Orbital-resolved parity plots for prediction accuracy analysis

[k-yoshimi]

Summary

Create parity plots (predicted vs. true values) for each orbital (s, px, py, pz, dxy, ...) to analyze which orbitals are easier or harder to predict.

Background

• NN output y contains spherical harmonic coefficients per (l, m) (y[:, :n**2])
• sph_harm.py already defines the orb_symbol mapping from (l, m) to orbital names (s, px, py, ...)
• Orbital-resolved accuracy visualization will guide model improvement

Tasks

• Generate per-orbital parity plots (predicted vs. true values)
• Display per-orbital MSE / R² scores
• Output a combined figure with subplots for all orbitals

Related files

• src/seap/prediction/sph_harm.py — orb_symbol dict, quantum_number()
• models/encoder_1/encoder.py — NN model
• src/seap/prediction/utils.py — plotting utilities