High-precision prediction of nuclear mass, or nuclear binding energy ($E_b$), remains a key goal in nuclear physics research. Many AI-based tools have recently shown promising results on this task, some achieving accuracies exceeding the best physical models. However, the usefulness of AI models is questionable, given that predictions are only useful in the absence of measurements. This inherently requires extrapolation from training (and test) samples. Because AI models are often black boxes, the reliability of these extrapolations is difficult to assess. In this study, we present an AI model that not only achieves state-of-the-art accuracy for $E_b$ but also does so in an interpretable manner. For example, we demonstrate that the most important dimension of the model's internal representation is the double helix, linking the numbers of protons and neutrons found in the most stable nuclei in each isotope chain, analogous to hydrogen bonds in DNA. We explain why. Furthermore, we demonstrate that AI predictions of $E_b$ can be hierarchically factored and sorted, with the most significant terms corresponding to well-known symbolic models (e.g., the famous liquid droplet model). Remarkably, the AI model's improvement over the symbolic model can be attributed almost entirely to Jaffe's 1969 observations of the structure of most known nuclear ground states. This results in a fully interpretable, data-driven nuclear mass model based on the physics inferred by AI.
