MarsWatch forecasts life-support sensor faults seconds to minutes before they happen — O2, CO2, suit pressure, cooling, battery, and more — giving astronauts and mission control real lead time no current suit provides.
To give astronauts the earliest possible warning of life-threatening suit failures — so that no fault goes undetected until it is too late to respond.
An astronaut is not a replaceable component. Losing one on the surface means losing the years of training, the mission, and every discovery that person was sent to make. The first Martian generation is projected for the 2030s, and the suits meant to carry them are the thinnest margin between a crew and an environment that will kill them in seconds.
NASA's EMU, Axiom's AxEMU, and every other EVA suit in development detects a life-support failure only after a threshold has already been crossed — not before. Suits also degrade far faster than the missions they serve, and the sensing needed to see that degradation coming is largely absent. On Mars, where you cannot call home for help in time, that gap is the whole problem.
Billions are moving into commercial and national EVA programmes right now. The suits are being redesigned; the fault-detection layer inside them mostly is not. That makes this a solvable problem with a narrow window to solve it in.
This is a simplified, illustrative animation of the core idea — the same early-warning behavior demonstrated in our real trained models. Trigger a fault, adjust its severity, and watch the forecast line cross into Critical before the actual sensor value does.
Each sensor runs the same fault dynamics and thresholds its trained model
uses — real horizons, real NASA-sourced limits. The forecast shown is a lightweight
in-browser approximation for illustration; the production forecasts come from the
gradient-boosted models served by server.py.
A pressure suit's entire job is holding a ~35× pressure gap against Martian ambient. Pairing an external sensor with each internal one turns "is this reading bad?" into "is the suit still separating me from Mars?" — a much harder signal to fool. Drag the slider to open a leak.
External ambient figures come from 16 real sols of NASA REMS telemetry (Curiosity's surface weather station), not simulation. Water vapour is compared as absolute mixing ratio, not relative humidity — Martian RH swings from under 1% at midday to near-saturation at night purely from temperature, so an RH-to-RH comparison would invert and fire false alerts.
Continuous, fast-changing signals get a forecasting model. Slow or binary signals get deterministic threshold logic instead — the right tool for each signal, not one hammer for everything. Click a card for details.
Recent readings become five signals: current value, rate of change, trend slope, rolling volatility, and acceleration.
A gradient-boosted model predicts a conservative percentile of the value ahead — deliberately biased toward the risk direction, not just the average outcome.
The forecast (and, for duration-sensitive signals, how long a value has persisted) is compared against NASA-sourced or engineering-spec thresholds.
Astronaut and mission control get a warning while there's still time to act — not just a confirmation that something already failed.
A suit is one of the most constrained engineering environments there is. These are the capabilities the system has to deliver inside those constraints.
What the system must deliver to be worth wearing.
The interactive demonstrations here are simulations, not real telemetry. No Mars EVA suit sensor feed exists to connect to. Sensor values shown in the live demo and the suit-integrity panel are generated in your browser to illustrate system behaviour. They are not measurements from any suit, test article, or flight hardware.
This site does not contain proprietary or production material. It is a public overview of an early-stage research prototype. Implementation details, trained model artefacts, and internal engineering data are deliberately not published here, and nothing on this page should be treated as a specification, a validated performance claim, or a description of a flight-qualified system.
Forecasting models are trained on synthetic fault episodes, because no public dataset of real EVA suit sensor faults exists. Physical thresholds are drawn from NASA standards where a citation exists and from our own engineering specification otherwise. Replacing synthetic data with real measured sensor data is the next milestone, not a completed one.