Physics-informed neural networks for predicting laser-tissue interaction in maxillofacial reconstruction surgery

This study employs Physics-Informed Neural Networks (PINNs) to simulate the thermal dynamics of biological tissue under laser irradiation by embedding the heat transfer and radiative transport equations into the training process. The PINN architecture, comprising three hidden layers with 50 neurons each and Tanh activation, accurately predicts temperature distributions, achieving close agreement with the analytical solutions (final MSE [Formula: see text]). The thermal penetration depths, calculated under uniform conditions of P = 5 W and beam diameter = 1 mm, were CO[Formula: see text] (0.11 mm), Nd:YAG (0.077 mm), Er:YAG (0.063 mm) and Diode (0.055 mm), lasers. These values are strongly dependent on laser power density and absorption coefficients, and thus represent reference conditions rather than generalizable results. CO[Formula: see text] lasers concentrated energy at the surface, enabling precise incisions; Nd:YAG achieved deeper subsurface heating suited for coagulation; Er:YAG provided highly efficient superficial ablation; and Diode lasers offered balanced heating for minimally invasive procedures. The analysis further showed that laser-tissue interactions are strongly influenced by fluence, power, and pulse duration. Higher power leads to more superficial heating, while lower power favors deeper diffusion. Simulations in age groups revealed that elderly tissue, with lower absorption coefficients, exhibits greater penetration, whereas adolescent tissue shows more superficial confinement, underscoring the importance of patient-specific parameter adjustment. Critical thermal thresholds for coagulation, vaporization, and irreversible cellular damage were identified, providing clinically relevant safety margins. These findings demonstrate that PINNs provide a robust, physics-based framework for predicting laser-tissue interactions, showing close agreement with analytical benchmarks and offering a computationally efficient alternative to traditional solvers, although the current model is one-dimensional and still not experimentally validated, which will be the focus of future work. Beyond clinical optimization, this work establishes a computational foundation for future extensions, such as incorporating Arrhenius damage integrals, heterogeneous tissue layers, nonlinear optical effects, and real-time feedback modalities.