AI & Materials
ISSN: 3006-7588 (Print)
ISSN: 3006-7596 (Online)
CODEN: AMABGR
High-precision automated nanoparticle segmentation using a deep learning framework with boundary-aware and attention networks
Accurate characterization of nanoparticle geometry and morphology is essential for understanding their structure–property relationships. However, in most electron microscopy images, nanoparticles are densely distributed and are often affected by strong background noise and particle overlap, making conventional manual analysis time-consuming and inefficient. To address this issue, this study proposes Nanoparticle Segmentation You Only Look Once (NSYOLO), an enhanced deep learning–based instance segmentation framework for the automatic and high-precision recognition and segmentation of nanoparticles in electron microscopy images. The framework is trained on a multi-type dataset comprising nanocubes, nanospheres, and nanorods, and introduces a boundary-aware dynamic snake convolution (BADSConv) module to enhance boundary feature representation, along with a bi-level routing attention (BRA) mechanism to improve global feature modeling. Experimental results demonstrate that NSYOLO increases mean Average Precision (mAP)@0.5 from 0.906 to 0.957 and outperforms open-source automated tools, such as ImageJ and ImageDataExtractor, particularly in images with complex backgrounds and overlapping particles. In addition, the NSYOLO-based analysis system is developed to enable automated nanoparticle segmentation, size statistics, and the generation of editable Word reports without requiring any programming experience, thereby providing an efficient, reliable, and user-friendly solution for high-throughput nanoparticle morphology analysis.
Towards signal-based estimation of contact tip–to–workpiece distance (CTWD) in wire arc directed energy deposition via 1D convolutional deep learning
In open-loop Wire Arc Directed Energy Deposition (WA-DED), the contact tip–to–workpiece distance (CTWD) is commonly adjusted between layers using expected average values of layer height derived from experimental data. However, due to complex thermal effects, mismatches between expected and actual process conditions lead to uncontrolled CTWD fluctuations, which can cause process instabilities, arc extinction events, and geometric defects. Reliable online estimation of CTWD is therefore essential for process monitoring and for enabling future closed-loop control strategies. Existing CTWD monitoring approaches typically rely on vision-based or acoustic emission systems, which are costly or sensitive to industrial noise. This work proposes a data-driven method for online CTWD estimation using only welding current and voltage, signals already available in industrial WA-DED systems. By avoiding additional sensors, the approach enables low-cost, robust, and real-time CTWD estimation suitable for direct industrial deployment. Several machine learning and deep learning models are evaluated, with particular focus on Ridge Regression, Support Vector Regression and a 1-dimensional convolutional neural network. Experimental results show that the deep learning approach provides higher estimation accuracy and robustness compared to conventional machine learning methods, while remaining suitable for real-time implementation. The proposed method offers a practical solution for CTWD monitoring in WA-DED and represents a step towards intelligent process supervision and control in wire-based additive manufacturing.
Advances in adsorption processes driven by machine learning
Machine learning, as an advanced data processing method, has become one of the key technologies in the research of adsorption processes due to its outstanding nonlinear modeling capabilities. Its significance lies in that machine learning not only can accurately predict the adsorption process but also plays an important role in the selection and optimization of material synthesis pathways. Currently, research in the field of adsorption mainly focuses on the design of adsorbents, the optimization of adsorption processes, and the development of reactors. This paper systematically classifies the role of artificial intelligence algorithms in adsorption research and reviews the specific applications of these algorithms in the adsorption process, including the screening and design of adsorbents, the prediction and modeling of adsorption parameters, and the design and manufacture of reactors. Machine learning models are classified according to different application scenarios, covering various algorithms such as data modeling, image processing, and sequence analysis. At the same time, this paper also emphasizes the progress made in developing interpretable models for adsorption processes. Finally, the paper discusses the future potential and challenges of artificial intelligence in the field of adsorption.
Research progress on SiCf/SiC composites and their cladding components
Owing to their outstanding high-temperature resistance, oxidation resistance, radiation tolerance, and corrosion resistance, silicon carbide fiber-reinforced silicon carbide (SiCf/SiC) composites have shown extensive potential in advanced applications, including aerospace and nuclear industries. SiCf/SiC composites are considered among the most promising accident-tolerant fuel cladding materials, especially in the context of fourth-generation fission reactor development, owing to their stability under extreme conditions. However, the complex processing and structural characteristics of the material leave room for further research, especially with the emergence of new technologies like artificial intelligence (AI). Therefore, this work will provide a review of various processes, including chemical vapor infiltration (CVI), polymer infiltration and pyrolysis (PIP), nano impregnation and transient eutectic method (NITE), and reactive melt infiltration (RMI), focusing on improving material density, mechanical properties, and irradiation stability. Additionally, an in-depth review of the mechanical properties and microstructural changes of SiCf/SiC composites and their cladding components under extreme conditions, such as high temperatures, irradiation, and corrosion, is provided, as these factors directly affect their long-term stability in nuclear reactors. Notably, numerical simulation technology has become a crucial tool for predicting the service performance of materials. Integrating advanced technologies like AI is expected to further promote the application of SiCf/SiC composites in future high-temperature structural materials. In summary, significant progress has been made in the study of SiCf/SiC composites as next-generation nuclear fuel cladding materials. However, further research is needed in areas such as fabrication process optimization, interface modification, service behavior evaluation, and integration with AI to meet the higher performance demands of future nuclear energy systems.
Applications of artificial intelligence in materials research for fuel cells
The increasing global energy demand and the growing environmental problems have intensified the pursuit of clean and sustainable energy solutions. Hydrogen, with its high energy density and clean by-products, is a promising candidate as an energy source. Fuel cells play a key role in harnessing hydrogen energy, but this technology faces challenges such as the trade-off between material stability and ion conductivity, which limits its widespread application. To address these challenges, designing material properties and adjusting system parameters are highly desirable. However, the traditional trial-and-error approach is no longer feasible when dealing with the vast array of possibilities. Fortunately, the advancement of artificial intelligence (AI) offers a new approach which can dramatically speed up the material design and parameter control. This article reviews the application of AI in fuel cells, especially its ability to accelerate material development. The review begins by outlining the mechanisms and classifications of fuel cells, as well as the property requirements for each part of the fuel cells. Subsequently, the article introduces the basic concepts of AI and its application in materials science, including the workflows of data aggregation, feature construction, model training, and experimental validation. Importantly, the applications of AI in predicting fuel cell material performance are highly emphasized and discussed. In addition, the challenges encountered in AI applications are introduced, including sparse datasets, complex feature engineering, the limitations of general models, and the weak interpretability of AI models, along with their respective development blueprints.
Materials discovery through reinforcement learning: a comprehensive review
Reinforcement learning (RL) is emerging as a powerful tool in materials science, delivering a paradigm shift in how we find and optimize high-dimensional chemical and structural spaces. Unlike traditional methods, RL agents are able to learn to explore complex energy landscapes in an adaptive manner, instantaneously making decisions that guide the discovery of novel materials with certain properties. However, the application of RL to materials discovery faces unique challenges, including data scarcity, computationally expensive, and the challenge of designing reward functions that can balance multiple material objectives optimally. In this review, the current challenges and difficulties in applying RLtechniques in materials science and recent advances combining RL with machine learning, generative models, and domain knowledge are emphasized. We also outline promising future directions, such as transfer learning, hybrid models, and the creation of collaborative, open-access data infrastructures. By addressing these challenges, RL has the potential to transform the discovery and design of functional materials for catalysis, energy storage, and sustainability applications.
