degradation underscores the importance of predictive capabilities to develop optimal maintenance strategies that minimize losses. In this study, we develop a Temporal Convolutional Network (TCN) model to predict the RO system states using the primary indicator for RO analysis: the fluctuations in differential pressure across the RO vessel. Specifically, data from a real desalination plant for the period 2015 to 2020 are used. The dataset encompasses 14 RO train operations, including routine operations, significant maintenance events, temporary shutdowns, and element replacements. The proposed approach uses temporal convolutional operations to capture the dynamic pressure behavior at both ends of the membrane, enabling faster, more accurate anomaly detection. A key challenge in applying deep learning to this domain is the heavy reliance on real-world operational data. The approach involves a strong data preprocessing strategy that reveals subtle relationships between operating time and pressure dynamics. Accurate prediction of membrane degradation also ena-bles preventive and recovery actions, which reduce maintenance expenses. The proposed method is evaluated against con-ventional models, including LSTM, CNN-LSTM, and GRU, using data from the real desalination plant. Experimental results demonstrate that the proposed model achieves the lowest predic¬tion error and shows strong potential for deployment in practical desalination operations.

In the context of accelerating digital transformation in education, developing artificial intelligence (AI)competencies for both pre-service and in-service primary school teachers has become an urgent priority.This study aims to explore the necessity and effectiveness of integrating AI competency development intoprimary teacher training through Project-Based Learning (PBL). A quantitative survey using a structuredquestionnaire with a five-point Likert scale was conducted with 702 participants, including teachereducators, pre-service teachers, and school-based mentors. Data were analyzed to assess perceptions,needs, and readiness related to AI-integrated PBL. The results reveal that the vast majority of respondentshighly value the importance of enhancing teachers’ capacity to use and apply AI tools in teaching andprofessional development. Participants expressed strong expectations for diversified and practicaltraining models that combine AI applications with authentic project experiences. Based on these findings,the study proposes an optimized AI-driven PBL model that fosters teacher autonomy, real-time feedback,and adaptive learning personalization. This model emphasizes iterative project design, AI-supported datacollection and analysis, and reflective evaluation as core mechanisms for professional growth. Theresearch contributes to the theoretical framework of AI competency development in teacher education byhighlighting PBL as an effective pedagogical strategy for bridging AI theory and practice. Implicationsare drawn for curriculum innovation, teacher training policies, and further empirical studies to validatethe AI-integrated PBL model in different educational contexts. Received: 14 November 2025 / Accepted: 26 February 2026 / Published: March 2026

This study focused on the effect of a two-solvent immiscible system consisting of 0.2M H2SO4 solution and ethyl acetate - hexane mixture for rapid extraction and recovery of astaxanthin and its derivatives from Haematococcus pluvialis (H. pluvialis). The effect of changing the composition of solvent in the two-phase liquid system on the ability to extract astaxanthin was evaluated and then, the ratio of the organic solvent mixture was optimized. The results under the survey conditions showed that the maximum extraction efficiency reached 95.02 ± 1.88% with the organic solvent volume/microalgae mass ratio was 12 mL/g when hexane/ethyl acetate ratio was 1/1, showcase significantly higher than the single solvent extraction efficiency. The results of the optimal function calculation using the centered combination method gave an expected organic solvent volume/microalgae mass ratio was 9.563 mL/g with the hexane ratio in the organic solvent was 61.31%, the expected efficiency achieved was 83.26%. In addition, the study also showed that the concentration of H2SO4 only affects the process of cell wall disruption and have none affecting the dissolution process. Plus, the dissolution reaction of astaxanthin from H. pluvialis microalgae is a first-order chemical reaction.

Generalizing deepfake detection across diverse manipulation techniques remains a critical challenge. In this paper, we propose a hybrid deepfake detection framework that combines global and patch-based local modeling to improve generalization. Our architecture integrates a Global Blending Detector (GBD), trained on self-blended images to capture coarse blending artifacts from full-face inputs, with a Patch-based Artifact Detector (PAD) that focuses on five strategically selected overlapping 2×2 facial patches. These patches are designed to extract subtle and localized forgery cues, particularly in semantically significant facial regions. The PAD module leverages both a frozen CLIP-ViT encoder for semantic context and a trainable EfficientNet-B4 to capture fine-grained visual anomalies. Features from both global and local branches are concatenated and fed into a unified classifier. Extensive experiments conducted on five benchmark datasets demonstrate the effectiveness of our approach. Our method achieves an average AUC of 91.4%, with top performance on CDF-v2 (95.6%), DFDCP (93.7%), and FFIW (93.9%), outperforming or matching state-of-the-art models in cross-dataset evaluations. These results confirm that the combination of global and region-specific features significantly enhances the robustness and generalizability of deepfake detection.