Physics-Informed Mamba Network for Mid-to-Low-Temperature Metrology via Raman Spectroscopy under Sparse Calibration

<jats:p>&lt;p dir="ltr"&gt;In the medium-to-low-temperature regime, sparse calibration and near-degenerate longitudinal-optical (LO) peak positions limit Raman thermometry. We propose a physics and machine-learning hybrid. A Residual-Spline physical backbone, based on quadratic phonon inversion and piecewise cubic Hermite interpolating polynomial (PCHIP) residual calibration, anchors the temperature readout. A residual Mamba (RaMamba) ensemble then applies only LO lineshape corrections under leave-one-out supervision and adaptive blending. On 22 GaAs spectra spanning 83.15-823.15 K, grouped five-fold cross-validation yields a mean absolute error (MAE) of 3.82 K (R&lt;sup&gt;2 &lt;/sup&gt;= 0.9996), improving the physics-only Residual-Spline baseline by 1.11 K, with 0.78 K, at the four lowest anchors. The largest gains at the lowest anchors indicate that LO lineshape asymmetry provides temperature cues beyond peak position alone. This physics-first residual design offers an interpretable route to sparse-anchor Raman thermometry when peak-shift information is limited.&lt;/p&gt;</jats:p>