Thermodynamically informed and balance-preserving neural networks for direct vapor–liquid flash calculations

Pressure-temperature vapor-liquid equilibrium (PT-VLE) flash calculations are performed repeatedly in compositional-flow simulations and can dominate runtime. To reduce this cost, we propose three thermodynamically informed GibbsFlashNN models that directly predict phase state, compositions, and vapor fraction. A common allocation decoder algebraically enforces nonnegativity, composition normalization, and component-wise material balance. Training incorporates differentiable conditions derived from phase equilibrium and constrained Gibbs-free-energy minimization, supplied by a fixed equation of state or activity-coefficient model, or by a neural thermodynamic potential trained jointly with the flash network or independently pretrained and frozen. Evaluation spans eleven studies involving petroleum fluids, CO2-containing mixtures, nonideal organic and aqueous systems, and a five-component mixture. Across 3,090 held-out experimental composition values from nine systems, the proposed models achieve pooled major-component mean absolute percentage errors of 6.63–7.63%, compared with 6.62% for classical flashes using equations of state or activity-coefficient models and 8.01% for a supervised balance-preserving baseline with the same architecture. Relative to this neural baseline, the proposed models also reduce the pooled 95th-percentile absolute percentage error from 30.3% to 25.8–29.6%. Across the benchmark cases, median end-to-end inference speed-ups over the corresponding classical flashes range from 5.2 to 241.7 times, including confidence screening and reference fallback. These results support incorporating thermodynamic conditions into neural flash training to reduce aggregate prediction errors while approaching classical-flash accuracy at substantially lower inference cost.