Machine learning of microstructure--property relationships from data is an emerging approach in computational materials science. Most existing machine learning efforts focus on the development of task-specific models for each microstructure--property relationship. We propose utilizing pre-trained foundational vision transformers for the extraction of task-agnostic microstructure features and subsequent light-weight machine learning of a microstructure-dependent property. We demonstrate our approach with pre-trained state-of-the-art vision transformers (CLIP, DINOV2, SAM) in two case studies on machine-learning: (i) elastic modulus of two-phase microstructures based on simulations data; and (ii) Vicker's hardness of Ni-base and Co-base superalloys based on experimental data published in literature. Our results show the potential of foundational vision transformers for robust microstructure representation and efficient machine learning of microstructure--property relationships without the need for expensive task-specific training or fine-tuning of bespoke deep learning models.

Superalloys are a special class of heavy-duty materials with excellent strength retention and chemical stability at very high temperatures. Nickel-based superalloys are used commercially in aircraft turbines, power plants, and space launch vehicles. The optimization of mechanical properties of alloys has been traditionally carried out using experimental approaches, which demand massive costs in terms of time and infrastructure for testing. In this paper, we propose a method for mechanical property prediction of Ni-based superalloys by learning from past experimental results using machine learning (ML). Five highly accurate ML models are developed to predict yield strength (YS), ultimate tensile strength (UTS), creep rupture life, fatigue life with stress, and strain values. We have developed an extensive database containing mechanical properties of over 1500 Ni-based superalloys. Basic material parameters such as the composition of the alloy, annealing conditions, and testing conditions are also collected and used as features for developing the ML models. The prediction root mean squared errors for the YS, UTS, creep, and fatigue life models are 0.11, 0.06, 0.19, 0.22, which are minimal, leading to a highly accurate estimation of the target values. These ML models are highly transferable and require a minimum number of input features. In addition, feature analysis performed by SHapley Additive exPlanations (SHAP) for individual properties reveals the relative significance of each descriptor in deciding the target property. We demonstrate that a unified and highly accurate ML framework can be developed using common features for all mechanical properties. The models are developed on experimental data, making them directly applicable for industries.