Hemi airflow and fuel: neural network, VE and injectors

The Gen III Hemi has no mass-air sensor. The file calculates air mass from a VE estimate and a density conversion. Early files can expose a direct VE table. Later files can run a neural network that estimates VE from rpm, pressure ratio, intake cam angle and exhaust cam angle.

Intake air temperature is used in the density conversion after VE, not as a network input. The year narrows the field, but the file decides which strategy is active.

• On a network file, VCM Editor can show a Neural Network section and a training command. The network is a set of weights and biases. It is not a normal VE table that you paste into. The Trainer builds editable surfaces from the model, you change those surfaces, and the tool fits new weights and biases back to them. It does not install an ordinary VE lookup table in place of the active network, and it does not learn on the road.
• Trainer eligibility is listed for select 2010 and later FCA vehicles. A network can be present in a file that the Trainer will not accept. Licensing is separate from vehicle credits. Check the command in the actual file before you plan the job.

What the file has to prove

1. Read the file and save the as-found copy. Record the OS, controller, network switch and VVT status before any edit.
2. Write down the exact labels and units. Airflow, pressure ratio, fuel mass, pulse width and cam angle all need the unit from this file.
3. Check the training command. If the command is absent, do not assume a 2010 and later file is trainable.
4. Pick a small region with good log coverage. Make a small reversible change in the table you believe is active.
5. Repeat the same steady condition. Watch aircharge, pulse width and trim response. Closed-loop correction can hide the mixture change. If nothing responds, check the write and strategy before continuing.
6. Restore the file if the test is inconclusive. Do not stack a second change on a path you have not proved.

The calculation runs in one direction. MAP and barometric pressure give pressure ratio. Rpm and cam position select the region. The airflow model returns an aircharge. The fuel side multiplies that aircharge by a commanded fuel-to-air ratio to get fuel mass. The injector table turns fuel mass into pulse width.

Calculated load is the output that matters beyond fuel. Spark, torque request, transmission pressure and shift feel can use the load estimate. If you correct fuel while the aircharge estimate is wrong, the engine can run at the commanded mixture with the wrong load.

That is why a fuel-only fix can look successful and still drive badly.

1. Check the pressure inputs. Key-on engine-off MAP should match local barometric pressure within the sensor resolution.
2. Check intake air and coolant temperature. A sensor that reads high or low changes air density and warm-up fueling.
3. Check cam angle on a VVT engine. Desired and actual should track on a stock cam and stock file.
4. Log fuel mass and pulse width on the same axis as the injector table. Confirm grams or milligrams before building a histogram.
5. Compare commanded fuel with measured lambda. Keep the wideband channel and the command in the same log.
6. Record calculated load, spark, throttle and torque request beside fuel. An injector-only correction should not be used to disguise a wrong load estimate.

The full channel list is on the logging guide. This page assumes you already log both banks, commanded fuel, wideband lambda, fuel mass, pulse width, MAP and baro.

Chrysler files can express mixture as fuel-to-air, not air-to-fuel. The labels vary by operating system, and some tables are additive. Read the help and the units in the file before you type a number. The math below is a calculation, not a claim that every file uses the same constants.

An additive example makes the unit failure visible. With F/A stoich 0.0688, stoich AFR is 14.5349. Using lambda 0.85 as an arithmetic example, F/A is 0.080941. If the file uses an additive power-enrichment value, the adder is 0.012141.

Typing 0.85 into that adder is a unit error, not a lambda 0.85 command. It adds 0.85 to the stoich F/A and commands F/A 0.9188, or AFR about 1.09:1. That is far richer than stoich.

• A second arithmetic example adds 0.0110 to a stoich F/A of 0.0722. The result is 0.0832, or AFR 12.02. Its lambda of 0.868 only holds if your file's stoich F/A is also 0.0722. Read the stoich value in your own file first. The equivalence ratio is 0.0832 / 0.0722 = 1.1524. Calling 0.868 an equivalence ratio reverses the rich/lean meaning. These are arithmetic inputs, not verified stock constants.
• Whether your file uses additive enrichment is a separate question. The additive rule is not confirmed for every OS. Read the table description. If the file uses a multiplier or a lambda target, the adder arithmetic does not apply.

Correct with the multiplier, then check the direction

• At a steady accepted point, fuel multiplier = measured lambda / commanded lambda. If the gauge shows 0.90 and the command is 0.85, the multiplier is 1.0588, or 5.88% more delivered fuel.
• Direction depends on the table. On an active airflow model, a lean error raises the aircharge estimate in that region, which raises fuel. On an injector table that maps fuel mass to pulse width, a lean correction adds fuel and lengthens the pulse. The scalar moves the other way: pulse width is fuel mass divided by flow rate. Reducing that scalar adds fuel in the ideal linear model. Keep the real injector characterization when the fault is airflow. Read the table and the scalar before a paste.
• Separate short and long trim fractions combine multiplicatively: (1 + S) × (1 + L) − 1. At +10% each, the combined correction is +21%, not +20%. Confirm the channel semantics first. If the scanner already reports a combined trim, do not combine it again.

A flow rating without pressure, fluid, temperature and test conditions is not enough data to enter. The same injector can be described at 3 bar, at the vehicle rail, or at a shop reference pressure. The table needs a curve, not one number.

• A reported 2011 Charger file uses InjPW vs Fuel Mass and Fuel Mass vs InjPW. Labels and available controls vary by OS. The first maps a fuel mass to a pulse width. The second maps pulse width back to mass. Confirm whether mass is grams or milligrams in this file. The breakpoints, offset and minimum pulse width matter as much as the peak flow.
• Data written for a different processor can have different breakpoints and units. A nominal cc/min rating is not a data set. Use the characterization for the exact injector and the exact controller format.

An injector swap can be mechanically correct and still run badly. A 1000 cc/min injector at idle opens for a short time. If the offset and minimum pulse width are wrong, the engine idles on the wrong part of the curve.

Static flow at a long pulse width does not confirm the small-pulse region.

• Stock injector ratings circulate for the early 5.7 and the 6.4. Those numbers do not have a matched part-number, fluid, pressure and low-pulse data chain in this pass. Do not treat a circulating rating as universal. Read the part number and the injector data in the file, or get the data for the exact part.
• The ID1050-XDS manufacturer rating is 1065 cc/min at 3 bar on gasoline at 52 °C. The ID1300-XDS manufacturer rating is 1335 cc/min at 3 bar on iso-octane at 52 °C. The two fluids are not the same. Those are component ratings, not the flow at 4 bar and not a stock value.
• Pressure scaling uses Q2 = Q1 × √(ΔP2 / ΔP1) as an ideal approximation, where each pressure is the differential across the injector. From 3 bar to 4 bar, 1065 cc/min becomes about 1230 cc/min. The manufacturer's characterization table takes precedence. Raising pressure also changes offset and pump delivery.

The historical injector-table workaround

• A worked 2011 Charger example describes correcting pulse width by fuel mass and rebuilding the inverse table. That can improve trims while leaving calculated airflow and load wrong. Treat it as a diagnostic clue in an inherited file, not the preferred correction for a cam or a blower.
• If the forward table maps mass m to pulse width p, the inverse must map p back to m using corresponding axis points. Multiplying both tables by the same factor is not inverse preservation. The two lookups can disagree even when both numbers look plausible.
• For example, suppose the forward curve maps 0.020 g to 3.00 ms. After a characterization correction, that point becomes 0.020 g to 3.20 ms. The inverse must then return 0.020 g at 3.20 ms, not at the old 3.00 ms point. If its axis is fixed, resample the inverse onto that axis. This illustrates matching lookups; it is not an injector setting to copy.

On a car with no rail-pressure transducer, the calibration works from an assumed pressure. A pump change, a restriction or a rising-rate regulator can move the real pressure away from that assumption. Check whether this file has a pressure channel before you plan the diagnosis.

• The published service test for a 2011 Challenger 5.7 is 407 ±34 kPa, or 59 ±5 psi. The test uses ignition on, engine off, with the pump actuated by the scan tool. That is an application-specific service range. The common 4 bar or 58 psi shorthand is not a more precise factory specification for every vehicle. Read the applicable test for the car in front of you.
• Rail gauge pressure is not the same as differential pressure across the injector. On a fixed-pressure system, subtract manifold gauge pressure from rail gauge pressure, or subtract absolute pressures on both sides. At 10 psi boost with a fixed 58 psig rail, the differential is 48 psi. The ideal flow ratio against 58 psi is √(48 / 58) = 0.9097. The injector loses about 9% of its pressure-driven flow at that condition.

A bigger pump can raise idle pressure if the regulator cannot bypass the extra flow. Measure before you scale. If rail pressure is higher than the file assumes, the engine runs richer at the same pulse width. Scaling the injector data to hide that error leaves the pressure problem in place.

1. Record the pressure specification for this engine and model year. Do not transfer the 59 ±5 psi test to every Hemi.
2. Measure at a stable condition. Key-on engine-off, warm idle and a steady loaded point are different tests.
3. Compare rail pressure with the value the calibration assumes. If the file has a pressure sensor, check the scale and the reading.
4. Compute differential pressure with manifold pressure in the same unit. A boost gauge in psi and a MAP in kPa need a conversion before subtraction.
5. Fix the hardware or the regulation first. Then recheck injector data and trims.
6. Repeat the measurement after any pump, regulator or fuel-line change. A single idle reading does not prove the rail holds under load.

Short and long fuel trims are the closed-loop correction. They show how much fuel the PCM added or removed to reach the commanded mixture at that condition. They do not prove the airflow model is right. They prove the loop is correcting something.

• At idle and part throttle, the narrowband sensors can hold stoich and the trims are useful. In open loop, the trim may freeze, carry over or stop updating. Its behavior is operating-system dependent. A displayed number is not proof that the correction still applies.
• Build the correction only from steady, accepted points. Exclude deceleration fuel cut, misfire, shifts, MDS transitions and cam phaser transitions. A histogram that mixes those states produces a correction that fits none of them.

In closed loop, measured lambda can match command because the trims already corrected the error. For example, lambda 1.00 against command 1.00 gives a wideband ratio of 1.00 even with a verified +10% applied trim. That ratio is only the residual error.

Use the valid applied trim for that condition, and do not count it twice. In open loop, establish whether any learned correction remains before using lambda error to change the base model.

• Trims do not have to be exactly zero. Seek a repeatable low error across the conditions you accept, with both banks close and no temperature-dependent split. A single frame at idle is not an acceptance test.
• An inherited file can have the injector data bent to make trims look good. Check the injector tables against the real part before you add another correction. If the previous tuner changed both the forward and inverse tables by the same percentage, the inverse may no longer be the inverse.

1. Warm the engine to operating temperature. Record the mode, the gear and the cam state.
2. Log both banks, commanded fuel, wideband lambda, fuel mass, pulse width, MAP and baro. A missing bank is a gap.
3. Filter to steady points. Remove fuel cut, shifts, misfire and MDS or cam transitions.
4. Separate closed-loop trim from open-loop lambda error. Account for active corrections before computing a base-model change.
5. Check the bank split before a global change. A single bank error is not an airflow model error for both banks.
6. Apply the smallest correction that moves the region. Repeat the same condition.
7. Stop when the error repeats within your tolerance. Do not chase a single frame.

The correction path follows the evidence. If the injector part or data is wrong, fix the injector characterization. If the injectors and pressure are right and the aircharge estimate is wrong, correct the active airflow model. If pressure is wrong, fix pressure.

Do not use one path to cover a fault in another.

The network path

• The network path starts with the model ranges, not the surfaces. If a boost or load range is not represented, a larger MAP sensor does not expand the model. Check Engine > Airflow > Neural Network, under Training Values. Set the required ranges before creating the initial training file. If those ranges change later, use Create New File for Training from Tune; do not keep training the earlier export.
• Choose only the surfaces the engine can physically reach. A linked-cam Hemi does not use every combination in a generic dual-independent-cam grid. The five-surface description is tied to a Hellcat example, not every file covered here. Use Table for Training controls which surfaces enter the fit. Exclude unreachable cam combinations only after matching the generated grid to this engine's linked cam travel. Training ranges describe model coverage; they do not command the phaser or replace a mechanical limiter.
• Filter the log to the surface you are correcting. Match the histogram axes to pressure ratio and rpm, and filter both actual cam-angle inputs near the selected slice. The 104–106° filter is an example, not a Hemi target.
• Apply measured error only to cells with enough samples. Leave low-coverage regions visible and unvalidated. Inspect neighboring cells for continuity, but do not smooth away a real feature to make the surface look neat.
• After the import, the fit is an approximation. It will not reproduce every requested cell exactly. Repeat the original conditions and check trims, wideband, calculated load, spark and throttle behavior. An airflow correction should change calculated aircharge and load where the old estimate was wrong. Check that the resulting spark lookup and torque behavior make sense.
• We have not run the Trainer and read the result back on our own Hemi file. Exact write-back behavior and fitted cells still need verification on the target OS.

The direct VE path

• The direct path is only available when the file exposes an active VE table and the switch or mode selects it. A table that is visible in the editor can be inactive. The 2009–2010 Dodge files carry a known dependency: disabling the network requires disabling VVT, and disabling VVT deactivates the network. A software VVT disable is not a mechanical cam lock. Confirm the phaser's physical behavior and the cam's clearance requirements. Decide whether the loss of phaser operation is acceptable before you change the strategy.
• Keep a network-on baseline. On an affected 2009–2010 file, check current VCM Suite support guidance and every VVT enable switch. Changing only the ANN checkbox is the reported nonworking approach. Change only the documented strategy controls. Initialize the direct table from a valid same-OS basis. Verify low-load operation before any loaded work. Then run the activation test: make a small localized change in a known region and confirm the fuel and load response. If nothing moves, stop and resolve the strategy before you edit more cells.
• Recheck startup, transient fueling, MDS state and the full temperature range. Document what happened to VVT. A car that runs at one steady point is not validated.

The injector path

• The injector path is for a known injector change or a known data error. Enter the characterization before the first start. If a large injector idles badly, check the offset, minimum pulse width and low-pulse data before the airflow model.
• An injector scalar can change fuel delivery without correcting aircharge. That is the distinction that matters on a load-based controller. Use it to fix delivery, not to hide a model error.

Inherited files

Check for a previous injector fudge, a maxed axis, a clamped table and a bank-specific correction. A correction past the axis limit may clamp at the last cell; check the axis before you assume the change applies.

If the previous file changed both injector tables by the same percentage, rebuild the inverse before you trust it.

The pedal is a request, not a throttle angle. The PCM turns pedal position and vehicle speed into a torque request. It then chooses throttle, spark, fuel and transmission state to meet that request. The airflow model sits inside that loop because the torque estimate depends on aircharge.

• When the estimated torque runs ahead of the request, the blade can close even with the pedal down. That looks like a throttle problem. If the airflow model is wrong, the torque estimate is wrong, and the blade may close for a reason that started in the air calculation.
• Separate the channels before you decide. Pedal, commanded throttle, actual throttle, torque request, estimated torque, MAP, calculated load, spark and trims tell different parts of the story. If the blade closes while the pedal and request stay up and knock is quiet, review the torque intervention in the spark and knock guide. If calculated load is wrong and the fuel correction was made in the injector tables, fix the airflow model first.

A pedal box changes the request. It does not correct the airflow model, the injector data or a torque limit. Zeroing a torque limit removes a protection and does not fix the estimate. The throttle page owns that diagnosis. This page owns the air and fuel side that feeds it.

Validation repeats the conditions that produced the error. A correction made at warm idle is not confirmed at cold start. A correction made at one load point is not confirmed at another. The file is finished when the error repeats low, both banks agree and the load-based channels still make sense.

1. Save the as-found file and the changed file. Record every table changed and the direction.
2. Repeat the same route and the same conditions. Label the mode, gear and cam state.
3. Check trims, wideband, calculated load, spark and throttle in the same log. One channel cannot carry the conclusion.
4. Compare both banks. Check bank-specific hardware, sensors and calibration before a global correction.
5. Check cold start, hot restart and a heat-soaked idle. Warm cruise alone is not the whole file.
6. Restore the file if the result does not repeat. Recheck the measurement before the next change.