这篇文章是部分 用于应用材料设计和开发的人工智能/机器学习 特刊。机器学习 (ML) 和其他人工智能 (AI) 方法必将改变科学和技术，因为这已经发生在人类活动的许多其他领域。机器学习在化学、材料科学及其应用中的应用已大幅增长。自 2019 年以来，关于这些主题的所有文章中有 60% 以上已发表。 除了对用于各种类型诊断的图像、传感和生物传感数据进行分析之外，最近的研究主要集中在材料设计和发现上。由于材料基因组计划和材料特性的大型数据库的可用性，在某些情况下，通过高通量实验和理论方法，已经观察到了提升。这并不奇怪，由于基本原理完全通用，因此已对各种材料进行了这些程序的应用。这个关于应用材料设计和开发的人工智能/机器学习论坛反映了这种多样性。15+篇论文集涵盖人工智能（AI）/机器学习（ML）应用于电热陶瓷和压电陶瓷、玻璃、热电材料、电池电极材料、介电和半导体聚合物、聚合物纳米复合材料、铜基催化剂、金属有机框架、熔盐和二维材料。尽管材料和预期应用各不相同，但报告的研究共享了一个整体的主题方法：使用 ML 算法的模式识别策略使用从实验结果和/或计算机模拟生成的大型数据库对材料（或材料特性）进行分类。论坛中的论文重点介绍了 AI/ML 在预测、筛选和优化属性和/或材料方面的能力。例子包括绝热温度变化的预测、高热导率、玻璃中的硫溶解度、新型二维材料、导电聚合物、热电材料和锂硫电池材料。此外，还对电池电极材料和金属有机骨架进行了筛选，并针对 3D 打印优化了墨水书写条件。论坛中展示的结果和应用说明了 ML 的优势和局限性。具体来说，它被用来解决长期存在的科学问题或促进数据采集和解释，否则这些问题将变得乏味、繁琐或完全无法管理。正如论坛中的一些论文和文献中所讨论的那样，发现和设计新材料或特性的预测能力取决于所用数据集的广度和覆盖范围。换句话说，机器学习的有效性和有用性取决于数据集的内在质量和范围。如果有足够的质量数据可用于训练，监督式机器学习适用于分类任务，在许多应用中它已经可以胜过人类。典型的例子是基于图像分析的面部识别和诊断，其中高精度，有时接近 100%，如果数据集足够全面并且有必要的计算能力，则可以获得。然而，对于需要相当程度的解释的任务，目前的技术并没有使 ML 高效。此限制适用于 ML 或 AI 的任何应用程序；在材料发现和设计的背景下，这意味着必须确定精确的科学或技术问题，可以从所分析数据的模式中推断出其响应。例如，今天有可能优化实现高性能材料的条件，但开发基于 ML 的策略来学习物理基本定律和此类材料的潜在功能机制是一项更具挑战性的挑战。人工智能对化学和材料科学的影响主要与机器学习有关。在人工智能的其他子领域，例如神经网络、专家系统、进化计算、机器人技术、计算机视觉、语音和自然语言处理以及规划，几乎没有做过什么。例外情况涉及在深度学习和其他机器学习方法中使用遗传算法和神经网络，以及在知识发现中使用自然语言处理，其中挖掘特定类别材料的文献。尽管人们认为知识发现是人工智能应用的核心，但由于机器无法解释文本，迄今为止的结果有限。对于短期内与 ML 相关的应用程序，我们相信加强计划以生成和管理具有材料特性的大型数据库。尤其是——但不限于——催化、能量存储与转换、传感、生物传感等领域，仍然缺乏大型、准确、包容的数据库。这些进展可以与材料基因组学项目和现有医学图像数据库的最新进展相结合来实现。可靠、全面的数据库的建立将使机器学习扩展到更广泛的应用领域成为可能。论坛中涵盖的主题都与使用机器学习帮助解决材料及其相关应用的特定问题有关。然而，有一个整个发展领域的重点恰恰相反，即使用材料使能人工智能应用。的确，虚拟现实环境和物联网依赖于各种可穿戴设备，从传感器和生物传感器到带有嵌入式电子电路的执行器和仪器。此外，机器人技术（尤其是软机器人）需要另一套基于纳米技术的设备，例如用于人工视觉和模仿人类感官。一旦具有应变传感能力的可穿戴设备达到足够的性能，语音处理和语音识别可能会发生革命性的变化。具有记忆效应的定制材料已经用于进化计算的“原理证明”实验中，人们可以设想通过硬件进行机器学习。尽管很少提及，但当今所有 ML 应用程序都是通过软件执行的。当机器学习用硬件有效地完成时，这可能需要几十年的时间，人工智能将彻底转变，使其可以成功用于设计创新材料类别，不仅可以满足目标需求，还可以解决与开发、理解和优化特定应用相关的细微差别。然而，只有在计算和材料科学方面取得相应的重大发展，才有可能取得这种突破性进展。这篇文章还没有被其他出版物引用。 This article is part of the Artificial Intelligence/Machine Learning for Design and Development of Applied Materials special issue. Machine learning (ML) and other artificial intelligence (AI) methods are bound to transform science and technology, as this is already happening in many other areas of human activity. The use of machine learning in chemistry, materials sciences, and their applications has grown enormously. More than 60% of all articles on these topics have been published since 2019. Recent research has been focused mostly on materials design and discovery, in addition to the analysis of images, sensing, and biosensing data for various types of diagnosis. A boost has been observed as a result of materials genome initiatives and the availability of large databases of materials properties made possible, in some cases, with high-throughput experimental and theoretical methods. Not surprisingly, applications of these procedures have been carried out for a wide variety of materials, as the underlying principles are entirely generic. This forum on Artificial Intelligence/Machine Learning for Design and Development of Applied Materials reflects this diversity. The collection of 15+ papers covers artificial intelligence (AI)/machine learning (ML) applied to research on electrocaloric ceramics and piezoceramics, glasses, thermoelectric materials, electrode materials for batteries, dielectric and semiconducting polymers, polymer nanocomposites, copper-based catalysts, metal–organic frameworks, molten salts, and 2D materials. Despite the diversity in materials and intended applications, the studies reported share an overall thematic approach: pattern recognition strategies with ML algorithms are applied to classify materials (or materials properties) using large databases generated from experimental results and/or computer simulations. The papers in the forum highlight the capabillity of AI/ML to enable the prediction, screening, and optimization of properties and/or materials. Examples include prediction of adiabatic temperature changes, high thermal conductivity, sulfur solubility in glasses, new 2D materials, conducting polymers, thermoelectric materials, and materials for Li–S batteries. Moreover, screening has been performed on battery electrode materials and metal–organic frameworks, and ink writing conditions were optimized for 3D printing. The results and applications showcased in the forum are illustrative of the strengths and limitations of ML. Specifically, it is being used to address either long-standing scientific issues or to facilitate data acquisition and interpretation that otherwise would be tedious, cumbersome, or unmanageable altogether. As discussed in some of the papers in the forum, and in the literature, the predictive power in discovering and designing new materials or properties depends on the breadth and coverage of the data sets employed. In other words, the effectiveness and usefulness of ML are dependent upon the intrinsic quality and scope of the data sets. If sufficient quality data is available for training, supervised ML is suitable for classification tasks, where it can already outperform humans in many applications. Archetypical examples are facial recognition and diagnosis based on image analysis, in which a high accuracy, sometimes bordering on 100%, may be obtained provided that the data set is sufficiently comprehensive and the necessary computational power is available. However, present technology does not render ML efficient for tasks that require a considerable degree of interpretation. This limitation applies to any application of ML or AI; in the context of materials discovery and design, it means that precise scientific or technological questions must be identified, whose response can be inferred from patterns in the data under analysis. For instance, today it may be possible to optimize conditions for achieving high-performance materials, but developing ML-based strategies to learn the fundamental laws of physics and the underlying functioning mechanisms of such materials is by far a more ambitious challenge. The impact of AI on chemistry and materials science has been mostly associated with ML. Little has been done in other subareas of AI, e.g., neural networks, expert systems, evolutionary computation, robotics, computer vision, speech and natural language processing, and planning. The exceptions involve the use of genetic algorithms and neural networks within deep learning and other ML approaches, and natural language processing in knowledge discovery, where the literature in specific classes of materials was mined. Though knowledge discovery is believed to be central for AI applications, the results so far have been limited, because of the inability of machines to interpret text. For ML-related applications in the short term future, we believe in the strengthening of initiatives to generate and curate large databases with materials properties. In particular, large, accurate, and inclusive databases are still lacking in fields such as─but not limited to─catalysis, energy storage and conversion, sensing, and biosensing. These advances can be achieved in conjunction with recent progress made with the materials genomics projects and the databases for medical images already available. The establishment of reliable, comprehensive databases will make it possible to extend ML to an even broader range of applications. The topics covered in the forum are all related to using ML to assist in solving particular problems of materials and their associated applications. However, there is a whole developing field in which the focus is exactly the opposite, i.e., using materials to enable AI applications. Indeed, virtual reality environments and the Internet of Things rely on various kinds of wearable devices, from sensors and biosensors to actuators and instruments with embedded electronic circuits. Furthermore, robotics (especially with soft robots) requires another set of nanotech-based devices, e.g., for artificial vision and in mimicking human senses. Speech processing and voice recognition may be revolutionized once wearable devices with strain sensing capability reach a sufficient performance. Tailored materials with memory effects are already used in “proof-of-principle” experiments for evolutionary computation, and one may envisage machine learning via hardware. Though seldom mentioned, all of the ML applications today are performed with software. When ML is effectively done with hardware, which may take decades, AI will be transformed entirely such that it can be used successfully in designing innovative classes of materials not only to tackle targeted needs but also address nuances associated with developing, understanding, and optimizing specific applications. Such breakthrough advances will, however, only be possible with correspondingly significant developments in computational and materials sciences. This article has not yet been cited by other publications.