0D IDT CH₄
1. Overview
Section titled “1. Overview”This model predicts the zero-dimensional (0-D) homogeneous ignition delay time (IDT) of a methane (CH₄) / air mixture. The fuel is fixed as methane; there is no mixture or SMILES input. For a given initial thermodynamic state and dilution, the model returns how long the well-mixed charge takes to ignite.
Ignition delay here is the induction time of a single, spatially uniform control volume: no spray, no walls as a heat sink, and no 3-D engine geometry. The charge is closed and adiabatic at constant volume. Its initial state is , , , and EGR.
Evaluating a detailed kinetic mechanism at every query is too slow for interactive sweeps. Scimlate uses a trained ignition-delay ML model instead: the four inputs below map to a fixed feature order, the model returns ignition delay in seconds.
-
Input: pressure (bar), temperature (K), equivalence ratio , and exhaust-gas recirculation (EGR) in percent.
Single point (importable Scimlate input bundle):
{"format": "scimlate-inputs","version": 1,"model_type": "zero-d-ignition-delay-time-ch4","parameters": {"pressure": "20","temperature": "1000","phi": "1.0","EGR": "0"}}Multi-sweep (
start:end:stepon each axis insideparameters; inclusive ranges expand as a Cartesian product — evaluations):{"format": "scimlate-inputs","version": 1,"model_type": "zero-d-ignition-delay-time-ch4","parameters": {"pressure": "20:30:5","temperature": "900:1100:50","phi": "1.0:1.6:0.2","EGR": "0:15:15"}} -
Output: ignition delay in milliseconds (
ignition_delay), within_rangewhen a row falls outside the training envelope.
| Quantity | Symbol | CLI key | Unit |
|---|---|---|---|
| Pressure | pressure | bar | |
| Temperature | temperature | K | |
| Equivalence ratio | phi | — | |
| Exhaust-gas recirculation | EGR | % | |
| Ignition delay | ignition_delay | ms |
2. Operating range
Section titled “2. Operating range”The training envelope is a soft box: values outside it warn and set in_range to false, but
inference still runs. Predictions outside this box are less reliable.
| Input | Symbol | Unit | Working range |
|---|---|---|---|
| Pressure | bar | ||
| Temperature | K | ||
| Equivalence ratio | — | ||
| Exhaust-gas recirculation | EGR | % |
Structured Arrhenius sweeps in the training set were denser at –; the claimed envelope is the full – interval.
3. Worked example
Section titled “3. Worked example”Stoichiometric methane/air at , , and no EGR:
| Quantity | Value |
|---|---|
| bar | |
| K | |
| % | |
| Predicted | ms |
Raising temperature to at the same pressure, , and EGR shortens the delay to , as expected for an Arrhenius induction time.
A two-pressure, two-temperature, two-EGR sweep at produces predictions (Cartesian product). Each row is one evaluation of the same forward map.
4. Credits
Section titled “4. Credits”Scimlate bills inference from the model’s cartesian_weighted usage contract, not from
analytics counters or output format. Sweep syntax is scalar_csv_range: a scalar is one value,
10,20,30 is a list, and 10:40:5 is an inclusive start:end:step range (10:15 defaults to
step ).
is the Cartesian product of the four sweep lengths. A dimension is active only when its expanded length is greater than one:
| Sweep axis | Weight |
|---|---|
| Pressure | |
| Temperature | |
| Equivalence ratio | |
| EGR |
Missing weights default to . With multi_sweep_factor: 2, a one-axis sweep has
, two axes , three axes , and four axes . After a
successful run Scimlate uses the run’s prediction count to settle .
The rows below match common requests: a lone state; 700:1500:50;
10,30,50 × 0,15,30; the 8-point worked sweep; and 20:30:5 ×
900:1100:100 × 1.0,1.5.
| Run | Active | Unit | Credits | ||||
|---|---|---|---|---|---|---|---|
| Single point | — | ||||||
: 700:1500:50 | |||||||
| × EGR arrays | , EGR | ||||||
| Worked 8-point sweep | , , EGR | ||||||
| × ranges × pair | , , |
5. Accuracy
Section titled “5. Accuracy”No per-prediction interval is emitted. The table is a global residual summary on the 43030-point training set used to fit the model.
| Metric | Value |
|---|---|
| Median absolute relative error | 0.99% |
| MAPE | 1.40% |
| 90th-percentile absolute relative error | 3.12% |
| MAE() | 0.006 |
| Maximum absolute relative error | ≈16% |
These residuals describe hold-in training-set agreement, not a withheld test split or a propagated kinetic-mechanism uncertainty.
6. References
Section titled “6. References”- Model implementation:
carnotai/src/mechanical/fuel_chemistry/scripts/0d_idt_ch4_fwd(CarnotAI repository). - ONNX artifact:
carnotai/src/mechanical/fuel_chemistry/models/0d_idt_ch4_fwd/0d_idt_ch4_fwd.onnx. - Credit contract: CarnotAI
SCHEMA.md(usage) andwiki/model_usage.md(cartesian_weighted). - Display symbols and Arrhenius chart: CarnotAI
wiki/entities/0d_idt_ch4.md. - Training and residual packaging: EulerTrain methane SIDT benchmark (43030 clean points).