Plastic inspection has emerged as an important component of industrial manufacturing processes, quality control, and recycling, driven by a growing emphasis on sustainable, circular, and efficient modes of production. This systematic narrative review focuses on three key areas: (i) a review on the imaging techniques used in the plastic industry for creating training datasets for artificial intelligence (AI) models; (ii) an evaluation of various AI approaches, including support vector machines (SVMs), deep reinforcement learning (DRL), convolutional neural networks (CNNs), and hybrid/multimodal techniques; and (iii) the integration of these techniques within robotic cyber-physical systems (CPS) for the automation of plastic defect identification, material classification, and sorting for recycling/remanufacturing, supplemented by circular business aspects. CNNs demonstrate exceptional performance in feature extraction and in detecting surface defects, such as scratches, cracks, and inconsistencies in plastic materials. SVMs, with their robustness to small, noisy datasets, provide accurate classification and quality control, making them a valuable complement to CNNs. Hybrid approaches that combine CNNs and SVMs leverage the strengths of both methods for complex tasks, thereby maximizing the advantages of each. DRL enhances the inspection and sorting capabilities of robotic CPS when integrated together. Despite these advancements, challenges remain, including high resource costs, data-intensive requirements, and constraints on real-time implementation. Potential solutions include adopting efficient architectures and lightweight frameworks. A pilot application of these AI strategies within a robotic CPS demonstrates their transformative potential to automate large-scale remanufacturing and recycling systems efficiently, accurately, and in an eco-friendly manner, supporting circular economy principles and sustainability.

Artificial intelligence (AI) is becoming deeply integrated into additive manufacturing (AM) workflows, reshaping how designers approach geometry, materials, and process constraints. AI holds significant potential by accelerating design exploration, revealing complex patterns in AM behavior, and supporting earlier assessment of manufacturability. At the same time, it introduces new risks related to model transparency, data quality, physical validity, and the potential for overreliance by students and practitioners. This perspective examines these issues through four guiding questions that address the role of AI in AM-enabled design, the gaps that limit or enable AI contribution, the implications for engineering education, and the responsibilities of the research community in ensuring trustworthy and secure AI-AM integration. The main contributions of this perspective include: (i) Highlighting AI and AM as a coupled inference-fabrication system rather than independent tools; (ii) identifying zones of strong interdependence where inference and manufacturability interact; and (iii) articulating implications for design reasoning, education, and responsible research practice.

The design of alloys and their manufacturing processes requires extensive exploration of a broad design space comprising various compositional and processing variables, many of which remain inadequately explored in practice. The existence of multiple viable processing routes for achieving desired alloy properties further complicates the design process. This paper presents a multi-agent deep reinforcement learning (DRL) framework for the in silico design of alloys and their processing routes/conditions tailored to specific property targets. The framework consists of distinct decentralized DRL agents, each responsible for making decisions regarding composition selection and the individual manufacturing steps involved in the process. These agents interact with their respective environments, which represent the assigned processes, and share responsibilities related to both process-specific outcomes and overall property satisfaction, as governed by the reward functions. The reward functions integrate considerations of sustainability, cost, and manufacturability into the decision-making process. A generative design step is proposed to leverage the capabilities of the trained DRL agents to produce multiple design alternatives for a given requirement. The framework is applied to the design of a hot-rolled steel sheet, exploring two feasible processing routes: Conventional casting and thin slab casting, resulting in several alternatives for each route. The framework’s performance is evaluated on two experimental cases from the literature, indicating its success in biasing the sample toward the preferred solution space. A benchmark study is conducted to evaluate the framework’s performance against designs produced by materials engineers for three distinct use cases, demonstrating the superior performance of the proposed framework.

Optimizing perovskite solar cells (PSCs) requires precise control of solution chemistry and functional additives. However, limited experimental data hinder systematic discovery. Here, we integrate 1,540 carefully selected experimental device records with 4,000 synthetic data points generated by a beta-variational autoencoder to investigate solution parameters and organic additives governing device performance. A residual neural network trained on this hybrid dataset achieves strong predictive accuracy with an R² of 0.87 for power conversion efficiency. Even when trained solely on synthetic data, the model attains an R² of 0.785. Within this framework, 733 organic additives with diverse functional groups were evaluated to identify molecular features that enhance absorber quality. High-efficiency PSCs are associated with solution concentrations above 1.3 molar and elevated formamidinium iodide (FAI) ratios, in combination with additives containing benzene rings, methylene, and amine groups. Notably, a composition comprising FAI (1.05), cesium iodide (0.03), methylammonium chloride (0.3), lead(II) iodide (1.5), and a molybdenum trioxide interlayer, combined with 1,3-dihydro-1-[1-(phenylmethyl)-4-piperidinyl]-2Hbenzimidazol-2-one as an additive, yields a PCE of 25.66%. This additive was absent from the training data, demonstrating the capability of the proposed framework to discover novel and effective organic additives for PSC optimization.

Traditional bone-implant lattices are commonly based on periodic unit cells, which simplify design and fabrication but limit geometric diversity and the fine, directional tuning of effective stiffness. Herein, a data-driven framework for multicell tessellations inspired by Euclidean-tiling (ET) is proposed to balance geometric freedom with controllable in-plane elastic modulus. Using three tiling-compatible unit cells (tris, quad, and oct), 10,000 randomly assembled 5 × 5 tiling structures were generated to construct a finite element database of structure-property relationships targeting the in-plane equivalent modulus (E_x and E_y). A fully connected neural network (FCNN) was trained with two input feature representations, namely a unit cell arrangement encoding (UCAE) and a unit cell frequency statistic (UCFS). As the FCNN input, the UCAE consistently outperformed the UCFS, achieving R² values of approximately 0.99 for both E_x and E_y, with a mean absolute error of about 1.4 MPa. Based on the FE database, a k-nearest neighbors strategy was applied to retrieve the five structures most closely matching the target modulus. The selected designs were subsequently fabricated by high-precision 3D printing. The elastic modulus values obtained from machine learning prediction, finite element analysis, and experimental testing showed close agreement. Overall, this framework enables rapid prediction and inverse screening of elastic modulus in ET-based structures with high degrees of freedom.