When Do PINNs Beat Classical Numerical Methods? A 1D vs 5D Experiment
AI Summary: The author compared Physics-Informed Neural Networks (PINNs) against the finite differences method to solve the quantum harmonic oscillator problem in one and five dimensions. In 1D, finite differences outperformed PINNs by several thousand times, reaching final accuracy in under one millisecond. In 5D, PINNs were significantly more efficient, requiring 89 seconds of CPU time, whereas matching the PINN's 0.1% error with a grid would require over 1 TB of memory and 5 billion points. PINN performance was optimized by combining Adam and L-BFGS optimizers and enforcing a positivity constraint on the wavefunction to prevent convergence to excited states.