1. Introduction
Our vision is to accelerate high-fidelity validation of new, more durable materials, reducing material usage, increasing resource efficiency, minimising premature failures, and lowering operational costs for high-value manufactured products.
High-temperature materials performance involves a complex interplay of factors, including creep, fatigue, oxidation, and microstructural stability [1]. This critically limits the longevity of many high-value technologies, infrastructures, and engineered products, from electricity generation to transportation and industrial processes. The development cycle for new materials, especially for safety-critical applications, typically spans 10-20 years. This process requires rigorous validation through testing at the lab scale before scaling up to production. However, the efficiency of validation is significantly hindered by existing testing methodologies, which are laborious, time-consuming, and costly [2]. These methods are low-throughput, often evaluating one alloy composition at a time [3].
This project aims to pioneer a data-rich, high-temperature testing protocol by integrating heterogeneous testing (evaluating multiple compositions within a single test) with predictive and generative AI techniques. The goal is to accelerate the qualification of sustainable stainless steels with enhanced durability. The key objectives are:
Develop a high-throughput testing procedure integrated with full-field strain measurements at high temperature, to efficiently evaluate material performance.
Create AI models (e.g., an established AI model based on 2-point statistics and principal component analysis techniques [4], as well as a more exploratory AI model called generative adversarial networks [5]) to automatically quantify key microstructural features and their evolutions before and after testing.
Establish data-driven models with enhanced predictive capabilities to assess material damage sensitivity and recommend the next test to further refine accuracy.
The methodology and associated tasks are described below.
Task 1: We will fabricate a stainless steel sample with varied compositions. Using scanning electron microscopy (SEM), we will capture high-resolution images, focusing on the grain-level microstructural features at 50 distinct points, each representing a unique composition. After subjecting the sample to high-temperature testing for a given time, we will re-image the same areas to observe microstructural changes. These images will be divided into smaller sections, resulting in approximately 50,000 paired images before and after testing.
Task 2: We will run established algorithms to identify regions where cracks or other forms of damage have occurred. Using these data, we will train AI deep-learning methods to predict damage based on the initial microstructure and composition. By analysing how the microstructure evolves, we aim to understand the relationship between composition and durability.
Task 3: We will employ advanced AI techniques, such as generative adversarial networks (GANs), to predict the behaviour of new, untested microstructures. This approach not only accelerates material development but also aligns with sustainable AI practices by minimising the need for extensive physical testing.
AI will streamline the entire material testing process, including the collection and processing of large volumes of microstructure data, ultimately delivering faster and more accurate prognostics. While the focus is on stainless steel, the AI-enhanced end-to-end material testing protocol developed by this project is designed to be adaptable, offering benefits well beyond this material type.
References:
1. Eswarappa Prameela, S., Pollock, T.M., Raabe, D. et al. Materials for extreme environments. Nat Rev Mater 8, 81–88 (2023). https://doi.org/10.1038/s41578-022-00496-z
2. Materials 4.0 – Predicting and Controlling Materials’ Microstructures and Performance, Henry Royce Institute, https://www.royce.ac.uk/collaborate/roadmapping-landscaping/materials-4-0/
3. Couet, A., Integrated high-throughput research in extreme environments targeted toward nuclear structural materials discovery. J Nucl Mater 559, 153425 (2022)
4. Cameron, B.C., Tasan, C.C. Microstructural damage sensitivity prediction using spatial statistics. Sci Rep 9, 2774 (2019). https://doi.org/10.1038/s41598-019-39315-x
5. Khan, A., Lee, CH., Huang, P.Y. et al. Leveraging generative adversarial networks to create realistic scanning transmission electron microscopy images. npj Comput Mater 9, 85 (2023). https://doi.org/10.1038/s41524-023-01042-3