Search papers, labs, and topics across Lattice.
University of Electronic Science and Technology of China
3
0
3
Standard residual-based refinement routinely falters on steep gradients and phase transitions, but combining diffeomorphic coordinate mapping with Fourier spectral preconditioning cuts PINN errors down to $10^{-5}$ across stiff non-linear PDEs.
A dynamic selection approach among microarchitectural policies can recover over 70% of performance potential without executing inactive policies.
Machine-learning predictors can mislead design decisions, with over 22% of configurations showing unexpected performance reversals that traditional simulation still captures best.
