Fix Targets AI Physics Errors

LondonNeural networks were quietly breaking the physics they modeled.

A new regularization method forces physics-informed neural networks to respect one-way processes like melting and corrosion, cutting predictive errors by over tenfold, researchers report.

The fix needs only minor changes to existing models, easing adoption in fluid dynamics and materials engineering.

Sources: Nature