An autoencoder plus neural ODE surrogate for ASTEC vessel physics compresses 1913 dimensions to 6 and cuts one simulation from about 4.7 hours to about 19 seconds

The authors build what they describe as the first purely data-driven (non-intrusive) surrogate of ASTEC vessel physics by decoupling the vessel from the primary and secondary circuits and predicting the ASTEC boundary variables sB1 and sB2, leaving an interface for later coupling to a primary-circuit model. Previously the ASTEC vessel physics could only be computed by the coupled CESAR-ICARE mechanistic solver; this work replaces that coupling with a deep-learning surrogate and provides couplable boundary-variable predictions. Data come from ASTEC 3.1.2 SBO and LOCA trajectories, with 700 training, 100 validation and 40 test trajectories for LOCA and 286 training, 50 validation and 27 test trajectories for SBO; results are reported per variable via RMSEmean, RMSEmax and RMSEstd on the test sets.

The autoencoder compresses the vessel physics from 1913 degrees of freedom to 6 latent variables, a factor of about 332, letting the full spatio-temporal prediction run in under a minute on both CPU and GPU. Relative to evolving the original high-dimensional physical state, this compression moves time stepping into latent space and turns ASTEC simulations that take days into seconds. For LOCA, ASTEC's ICARE module alone averages 16879.8 s (median 13082.4 s), while AE-NODE averages 26.5 s (median 25.8 s) on CPU and 18.9 s (median 18.3 s) on GPU; the authors report roughly 640x mean speedup on CPU and 890x on GPU.

The neural ODE rolls out stably in latent space, spanning 10k to 50k time steps (about 4 to 20 hours) across different test trajectories without diverging, and responds to boundary inputs from the primary circuit by exhibiting different dynamics. Relative to training only with teacher forcing, the work combines teacher forcing with autoregressive training and uses an adaptive-window strategy that grows the time window, mitigating distribution shift and training instability. Latent-trajectory comparisons show the NODE prediction matching the encoder's true latent trajectory, with different dynamics across test trajectories; the authors also note that the autoencoder, not the NODE, is the main source of AE-NODE error.

Prediction quality varies by variable: trends and differences across trajectories are reproduced reasonably, but strong nonlinearities, sharp jumps and high-frequency oscillations are hard to match pointwise, and some scalar variables show large RMSEmean and RMSEstd. This work is the first to systematically characterize the capability boundary of a deep-learning surrogate on ASTEC's strongly nonlinear, discontinuous severe-accident physics rather than reporting only aggregate accuracy. Per-variable test errors show large errors for x alpha, P H2, m gas, Q liq vap, porosity, Q H2O ptv, m H2O ptv, Q steam vtp and fission-product-related variables; the authors use AE error as a lower bound, so AE-NODE cannot outperform the AE.

The result targets surrogate modeling of the ASTEC vessel domain (the CESAR-ICARE coupling), for SBO and LOCA scenarios in which operator actions are the only source of variation, the initial condition is fixed at nominal power, and the simulation ends at vessel rupture. It enables the vessel surrogate to be coupled to a primary-circuit model (provided by ASTEC or another surrogate), laying the interface groundwork for a real-time severe-accident simulator for operator training; the authors also state the methodology can transfer to other multi-module coupled multi-physics simulations.

The authors state the surrogate cannot yet be treated as a reliable substitute for ASTEC vessel physics because of the model's failures on some variables; whether those failures stem from modeling or from insufficient data requires generating more training data. Which spikes and nonlinearities in the data are numerical noise versus real physics needs variable-by-variable study with severe-accident experts; smoothing makes the latent space more regular but introduces artifacts such as large negative masses. In addition, the simulation end time is set by the end of the ground-truth trajectory, so the model does not itself predict vessel rupture, and the feedback effect at the primary-circuit volumes is not included, so prediction deviations could be amplified further. This is a full-text reading, with figures and numbers taken as reported in the main text and appendices.