DFT Labs investigates whether geometry-aware physical models with learned corrections, developed under Data Field Theory (DFT), can improve aerospace computation under explicit accuracy and resource constraints. Current work covers storm-time atmospheric density, synthetic reentry heating and numerical fault monitoring. The program has built benchmark and verification infrastructure; an incremental benefit from DFT has not yet been established. We seek engineering partners to define operational decisions, tolerances and comparisons against existing methods.
The workloads this program studies can face four constraints together. Observations may be limited: one mission's telemetry, or a few hundred trajectories. The operating point can shift away from the data by design, because a storm, a new entry corridor or a component count never flown may be absent from the record. Compute may be constrained, on the ground by latency and on board by a power budget that can be a few watts under radiation. And errors must be handled: a wrong field value has to be caught by something an engineering review can inspect. Much of this computation runs on the ground today; whether any of it should run on board is a question for a later evaluation.
The first case comes from a documented event. On February 3, 2022 a batch of 49 newly launched low-orbit satellites met two minor-to-moderate geomagnetic storms during orbit raising, and 38 were lost to storm-driven drag. Published analyses attribute the loss to a density increase at the deployment altitude that the empirical atmosphere models in common use under-predicted (Fang et al., Space Weather 20, e2022SW003193, 2022; Kataoka et al., J. Space Weather Space Clim. 12, 41, 2022; Berger et al., Space Weather 21, e2022SW003330, 2023). Forecasting density along an orbit from the indices known at issue time is the computation those operations depend on.
Active research, illustrated by the density and heating benchmarks below. The question is whether a field model on the problem's own geometry, with a learned correction, beats generic learned models where data are scarce and the operating point shifts.
A measured objective: accuracy, latency, memory and energy are assessed together, against the small models that already fly. No flight-device measurement exists yet.
A separate feasibility direction with its own incumbents and its own evidence requirements. A bounded receiver study is in progress; nothing in the benchmarks below bears on it.
The cases investigate geometry-aware models and learned corrections through case-specific equations and numerical schemes: a field on the problem's own geometry, an orbit shell or a capsule surface, evolved by a classical operator plus a learned correction. In each case the comparison isolates the proposed correction and tests it against classical and generic learned alternatives; with the correction removed, each model is its own classical baseline.
Two hypotheses are under test:
That the field structure gives better accuracy than generic learned corrections where observations are limited and the operating point leaves the data, at a cost within the resource budget of the host.
That a curvature-based numerical indicator, derived from the integrator's energy law, adds useful warning information beyond conventional checks. Energy and spectral checks are known to accept some wrong trajectories; whether this indicator does better is the question, and it has no demonstrated assurance role yet.
Each case names a proposed operational decision and quantity, and baselines that receive the same data: the operational model, the classical linear model, and a generic learned correction. Decision margins and resource budgets will be fixed before confirmatory evaluation; development results remain exploratory. A baseline that matches the candidate counts against the candidate.
DFT-specific benefit has not yet been demonstrated. Candidate comparisons on the density and heating benchmarks remain incomplete. The first fault-monitoring pilot, on the classical arm, failed its performance and cost criteria. Numerical verification tests have been run; a documented timestep-tolerance issue remains to be resolved before accuracy-matched cost conclusions, and independent verification of the solver and derivations is pending.
Proposed decision: a density forecast along the orbit from the indices known at issue time, for launch holds, drag make-up burns and conjunction screening; approval of the decision and its tolerance by a workload owner is outstanding. Data: accelerometer-derived densities from a public three-satellite mission and public geomagnetic indices. The February 2022 storm is the development case. A later storm is the holdout, held in a separate account under a custodian role, with access revoked from the program's working and administrative roles and every access logged; administrative recovery of access remains possible and would itself be logged. Incumbents are trained at scale on the development data; candidate comparisons are running.
Proposed decision: the heating field over a thermal-protection surface from sparse sensors. We have not secured a suitable measured dataset for this benchmark, so it is synthetic with stated classical physics: a sphere-cone forebody, ballistic entries, classical heating with thin-shell conduction, a hidden coupling field, and sparse noisy observations. 320 entries over development, in-distribution and out-of-distribution splits. Incumbents are trained; the candidate comparison is in progress.
Proposed decision: flag a wrong field computation before it is acted on. Software fault injection into the classical sphere solver: 11 fault families in 35 variants at 8 injection points, 5 detectors against 3 conventional baselines (duplicate execution, checksums, range checks). Scored on miss rate at a bounded false-alarm rate, warning time and overhead. The first pilot, with the classical arm only, is complete and failed its declared criteria: the indicator added no detections beyond the conventional checks and its overhead exceeded the cap. The candidate arm has not been tested. An injected-fault frequency is not an operational failure rate.
| Case | Proposed decision | Data | What exists today | Outcome so far |
|---|---|---|---|---|
| A. Storm-time density forecasting | Density forecast along the orbit for launch holds, drag make-up and conjunction screening (owner approval outstanding) | Real; public mission and indices | Benchmark; development split; holdout under restricted, logged custody; incumbents at scale | Partial development results; no completed confirmatory comparison |
| B. Reentry surface heating | Heating field over a thermal-protection surface | Synthetic; stated classical physics | Verified generator; 320 entries over three splits; incumbents trained | Candidate comparison in progress; no completed confirmatory comparison |
| C. Numerical fault monitoring | A wrong field computation flagged before it is acted on | Software fault injection | Fault catalog; injection harness; first pilot complete | Classical-arm pilot failed its criteria; candidate arm untested |
Numerical verification tests (manufactured solutions and energy laws) have been run on the classical schemes and reproduced in a fresh environment; a documented timestep-tolerance issue remains to be resolved before accuracy-matched cost conclusions. The problem statements rest on a graded evidence base:
1 / 26
2 / 26
3 / 26
4 / 26
5 / 26
6 / 26
7 / 26
8 / 26
9 / 26
10 / 26
11 / 26
12 / 26
13 / 26
14 / 26
15 / 26
16 / 26
17 / 26
18 / 26
19 / 26
20 / 26
21 / 26
22 / 26
23 / 26
24 / 26
25 / 26
26 / 26The running comparisons are reported whichever way they come out. If a workload owner approves the decision and sets the tolerance, one comparison is run on the holdout.
The candidate comparison is reported. The benchmark moves to flight data only if a data agreement is secured.
The candidate's indicator is tested on the monitor under a rule reviewed by an assurance specialist. Any assurance role for it would need that evidence, separately from the accuracy results.
Accuracy, latency, memory and energy of the solver measured together on a flight-class device against the flight baseline. Until then there is no onboard suitability result.
The receiver feasibility study reports against its own incumbents; it is not inferred from the cases above.
The benchmark, the solver, the evidence map and the code, on clearance.