Closed-loop inverse design of titanium alloys with multi-objective performance via multimodal data-driven modeling

This study proposes a high-throughput, automated design framework for high-performance titanium alloys, integrating multimodal data fusion, deep learning prediction, and closed-loop genetic algorithm (GA) optimization. Unlike varying contrast methods that struggle with heterogeneous data, we constructed a standardized multimodal database fusing alloy composition, processing parameters, and ResNet50-extracted microstructural features. A long short-term memory (LSTM) neural network model was developed for full-range non-destructive prediction of 604 experimental stress-strain curves of titanium alloys. The mean absolute error (MAE) of stress prediction was 38.4 MPa, and the mean absolute percentage error (MAPE) of fracture strain was only 0.438 %. Furthermore, a closed-loop inverse design framework, integrating the GA with the deep learning model, simultaneously optimizes composition, processing, and microstructure to automatically identify new titanium alloys with superior ultimate performance across diverse service environments. Experimental verification on five representative new alloys shows that under room temperature, high temperature, and dynamic compression conditions, their highest ultimate compressive strength has been increased by up to 59.7 % compared with the original reference alloy. These results comprehensively verify the effectiveness of data-driven closed-loop material design in rapidly discovering and optimizing the extreme properties of advanced alloys, providing new theoretical and technical insights for the intelligent design and engineering application of complex alloy systems.