Recursive swing-bed control tested against NASA-published capture data.
NASA holdout testrecursive bed memoryexact species ledgers
A fixed timer treats every adsorption cycle as identical. Flow carries the recovered capacity of each
bed from one half-cycle into the next. Injection spends the active bed; vacuum exposure restores the
opposite bed; the transported state determines when the valve should route. The result is a controller
that responds to the condition of the capture system instead of merely waiting for the clock.
NASA TEST The recursive law was frozen before it saw
eight steady operating points and an eight-hour variable-load experiment from two published RCA test articles.
RESULT Predictive bed-switch planning from measurements
the model was not allowed to fit:
4.32% mean prediction error across eight untouched steady operating points.
85.99% less error than the registered constant-capacity baseline.
122 predicted routes versus 127 measured across the untouched eight-hour variable-load run.
11 / 11 load changes moved switching frequency in the measured direction.
Prediction on NASA measurements excluded from fitting
NASA-published RCA holdout test
86% less prediction error
Mean error across eight operating points the model was not allowed to fit. Less is better.
Constant-capacity baseline30.82% error
Flow recursive model4.32% error
8 hoursuntouched variable-load test
122 / 127predicted / measured routes
11 / 11load-response directions correct
The law was calibrated on three published markers per article. All eight remaining steady
markers and every event in the variable-load experiment were withheld from fitting. Source simulation
curves were excluded and no extracted point was manually corrected. Source:
NASA-20110011193. Exact readings:
public evidence receipt.
What that predictive advantage does in control
Registered two-bed control stress test
Capacity fell. Flow held the CO₂ peak at 2.000 mmHg.
The same simulated swing-bed plant was run under a permanent capacity loss. The fixed schedule rose
to 2.629979 mmHg. Flow observed the lost capacity and changed the route, holding the registered result
at 2.000000 mmHg.
Fixed schedule2.630 mmHgcapacity-fault peak
Flow control2.000 mmHgcapacity-fault peak
8 fewer swapsat the same nominal 2.000 mmHg peak
< 4.5×10−14 molworst CO₂/H₂O accounting residual
one-hour case
controller
peak CO₂ (mmHg)
swaps
mean load
nominal
timed reference
2.000000
29
0.6690
nominal
Flow
2.000000
21
0.8221
metabolic step
timed reference
2.467325
29
0.7294
metabolic step
Flow
2.467325
27
0.8192
capacity fault
timed reference
2.629979
29
0.8019
capacity fault
Flow
2.000000
29
0.8230
The benchmark drew a precise line
The NASA holdout established steady predictive authority, near-total route-count
recovery, and the correct response direction at every load change. It did not establish precise
transient route timing after every sharp load change: the worst segment missed by 67.86%. Flow therefore
uses the recursive model to carry the plan and measured receptor state to authorize the route. The
unsuccessful transient-authority gate remains part of the public audit rather than being hidden.
Closed material accounting
Every CO₂ and H₂O simulation closes the same ledger:
initial cabin + generated = final cabin + final sorbent + vacuum exhaust. The permanent fault is
recorded as lost bed capacity and changes the route; it is never recorded as completed capture work.