Physics-Inspired Machine Learning for Nanomaterial-Based Voltammetric Biosensors: A Data-Efficient Feature Engineering Approach

Abstract

This study introduces a physics-inspired machine learning (ML) framework for the calibration of voltammetric nanosensing platforms. By explicitly embedding electrochemical domain knowledge, specifically dynamic resistance (R = V/I), instantaneous power (P = V × I), and logarithmic kinetic descriptors, into the ML pipeline, we enable a peak-agnostic sensing paradigm that avoids manual peak selection and baseline correction. The framework is validated across three distinct nanopolymeric biosensing systems: reduced graphene oxide–incorporated molecularly imprinted polymer (rGO-MIP) nanostructures for C-reactive protein (CRP), histidine-functionalized poly(2-hydroxyethyl methacrylate-co-ethylene glycol dimethacrylate) (HEMA-co-EGDMA) nanoparticles for vitamin C, and vitamin E acetate (VEA)-imprinted poly(2-hydroxyethyl methacrylate-co-methacryloylamido phenylalanine) (HMA-co-PA) nanoparticles for vitamin E analysis. Across 13 evaluated ML algorithms, physics-inspired features yielded a real but model-dependent improvement in predictive performance, averaging an 8.7% reduction in Test root mean square error (RMSE); the benefit was most pronounced for linear and kernel-based regressors and more limited for the strongest tree-based ensembles, which already captured substantial nonlinear structure directly from raw voltage–current signals. Permutation feature importance indicated a shift in predictive attribution toward electrochemically interpretable descriptors, though collinearity among the engineered features warrants caution in attributing this shift to any single descriptor. Grouped cross-validation and prospective external validation on independently collected replicate measurements further supported the reliability of the proposed models for new, previously uncharacterized measurements within their calibrated concentration range. A comparison with classical peak-current calibration indicated that the proposed framework performs competitively with this established approach, with its principal practical advantage lying in avoiding manual peak selection and baseline correction rather than in superior raw predictive accuracy.