First-principles simulations reveal the competition of defect formation and healing at grain boundaries in lead-halide perovskites. Fast halide migration at GBs mediates structural healing, but also gives rise to enhanced Frenkel defect formation.
21 January 2020 in The Journal of Physical Chemistry C
Artificial Intelligence (AI) has emerged as a transformative force in modern agriculture, revolutionizing traditional farming practices into smart agriculture ecosystems. This paper presents the ideas and uses of AI in smart agriculture, therefore highlighting its great influence on improving farming efficiency, sustainability, and production. Consisting of several layers that enable data collecting, analysis, and decision-making in farming operations, we suggest in this paper an AI-enabled Internet of Things (IoT) architecture for smart agriculture. This paper also investigates several AI-driven technologies like Machine Learning (ML), computer vision, and IoT integration, which enable farmers with real-time data insights, predictive analytics, and autonomous decision-making capability. We also look at how AI might solve important agricultural problems, including resource optimization, climate resilience, insect control, and crop monitoring. This paper clarifies the bright future of smart agriculture driven by AI in guaranteeing sustainable farming and world food security.
5 March 2021 in The Journal of Physical Chemistry C
ML models heavily rely on the quality of the data, but researchers are more focused on model optimization rather than on data refinement. The paper assesses the effect of mislabeled information on the accuracy of image classification with the Convolutional Neural Network (CNN) and Artificial Neural Network (ANN) models. Our dataset consisted of 1,000 images of plants in five classes, and we progressively mislabeled the dataset with different ratios (5%, 10%, 15%, and 20%). CNN had a 91% accuracy on clean data compared to ANN, which had 84%. The addition of only 5% mislabeling decreased CNN accuracy by 8% (to 83%) and ANN accuracy by 5% (to 79%). CNN accuracy decreased to 73% and ANN to 67% at 20% mislabeling, which indicates that CNN is more robust to label noise. The AQUAVS method of mislabeled instance detection and removal enhanced accuracy (e.g., 83% to 86% on 5% mislabeled), but did not restore baseline performance. Moreover, the recovery of the baseline accuracy involved the addition of considerable clean data: 100 images for 5% mislabeling and 200 images for 20% mislabeling. These results indicate that data quality is far more important than data quantity, and it is important to have strict data quality criteria in ML processes.