Below is the text version of the Co-Optima Capstone webinar, “How Can New Simulations and Modeling Tools Accelerate Fuel-Engine Co-Optimization?,” held on October 26, 2021.
Sibendu Som, Argonne National Laboratory: All right. So let’s get started. Thank you for the introduction. My name is Sibendu Som. Today I am presenting the toolkit in here. And we have a lot of results to showcase on some of the work that we have done in the past couple of years on using modeling and simulation tools to accelerate fuel-engine co-optimization. So, customary slide for Co-Optima. I think all of you have been attending the Co-Optima calls for some time now, so you understand the overview of Co-Optima. The idea is to have better fuels, better engines sooner by combining engine R&D and fuel R&D. Specifically for the toolkit team, we’ll present what were the goals of the toolkit team. We’ll talk about some of the models that we have developed, the technical approaches that we have taken, and we’ll spend a lot of time in highlighting some of the main results. And again, the main results were derived from multiple national labs. So by no means am I an expert on all the results that we’ll show, but I’m hoping that some of the members of the toolkit team are actually present today, so in case there are questions they’ll be able to help out with the answers as well. And then I’ll have a couple of last slides on key takeaways from a toolkit team perspective.
So you have seen this slide before multiple times. The focus is on on-road transportation starting from light duty moving to heavy duty. The program initially started with light-duty work, and then the last 3, 4 years we have transitioned into doing more medium-duty/heavy-duty type of work. And as we started the program, we quickly realized that we do not have physics-based models that are robust enough to predict physics that’s happening at the fuel-engine interface. So we have developed a lot of models that we talk about. We have accelerated these models using HBCMAI. And then we have also collaborated very closely with our colleagues in the AED team and the FP teams, who have given us invaluable data for validating our models. So most of the presentation today is going to focus on medium-duty/heavy-duty applications, but we’ll also show you a couple of slides on combustion concepts and modeling of those combustion concepts that are sort of cross-cutting.
Toolkit team goals. Essentially, we have three goals. We want to provide insights on fuel-engine interactions through accurate CFD simulations, CFD science for computational fluid dynamics. We want to use this data to train lower-fidelity yet accurate models that enable fuel-engine co-optimization leveraging HPC and machine-learning techniques for acceleration. And then we also want to provide feedback to experimentalists around these new fuel-engine configurations for hypothesis testing and cross-validation. We are somewhat successful in doing this from the light-duty standpoint, and we were just about getting to do the last goal in the medium-duty/heavy-duty standpoint. And we’ll talk about this as we go through the presentation.
Before I go too deep, I wanted to acknowledge the toolkit team folks who have helped me put together these slides. And the toolkit team has been larger. Many people have come, contributed to the toolkit team, and then gone on to do other things. So for this specific slide deck I wanted to acknowledge Matt McNenly, Nick Killingsworth, Simon Lapointe, and Russel Whitesides from Livermore; our colleagues from NREL, Shashank and Bruce; our colleagues from Sandia, Marco and Everett; colleagues from Oak Ridge, Flavio and Dean Edwards; Juliane Mueller from Lawrence Berkeley, and she’s also the toolkit team deputy. And then of course my colleagues at Argonne, Riccardo, Pinaki, Chao, Roberto Torelli, Hengjie, Sayop, Krishna, Ram Vijayagopal, and myself. So these folks have provided me some slides, also some talking points that we’ll be going through.
This is a rather busy slide. I don’t want you to read any of the content of this slide except that I’ll say that today’s presentation will focus on some of the new multiphase flow models that we have developed across the labs. We’ll talk about some of the ignition and combustion modeling that has been carried out for medium-duty/heavy-duty applications, and also some work on light-duty applications. We’ll talk a little bit about our neural network predictor for octane predictions. We’ll talk about laminar flame speed solvers that have been used to accelerate laminar flame speed calculations for Co-Optima mechanisms, and then we talk about some of the zero-RK accelerations and addition of engine-relevant models zero-RK so that we can do bilevel optimization.
So let’s dig deep. So if you want to perform good computational fluid dynamic simulations, reliable computational fluid dynamic simulations and understand the interface between fuel and engine interactions, we need several inputs to our CFD simulations. Here I’m just showing you a typical video of temperature contours that you’re able to predict inside an engine. But in order to predict this we need several—a lot of information, starting from real geometry of the engine, not only the combustion chamber but also understanding of fuel injectors, intake ports, exhaust ports, real geometric information. We also need to have some understanding on what’s the combustion mode, whether it’s spark ignition, spark-assisted compression ignition, gasoline compression ignition, regular diesel combustion, or HCCI, or MCCI. In some ways this is important for us as we choose our combustion work. We need to understand the operating conditions, the speed, load, etc. And these are all inputs to our simulations. The initial conditions, the boundary conditions, and of course once we are able to perform all simulations, we also need experimental data or some level of value.
From the fuel side we understand that all fuels are multicomponent, and we typically are restricted in terms of simulations with number of components that we can use, as typically we use somewhere between five to nine component surrogates to represent a very complicated fuel.
Dan Gaspar, Pacific Northwest National Laboratory: I’m sorry to interrupt you. Your voice kind of fades in and out every once in a while, so maybe you want to try with your video off. I don’t know if that’s what it is.
Sibendu Som: Yeah. I was in a meeting earlier and I had this issue. So let me try to turn off my video and I’m hoping that it will be better. Thanks for letting me know—and please interrupt me again if you see issues. Thank you. So let me continue. So, in terms of representing the fuel as a surrogate, we need to have some understanding of the fuel physical properties, density, viscosity, surface tension, vapor pressure—not only at a fixed temperature, but as a function of temperature and pressure. Essentially the thermophysical conditions that we would encounter in an engine. We would need the fuel thermal chemistry, the number of species, number of reactions, and we also need a reduced kinetic mechanism as we are trying to perform some of these simulations. All this leads to prediction of ignition delay, laminar flame speed, phi sensitivity, and RON/MON, etc. for the fuel, which are actually used in the simulation.
And then last but not the least, but perhaps the most important, is physical sub-models of physics that’s happening inside—for example, spray dynamics, fuel injection, fuel breakup, fuel vaporization. We need to have someone sending physical models of heat transfer, turbulence mixing, turbulent combustion. As we talk about turbulent combustion, this whole idea that I talked about whether it’s SI, whether it’s SACI, whether it’s GCI, this actually helps us select our combustion models. And all this, the choice of the models, the choice of the chemistry mechanism, all this feeds into what’s the computational cost of the simulation and leads to what’s the computation efficiency of the simulation. So as you can see it’s a very complicated process and there are a lot of uncertainties in this process. There are a lot of things that can actually go wrong. So as toolkit team we pay a lot of attention to ensuring that our input conditions, our boundary conditions, our physical models are as accurate as possible as we try to perform this simulation. And a shout-out to our experimental colleagues who have provided us a lot of data for validating our engine simulations. So in the next few slides we’ll talk about a lot of these aspects and also show you many, many engine simulations.
So let’s start off by some work that Matt McNenly and his team has done at Livermore looking at neural network and prediction model. So essentially this was created using neat and multicomponent blends in the fuel properties database created by Gina Fioroni at NREL, and this was augmented by an extensive literature review of last 2 decades to obtain more than 700 octane readings for well-characterized blend compositions. The neural network model was used to create five new research BOBs that had a matched octane rating as the original four-component BOB that was proposed for Co-Optima. For each of the five new BOBs, one of the PIONA classes was maximized for its content while adjusting the other classes to maintain the same predicted octane number. These five new BOBs were then blended with six high-performance fuels at different blend levels starting from 10%, 20%, to 30%. And they were blended with ethanol, isobutanol, 3-pentanol, methyl acetate, 2-methyl furan and di-isobutane. This work has also been published, and this octane predictor model has been very helpful for us.
As we talked about in my previous slide, we also needed to have surrogates that would present the real fuel towards this goal. Again, Matt McNenly and his team at Livermore also developed an automated surrogate fuel designer. It’s essentially a new optimization tool was created that specifically that zero-RK solver—and we’ll talk a little bit more about the zero-RK solver in the next slide. But the zero-RK solver with Python to enable surrogate blends to be designed using available chemical kinetic components. So we looked at real fuel blends that actually contain hundreds to thousands of distinct chemical species that can be identified with gas chromatography. The surrogate optimizer allows a user to specify over 30 fuel properties that can be specified from experimental measurements of fuel boiling-range fuels. The optimizer then minimizes the error of the predicted fuel properties of the fuel blend from this available pallet. This allows a user to create a fuel surrogate that can actually be simulated using CFD or other chemical kinetic models while capturing as many critical properties of the real fuel as possible. And as you can see, this optimizer also uses inputs like flame speed, liquid properties, etc. to come up with an ideal surrogate.
The next thing that the Livermore team worked upon was accelerating and adding more features to the zero-RK solver. So, the Co-Optima actually funded numerous improvements to the zero-RK solver, and this new zero-RK solver is going to be released next week. And these new improvements accelerated a number of canonical reactor flow configurations, which are essentially used to develop new mechanisms, validate new mechanisms. And this was critical to expand the gasoline pallet to include more than two dozen new high-performance fuels in such a short amount of time. A number of lightweight engine models were also created to optimize a blend for octane synergy, phi sensitivity, and spark-assisted compression ignition. And in a couple of slides from now we’ll see how we use this zero-RK engine model.
Now these models essentially deliver linear scaling of computational costs with mechanism size and delivering an order of magnitude or more speed-up compared to the other available commercial tools. And this is key. It’s well known that the computational time scales with cube—square cube number of species. A lot of the work that the Livermore colleagues have done. Now computational time scales about linear with the number of species. So it’s a huge acceleration that we have been able to use. And in some cases, as Matt notes, commercial tools like Chemkin Pro could not even complete a single calculation. Using the large Co-Optima mechanism that has more than 4,500 species you do memory and solver issues. So with this initial set of tools available we then try to use these tools, these models and tools, as we try to perform computational fluid dynamic simulation, so full-scale simulations.
So next few slides will focus on developing full-scale CFD simulations for different engine concepts and getting some key insights on fuel-engine interactions as we think about these fuels and these engine concepts. So, this is work done by Chao Xu at Argonne National Lab, where he has collaborated with Magnus Sjoberg at Sandia where experimental data was available, and essentially development an engine CFD model based on large-eddy simulations and a hybrid combustion modeling approach or well-mixed and partial-fuel-stratified assisted SACI combustion. The hybrid combustion models allows to simultaneously predict both flame propagation and autoignition as you see in this video. And we actually developed this new hybrid combustion modeling approach so that we have a tabulation of flame speed to predict the turbulent propagating flame, and we also have ways to capture the end-gas autoignition.
As Chao shows in this slide, the use of LES was critical in capturing the cyclic variability that’s seen in this process. And the LES was also important to capture the two peaks in ignition in each of these that was observed. The engine model validation was performed against experimental data, as you can see here provided to us by Magnus. And our modeling approach was able to capture these key things that were of interest. Now, for this simulation, we were also able to leverage some first-principles-based direct numerical simulation calculations of the combustion concept. And this also helped us validate our combustion model. So this was a unique case where we were actually able to not only use experimental data, but also use direct numerical simulation data to validate our models.
For this condition, for this engine platform specifically, we were fortunate that we had a lot more data. We had additional optical engine data that was available to validate the spray models as you can see here. And we were also able to validate the flame structure for the PFS-SACI condition. And essentially for the spray model we were able to predict the liquid and vapor penetration and quantitatively and qualitatively agreed well with experimental data. For the flame, we observe—we perform detailed comparison of the flame structure and shows that the LES-based combustion modeling approach is able to capture the stretch and the wrinkle flame front and the transition from the diffusion flame, which is sooting, to the premixed flame, which is non-sooting. And again, we saw that the use of LES allowed us to capture the flame structure a little bit better. Chao also looked at emission predictions and he looked at detailed chemistry-based NOx model and found that it better captures NOx emissions compared to some of the standard use models like the Zeldovich-based approach. And those results are shown here for both well-mixed and PFS-based—PFS-assisted SACI conditions.
Now as the tools were developed, available, validated gives us confidence then to use this tool to study some fuel property effects. So here we essentially perturbed heat of vaporization, or HOV, and laminar flame speed, and we perturbed them to understand the effect of these property variations on things like CA50, combustion efficiency, etc. So the figure on the left shows heat release rate profiles with different HOV perturbations and it is seen that the reduction of HOV slightly promotes deflagration and significantly accelerates autoignition, which results in actually advancing CA50. This is mainly due to the cooling effects, as seen in the change in IBC gas temperature that you also see here. These observed HOV sensitivities from PFS operation are very similar to the well-mixed charge operation as well.
We also looked at the effect of flame speed sensitivity, and that’s shown in the figure on the right. And it seemed that the magnitude of flame speed is directly associated with the initial slope of the apparent heat release rate in the deflagration phase. The enhanced deflagration further promotes autoignition, which makes sense. And compared with the well-mixed charge operation, the PFS operation shows a smaller flame speed sensitivity, indicating that perhaps the PFS-assisted SACI is more tolerant of fuel property changes than well-mixed SACI. The changes in the slopes of CA50 and heat release rate versus flame speed coincides with the change in dominant combustion mode, as we can see here.
So as this tool—as these CFD simulations are validated, used for sensitivity analysis of fuels, we also use the CFD tools to study—to accelerate zero-RK so that zero-RK would have additional features for performing engine simulation. So this was a collaboration between Livermore and Argonne, and we understand that fuel-air mixing preparation is a critical factor affecting engine conference, engine performance. So simulating a priori that impact and evolution of the fuel-air mixture on engine performance requires detailed turbulency simulations of the intake, exhaust opening valves, closing, spray breakup, evaporation, heat lost to the wall, etc., etc. So again, zero-RK would need this data to be improved. So without some means to represent the temperature and fuel-air mixture stratification, real engine performance that is affected by fuel chemistry like sequence of ignition and then gas and spark-assisted compression ignition would not be possible.
So previous virtual searches that were performed anywhere from 6 to 300 independent reactor simulations to estimate the fuel performance with respect to synergetic octane blending and primary phi sensitivity for partially stratified ACI was also performed. And SACI was performed up to 312 operating conditions. And those were enhanced here with the fully coupled multi-zone reactor model and zero-RK train based on CFD. And while these light-grade models, multi-zone with and without flame propagations, stochastic reactor models have been around for almost as long as detailed fuel chemistry mechanisms have been around. With the Co-Optima funding, what’s new is that we are improving that performance to be fast enough to run in a fuel-engine co-optimization loop that may evaluate hundreds of thousands of blend combinations.
So with this understanding, we continue to work with experimentalists, and one of the things that was brought to our attention was this whole idea of pre-spark heat release, and we had this understanding that while a lot of—while some researchers have been looking at pre-spark heat release, there’s really not a good computational model that exists for predicting pre-spark heat release. So here this is work led by Hengjie Guo at Argonne National Lab in collaboration with Jim Szybist at Oak Ridge, and spray data was obtained from Lyle Pickett at Sandia. And the hybrid combustion modeling approach that was developed and shown in the last slide that uses equation for tracking the flame front and a reactor model-based approach to capture the end-gas autoignition was also used for this approach. And we also used the detailed chemistry mechanism from Livermore and spray measurements from Sandia as I mentioned earlier.
For this work, at least for the validation phase, we use the Co-Optima alkylate and the E30 fuels at operating conditions characterized by different pre-spark heat release intensities, as you will see here. The predictor in the pressure heat release rate were found to agree well with experiments. And it was found that the estimate of previous cycle trap residuals is of utmost importance for capturing pre-spark heat release correctly. So in order to track the evolution of the intermediate species from one cycle to the next, we had to keep the detailed chemistry solver on and active for the entire simulation. So this was a key finding from our work. Otherwise we were not able to get the extent of pre-spark heat release accurately predicted.
So, with some initial validation, then we analyzed the dynamics of pre-spark heat releasing in detail, employing the pressure-temperature trajectory framework. And here the pressure-temperature trajectory was obtained using CFD simulations. It was shown that the pre-spark heat release correlated well with the first-stage ignition of the fuel, ensuring close correlation to the in-cylinder heat trajectory and the chemical kinetics. The 3D analysis also showed that the pre-spark heat release begins in the fuel-lean region near the intake valves, but its effects become more significant in the fuel-rich regions as the time progresses, as you can see in the pictures here. We also found that the low-temperature heat release is a self-limiting process that has the effect of attenuating the thermal stratification in the combustion chamber.
Further analysis was carried out on the effects of fuel properties, including laminar flame speed, heat of vaporization, saturation pressure, liquid fuel specific heat, etc. And the results indicate that pre-spark heat release phasing is slightly advanced with lower laminar flame speed fuels as more trapped unburned fuel is made available for the next cycle. A similar trend was also observed for lower HOV as less-intense precooling led to higher mixture temperatures, that is increasing the mixture’s sensitivity. And the team has also published multiple papers on this topic as well.
So now we’ll move more into the medium-duty/heavy-duty space, again leveraging some of the spray modeling, some of the combustion modeling approaches that were developed in the initial part of Co-Optima. This is work done by Flavio at Oak Ridge National Lab, where he’s modeling the common 6.7-liter single-cylinder medium-duty engine with the high compression ratio of about 20 to 1. The goal here for Flavio was to link fuel properties to NOx emissions. And the physical property effects were initially explored through the development of three diesel fuels—diesel number two, diesel number one. And these fuels had increasing volatility but the same reactivity, as you can see from this distillation curve. And Flavio collaborated with Livermore colleagues to develop reaction mechanisms relevant for these fuels. And these five surrogates represent the distillation curve, another component limitation of the five.
So some validation was performed, and some of the key findings are shown here where Flavio shows that large changes in distillation curve resulted in only small changes in NOx levels. However, larger changes were seen for later injection timings due to later combustion phasing and sensitivity of the mixture, as you can see here. And Flavio also notes that changes in physical property had small impact on the control lever for NOx at low-load engine operation conditions. And the conclusion here was that it’s unlikely that physical properties alone can help reduce NOx significantly. Now one of the key advantages of performing computational fluid dynamics is that with CFD you can actually tweak NOx and you can understand the effect of individual properties on NOx emissions.
And this is what was done by Flavio, where he individually looked at the effect of density, surface tension, viscosity, thermal conductivity, heat of vaporization, etc. And these all were modified independently by large amounts. And essentially what he found was the effect on heat release rate was observed for HOV, vapor pressure, density, and specific heat changes. He saw that changes in mixture formation were substantial, where low-density fuel resulted in richer mixtures and higher NOx. However, he notes that changes in NOx were minimal for all property changes even when property changes were actually large. So essentially, changes in physical property alone are likely not enough to enable low NOx, and this was a finding from Flavio’s work.
So, moving on from medium-duty work, the team also looked at a lot of heavy-duty work. Here we are presenting some work done by Pinaki Pal and his team at Argonne looking at Caterpillar single-cylinder heavy-duty engine at low-load GCI conditions. The experimental data here comes from Chris Kolodziej at Argonne National Lab. This slide is busy. There are a lot of details about the species, the chemical mechanism that’s used, the NOx mechanisms that are used, and the soot models that are being used for the simulation. And these are fairly detailed models that Pinaki employs. So, what all the models displayed good predictive accuracy in estimating the in cylinder pressure, heat release rate profiles, CA10, CA50, NOx. A lot of these things are not shown in the plot. Some of the key results are actually shown in the plot. For all the SOI conditions we had a global lambda of 3.2, injection pressure of 500 bar. Initial pressure and initial temperatures of 1 bar and 145°. And here we are looking at a no EGF condition. Four levels of fuel stratification were considered. And these fuel stratifications were obtained based on different injection times.
The animation shows that the spray development and evolution of the flame temperature equivalence ratio and soot mass as it’s formed inside the cylinder. We found that the ignition delay times increased with advancing SOI. This is essentially due to cooler in-cylinder conditions at earlier SOI slowing the fuel vaporization process. We also found that the cooler in-cylinder conditions results in higher fuel mass accumulation. And fuel films on an operation from locally rich regions does increase the propensity of soot formation. We also found that the soot formations follow a non-monotonic trend with respect to SOI, and we were able to explain the causes for this non-monotonic trend as Pinaki notes here. For example, like minus 27 crank angle degree SOI timing, the condition has the lowest amount of soot emissions, suggesting that the mixing time available in in-cylinder conditions for this SOI are optimum for minimizing soot emissions.
Now, once we validated this model, the modeling approach I should say, we then looked at different blends of gasoline with ethanol. So here we are looking at E10, E30, and E100. And we were comparing with E10, E30, and E100 and we were comparing with RD587 gasoline surrogate that has 10% ethanol. So let’s look at some of the results here as well. So here we are looking at autoignition timing is retarded with the ethanol content due to poor reactivity of ethanol. And we found that the soot emissions also do not follow a monotonic trend. We found that E30 tends to have more soot compared to E10 and then E100. And we were able to use the simulations to explain why this is the case. So HOV and viscosity of the fuel blend increases with ethanol content, and a high HOV tends to lower the in-cylinder temperatures, meaning poor vaporization conditions. As a consequence, what we found was that the higher fuel fill mass accumulation is noticed for E30 and E100. Moreover, we found high viscosity also slows down the vaporization of the fuel. We were also able to explain why soot in E30 was higher than E10, for example. And the team has publications in the pipeline that provides a lot more details as well.
So then moving on from engine simulations, in parallel, we had a lot of work going on in improving our predictions of fuel injection processes and fuel injection sprays and mixing as well. Here I’m highlighting work done by Marco Arienti at Sandia National Lab. And Marco notes that, as we all know, primary atomization is a complex process and the use of nontraditional blends requires an approach that does not only rely on calibration from traditional fuels. So if you look at the top two diagrams, it shows an example of fuel properties that cannot be derived by linear superposition of the component species properties. The two bottom plots are from two GDIs cases for spray G1 conditions for iso-octane and E30. And this was—these simulations were performed with the Sandia code called CLSVOF.
All the eight holes of the injectors are included based on CT scans performed at Argonne National Lab. The snapshots correspond to fully open pintle position, and the solid surface in the plot corresponds to the interface separating the liquid from the vapor non-condensable gas mixture. The vapor is presented by the transparent iso surfaces, a vapor mass fraction with increasingly darker shades of green. The essential conclusion here was that Marco and his team were able to perform the first-ever simulations of surface vaporization for interface of liquid jets in high-temperature gases. And then they were able—they are now in the process of using this data to develop relevant statistics that can also be used for model development.
Marco and his team also developed first-principles-based approaches for vapor generation rate. Excuse me. Marco and his team were also able to develop first-principles-based model for vapor generation rate. And it was implemented in CONVERGE code that is traditionally used by industry by a user-defined function to demonstrate its greater sensitivity to fuel properties compared to the commonly used empirical correlations. What you can see is an example on the right, the first to flare flash boiling spray G configuration defined by the engine combustion network. The four-case matrix is organized by fuel type and by vapor generation rate model. The plots of temperature in the cross-sectional slice are indicative of flash boiling intensity. Let me play the videos again. The key result here is that the difference between the lower two plots is larger than the difference between the upper two plots, showing more sensitivity to the addition of ethanol. This result is also corroborated with experiments performed at Argonne. Moreover, the comparison of the radial profile of density with the measurement demonstrates that the spray cone angle is captured more correctly with this new model that is now available as a UDF in CONVERGE.
Marco continued his work, and here we are showing plots that were obtained from two end-of-injection simulations using the CLSVOF code that Sandia and Marco have developed. The schematics on the left show the conic surface represented as a dashed line here that is used to find the cross section of the eight counterbores. The example of one counterbore cross section is shown on the left-hand side. The main point to take away here is that we can quantify residual fuel and its temperature at the end of an injection. This information is then used to predict the film drying for the remaining engine cycle of the dribbling droplet, so really quantifying the amount of fuel dribble that may happen and the film that may be formed, and the effects of fuel on that film that’s formed. And again, several publications that Marco notes based on his work.
Moving on, I will talk a little bit about some of the co-optimization tools that Juliane Mueller has developed at Lawrence Berkeley. Juliane provided me with many slides, and I had to trim down all that information into one slide that really talks about the co-optimizer. So in order to find the most promising fuels to try in lab experiments, an optimization problem needs to be modeled and solved. So, Juliane—basically Juliane Mueller has actually developed this approach. This is based on Gaussian process-based optimization. She applied this Gaussian process-guided optimizer to a problem where we seek to maximize the robustness of the fuel mix. For each fuel mix, the robustness is evaluated using zero-RK chemistry solver that was discussed earlier. And uses the simulation of 288 engine operating conditions. The robustness then represents how many of these operating conditions are feasible. In the optimization, we have constraints imposed on the RON value that must be satisfied. So, this Gaussian process optimizer does also have stochastic components. Therefore, we do five optimization trials to get an idea of the expected outcome.
We optimized—Juliane actually optimized this over nine components, but that nine component is really not the limit. It can be extended to additional components. And we find that different fuel mixes lead to similar performance—that is, similar robustness value. Now the bottom plot is really important here. It’s really the convergence graph. The blue curve is the result from the Gaussian process model. The red plot indicates the result obtained from an evolutionary algorithm that may not necessarily find the optimum. And the key point here is that the evolutionary algorithm run until 2,500 evaluations and its best solution after 2,500 evaluations is still worse than this new Gaussian process model that Juliane has developed, which converges after 300 evaluations.
So this optimizer is then used as we then try to wrap everything up. This is a collaborative work led by Nick Killingsworth at Livermore. And here a framework was developed to predict the fuel economy of a midsize sedan based on fuel properties. So we looked at engine experimental data, which was combined with a validated engine cycle simulation model in GT-Power. Secondly, the GT-Power model is used to train the Gaussian process regression models. Then these Gaussian process regression models are combined with the fuel property information to quickly generate the engine maps. And then the multimode engine maps were created by combining maps from stoichiometric operations and SACI operation like we talked about earlier. And then the engine maps are fed into a drive-cycle simulation tool called Autonomie to essentially determine the fuel economy. All this is shown in this workflow chart here.
So let’s look at some results. Here we used the simulation framework to explore the effect of engine’s thermal state on fuel economy over several drive cycles and for four different fuels with different values of RON and sensitivity. Here we look at the NOx limits based on two thermal states of the engine, and the NOx limits incorporate fuel properties including RON and sensitivity. We find that the high-sensitivity fuels are more sensitive to the thermal state of the engine. If you look at the plot here, the MPG plot shows that due to downsizing as hot operation we see benefits for FTP-75 and HWFET but not for the US06 cycle. As we downsize the engine, it is forced to operate at higher loads as well. The left plot shows that the three-cylinder engine, which operates at higher IMAPs, tend to be more NOx-limited, so CS50 must be retarded as a consequence. We also find that the high-RON fuels perform better when we are downsizing.
We then explored the effect of multimode operation on fuel economy over several drive cycles for different fuels. The top plot shows the gain in MPG going from stoic SI to using a map with multimode operation for different drive cycles. The benefit of the multimode depending on the drive cycle itself—the bottom plot shows the map of engine efficiency for multimode. So basically looking at SI SACI together with stratified charge operation. The green line denotes the border between the SI and the SACI SC operation, and essentially the stratified charge operation is used at the lower loads. The dots indicate the location of operating points over the US06 drive cycle sample every 0.1 second. The blue dots indicate stoichiometric operations, and the red dots indicate the lean SI operation. You’ll see that for US06 drive cycle much of the operation lies outside of the SACI SC operation window, leading to little benefits.
I have one more plot here looking again at the gain in MPG going from strike SI to using a map with multimode operation for different cycles. And the multimode operation provides anywhere from 9% to 14% MPG gains while the HWFET cycle and the UDDS cycles. The bottom plot shows map of engine efficiency for multimode operation over the UDDS drive cycle. For the UDDS drive cycle, much of the operation lies within the SACI plus stratified charge operation window leading to fuel economy benefits. Here, the higher SACI load limit for the high-RON, high-sensitivity fuels provides benefits. The bottom plot shows map of engine efficiency for the map multimode operation. And the green line denotes the border between SI and SACI SI SC operation again. The dots indicate the location of the operating points over the US06 drive cycle, sampled every 0.1 second. The blue dots indicate the stoichiometric SI operations, and again the red dots indicate the lean SACI plus stratified operation. And you can see that the higher SACI load limit of high-RON is high-RON, high-sensitivity fuels provides benefits here. So clearly the last three slides were not in my area of expertise, and I’m hoping that Nick would be available or Ron would be available with any questions.
So with that I’ll end here. I think I went a little bit over time. I do want to end with some key takeaways. I want to stress that initially as we started this program, the computational tools to study physics at the fuel-engine interface for both the light-duty and the heavy-duty platforms were really not available. So, within the toolkit team we worked towards developing physics-based models essentially for improving predictions of fuel-engine phenomena. And a lot of these models as you saw we were presenting today were available—are being made available to UDFs or improved best practices into industry standard use codes like CONVERGE. Detailed chemistry mechanisms, reduced chemical kinetic mechanisms of several fuels of interest have also been tested with CFD and they’re available. Very importantly, a lot of these engine models that we have developed for multiple platforms such as CFR, CAT, Navistar, Cummins, Ford, GM—talk about any OEM in Co-Optima, we probably have worked with an engine and developed, and to a certain extent validated the models as well for these engine platforms. And some of them may be available for public dissemination.
Lower-order, open-source tools have also been developed, and they are being made available as well. And then I like to stress that initiation of PACE, which was computationally focused, really helped us. It really accelerated TK team’s prediction capabilities. And if we had additional 3 years, we could have enabled us to perform true co-optimization. As you can see for some of the SACI multimode operating conditions, we were able to do some co-optimization. But we were not able to get to it for the heavy-duty platforms. With that, I’ll have this acknowledgement slide, and I’ll be very happy to take any questions, suggestions, etc. Thank you.
Dan Gaspar: Thank you very much Sibendu. Quite a comprehensive body of work you presented there. So that was quite a challenge to get it into the time that you had allowable. We do have a few minutes. We’ll open the floor up to questions. So please raise your hand or use the chat and type the word question in or type your question in the chat and we can have some dialogue here.
So, you kind of answered the one question I was going to ask, Sibendu. And I’ll just add on—so this last slide was really what I was looking for. And you mentioned that a lot of these tools are available. Is there a way to—is there a gatekeeper for this? How would someone contact—or who would they contact to find out how to get access to some of these tools?
Sibendu Som: Yeah. That’s a good question Doug, and me and Bob are on the hook to have a list of all these tools that are open-source that can be made available. So, we are working on the list, and I think it’s going to be uploaded on the Co-Optima website if I’m not wrong. More specifically, if you are looking at some of these engine models like the CFR engine, I mean those things can be made public—available for public dissemination. Some of the other engine models that we have listed here we probably need to have appropriate levels of permissions before they are disseminated. But in general a good starting point would be to reach out to the lab leads. If you do not know anyone, reach out to me and I can certainly ensure that you’re talking to the right person at the labs to get access to these models. I’m not guaranteeing that these models—these engine models will be available because of some of the tags associated with them. But at least we can have a conversation. The open-source tools will be available definitely.
Dan Gaspar: Okay. Great. Great answer. Thanks. I see Tom has a question typed into the chat. So, Tom would you like to unmute yourself? Looks like he sent the question directly to me. I can read it out loud or you can unmute yourself and ask it Tom.
Tom, Webinar Participant: There we go. Okay. So can you hear me now? Can you hear me?
Sibendu Som: Yes.
Tom: Okay. Yeah. I thought it was an excellent presentation. Could you say a few words about the soot model that was used in simulations?
Sibendu Som: Yeah. So yeah. I can go to that slide actually. So we used this hybrid method of moments model. But we also looked at a couple of other models that were available in CONVERGE code. So specifically we did not develop a soot model. The development of a soot model was part of tasks in PACE. In Co-Optima, we used the existing soot models that are available through CONVERGE software but looked at different chemistry mechanisms together with these soot models. So for example, in this case we were looking at the HMOM model, and although it’s not mentioned here, not mentioned in the slide, we were—our best efforts we were able to get to close to a factor of two in terms of predicting. So qualitatively we were able to get the trends, but quantitatively we were still about a factor of two away in terms of predicting the right soot models. And that can be quite typical, actually, based on some of the other work that we have done as well.
Dan Gaspar: Great. Thank you for the question, Tom. Andre, I see you’ve typed something in here. I’m not sure if this is a question or a comment, but go ahead and unmute yourself and we can have some conversation on this.
Andre, Webinar Participant: Okay. Can you hear me okay?
Sibendu Som: Yes.
Andre: Okay. Good. Yeah. It’s been a curiosity that repeatedly in the literature over the years, and again just recently in Combustion and Flame, people have shown in the fundamental flame studies you add a bit of ethanol and the first thing that happens is soot formation goes up. And I immediately thought about because I tagged one of the most recent papers and set that on my desk. This is a paper I ought to read. But your presentation of the impact at E30 yielded the most soot seemed to go right back to that same observation. Obviously very different combustion systems and processes at effusion plane versus that you have going on in this GTDI engine. But have you guys related anything back to those more fundamental flame papers as you’ve written things up on this observation?
Sibendu Som: Yeah. So Pinaki actually led this work. You were able to understand why E30 has more soot than E10, and we were able to explain it based on the higher fuel film mass that was being formed, which led to increasing soot. And we also found that under which conditions ethanol produced higher amounts of acetylene, which was the case where we saw that E30 was actually providing more soot than E10. I think we have an ASME paper on this if I’m not wrong. So Andre I’ll be happy to look it up and pass it along.
Andre: Okay. Okay. Great. Thank you.
Sibendu Som: Thank you Andre. Thank you for your question.
Dan Gaspar: Okay. We have just a couple minutes left here. We could take another question or two. I’m not seeing any in my view, anyway, so let me ask another one. This gets a little bit into details. And maybe Flavio, who is on the line, can answer this. It was his slide. The slide that was talking about basically turning the knobs and varying the fuel properties by a large amount. The density one was the one that caught my eye. I feel like I’ve seen this before. How much is a large amount where its being changed, I guess is my curiosity.
Flavio Dal Forno Chuahy, Oak Ridge National Laboratory: Yeah. Hi. So yeah. I just chose to say a large amount in the slide because the true answer is a little bit more complicated. But basically, I would have to check what the real—what the actual number is. But what I did is I got the ballots of all the different species that could be in diesel fuel and then I picked the maximum and the minimum of that, of a particular species, and applied to it to all the other species. So for example, if I had hexadecane and methane, for example—I’m not saying methane was in there but just thinking two fuels with very different properties. So density is going to be quite different between them. So it would be, the maximum would be hexadecane and the minimum would be methane. And that’s applied to all the other species in the surrogate. So that’s how I define large amounts for this.
Dan Gaspar: Okay. Great. That’s helpful to give me some perspective. Thank you. Any last questions? We’ll take one last one and we’re a little bit minute over here. Okay. Well if not, I’d like to thank Sibendu once again. That was an incredible presentation, and thanks to all the people that helped you put that together, too. A lot of contributions there. So with that, I’d like to close the meeting for today and thank you for joining and join us again next month. Like I said, we’re going to do a lightning round of the DFO, the directed funding opportunity awards, the first round. A lot of those projects are nearing completing or complete, and we’ll get a nice readout on those. So anyway, have a great day everyone.
[End of Audio]