Because the model is constrained by physics as well as data, it can produce results that are more trustworthy in engineering contexts than purely data-driven models alone.
How it works
The model is trained not only to match historical data but also to respect physical laws, which helps reduce unrealistic outputs and makes the results more trustworthy in engineering applications.
Application here
Models with built-in physical constraints deliver better predictions when there is not enough training data to rely on machine learning alone.
Business impact
This is particularly useful when engineering data is scarce or expensive, because physical knowledge helps compensate for limited data.
Limitations
It requires strong domain expertise to encode the right physics. Wrong assumptions can make the model worse rather than better.
In production
This is already valuable when pure machine learning is too unreliable but full first-principles simulation is too slow or difficult to operationalize.
Research
The frontier is toward systems that can recover or refine the hidden rules of a process from measurement data, making advanced modeling possible even when the full physics is not already written down.
Examples
NVIDIA Modulus (now PhysicsNeMo) trains physics-informed neural networks (PINNs) for engineering tasks: the neural net incorporates physics equations as part of the loss function, reducing data requirements and improving interpretability. Siemens uses PINNs for turbine thermal modelling (Siemens Technology, 2023) — .