Sasidhar Kasturi Narasimha, Siboni Nima Hamidi, Mianroodi Jaber Rezaei, Rohwerder Michael, Neugebauer Jörg, Raabe Dierk
Max-Planck-Institut für Eisenforschung GmbH, Max-Planck Straße-1, 40237 Düsseldorf, Germany.
Ergodic Labs, Lohmühlenstraße 65, 12435 Berlin, Germany.
Sci Adv. 2023 Aug 11;9(32):eadg7992. doi: 10.1126/sciadv.adg7992.
We propose strategies that couple natural language processing with deep learning to enhance machine capability for corrosion-resistant alloy design. First, accuracy of machine learning models for materials datasets is often limited by their inability to incorporate textual data. Manual extraction of numerical parameters from descriptions of alloy processing or experimental methodology inevitably leads to a reduction in information density. To overcome this, we have developed a fully automated natural language processing approach to transform textual data into a form compatible for feeding into a deep neural network. This approach has resulted in a pitting potential prediction accuracy substantially beyond state of the art. Second, we have implemented a deep learning model with a transformed-input feature space, consisting of a set of elemental physical/chemical property-based numerical descriptors of alloys replacing alloy compositions. This helped identification of those descriptors that are most critical toward enhancing their pitting potential. In particular, configurational entropy, atomic packing efficiency, local electronegativity differences, and atomic radii differences proved to be the most critical.
我们提出了将自然语言处理与深度学习相结合的策略，以增强耐腐蚀合金设计的机器能力。首先，用于材料数据集的机器学习模型的准确性常常受到其无法纳入文本数据的限制。从合金加工描述或实验方法中手动提取数值参数不可避免地会导致信息密度降低。为了克服这一问题，我们开发了一种全自动自然语言处理方法，将文本数据转换为适合输入深度神经网络的形式。这种方法使点蚀电位预测准确率大幅超越现有技术水平。其次，我们实施了一个具有变换输入特征空间的深度学习模型，该特征空间由一组基于合金元素物理/化学性质的数值描述符组成，取代了合金成分。这有助于识别那些对提高其点蚀电位最为关键的描述符。特别是，组态熵、原子堆积效率、局部电负性差异和原子半径差异被证明是最关键的。
