Accurate representation of the direct radiative effects of atmospheric aerosols is a crucial component of modern climate models. Direct computation of the radiative properties of aerosol populations is far too computationally expensive to perform during climate simulations however, so optical properties are typically approximated using a parameterization. This work develops artificial neural networks (ANNs) capable of replacing the aerosol optics parameterization currently used in the Energy Exascale Earth System Model’s (E3SM) Atmosphere Model (EAM). The ANNs are trained using a large dataset of aerosol optical properties computed using accurate Mie scattering code. Two ANN models are developed, one that directly replaces the current parameterization while achieving significantly higher accuracy, and a second that represents a more sophisticated core-shell scattering model, a scattering model that could not feasibly be represented by extending the existing parameterization. Optimal neural architectures for this problem are identified by evaluating ANNs with randomly generated wirings. We show that randomly generated deep ANNs that include many skip connections can consistently outperform conventional multi-layer perceptron style architectures with comparable parameter counts. This style of neural architecture search has substantial potential to improve performance in future ANN-based parameterization development.