Shape memory alloys (SMAs) are materials that can generate and recover moderate inelastic strains through temperature modulation. This characteristic enables SMAs to function as lightweight and compact thermomechanical actuators. The synthesis of SMA actuators for systems with specific requirements on their actuation path (displacement stroke vs. temperature) currently relies on complex and expensive material processing and characterization. This paper first presents a geometric approach for synthesizing novel SMA axial actuators, termed as hybrid SMA actuators, whose dimensions (length and cross-sectional area) and material distribution are modulated to approximate a target actuation path. Through the combination of multiple SMA wire sections in series, the hybrid SMA actuators can exhibit actuation paths not achievable by using single monolithic SMA wires. The paper then presents a machine learning-assisted framework for the surrogate modeling of hybrid SMA actuators. This approach allows for the prediction of their actuation path without the use of structural simulations leveraging numerical implementations of constitutive models, allowing for simplified and computationally efficient modeling and circumventing convergence issues. A surrogate model consisting of an ensemble of binary decision trees is trained using data obtained via a design of experiments performed using structural simulations. A validation test using 5000 design samples for hybrid SMA actuators with two sections demonstrates R2 values of 0.99983 and 0.99979 for the actuation displacement during heating and cooling, respectively. The evaluation time for the validation samples using the trained surrogate model is less than 8 minutes, while the evaluation time using structural simulations is 59 minutes. Finally, a surrogate-based optimization approach is demonstrated through the synthesis of hybrid SMA actuators capable of exhibiting prescribed displacement vs. temperature target actuation paths.