Timezone: »

 
Provably Correct Physics-Informed Neural Networks
Francisco Girbal Eiras · Adel Bibi · Rudy Bunel · Krishnamurthy Dvijotham · Phil Torr · M. Pawan Kumar

Fri Jul 28 06:20 PM -- 06:50 PM (PDT) @
in 2nd Workshop on Formal Verification of Machine Learning »

Physics-informed neural networks (PINN) have been proven efficient at solving partial differential equations (PDE). However, previous works have failed to provide guarantees on the worst-case residual error of a PINN across the spatio-temporal domain – a measure akin to the tolerance of numerical solvers – focusing instead on point-wise comparisons between their solution and the ones obtained by a solver at a set of inputs. In real-world applications, one cannot consider tests on a finite set of points to be sufficient grounds for deployment. To alleviate this issue, we establish tolerance-based correctness conditions for PINNs over the entire input domain. To verify the extent to which they hold, we introduce ∂-CROWN: a general and efficient post-training framework to bound PINN errors. We demonstrate its effectiveness in obtaining tight certificates by applying it to two classical PINNs – Burgers’ and Schrödinger's equations –, and two more challenging ones – the Allan-Cahn and Diffusion-Sorption equations.

Author Information

Francisco Girbal Eiras (University of Oxford)
Adel Bibi (University of Oxford)
Rudy Bunel (Deepmind)
Krishnamurthy Dvijotham (Google DeepMind)
Phil Torr (Oxford)
M. Pawan Kumar (DeepMind)

More from the Same Authors