Physics‐Guided Transfer Learning for Predicting Stress–Strain Curves With Limited Experimental Data

Many measurements in materials science are expressed as full response curves rather than single scalar values. Nevertheless, predicting entire curves reliably from limited experimental data remains challenging. Here, we address this problem for hot‐deformation stress–strain curves using a physics‐guided transfer learning (PGTL) framework that integrates constitutive modeling with deep learning. The strain‐compensated Sellars model serves as a data engine that generates synthetic curves over a wide range of temperatures and strain rates. These simulations are used to pretrain a bidirectional long short‐term memory encoder that maps processing parameters and strain history to the flow stress. Subsequently, only the output layer is fine‐tuned on a small set of experimental curves. For a nickel‐based FGH98 superalloy, the PGTL model reduces the percentage error to approximately 2%–4% in leave‐one‐curve‐out validation compared with 9%–19% for a purely data‐driven network, and it remains accurate under independent test conditions beyond the training domain. When applied to a Mg–3Sn–1Mn alloy and a Ti–Al intermetallic, the same PGTL framework captures distinct curve shapes while maintaining low errors. Collectively, these results demonstrate that PGTL generalizes across both processing spaces and alloy families, providing a reusable surrogate for hot‐forming process design and a template for modeling other curve‐type responses, including battery voltage profiles, impedance spectra, and polarization curves.