In volume visualization, a transfer function tailored for one volume usually does not work for other similar volumes without careful tuning. This process can be tedious and time-consuming for a large set of volumes. Leveraging the differentiability of volume rendering, this work presents a novel approach to transfer function optimization based on a reference volume and its transfer function. Using two fully connected neural networks, our approach learns a continuous 2D separable scalar-gradient transfer function that visualizes the features of interest with consistent visual properties across volumes. The resulting optimized transfer function is exportable for further interactions in domain-specific applications. Through two compelling use cases—tracking the aftermath of an asteroid blast near the ocean surface and visualizing brain white matter, grey matter, and cerebral fluid in magnetic resonance (MR) images—we demonstrate the effectiveness of our approach, validated through collaboration with domain experts.