Physics-informed machine learning

Physics-informed machine learning is a class of algorithms and mathematical models that build known physical laws or constraints, such as governing differential equations, directly into how the model learns and predicts, rather than relying on data patterns alone. This can improve accuracy and generalization, especially when training data is limited, since the model is constrained to respect known physics rather than free to fit any pattern in the data.¹ Some plain-language writing calls this physics AI as a broader category label, to distinguish it from physical AI (embodied AI in robots and autonomous machines). The two categories sound alike; they describe fundamentally different things. This page covers physics-informed machine learning specifically, one canonical approach within physics AI.

Why physics-informed approaches can need less training data

Physics-informed approaches can perform well with less training data than a purely data-driven model, since physical constraints reduce the space of solutions the model needs to search, which matters in manufacturing contexts where extensive labeled data may not be available.

Where physics-informed machine learning applies

Not on the drawing; this is a modeling technique used within engineering and manufacturing software, not something specified by a part's requirements.

Common mistakes

Assuming physics-informed methods eliminate the need for training data entirely; they still require data, but incorporate physical constraints to make better use of the data available, especially when it is limited.

Does physics-informed machine learning need less data than standard ML?

Often yes, since incorporating known physics reduces the space of possible solutions the model must learn purely from data.

Is this the same as a physics-based simulation?

No, it's a hybrid approach: a machine learning model trained with physical constraints, not a full traditional physics simulation on its own.

Where might this apply in manufacturing?

In modelling processes governed by known physical relationships, such as cutting mechanics, heat transfer or material deformation, where pure data-driven models can struggle with limited data.

Is this the same as physics AI?

Physics AI is a broader category label; physics-informed machine learning (this page) is one canonical approach within it. Its contrast is with physical AI: embodied AI in robots and autonomous machines.

How does this relate to physical AI?

It does not. Physics-informed machine learning is about how a model reasons, with physical laws built into the algorithm. Physical AI (embodied AI in robots and autonomous vehicles) is a separate category; see /glossary/physical-ai.