This study aims to develop and evaluate the capability of a machine learning (ML) framework as a faster alternative to the more widely used finite element (FE) modeling in the prediction of the bone fracture healing process. Seven machine learning algorithms (SVR, RF, XGBoost, MLP, CNN, RNN, LSTM) were trained using 648 simulated healing processes from an FE model, which contained differing implant, loading, and biological data. Each sample was evaluated for central stiffness, intermediate stiffness, strain energy, and outer stiffness. ML models use two types of parameters: learnable (optimized during training) and hyperparameters (set before training). Hyperparameters were optimized via a Bayesian optimization with a tree-structured Parzen Estimator (TPE). To evaluate performance of each model, mean squared error (MSE), emphasizing larger errors, and mean absolute error (MAE), offering robustness to outliers, were calculated. The study concludes that the LSTM (Long Short Term Memory) model performed the best, securing up to a 98% error reduction compared to baseline models in all four categories. This model’s performance was found in accordance with major mechanobiological principles, such as increasing screw number increases central stiffness and that excessive loading negatively impacts outer callus formation. It also was able to produce accurate results even with significantly less data for training, exemplifying its data efficiency.