Case study · Space / microgravity · SA Space Collaboration and Innovation Fund
One scarce dataset in. A validated engine out.
Round 3 of the South Australian Space Collaboration and Innovation Fund backs the engine behind the ALTDATA Digital Twin: the validation work that proved it holds up, and the flight that will test the one thing it still can’t certify.
Challenge
Microgravity research has an accessibility problem: flight missions are rare, expensive and high-risk, and for most teams the gap between 1G on Earth and microgravity in orbit is a black box.
The tempting move is to train an AI on whatever biological data you can find and call the output a simulation of space. We think that produces confident nonsense. The harder question this grant funded was whether you could build a twin anchored in physics, from genuinely scarce public data, and then prove it against reality rather than against itself.
Solution
We built the engine in two halves. A high-fidelity physics simulator acts as an oracle, generating rigorous ground truth for 3-D biological growth. We then distilled that oracle into a fast AI surrogate, which is the part you actually run.
The critical design decision was the anchor. The surrogate is calibrated against real, openly licensed biological growth data, not against the simulator alone. Without that, a twin can only ever prove it agrees with itself.
Technology
One scarce real biological dataset seeded a calibrated physics oracle. The oracle generated a corpus of virtual experiments across a systematic sweep of conditions, and from that corpus we trained the surrogate. It came out small. Small enough to matter, as it turned out.
Gravity enters as a tunable settling force: remove settling and cultures self-assemble round and cohesive; let it dominate and they flatten into a disk. That is the Gravity Knob, and it is the part of this work that points at space.
Results
Measured, not asserted:
- ~88% fidelity reproducing its physics oracle, head to head.
- 96% of the accuracy of a model trained directly on the real growth data, on a train-on-synthetic, test-on-real benchmark.
- Simulation-to-reality anchor error a small fraction of the growth it predicts, measured against real biological data rather than against the simulator.
- Lean by design. Orders of magnitude faster per sample than the physics simulator it learned from, and small enough to run on a laptop, or on low-power hardware in orbit.
- Calibrated uncertainty on every prediction, with confidence bands that run conservative rather than optimistic. That is the direction you want to be wrong in.
The number we’re proudest of is one that started badly. An early benchmark came in at 68%, well under target. We investigated it rather than tuning it, traced it to a root-cause limitation in the growth law, and redesigned that law. The benchmark rose to 96% while physical calibration simultaneously improved more than fourfold and the necrotic-core structure stayed intact. A metric-hack cannot do that; only a real fix moves all three at once.
Business impact
The grant took ALTDATA from a plausible pitch to a validated engine: the difference between describing a digital twin and running one. That engine is now the foundation under our biomedical Digital Twin work, and its deliberately small footprint keeps a data-sovereign, edge and in-orbit deployment path open rather than forcing everything through a cloud.
It also produced something less tangible and more useful: a validation discipline. We keep a three-bucket ledger that never blurs. What is validated against reality, what is corroborated premise, and what is an open roadmap gap. We ship the gaps too.
Future roadmap
Be precise about the limit. The settling mechanism is validated against real 1G suspension-culture data, which shares the round, low-settling regime with true microgravity. The microgravity-specific delta (3-D biological growth under real µg against a matched 1G control) is not yet acquired, and not yet validated. The engine is built to report that it cannot self-certify there, and it does.
Closing that gap needs real flight data, and that is now funded. The Round 3 project is led by ResearchSat, with ALTDATA and AICRAFT as partners, and takes a proof-of-concept mission to the International Space Station, combining space-flown biological data with edge computing and our models. When those results land, this section gets the answer. Until the mission flies, the gap stays open and we’ll keep saying so.