Physics-informed machine learning

Machine learning constrained by the governing equations of the system.

Rather than learning relationships purely from data, physics-informed approaches embed known physical laws as constraints. Conservation of mass and energy, thermodynamic limits and established rate expressions restrict what the model is permitted to predict.

Two properties follow. The model extrapolates more safely, because the constraints continue to hold outside the training distribution. And its outputs are attributable to a mechanism, which is what allows a result to be explained to an operator or to an auditor rather than presented as a score.

It also permits estimation of quantities no instrument reports directly, by inferring them from measurements that are related through the physics.