Shengkun XI, Jiahui LI, Qiuling TAO, Haijun ZHANG, Cuiping WANG, Xiaoyu CHONG, Rongpei SHI, Xingjun LIU
Under extreme service conditions involving long-term high temperatures, thermo-mechanical cycling, and concurrent oxidation/corrosion, the design of aerospace structural alloys is simultaneously constrained by the exponentially expanding compositional space, the scarcity and high cost of high-fidelity property labels, and the limited transferability of strongly coupled multi-scale mechanisms. Along the “composition–process–microstructure–property–service” chain, a materials intelligent design paradigm is constructed with physics-based constraints at its core: multi-modal and multi-fidelity data are standardized, aligned across domains, and stored in a unified database; conservation laws, crystallographic symmetry, and phase-diagram consistency are embedded into classical machine learning models, convolutional neural networks, graph neural networks, and Transformer/pre-trained architectures; microstructural intermediates such as segmented phase maps and size distributions are explicitly introduced to strengthen the mapping among processing, microstructure, and properties; and uncertainty quantification, domain adaptation, and out-of-distribution detection are employed to control the risk associated with model extrapolation. At the decision-making level, generative design and multi-objective Bayesian optimization are incorporated to form a closed-loop “generation–screening–validation–update” workflow. For γ–γ′-strengthened Ni/Co-based superalloys, L12-strengthened heat-resistant/high-temperature Al alloys, and multi-principal/high-entropy alloys, multi-objective trade-offs are performed with respect to γ′ volume fraction and solvus temperature versus lattice misfit, precipitation and coarsening kinetics versus the synergy between thermal conductivity and strength, and sublattice occupancy versus long-range order. Overall, this physics-informed intelligent framework enables robust extrapolation that balances performance and confidence under small-sample, cross-domain, and multi-modal data conditions, and provides a unified feature space and evaluation criterion for the continuous iteration of long-life high-temperature alloys.