LSTM -based prediction of thermodynamic evolution for supercritical fluids

Accurate reconstruction of thermophysical–property evolutions across the pseudo–critical region is a key requirement in the design and analysis of supercritical CO2 components, yet high–fidelity equations of state can become computationally expensive and numerically delicate when repeatedly queried in large–scale or real–time workflows. This work introduces a novel approach by framing supercritical property evaluation as a sequence—learning problem, leveraging a sequence–to–sequence Long Short—Term Memory regressor to predict entire isobaric property evolutions, rather than isolated state points. Reference datasets for pure CO2 are generated from a high–accuracy multi–parameter equation of state over a supercritical domain spanning different temperature and pressure ranges. Each isobaric curve is encoded as an input sequence with pressure and temperature channels, and three specialised Long Short—Term Memory networks are trained to reconstruct the full evolutions of density, enthalpy, and speed of sound. On held-out tests, the models reproduce the coupled curve shapes with MAPEs below 0.2% for density and speed of sound. With compact architectures ranging from 10 000 to 30 000 parameters, the networks allow for fast inference on the order of 10−4 s per sequence, outperforming baselines and supporting integration into real–time CFD and control settings. An ablation study on networks’ complexity further characterises the accuracy–cost trade–off, supporting the adoption of lightweight models for rapid property table generation, cycle analysis, and integration into CFD, digital twin, and control settings.