Hot fire, simulators and glossary¶
In short
- A hot-fire test is a scripted sequence: chill-down, ignition, an open-loop ramp-up, a closed-loop phase lasting seconds, and shutdown. Real interaction is measured in minutes per year.
- DLR's LUMEN model is EcosimPro with ESA's ESPSS library, validated against hot fire to a few percent. ESPSS needs ESA's approval to use. The challenge's generalised simulator needs no EcosimPro licence and runs at about real time (organisers, Oct 2026).
- The glossary at the end links each term to its chapter.
What a hot-fire test looks like¶
A test is a scripted sequence (thesis p. 9–10):
- chill-down of lines and pumps;
- ignition;
- an open-loop ramp-up to a stable point;
- an optional closed-loop phase;
- a scripted shutdown.
Closed loop is only handed over once pressures and flows are established. In DLR's two RL hardware tests the first 15 s and 18 s were open-loop (Dauer et al. 2025, Fig. 4).
Real data is scarce:
- Across the 2024 and 2025 engine campaigns LUMEN had 17 ignitions and "an accumulated hot-fire test duration of over 20 minutes" (Kurudzija et al. 2025, p. 4).
- The successful RL hardware run was "over 16 s" of closed loop (thesis p. 117).
- The run shown at the RL Bootcamp 2026 tracked for about 15 s before a sensor failure ended it (livestream, §7.2).
In RL terms, the real environment offers on the order of minutes of interaction per year, with episode termination controlled by safety logic you do not own.
flowchart LR
A["Chill-down"] --> B["Ignition"] --> C["Open-loop<br/>ramp-up"] --> D["RL closed loop"] --> E["Shutdown"]
C -. redline .-> X["Safe shutdown"]
D -. "redline, sensor fault" .-> XOne hot-fire test as an RL episode. In DLR's RL tests the open-loop ramp-up took 15–18 s and the best closed-loop phase lasted over 16 s. Only the closed-loop block is the agent's; everything else is scripted by the test team.
How DLR's simulator is built, and how good it is¶
The LUMEN model is written in EcosimPro, a commercial differential-algebraic modelling tool, with ESA's ESPSS propulsion libraries: versions 6.4.0 and 3.6.0 in the thesis (p. 60). The components (pipes, pumps, turbines, valves, cooling channels, chamber) are calibrated individually on sub-system tests, and then the whole engine is validated on hot-fire data. Against a single post-start-up test run the validated model reaches (Kurudzija et al. 2025, Table 2):
| Quantity | MAPE over test | Max steady-state error |
|---|---|---|
| Chamber pressure | 2.4 % | 3.3 % (2.4 bar) |
| Mixture ratio | 3.5 % | 5.2 % (0.2) |
| Fuel pump outlet pressure | 2.8 % | 4.8 % (5.2 bar) |
| Oxidiser pump outlet pressure | 4.1 % | 7.4 % (3.2 bar) |
| Fuel injection temperature | 1.9 % | 4.8 % (15.1 K) |
| Cooling-channel outlet temperature | 2.1 % | 6.0 % (25.4 K) |
| Turbine inlet temperature | 2.0 % | 4.0 % (23.2 K) |
| Fuel turbopump speed | 1.5 % | 3.1 % (1292 rpm) |
| Oxidiser turbopump speed | 1.8 % | 2.6 % (684 rpm) |
{
"$schema": "https://vega.github.io/schema/vega-lite/v6.json",
"title": {"text": "LUMEN model against hot-fire data", "subtitle": "Validated EcosimPro model on one post-start-up test run (Kurudzija et al. 2025, Table 2)"},
"width": "container", "height": 270,
"data": {"values": [
{"q": "Chamber pressure", "mape": 2.4, "worst": 3.3},
{"q": "Mixture ratio", "mape": 3.5, "worst": 5.2},
{"q": "Fuel pump outlet pressure", "mape": 2.8, "worst": 4.8},
{"q": "Oxidiser pump outlet pressure", "mape": 4.1, "worst": 7.4},
{"q": "Fuel injection temperature", "mape": 1.9, "worst": 4.8},
{"q": "Cooling-channel outlet temperature", "mape": 2.1, "worst": 6.0},
{"q": "Turbine inlet temperature", "mape": 2.0, "worst": 4.0},
{"q": "Fuel turbopump speed", "mape": 1.5, "worst": 3.1},
{"q": "Oxidiser turbopump speed", "mape": 1.8, "worst": 2.6}
]},
"encoding": {"y": {"field": "q", "type": "nominal", "sort": null, "title": null, "axis": {"labelLimit": 260}}},
"layer": [
{"mark": {"type": "rule", "strokeWidth": 2, "color": "var(--viz-axis)"},
"encoding": {"x": {"field": "mape", "type": "quantitative", "title": "Error (%)", "scale": {"domain": [0, 8]}, "axis": {"tickCount": 8}}, "x2": {"field": "worst"}}},
{"transform": [{"fold": ["mape", "worst"], "as": ["kind", "err"]},
{"calculate": "datum.kind === 'mape' ? 'MAPE over the run' : 'Worst steady-state point'", "as": "measure"}],
"mark": {"type": "point", "filled": true, "size": 90, "stroke": "var(--md-default-bg-color)", "strokeWidth": 2},
"encoding": {"x": {"field": "err", "type": "quantitative"},
"color": {"field": "measure", "type": "nominal", "scale": {"domain": ["MAPE over the run", "Worst steady-state point"]}, "legend": {"title": null, "labelLimit": 300}},
"tooltip": [{"field": "q", "title": "quantity"}, {"field": "measure"}, {"field": "err", "title": "error (%)"}]}}
]
}On the specific run used for the RL hardware demonstration, model–experiment errors were larger: mean 4.1 % on \(p_{cc}\), 4.9 % on coolant flow, maximum 10.6 % (thesis Table 6.3). A 2026 preprint describes a "representative upper-stage rocket engine based on the LUMEN" model, with errors "below 5 % over the full operational envelope" (Kurudzija et al. 2026, abstract only). It is plausibly the generalised simulator behind the challenge.
ESPSS itself is ESA property: "Any entity interested in using ESPSS needs prior approval from ESA" (EcosimPro brochure).
How this compares with the surrogate in the Engine Lab:
| DLR's LUMEN model | This repo's surrogate | |
|---|---|---|
| Tool | EcosimPro + ESPSS (commercial, ESA-licensed) | about 250 lines of Python, mirrored in JavaScript |
| Components | 1-D pipes, maps for pumps and turbines, cooling channels with 1-D flow and 3-D walls, all valves | lumped: 10 states, 2 actuated valves, 4 frozen |
| Calibrated on | component tests and hot fire (17 ignitions) | DLR's published gains, settling times and overshoot at one operating point |
| Accuracy | 2–4 % MAPE against hot fire | against DLR's model at 40 bar: static gains within about 13 %, except TOV on coolant temperature (+20 K against +12 K); TFV settling times within 25 % |
| Speed | DLR trained at about 1 M steps per day on 10 instances (thesis p. 83) | about 3,000 control steps per second per CPU core |