A machine learning approach was developed to predict load-dependent power loss in gear pairs under mixed
elastohydrodynamic lubrication (EHL). Using a homogenized mixed EHL solver, training data were generated,
considering surface roughness and cavitation phenomena. A multimodal deep learning (MMDL) model improved
regression performance for multimodal inputs. Validation against experimental data confirmed the model’s
reliability, achieving a maximum gear efficiency error of 0.07%. The MMDL model with concatenation fusion
was selected for its highest R-square value of 0.99166, significantly accelerating simulation speed by 0.05%. This
method can be applied to optimize gear design and provides an efficient solution to reduce power loss in
automotive drivetrains, overcoming the limitations of conventional analytical and EHL methods.