Machine-assisted Adsorption Process Learning and Emulation: the trained
neural-network surrogate of a 4-step PVSA cycle for CO2/N2 separation
(Pai, Prasad & Rajendran) ported to JavaScript — cyclic-steady-state purity,
recovery, energy and productivity in microseconds, for any single-site-Langmuir
adsorbent inside the trained range. Fixed rig: 1 m column, presets from the
two source papers.
Cycle & adsorbent
Pressurization
Adsorbent
Process conditions
Feed CO2 fraction yF
Adsorption time tADS [s]
High pressure PH [bar]
Intermediate PI [bar]
Evacuation PL [bar]
Feed velocity vF [m/s]
Cyclic-steady-state performance
Isotherms at 30 °C (pure component)
Parametric sweep
Multi-objective optimization & sorbent comparison (NSGA-II on the surrogate)
Objectives
Constraints [%]
GA settings
Sorbents to compare
Decision-variable search ranges
Narrow the space the optimizer may search
(e.g. a pump that cannot pull below 30 mbar, or a fixed cycle time). Values clamp to the
trained support — you can restrict the search but not push it outside the data.
GA convergence — normalized hypervolume of the current
Pareto front vs generation (1 = final front's hypervolume):
Decision-variable trends along the front
Each optimal point's decision variables plotted
against the front's x-axis — shows how the optimizer trades them off along the Pareto.
Decision variables: tADS, PH (unless VSA), PI/PH,
PL/PH (log-scaled), vF at the yF and cycle chosen above.
NSGA-II: real-coded, SBX crossover (ηc = 15), polynomial mutation
(ηm = 20, rate 1/nDV), binary tournament, crowding-distance survival;
deterministic seed — same settings give the same front. The Pu/Re constraints are enforced
by penalty in every mode except Purity–Recovery (there the whole trade-off is the result);
the shaded box marks the constraint region (defaults = US-DOE 95/90). The front is redrawn live
every generation. Surrogate outputs outside the physical window (0–102 %) are rejected,
as in the published wrapper. PI and PL are additionally confined to the
absolute sampled windows of the training data (0.07–4 and 0.01–1 bar):
the LHS sampled absolute pressures, so the ratio ranges alone would admit joint data voids at
low PH where the surrogate is unsupported.