A transfer learning–enabled framework for rapid property prediction toward scalable and data-efficient metamaterial design

Metamaterials offer transformative potential across various engineering domains. Despite significant progress in building databases linking metamaterial architectures to their properties, conventional machine learning (ML) approaches still face notable limitations. They cannot be directly applied to predict the same properties when the geometry or size scale of metamaterials changes, nor can they be readily extended to predict new properties. This is because a substantial amount of new training data is typically required to retrain the model, which is essentially equivalent to restarting the learning process from scratch. We present a transfer learning-based framework that significantly reduces the amount of required training data while providing high accuracy and stability. This framework employs a size-independent Convolutional Neural Network (CNN) architecture for both the source and target models through the implementation of 1×1 convolutional transformation and global average pooling (GAP). The source model is well-trained to predict the Young’s modulus of any 8 × 8 lattice metamaterial composed of four given unit cells. The target model can predict the Young’s modulus even when new unit cells are introduced or when the size scale of the metamaterial changes. Moreover, the learned structure–Young’s modulus knowledge can be successfully transferred to predict other mechanical or non-mechanical properties. Compared to conventional ML approaches, this framework requires only 1% of training data in the studied scenarios and no modifications to the model architecture. This work provides new avenues for building scalable and data-efficient metamaterial design spaces for various applications.