An integrated approach for linking materials science and engineering, digitalization, and artificial intelligence enabled by semantic technologies

The sustainable and environmentally friendly manufacturing of materials and their transformation into products are among the major challenges currently facing industry in the pursuit of a greener society. In materials science and mechanics of materials, the past decades have been dominated by advances in modeling and simulation as well as characterization techniques. The determination of process–structure–property relationships plays an increasingly im- portant role. Notably, the information about the chemical composition of materials has been identified as a pivotal factor, with the potential to markedly accelerate the design of materials, processes, and products while substantially reducing experimental requirements. This article advocates a paradigm shift towards an integrated approach underpinned by the Integrated Computational Materials Engineering (ICME) framework, the semantic technology–driven digital transformation of heterogeneous data into knowledge, and the strategic application of artificial intelligence to accelerate computational workflows and support materials and process design decisions. Central to this paradigm shift are FAIR (Findable, Accessible, Interoperable, Reusable) and linked data, together with the application of machine learning to materials and process design challenges. The article demonstrates how these three central elements can be linked to accelerate the solution of materials science and engineering challenges, in particular for materials and process design. As a case study, we demonstrate how semantic technologies provide the integration and provenance layer in a parameter identification task for a material model capturing the pseudoelastic behavior of shape memory alloys, while machine learning performs the inverse mapping from experimental stress histories to nine material parameters. The quantitative evaluation shows that suitable machine-learning configurations reproduce the experimental stress response with coefficients of determination of approximately R 2 = 0.99 , while the identifiability of the individual constitutive parameters remains parameter dependent and benefits from additional loading–unloading information. Once trained, the inverse model replaces the repeated iterative solution of the parameter identification problem by a single neural-network evaluation, providing the basis for highly automated recurring calibration workflows in the semantic dataspace. Comparison with classical optimization further shows that iterative optimization can improve the stress-response fit for an individual experiment, while the machine-learning prediction provides an effective initial estimate for subsequent local optimization.