The problem
Open-circuit breathing apparatus exhausts every breath, wasting roughly two-thirds of the oxygen it delivers and limiting a firefighter to about 30 minutes under heavy exertion, in fires that often exceed 500 degrees Celsius. Recycling the breath, as military divers and spacecraft do, could potentially triple that, but closed loops carry risks of their own, carbon dioxide retention among them.
A positive-pressure suit adds a further problem. Its one-way exhaust valves leak gas outward, and topping up with pure oxygen slowly enriches the atmosphere towards a fire hazard. Managing that is the central control problem.
What the paper does
Part I models the loop from first principles. A soda lime scrubber removes carbon dioxide, a silica gel bed removes moisture, and pure oxygen is added from a finite supply — in that order, because the scrubber needs moisture and also produces it. The model covers the chemistry, the heat each bed gives off, humidity and the venting that drives enrichment.
Part II is the controller. An 18-state model with three controls — oxygen injection, fan speed and scrubber bypass — is estimated from sensors that work in a structural fire, including three oxygen cells with median voting. A learned model separates the part of the heart rate that reflects real work from heat strain. A predictive controller plans 15 to 20 seconds ahead, with a cost on venting that rises as the tank empties, so it grows more conservative as oxygen runs short. A reinforcement-learning policy only suggests starting points, and a control-barrier-function safety filter checks every command before it reaches the hardware.
What it shows
The consumables are 1 kg each of soda lime and silica gel and 3 kg of usable oxygen. The binding constraint in normal operation is the fire limit of 23.5 per cent oxygen, and in a worst case of heavy venting the paper calculates that the loop could reach it in about two and a half minutes.
In three simulated scenarios — steady work of unknown length, intervals of hard work and rest, and rising ambient heat — the predictive controller extended the time to oxygen depletion by 24.6, 33.7 and 18.7 per cent over a conventional PID controller, while keeping peak carbon dioxide and core temperature lower.
What it does not claim
The results are simulation only, described by the paper as preliminary, with no hardware or human testing. It is frank that a single shared atmosphere is marginal for anything beyond short deployments: the controller is necessary but probably not sufficient, and a separated breathing loop is the path to fielded hardware. It assumes one sortie with fresh consumables, and calls for a fail-safe fallback to fixed set points and for every decision to be logged.
Where it sits
Written at Galactic Bioware with Nicholas Johnston. It is the starting point for the later paper on exact viability and gas allocation, which takes a venting atmosphere in the abstract and solves it exactly.