Transfer Learning for Novel Material Property Prediction Using Pretrained AI Models

Abstract

Marek Grzesiak

Accurate prediction of material properties is essential for accelerating the discovery of innovative materials across applications such as electronics, energy, and structural engineering. However, traditional machine learning approaches require large, labeled datasets, which are often unavailable for novel or hypothetical compounds. This limitation significantly restricts their practical utility. To address this challenge, we propose a transfer learning framework that leverages pretrained deep learning models trained on large-scale materials datasets to predict physical, thermal, and electronic properties of new materials with minimal labeled data. Our method uses feature reuse and fine-tuning strategies to adapt neural networks to new prediction tasks across diverse material classes. We evaluate the framework on benchmark datasets from the Materials Project, OQMD, and AFLOW, focusing on three key properties: Young’s modulus, thermal conductivity, and electronic bandgap. The results demonstrate that transfer learning consistently improves prediction accuracy over models trained from scratch, particularly in low-data regimes. Mean absolute error reductions of up to 45% were observed, with faster model convergence and reduced overfitting. This work highlights the potential of transfer learning to make materials informatics more scalable and efficient. We further examine the role of domain similarity and transfer strategy in performance gains, providing practical guidance for model selection and application. By lowering the barrier to data-driven modeling, the proposed framework can support more rapid exploration of the materials design space, aiding in high-throughput screening and accelerating discovery pipelines. These findings suggest promising directions for integrating transfer learning with next-generation AI tools in materials science.

PDF