Machine learning-assisted prediction of non-Fourier multilayer bioheat transfer during laser–tissue interaction using ANN and XGBoost surrogate models

Accurate and rapid prediction of laser-induced heating in multilayer biological tissue is essential for optimizing therapeutic energy delivery while minimizing unintended thermal injury. However, conventional heat-transfer models may inadequately represent finite thermal-response effects, whereas repeated high-fidelity numerical simulations are computationally expensive for real-time parameter assessment. This study develops a unified physics-based and machine-learning framework for predicting transient laser-induced thermal transport and cumulative tissue damage in multilayer skin. The physical model represents the epidermis, dermis, and subcutaneous tissue and incorporates wavelength-dependent optical absorption and scattering, finite thermal relaxation, physiological perfusion, and generalized non-Fourier heat transport. An implicit finite-difference scheme is employed to generate high-fidelity temperature fields, which are subsequently used to train neural-network and gradient-boosted tree surrogate predictors. Cumulative thermal injury is additionally evaluated using a kinetic damage formulation that accounts for the complete temperature–time history. Numerical validation against published benchmark data yields errors of only 1.26% in peak temperature, 1.21% in average temperature, and 3.29% in thermal penetration depth. The gradient-boosted surrogate achieves a coefficient of determination of 0.993, with a root-mean-square error of 0.55 °C and a mean absolute error of 0.38 °C, while providing approximately 120-fold computational acceleration relative to the numerical solver. The thermal-damage analysis reveals a strong dependence on both wavelength and exposure duration; for example, under representative strongly absorbing irradiation, increasing exposure from 1.5 to 3.0 s raises the damage index from approximately 1.41 to 13.83, indicating that peak temperature alone may substantially underestimate cumulative injury. The principal novelty is the integration of scattering-corrected multilayer optical transport, physically constrained non-Fourier thermal relaxation, kinetic damage assessment, and interpretable machine-learning surrogates within a single computational framework. Unlike approaches focused solely on temperature prediction or computational acceleration, the proposed method simultaneously resolves transient thermal physics, quantifies cumulative tissue injury, and enables rapid prediction. These findings demonstrate its potential as an efficient model-based platform for laser-parameter screening, thermal-safety assessment, and future patient-specific treatment planning.