Synthetic flights
See the 2,200 simulated flights, with known truths and their loggers’ flight logs, on which readings and 90% intervals will be tested.
In short
- What it models: 2,200 flights of the ten public RocketPy designs that HPR Sim, FusionSpace HPR’s simulator, validates against, in three sets on fixed seeds. Each flight draws its motor, drag, wind, rail angle, site and air, and one of six simulated loggers like those FusionSpace HPR reads. The logger writes what its sensors would, each with its own noise terms. A fourth set of 2,000, the calibration set, fits the readings’ 90% intervals and isn’t in the census (The sets). Each flight’s truth is HPR Sim’s trajectory, every 0.01 s, with its dynamic pressure and Mach number.
- Sources: sensor datasheets and public logs (the sensor noise note). The port’s error is from a wind-tunnel test (the Mach error note). The ejection pulse is a public log’s, and the GNSS lag that of hobby receivers in boost (the GNSS note). The wind’s limit is the NAR’s high-power safety code. Some terms are assumed (The noise, term by term; evidence row G32).
- How well it is validated: not against real logs. The noise follows datasheets and a few public logs, and the flights are HPR Sim’s own. So a check on them shows how well a method recovers a known trajectory, not how real flights behave. Each noise term’s test holds it to its stated value and fails when it is changed. That shows the generator draws what it says, not that the terms are right. A flight drawn again is the same to the bit, on one platform. No test yet compares the terms together with a real log of the same family (#725; evidence row G32).
- What it leaves out: the bay’s vent lag, a temperature term, clock drift, vibration and the logger’s own speed column, and flights at Mach 1.18 or more, which none of the 2,200 reaches (What it leaves out).
Why simulated flights
A real log has no truth beside it. The altimeter’s own apogee comes from the same pressure trace that a reading takes it from. So the two agreeing shows they read the trace alike, not that either is right (Flight-log readings). To learn how often a 90% interval holds the truth, the truth must be known, on many flights.
Each reading and fit is to state a 90% interval. ADR-257, calibrated intervals asks that each be tested on at least 1,000 flights whose truth is known. It must miss about 5% of the time on each side. The sets were fixed before any interval was written. By rule, no interval is tuned on the flights it is judged on (ADR-280, the synthetic flights first).
A simulated flight knows its truth, but a check on it is only as good as its noise. A model that adds white noise alone flatters an interval. Real barometer noise is correlated from sample to sample, so its errors don’t average away as fast (the smoothing note §6). So most terms here come from a measurement or a datasheet. The few that no source states are said to be assumed where they are described.
A check on these flights doesn’t show how well HPR Sim predicts a real flight. That is M7.3, a flight against its simulation’s question, on real logs.
The flights
The sets
The census’s three sets, each on its own seed, written down before any method was tuned:
| Set | Flights | Seed number | Model flown |
|---|---|---|---|
| Tuning | 200 | 2026101700 | 100 of A’s and 100 of B’s, alternating |
| A | 1,000 | 2026101701 | HPR Sim’s |
| B | 1,000 | 2026101702 | its drag, wind and thrust reshaped |
By rule, anything fitted or tuned uses the tuning set, and sets A and B only judge. A fourth set,
the calibration set, fits the points that scale the readings’ 90% intervals
(The 90% intervals). Its 2,000 flights are drawn as the tuning
set’s are, alternating A’s model and B’s, on seed 2026102900. That seed was fixed on
October 29, 2026, after the readings and the smoother were tuned and before any reading’s interval
existed. The census and the smoother’s check don’t fly it. Flight
of a set draws each group of inputs (its design, motor, drag, wind, rail, site, model and
logger) from its own stream. The stream is keyed by the set’s seed, and the group. So a
set’s first flights stay the same if it grows, and an input added to one group moves no other group.
Set B flies a model that differs from HPR Sim’s on purpose, so that a method can’t pass by knowing HPR Sim’s model:
- The drag rises differently through Mach 1. The zero-lift drag is multiplied by , where . Here is between −0.2 and 0.2 and is between −0.05 and 0.05.
- The wind grows with height as a power law, its exponent 0.14 to 0.20, the drawn speed at 18.3 m (60 ft), after NASA/TM-2008-215633 §2.2.5.2. That report ties the larger exponents to lighter winds; here the exponent is drawn on its own. Set A’s wind is the same at every height.
- The thrust curve is tilted: its thrust times , with between −0.3 and 0.3, then scaled back to the same total impulse.
What a flight draws
| Input | Range | Why |
|---|---|---|
| design | one of the ten | the RocketPy designs, each with its example’s parachutes; Cavour’s has none, so it lands without one, as on a failed recovery |
| total impulse | 0.94 to 1.06 times | its half-width twice the 2.5 to 3.1% spread (1σ) of repeat certifications (the dispersion note §2); drawn evenly, its own spread is 3.5%, wider |
| burn time | 0.90 to 1.10 times | its half-width near the 10.7% spread (1σ) of the one burn-time batch found, of small black-powder motors (§2); drawn evenly, its own spread is 5.8%, narrower |
| zero-lift drag | 0.8 to 1.2 times | its half-width twice OpenRocket’s 10% default; drawn evenly, its spread is 11.5%, narrower than the 16% coast fits suggest (§4) |
| wind | up to 20 mph (8.94 m/s), from any direction | the safety code’s limit |
| rail | 0° to 10° from vertical, toward any direction | inside the code’s 20° |
| site | 0 to 1,500 m (0 to 4,921 ft) above sea level | sea-level fields to the high deserts |
| air | the 1976 standard, −15 to +15 K | a day off standard |
A range’s half-width is its largest departure from 1. Drawn evenly over ±h, a value spreads by h/√3 (1σ), the spread the rows compare with their sources’.
The loggers convert pressure to height by the standard atmosphere, so a day off standard is a real error in their heights. On a cold day a barometer reads high. The readings must carry it.
The truth is the flight’s center of mass, every 0.01 s, and its events: liftoff, rail exit, burnout, apogee, each charge and deployment, and landing. Its heights are above the ground; a reading’s truth is the height above where the center of mass stood at liftoff. An accelerometer would feel the specific force along the rocket’s axis, at the nose tip where HPR Sim records the acceleration. Under a canopy the rocket is taken to hang vertically.
The loggers
A flight draws one of six profiles, each like a family of real loggers, at that maker’s logged rate and step. Its log is tagged with that family’s format, its logger named as synthetic.
| Profile | Rate | Pressure written as | Accelerometer | GNSS |
|---|---|---|---|---|
| PerfectFlite | 20 Hz | height, whole feet | none | none |
| RRC3 | 20 Hz | pressure, 10 Pa steps | none | none |
| Eggtimer | 20 Hz | height, whole feet | none | none |
| AltOS | 100 Hz climbing, 10 Hz after | pressure, 1 Pa steps | high-g | 1, 5 or 10 Hz |
| 500 Hz IMU | 500 Hz | pressure, 2 × 10⁻⁵ atm steps | low-g, ±32 g | none |
| GNSS tracker | 1, 5 or 10 Hz | none | none | yes |
The height a profile writes is its pressure’s height in the 1976 standard atmosphere, less the mean of its samples on the pad. The RRC3 uses the NOAA formula instead, as its maker’s software does (the height formulas note). The two differ by about 0.83 m (2.7 ft) in 2 km (6,562 ft).
Two loggers and stages
Four more sets check two logs on one timeline. None joins the census:
| Set | Flights | What it flies |
|---|---|---|
| Two-logger calibration | 2,000 | drawn as the calibration set is, to fit the offset’s 90% interval |
| Two-logger A | 1,000 | set A’s model |
| Two-logger B | 1,000 | set B’s model |
| Staged | 200 | the design synthetic-two-stage-75mm-54mm, a logger in each stage |
Each flight carries two loggers, drawn independently from the table above, their noise on streams of their own, and each log’s clock moved by up to 300 s either way, as a board’s is by when it was switched on. A flight’s truth says where each clock stood, so an offset can be held to it.
A staged flight’s booster burns out, the stack parts 0.5 s later and the sustainer lights 0.5 s after that. The sustainer comes down under a canopy. The booster tumbles from the split, as HPR Sim flies one: a point mass with its tumble’s drag (staging). Its truth is that descent on the same 0.01 s grid, from the integrator’s dense output, so these flights say how well two stages’ liftoffs line their clocks up, not how well a booster’s fall is modeled.
The noise, term by term
The pressure
The logged pressure is the true pressure plus seven terms, then rounded to the logger’s step. The sections are those of the sensor noise note unless another note is named:
| Term | What it is | Drawn per flight | Source |
|---|---|---|---|
| offset | the sensor’s absolute error, which the pad’s zero takes out of the heights | ±100 Pa for an MS5607 and PerfectFlite’s; none for a Bosch-class part | §2, §6 |
| gain | an error proportional to the change from the pad | ±0.25% for an MS5607, ±0.1% for PerfectFlite, ±0.08% for a Bosch-class part | §1, §2, §6 |
| drift | a slow change over the flight | 1 to 3 Pa over 30 to 300 s, either way | §3 |
| fast correlated noise | the noise measured on real logs’ pads | 0.5 m (1.6 ft) at 50 Hz, correlation 0.5 | §3, §6 |
| slow correlated noise | the noise that averaging over a second keeps | 0.24 to 0.29 m (0.79 to 0.95 ft), time constant 0.65 to 1.26 s | §3 |
| white noise | the datasheet’s | 2.3 Pa (RRC3), 2.4 to 13 Pa (AltOS), 0.2 to 6.6 Pa (Bosch class); none for PerfectFlite | §6 |
| port error | the static port’s error near Mach 1 | the port 1 to 5 calibers behind the nose’s shoulder, assumed | the Mach error note §1–2 |
A charge adds an eighth, the ejection pulse, at each charge’s time.
What each profile’s part gives. An MS5607 zeroed on the pad holds ±1.0 mbar from 0 to 50 °C, stated over 700 to 1,100 mbar (§2). Read as a gain, that is ±0.25% of the change from the pad. That is an inference, as the datasheet doesn’t say how the error grows with the change. PerfectFlite names no part, so its profile takes an MS5607’s offset, which is an assumption. It takes its maker’s stated ±0.1% of the height as a gain (§1). It has no white noise beyond the noise measured on real logs. Eggtimer and Featherweight name no part on most of their boards (§1). So the Eggtimer and 500 Hz IMU profiles take a Bosch BMP3xx or DPS310-class part, an assumption. Its white noise is 0.2 to 6.6 Pa by its setting, and its gain is ±0.08% relative (§2, §6). The offset stays in the logged pressure, but a height zeroed on the pad loses it. What is left after the zero is the gain and the drift.
The fast correlated noise is a first-order autoregressive process, a red noise: each sample keeps part of the one before. On the pads of five Blue Raven logs at 50 Hz the spread is 0.43 to 0.62 m (1.4 to 2.0 ft). The correlation from one sample to the next is 0.40 to 0.54 across six windows of those logs. Those windows are on the pad or at rest (the sensor noise note §3). The spreads are upper bounds that mix in wind and handling. The model takes 0.5 m (1.6 ft) and 0.5 for every barometer, and adds each part’s white noise. It is drawn as an Ornstein-Uhlenbeck process, a red noise in continuous time, so the correlation follows the time between samples, whatever the rate:
With 0.5 at 0.02 s, . A 20 Hz logger’s samples then correlate by 0.177, a 100 Hz logger’s by 0.707 and a 500 Hz logger’s by 0.933. The spread in pascals is the height’s times . At a 500 m (1,640 ft) site in standard air, 0.5 m (1.6 ft) is 5.72 Pa.
The slow correlated noise is a second term of the same kind, but slower. Its spread is 0.24 to 0.29 m (0.79 to 0.95 ft) and its time constant 0.65 to 1.26 s, each drawn per flight. Ten minutes of a BMP085 at rest gave it (Sabatini and Genovese 2013, Tables 1–2; the sensor noise note §3). The fast term loses its correlation within a tenth of a second. So a reading that averages over a second would see it shrink almost away. The slow one keeps most of its spread over that second. One part measured on one bench is taken for every barometer: an assumption.
The port error is the static port’s error where the pressure along the body changes near the speed of sound. It comes from a wind-tunnel test of the static pressure along an ogive-cylinder from Mach 0.7 to 1.2. The test is Czarnecki and Monta’s (NASA TN D-3516, 1966, Fig. 5; the Mach error note §1–2). The port reads off the true pressure. is read from the note’s trace of the report’s figure, to about ±0.01. It is read at the port’s distance behind the shoulder, in calibers, and the flight’s Mach number. Outside the table it is held at its end values. One caliber back at Mach 1, .
The ejection pulse has roughly the shape of the one on the public Pnut log that Debrief ships. It has seven samples 0.05 s apart from the charge. The height reads 8 m (26.2 ft) low at the charge and 16 m (52.5 ft) high 0.1 s after it. From 0.25 s it reads 3 m (9.8 ft) low, then true again. The shape is scaled by 0 to 2 per flight, a range assumed, as one log gives one pulse. It is turned into pressure at the density where the charge fires. No test yet compares it with real logs’ pulses (evidence row G32).
The acceleration
The logged acceleration is the truth’s axial specific force times , plus a per-flight bias of up to 0.05 g and white noise. It is clipped at the part’s range and rounded to its step. No source states a bias left after the pad’s zero. So the 0.05 g is assumed (the sensor noise note, What this suggests; evidence row G32):
| Part | Noise density | Step (mg, g) | Range, g | Scale , % |
|---|---|---|---|---|
| high-g, ADXL375 class | 5 mg/√Hz | 49 mg | ±180 g | ±10% |
| low-g IMU axis | 0.16 to 0.4 mg/√Hz | 64/65,536 g | ±32 g | ±3% |
Each sample’s noise is the density times . That is the noise bandwidth of a single-pole filter at half the rate the logger samples at while climbing. That bandwidth is an inference in the sensor noise note §4. The AltOS profile’s high-g part at 100 Hz has 44.3 mg (0.434 m/s²) per sample. It has the same after apogee too, when it logs at 10 Hz. The part’s filter doesn’t change with what the logger writes. Past 32 g the low-g part clips: the high-range part that takes over on a real IMU isn’t modeled.
The tuning set’s 500 Hz IMU flights, 38 of them, are also logged with the low-g part set to ±8 g. The steps are 16/65,536 g. The BMX160 and the MPU-6000 offer that full scale on the same 16 bits, as their datasheets state (the sensor noise note). The part holds its top value past the range. The noise and the clock are the same as on its own log. Of these 38 flights, 28 have a true top that reaches 8 g and 10 don’t, as the synthetic report counts. So the readings’ rule for a top clipped at its part’s range is held on flights that reach it (Flight-log readings). At ±32 g none does.
The GNSS fix
A hobby GNSS receiver’s navigation filter follows the rocket like a second-order loop. Under a steady acceleration it settles behind by
where is the loop’s natural frequency, drawn from 0.25 to 1.0 rad/s, with damping . No receiver’s loop has been measured against a truth. About 0.5 rad/s is near the 0.47 to 0.69 rad/s of the GNSS note’s simulated filter. That filter fell 103 to 261 m (338 to 856 ft) behind in boosts of 5 to 15 g, an inference (the GNSS note §4). One flyer reported a 440 m (1,444 ft) lag, its acceleration not given. That is larger than any of the simulated lags (§3). Drawn evenly, two in three loops lie above 0.5 rad/s and lag less than 0.5 rad/s would. So the draws err toward a fix more accurate than a real receiver’s. The damping is the usual choice for such a loop, assumed (evidence row G32). At 0.5 rad/s, a steady 50 m/s² would leave the fix 200 m (656 ft) behind. A 2 s boost ends long before the loop settles, so the lag peaks after burnout. Each axis has 3 m (9.8 ft) of white noise, a fixed value, assumed. No fix is written for 0 to 10 s from liftoff and from burnout. Each span is drawn per flight, as the GNSS note suggests. That is an assumption. The fixes written carry no flag that they lag.
The clock
The log starts 2 to 10 s before liftoff and ends 2 to 10 s after landing, each drawn. The AltOS profile changes from 100 Hz to 10 Hz at its first sample after apogee.
A worked example
An AltOS-like logger is launched from a sea-level pad. It flies through Mach 1 at 1,000 m (3,281 ft) above sea level, a geometric height, on a standard day. Its port is one caliber behind the nose’s shoulder.
- The air. The standard atmosphere there has and density 1.112 kg/m³.
- The dynamic pressure at Mach 1 is , which is 62,913 Pa.
- The port error is , here , which is −5,033 Pa: the port reads low.
- The height it writes. 84,843 Pa is 1,473 m (4,833 ft) in the standard atmosphere. So the logger reads 473 m (1,552 ft) too high while the rocket is near Mach 1. Three calibers back the sign flips: reads 1.9 kPa high, and the height low.
- The noise beside it. The fast correlated noise is 0.5 m (1.6 ft) at the pad, 6.01 Pa at sea-level density. The slow noise is 2.9 to 3.5 Pa and the white noise 2.4 to 13 Pa. Near Mach 1 the port’s error is about 840 times the fast correlated noise and 387 to 2,097 times the white. That is why a reading that leans on a barometer through Mach 1 must say so.
Validation and limits
Checked against
Each noise term, on a truth built for it. The tests build a truth that stands still, or climbs at a constant acceleration, with only the term under test switched on. They hold the log to it. They check the offset as the residuals’ mean, the drift as their slope, and the white noise’s spread to within five standard errors. They check the fast and slow correlated noise’s spreads and correlations, and the gain on a 2 km (6,562 ft) climb. They check the port’s at a measured point and the pulse at 0.15 s. They check the accelerometer’s scale, bias, spread (after apogee too) and clip, and the GNSS lag and dropouts. They check the log’s span and rates, and its values lined up with the truth by the drawn pad span. Each drawn noise term is checked again changed, and must then fail. The offset, drift, white noise and gain are tried at half and one and a half times their values. The others are zeroed, halved, doubled or moved (the port to 3 calibers, the GNSS loop to 1 rad/s). The fix rate and the spans are set to other values, which the log must follow. Each part’s stated ranges are pinned to the values on this page, so a changed constant fails too. The fixed 3 m (9.8 ft) GNSS noise is held to its value on each axis. The census is a file in the repository (synthetic-flights.md). It counts each set’s flights by design and logger. It also gives the ranges of eight quantities: apogee, top Mach number and speed, flight time, wind, rail angle, site elevation and the air’s offset.
What the census shows. Every one of the 2,200 draws flew. Apogees above the pad run from 412 to 4,810 m (1,352 to 15,781 ft), and top speeds from Mach 0.269 to under 1.18. Values are kept to three significant figures, the digits every platform shares, each range’s ends rounded outward. Each night CI flies every flight again on macOS, Windows and Linux and holds the census to them.
What it leaves out
- Fast flights. None of the 2,200 reaches Mach 1.18 (the census rounds its greatest up). So from there to the Mach 3.5 HPR Sim is built for (the operating envelope) there is no synthetic flight.
- The bay’s vent lag. A bay’s pressure lags the outside air through its vent holes. In an invented flight, three invented bays vented by the makers’ rules moved the apogee by at most 2.1 cm (0.07 ft). They lagged the height by 0.4 to 135 m (1.3 to 443 ft) through Mach 0.9 to 1.07. That is an inference (the Mach error note §5), and no hobby bay’s lag has been measured.
- A temperature term. Not drawn. Below 4 km (2.5 mi) above the pad it moves the apogee by up to 0.72 m (2.4 ft) on every site and day drawn. The most is at that height above a 1,500 m (4,921 ft) pad on a +15 K day. From a sea-level pad on a standard day it is 0.60 m (2.0 ft) there. Both take a sensor 0.6 K off its pad value at apogee. No altimeter log that peaked that low moved more than 0.56 K. The census reaches 4,810 m (15,781 ft) above the pad. Logs that peaked higher warm by +2.48 K at their 90th percentile. At the census’s top, above a 1,500 m (4,921 ft) pad on a +15 K day, that makes the term about 3.2 m (10.6 ft). The largest logged warming, +6.73 K on a faster and higher flight, would make it about 8.8 m (29 ft). All of these take 8 Pa/K, which is no ceiling. It is the mean slope over 25 K that the MS5607’s error bands allow below 700 hPa, where each of these apogees sits. At higher pressure it is 6 Pa/K, and a shorter span allows a steeper slope (bay temperature §2, §5 and §7, an inference).
- Clock drift. The logger’s clock is taken as exact.
- The logger’s own speed column. Each logger filters its own speed its own way, so the synthetic logs carry none.
- The rocket’s attitude in the accelerometer. Only the axial axis is logged, at the nose tip rather than where the logger sits, and under a canopy the rocket hangs vertically. Each parachute opens at full size after its lag, as the RocketPy examples define them, so a main’s opening is a sharp spike.
- Vibration. The accelerometer’s noise is the same through the boost as on the pad; a motor’s vibration isn’t drawn.
- A GNSS error that wanders. Each fix’s noise is drawn afresh, so an average of fixes is quieter than a real receiver’s, whose error drifts over seconds to minutes.
- A sharp port jump. The port’s error is read linearly between the tunnel’s Mach numbers, 0.1 apart near Mach 1. So its jump is spread over that much speed, and a real port’s may be sharper.
- A part’s stated noise, where it is larger. The noise from the density can understate a part’s own stated figure. The MPU-9250’s 8 mg at 94 Hz is 2.2 times what the density gives (the sensor noise note §4). The low-g profile may be quieter than a real IMU.
Tests and references
Tests
hpr_forensics::synthetic: every draw inside its range and filling it, in every set. Each set is on its own seed. One flight drawn again is the same to the bit, whether flown alone or after another that shares its drag table.hpr_forensics::synthetic::noise: the correlated noise’s spread and correlation at 10, 50 and 100 Hz, and the port table at its measured points. It also checks the pulse’s shape and the GNSS loop’s steady lag.hpr_forensics::synthetic::logger: each drawn pressure, acceleration and GNSS term held, and failing when changed. It also checks the fast noise’s correlation at 100 and 500 Hz and the slow noise’s time constant. It holds the GNSS noise to 3 m (9.8 ft) on each axis. Last, it checks the log’s span, each part’s rate, each profile’s step and channels, and the height formulas.hpr_forensics::synthetic::held: the 500 Hz IMU’s part set to ±8 g, with its range stated to the readings. A true top past it is read as clipped at the range, a floor that holds. A part with no lower setting writes no such log.hpr_forensics::synthetic::truth: set B’s drag and tilted thrust, which keeps its impulse. It also checks that set B’s three changes reach its flights’ inputs while set A’s keep HPR Sim’s.cargo xtask synthetic --check, in CI on every platform each night (ADR-149’s amendment, the time budget for CI), and by hand: the census reproduced. CI’scensusjob flies the sets in two steps,--sets tuning,aand--sets b,calibration. Each compares what its sets make with their sections of the committed report, so every flight is flown and compared on every platform. A test readsci.ymland fails if a set isn’t on exactly one line of a job that runs on all five platforms, or if that job runs on any condition but a full run’s, nightly or started by hand.cargo xtask synthetic --slice, on pull requests on all five platforms and inscripts/gate.sh: every 21st flight of each set (202 of 4,200) flown again, held to the census’s rules for one flight and to thesynthetic-flights.slice.jsonthe whole run writes, by the byte on the Mac’s everyday build that writes it and by the word elsewhere (a flight’s numbers can move between builds). A test changes one flight’s truth, then its log, and checks that its lines change.
References
- K. R. Czarnecki, W. J. Monta, NASA TN D-3516 (1966), as the Mach error note cites it.
- Sabatini and Genovese, Sensors 13(11):15692 (2013), as the sensor noise note cites it.
- NASA/TM-2008-215633, Terrestrial Environment (Climatic) Criteria Guidelines for Use in Aerospace Vehicle Development, §2.2.5.2.
- The National Association of Rocketry, High Power Rocket Safety Code.
- The research notes this page draws on: sensor noise, the Mach error, GNSS in boost, height formulas, smoothing methods, dispersion sigmas.