Seismologists listen to the Earth the way a doctor listens to a heart. Surface waves from earthquakes travel across the planet, and their speed reveals what lies beneath our feet — solid rock, melt, or water. Usually, this requires building complex physical models. Now neural networks have come to the rescue: they quickly predict how a wave will travel through the interior, but it remained a mystery whether they truly understand the physics or just memorize the data.

The authors took a neural network surrogate trained on thousands of surface wave examples and looked "inside" it using automatic differentiation — a method that shows how much the network's output changes when input parameters are varied. They compared the resulting gradients with theoretical sensitivity kernels — the gold standard that tells at what depth a wave "feels" the rock. The match was striking: the network reproduced the depth structure of the kernels across a wide range of wave periods, from short ones (sensitive to the crust) to long ones (reaching the mantle).

But not everything is smooth. If the training data contain strong structural assumptions (for example, a constant crustal thickness) the network picks up these biases and introduces systematic distortions in sensitivity. In other words, the network learns not only physics but also the habits of its teachers. This is an important warning for those who will apply such models in practice: the training set must be carefully chosen to avoid a false picture of the subsurface.

Nevertheless, the result is encouraging. Neural network surrogates can be used not only for fast predictions but also for extracting physically meaningful information — for example, for inversion (recovering rock properties from wave data) or uncertainty estimation. It's as if a student not only solved textbook problems but also derived their own formulas that match the laws of nature.