Deep learning modeling of 4J36 invar alloy process-microstructure-property relationships based on multimodal data

4J36 alloy is a key material in aerospace, electronic devices and other fields. It has been an important challenge for the design, development and optimization of related materials to construct the complex correspondence between process, microstructure and properties. In this paper, different specifications of 4J36 Invar alloy plates were prepared by designing various heat treatment and rolling processes, and the microstructure and mechanical properties curves of the rolled plates were tested. Combining rolling process data (including heat treatment temperature, holding time, cooling method) and microstructure image data (four EBSD images), six different types of deep learning models were constructed, namely MLP and CNN models based on single-modal process data and MLP-ResNet18, MLP-ResNet50, MLP-VITb16, MLP-VITb32 models based on multi-modal data. The stress-strain curves from the yield point to the tensile point of the rolled strip were predicted by using the constructed models, and an effective correspondence between the process-microstructure-properties of 4J36 Invar alloy was achieved. The research results show that under the multimodal and small sample data set in this work, the ResNet network model has more advantages than the VIT Transformer model. The predicted stress-strain curve of the MLP-ResNet18 multimodal mechanical performance model has the highest fit to the actual stress-strain curves, and the R 2 value can reach 0.96. This research provides a new method for the intelligent design and process optimization of 4J36 alloy, promoting its reliable application under different process conditions.