Discovering Explicit Magnetic Core Loss Equations via Learnable Symbolic Sparse Identification

What the math gives to ML

The paper offers a constructive sparse symbolic-regression mechanism: represent a target as a linear combination of candidate nonlinear basis functions while jointly learning both their coefficients and structural parameters such as exponents. Its transferable asset is a compact, differentiable, interpretable head that can replace an overparameterized final MLP for positive, multi-scale regression problems. The most useful neural-network adaptation is a learnable sparse power-law library on top of a shared encoder, trained with logarithmic relative error and explicit coefficient sparsity. The predicted signature is that only a few basis terms remain active while preserving accuracy across orders of magnitude.