Scientists have developed a novel machine learning-based method to discover unknown physical equations from experimental data. This approach, utilizing fully differentiable finite elements, allows for the identification of underlying laws governing a system without the need for prior formulation. The technique has been demonstrated in recovering constitutive equations in materials mechanics, an area where the relationships between stress and strain can be complex and challenging to model explicitly.

The method integrates a finite element model within a deep neural network, making the entire system differentiable. This means that the parameters of the physical equations can be optimized directly from the data, adjusting the proposed laws until the model accurately predicts observed behavior. Unlike traditional approaches that require an initial hypothesis about the equation's form, this approximation can explore a broader space of possible physical laws, revealing non-intuitive or unexpected relationships.

The implication of this breakthrough is significant for physics and engineering. It could accelerate the discovery of new materials with specific properties by enabling the identification of their constitutive laws from experiments. Furthermore, it offers a powerful tool for fundamental science, where the formulation of new theories often depends on the ability to infer general principles from detailed observations. This approach opens the door to a new era of AI-assisted scientific discovery, where machines not only analyze data but also formulate the laws that govern them.