One universal interpreter for all dialects of sodium. Liquid sodium in a reactor behaves differently: near the pipe wall it is dense and fast, near the boiling surface it is rarefied and slow. Previously, each state required its own set of calculations, as if hiring a new translator for every phrase.

Now a single neural network potential, trained on quantum data with the r2SCAN functional, handles all the "dialects" at once. From room temperature, where sodium is a solid metal, to the critical point, where liquid and vapor become indistinguishable. The accuracy is close to experimental: the density of solid sodium differs from the measured value by only 0.06%.

To prevent the interpreter from making mistakes in "tone" (that is, energy), an additional model of electronic entropy was added. This allowed the internal energy to be recovered and caloric properties to be calculated. Without such a model, the potential would know only forces, not heat.

The verification was rigorous. Direct simulation of solid-liquid coexistence gave a melting temperature of 346 K, which is 25 K below the recommended value but within the method's sensitivity: the free energy error is only 3.5 meV per atom. That is enough to shift the phase transition.

The speed of sound, calculated directly in a cell of 8000 atoms, confirmed the model's thermodynamic consistency. The self-diffusion coefficient increased from 0.351 Ų/ps at 380 K to 30 Ų/ps in the critical region, while viscosity dropped from 0.653 mPa·s to 0.06 mPa·s. These are data from the dilute liquid, where experiments are almost nonexistent.

The universal interpreter for sodium is ready. It links the equation of state, phase boundaries, acoustics, and transport into one consistent picture. For fast reactors, this means that coolant properties can be predicted up to accident conditions with overheating, without waiting for dangerous experiments.