The booming demand for efficient, scalable optical networks has intensified the exploration of innovative strategies that seamlessly connect large-scale fiber networks with miniaturized photonic components. Within this context, our research introduces a neural network, specifically a convolutional neural network (CNN), as a trailblazing method for approximating the nonlinear attenuation function of centimeter-scale multimode waveguides. Informed by a ray tracing model that simulated many flexographically printed waveguide configurations, we cultivated a comprehensive dataset that laid the groundwork for rigorous CNN training. This model demonstrates remarkable adeptness in estimating optical losses due to waveguide curvature, achieving an attenuation standard deviation of 1.5 dB for test data over an attenuation range of 50 dB. Notably, the CNN model's evaluation speed, at 517 µs per waveguide, starkly contrasts the used ray tracing model that demands 5-10 min for a similar task. This substantial increase in computational efficiency accentuates the model's paramount significance, especially in scenarios mandating swift waveguide assessments, such as optical network optimization. In a subsequent study, we test the trained model on actual measurements of fabricated waveguides and its optical model. All approaches show excellent agreement in assessing the waveguide's attenuation within measurement accuracy. Our endeavors elucidate the transformative potential of machine learning in revolutionizing optical network design.