Phase-sensitive surface plasmon resonance biosensors are known for their high sensitivity . One of the technology bottle-necks of such sensors is that the phase sensorgram, when measured at fixed angle set-up, can lead to low reproducibility as the signal conveys multiple data . Leveraging the sensitivity, while securing satisfying reproducibility, is therefore is an underdiscussed key issue . One potential solution is to map the phase sensorgram into refractive index unit by the use of sensor calibration data, via a simple non-linear fit . However, basic fitting functions poorly portray the asymmetric phase curve . On the other hand, multi-layer reflectivity calculation based on the Fresnel coefficient can be employed for a precise mapping function . This numerical approach however lacks the explicit mathematical formulation to be used in an optimization process . To this end, we aim to provide a first methodology for the issue, where mapping functions are constructed from Bayesian optimized multi-layer model of the experimental data . The challenge of using multi-layer model as optimization trial function is addressed by meta-modeling via segmented polynomial approximation . A visualization approach is proposed for assessment of the goodness-of-the-fit on the optimized model . Using metastatic cancer exosome sensing, we demonstrate how the present work paves the way toward better plasmonic sensors.