A new neural network architecture turns raw material test data into human-readable physics equations — automatically.
Researchers introduced inelastic Constitutive Kolmogorov-Arnold Networks, or iCKANs, a framework designed to identify how materials deform and recover under stress. Unlike conventional neural networks that produce opaque numerical outputs, iCKANs generate closed-form mathematical expressions — the kind a materials scientist can actually read and verify. The team validated the approach on synthetic data and on two real viscoelastic polymer materials, VHB 4910 and VHB 4905, with results showing accurate capture of complex viscoelastic behavior. The architecture can also ingest additional material metadata, such as temperature-dependent properties, alongside mechanical test data.
Most machine learning models applied to materials science are black boxes: they predict well but explain nothing, which limits how much engineers can trust or generalize the output. iCKANs sidestep that problem by producing interpretable potential functions — outputs that slot directly into existing physics frameworks rather than replacing them with an inscrutable model.
Kolmogorov-Arnold Networks have been generating quiet excitement in scientific computing circles since their introduction as an alternative to standard multilayer perceptrons, and this application to solid mechanics is a credible use case. Whether iCKANs hold up outside polymer materials — in metals, composites, or biological tissue — remains the test worth watching.