|FHWA > Engineering > Pavements > HIF-11-026 > Chapter 8. Mepdg Model Calibration|
Local Calibration of the MEPDG Using Pavement Management Systems
Chapter 8. Mepdg Model Calibration
This chapter describes the research team’s efforts in calibrating the MEPDG performance models to North Carolina conditions. Calibration of NCDOT flexible and rigid pavements using the most current version of the MEPDG software available at that time (version 1.100) was conducted. The research team executed the MEPDG design software using the inputs provided within the MEPDG calibration database, although MEPDG default values were selected where NCDOT specific data elements were not available. A minimum of three NCDOT pavement sections were used in the calibration of each of the MEPDG performance prediction models.
MEPDG design inputs were prepared for all pavement sections used in the calibration process and the MEPDG was run to obtain the resulting pavement performance distress profiles. The MEPDG predicted pavement performances were then plotted against the field measured performance as noted in the NCDOT pavement condition surveys. Based on how well the predicted performance meet the measured performance determined whether or not modification of the calibration coefficients was necessary. Figure 3 illustrates the general procedure used in the calibration process.
The premise behind any mechanistic-empirical design procedure is the ability to relate key structural response variables (i.e., deflection, stress, and strain) to observed performance. This process hinges on the use of robust pavement performance models, which are typically regression equations that relate a material property, such as HMA stiffness, to an observed distress, such as rutting or cracking. The following briefly summarizes the pavement response models used within the MEPDG.
Quantifying Pavement Condition
The development of calibrated models for use in the MEPDG is highly dependent on the data contained within the LTPP database, primarily since it is the only database of its kind providing material properties, traffic, pavement condition data and so on, for a wide variety of pavement sections under a broad range of climate and traffic loadings. However, a survey conducted by McGhee (2004) determined that approximately 5 percent of respondents were using the LTPP Distress Identification Manualfor assessing pavement condition. Furthermore, although the majority of SHAs collect pavement smoothness, rutting, and cracking data, the collected data may be based on different distress definitions or data collection procedures from those contained in the LTPP Distress Identification Manual. The challenge, therefore, is to be able to convert SHA historical pavement condition data that has been collected in accordance with the different criteria to the definitions contained within the LTPP Distress Identification Manual. Each SHA should assess the differences between the LTPP and their state pavement distress collection protocols in order to determine how these differences may influence the MEPDG calibration activities.
NCDOT Pavement Condition Assessment Methodology
NCDOT assesses pavement condition through the use of windshield surveys and pavement profilers. Pavement condition surveys are conducted on all flexible and rigid pavement sections every 2 years. A 100 percent survey is conducted on all flexible pavement sections, while a 20 percent sample is conducted on rigid pavement sections. Pavement condition surveys are conducted by trained personnel traveling at 15 to 20 mi/hr who note the presence of a variety of observed pavement distresses. NCDOT also collects rutting and IRI data using a high-speed profiler outfitted with a three-sensor rut bar (one sensor in each wheelpath and one sensor centered between the wheelpaths). Faulting measurements are obtained either by a faultmeter, the profiler, or other hand measurement methods.
In relation to correlating the LTPP based pavement condition assessment to that of the NCDOT, Corley-Lay et al. (2010) conducted a study to determine if any disparities exist between the two data sets in North Carolina. For asphalt-surfaced pavements, NCDOT compared pavement condition data for all LTPP monitored sites (flexible pavement, general pavement study sites only) to those contained in the NCDOT pavement condition survey for corresponding roadway segments. Conclusions from this comparison included the following observations (Corley-Lay et al):
Discrepancies in the data collection process between LTPP and NCDOT were noted by the research team during the calibration process. Any resulting challenges due to differences in pavement condition definitions (or procedures) have been noted in the calibration section of this report.
The current NCDOT survey procedures report the presence of the following distress types on asphalt-surfaced pavements:
Though all of the above distresses have been included in the MEPDG calibration database, the following discussion will only include flexible pavement distress types that are considered in the MEPDG, which include rutting, load-related cracking, and thermal cracking.
Table 48 includes the definition for rutting for both LTPP and NCDOT. Rutting is measured as the actual rut depth for the LTPP method, while NCDOT categorizes rutting according to the three severity levels shown in table 48.
Table 49 includes the definition for alligator cracking for both LTPP and NCDOT. For the most part, the LTPP and NCDOT alligator crack definitions are very similar and only differ in that NCDOT provides a measure of crack width for each level of severity.
However, the procedures for measuring the extent of alligator cracking are significantly different between LTPP and NCDOT. For LTPP, the actual area of alligator cracking is determined, resulting in a square-foot measure of alligator cracking for each severity level. NCDOT measures the amount of alligator cracking as the percent of total area; however, as noted previously, the presence of alligator cracking in one wheelpath is considered to have the same extent as if the alligator cracking was in both wheelpaths. Corley-Lay et al. (2010) identified the need to evaluate the impact of the current NCDOT methodology for quantifying HMA alligator cracking in the MEPDG. The impacts of the NCDOT alligator cracking methodology on the calibration conducted is beyond the scope of this project.
Table 50 includes the LTPP and NCDOT definitions for thermal cracking. The LTPP and NCDOT thermal crack definitions are similar, but differ in that NCDOT includes block and reflective cracking, and that moderate and high severity levels differ by the width of the defined crack. In this case, NCDOT uses a slightly more stringent requirement in that moderate-severity thermal cracking occurs at a crack width of 0.25 to 0.50 in and high-severity cracking is defined as a crack width greater than 0.50 in, while LTPP defines moderate-and high-severity cracking as between 0.25 and 0.75 in and greater than 0.75 in, respectively.
Procedures for measuring thermal cracking extent also differ significantly between LTPP and NCDOT. LTPP recommends measuring the number and length of the thermal cracks at each severity level, while NCDOT rates only the condition that represents the majority of the segment (Corley-Lay et al).
Smoothness is quantified by LTPP according to IRI. Historically, NCDOT has quantified ride condition using a subjective rating scheme that includes:
NCDOT began collecting profile date for determination of IRI beginning in 2001.
As with asphalt-surfaced pavements, NCDOT reports the presence of a number of distresses for jointed concrete pavements (JCP) that include:
The following comparisons will only include rigid pavement distress types that are considered in the MEPDG, which for JCP include transverse cracking and joint faulting.
As shown in table 51, for the most part, the transverse cracking definition for LTPP and NCDOT are essentially the same. The difference in the transverse cracking definition is in terms of the allowable crack width that defines moderate and high severities. However, none of the pavement sections used in the analysis had cracking above the low severity level; the majority of the pavement sections reported no cracking. There is also a difference in the definition of allowable spalling; NCDOT specifies the spall width, while LTPP evaluates the percent of the joints spalled. This difference is not considered to be significant and the NCDOT measurement is considered to be the same as the LTPP measurement.
As shown in table 52 the faulting definition for LTPP and NCDOT are exactly the same. NCDOT obtains faulting measurements through the use of a faultmeter, profiler, or manual methods and LTPP manual uses the FHWA-modified Georgia Faultmeter. LTPP specifies that the fault should be recorded within the outside wheelpath; however, NCDOT provides no specific guidance on the location of fault measurement. While the accuracies of the collection methods are different, the results are considered to be equivalent.
Smoothness is quantified by LTPP according to IRI. Historically, NCDOT has quantified ride condition using a subjective rating scheme that includes:
NCDOT also began collecting profile date for the determination of IRI beginning in 2001.
As described previously, projects were selected for use in the calibration process based on pavement type (new HMA, overlaid HMA, and new JCP), uniformity over the entire pavement section (e.g., pavement thickness, material type, traffic), and availability of pavement data (e.g., pavement condition). Figure 4 illustrates the location of each of the pavement sections used in the calibration process. Detailed information on the HMA and JCP projects selected for use in the calibration process are included in appendix E of volume 2.
Climatic data for all pavement sections was interpolated from the two nearest weather stations using the NCDOT provided project coordinates (latitude, longitude, and elevation). Climatic files were obtained from the updated climate files located on the MEPDG website http://www.trb.org/mepdg/climatic_state.htm).
NCDOT provide traffic data in terms of AADTT and percent trucks for all pavement sections. MEPDG default values were used for all other inputs for all pavement sections.
For the most part, detailed material properties for all pavement layers were not available for any of the NCDOT pavement sections. Any provided material properties (e.g., layer material type and thickness, subgrade soil type, HMA material properties) have been included in the MEPDG calibration database and used in the calibration process. Due to the previous study conducted by the North Carolina State University, the following HMA material properties have been included in the MEPDG calibration database:
For all other needed inputs not included in the MEPDG calibration database, MEPDG default values were selected. All data inputs for all pavement sections used in the calibration process are included in volume 2 (appendix F for HMA sections and appendix G for PCC sections). In addition, tables 53 and 54 summarize the HMA and PCC pavement sections used in the calibration process, respectively.
1 Initial construction/overlay (where appropriate) 2 AB = aggregate base; SS = stabilized subgrade 3 overlay application
1 ATB = asphalt treated base; CS = crushed stone subbase; CSS = cement stabilized subbase; LSS = lime stabilized subbase
The steps for conducting calibration of the MEPDG pavement performance models to location conditions include (NCHRP 2009):
Since many of the NCDOT provided pavement sections either had very little distress or had not been in-service for a sufficient period of time (specifically the JCP section), full calibration of the pavement prediction models is limited. Though the number of submitted pavement sections is adequate to demonstrate the calibration process, a larger pavement section sample is required for both the calibration and validation process.
The following describes the results of the calibration process for the NCDOT HMA and PCC pavement sections.
For new flexible pavement design, the MEPDG performance parameters include rutting, load-related cracking (alligator and longitudinal), thermal cracking, reflective cracking (HMA overlays only), and IRI. The HMA pavement sections listed in table 57 were used in the calibration process.
Figures 5 through 7 illustrate the MEPDG uncalibrated predicted distress versus the NCDOT
As shown in figures 5 through 7, the predicted performance using nationally calibrated models under predicts the depth of rutting for all but two projects, over predicts alligator cracking on one project and under predicts alligator cracking on four projects, and under predicts the amount of thermal cracking as compared to the NCDOT measured distresses.
The residual errors (the difference between the predicted value and the actual value) for the calibration sites using the nationally calibrated models for rutting are shown in figure 8. The residual error for rutting on all pavement sections increases with age and are all the same sign (except for early age rutting on pavement sections R2211BA and R2232A). However, on three pavement sections (1006-3, R2211BA, and R2232A), the residual error is relatively low compared to the performance (or failure) criteria (0.75 in); indicating that model prediction may be improved through adjustment of the calibration coefficient. For pavement sections 1024-2 and 1817, the residual error is considered high when compared to the performance criteria, suggesting that some other factor may be influencing the prediction. Based on the data provided, no specific reasoning can be given for the higher residual error on pavement section 1024-2 and 1817.
The residual errors for alligator cracking are shown in figure 9. For all but two pavement sections (1802 and 1817), the residual error is negative, with a relatively constant slope, and a residual error that is low compared to the performance criteria (25 percent). This indicates that adjustment of the calibration coefficient may improve the performance prediction on these pavement sections. Again, based on the available pavement section information, no specific reasoning can be found for the high residual error for pavement section 1802.
The residual errors for thermal cracking are shown in figure 10. The residual errors for each pavement section are negative, the slopes are considered to be high, and the value of the error is high compared to the performance criteria (1000 ft/mi [189.39]). Adjustment of the calibration coefficients may improve the performance prediction.
The primary goal in the calibration process is to reduce the error between the measured and predicted distress. However, there are a number of limitations in the available data, including the relatively few data points, the required conversion of NCDOT’s subjective measure of rutting and thermal cracking to an estimated value, and the limited data available at levels approaching the established failure criteria or at the end of the performance period; all of these pose a significant challenge to the model calibration process. Nevertheless, the data limitations encountered are likely fairly common within other SHAs and some method of model calibration is needed in the interim until additional data can be collected. The MS Excel® solver routine, which employs linear programming optimization techniques, was used to minimize the root square error between the available NCDOT measured and MEPDG predicted values. With this process the beta coefficients were changed until a minimum square root error was reached. This procedure was repeated for each pavement section separately. The final beta coefficients were then obtained by averaging the resulting calibration coefficients for each of the pavement sections. This process was done for each of the key HMA distresses, as described below.
As indicated previously, the NCDOT subjective rut depth measurement had to be converted to an estimated measured value. In addition, progression of rut depth over time (e.g., for one year to the next), due to the limited amount of rutting data on the NCDOT pavement sections, was also needed. A number of studies have been conducted that document the development of rut depth development over time (Haddock 1999; Sivasubramaniam et al; White et al). These studies, in addition to the MEPDG rutting models, were used to predict rut depth over time for the NCDOT data.
With the absence of actual measured data for rut depth, the research team determined that it would be more realistic to base the rutting severity on the last NCDOT survey year for each section used in the calibration process. The only difference would be on those sections that received an HMA overlay; in those instances, the research team selected the rut condition prior to the applied overlay. In this manner, rutting was assumed to progress from zero to the assumed numeric value over the life of the pavement. The assumed values for rut depth are:
Rut depth progression was based on the number of NCDOT rut depth ratings and distributed over the measurement period to best reflect the slope of the MEPDG predicted rut depth over time. Table 58 and figure 11 (for low severity rating) illustrate the rut progression process.
Table 59 includes the estimated value for all projects used in the rutting model calibration process.
In addition, rutting was assumed to be totally contained within the HMA layers (i.e., no rutting in the unbound layers). However, this assumption should be validated through coring or more preferably through trench studies.
There are three calibration coefficients for HMA rutting: βr1, βr2, and βr3. Recommended calibration coefficients for the rutting model are shown in table 60 and the resulting calibrated rut prediction models are shown in figures 12 through 16 for the five NCDOT pavement sections.
While the adjustment to the calibration coefficients appears to better characterize the observed performance of NCDOT HMA pavements (see also figure 17), the calibration coefficients should be reviewed and revised to include a larger number of NCDOT pavement sections.
The MEPDG software version 1.100 includes a performance prediction model for longitudinal (or top-down) cracking. However, at the initiation of this study this model was considered to still be a work in progress. In addition, NCDOT only characterizes load-related cracking as alligator cracking. Therefore, calibration of the longitudinal cracking model would require significant field testing and evaluation and therefore is considered beyond the scope of this study. All load-related cracking was considered to initiate from the bottom up and so only the alligator cracking model was calibrated as part of this study. In addition, the calibration of the MEPDG model for alligator cracking took into account the following assumptions (NCHRP 2004; NCHRP 2009):
Due to the variability of the subjective rating of alligator cracking NCDOT reported over time, the research team decided to use only the most recent distress severity reported at each site in the calibration process. The extent of alligator cracking was then distributed over the age of the pavement section similar to the process used for rutting. In addition, the following assumptions were made:
There are three calibration coefficients for alligator cracking: βf1, βf2,βf3. Based on the analysis conducted as part of this study, the recommended calibration coefficients for the alligator cracking model are shown in table 61 and the resulting alligator cracking prediction for the NCDOT pavement sections are shown in figures 18 through 23.
Figure 24 illustrates the comparison of the predicted versus measured alligator cracking for the NCDOT pavement sections based on the calibrated model. The use of the adjusted calibration coefficients does provide a much better fit of the data, but additional pavement sections should be included in the analysis prior to implementation of the revised model.
Calibration of the thermal cracking model within the MEPDG requires four measurements: crack depth, crack width, crack length, and crack spacing. NCDOT only includes a subjective rating of crack width and crack spacing as part of their distress survey. In order to convert the NCDOT subjective rating into measured values for use in the calibration process, a number of assumptions were necessary, including the following:
The adjusted thermal cracking model coefficients are summarized in table 62 and the resulting thermal cracking prediction for the NCDOT pavement sections are shown in figures 25 through 27. For the uncalibrated condition, the MEPDG software predicts no thermal cracking for all NCDOT pavement sites used in the calibration process. The estimated thermal cracking length resulted in a range of 1,045 to 1,278 ft (318.73 to 389.79 m) for the NCDOT pavement sections. This results in a crack spacing of approximately 30 ft (9.15 m), which is considered conservative, but was selected to avoid the MEPDG thermal cracking model limit of 50 percent cracking over the project length.
Figure 28 illustrates the comparison of the predicted versus measured thermal cracking for the NCDOT pavement sections based on the calibrated model. The use of the new calibration coefficients provides a much better fit of the data; however, additional pavement sections should be included in the analysis prior to implementation of the revised model.
As with alligator cracking, the reflective cracking model within the MEPDG (version 1.100) was also considered to be a work in progress at the initiation of this project. In addition, NCDOT currently does not include reflective cracking in their pavement condition surveys. Therefore, calibration of the reflective cracking model was not performed.
The MEPDG IRI model consists of a regression equation that is calculated from other distresses. For HMA pavements, the IRI model is a function of the initial IRI, a site factor (which considers pavement age, subgrade soil plasticity index, freezing index, and precipitation), load-related cracking, thermal cracking, and rut depth (see also appendix D, table D-7, of volume 2).
IRI calibration requires an extensive number of pavement sections and years of data collection that would be challenging to obtain under this study. NCDOT had only been collecting IRI data for 8 years, which was considered to be insufficient to accurately develop a calibrated IRI models. Therefore, calibration of the HMA IRI model was not performed.
NCDOT PCC Pavement Sections
As previously noted, for new rigid pavement design the MEPDG performance parameters include transverse cracking, transverse joint faulting, and IRI. Three PCC pavement sections in North Carolina were selected for assessing local calibration coefficients. The three pavement sections selected for calibration include I-10CC, I-2511BB, and I-900AC. Figures 29 and 30 illustrate the MEPDG uncalibrated predicted distress versus the NCDOT observations for cracking and faulting, respectively.
As shown in figures 29 and 30, the predicted performance using nationally calibrated models over predicts the distress development as compared to the NCDOT measured distress. The use of the nationally calibrated models would result in an overdesign and more costly pavement sections than is otherwise suggested by the performance of the three pavement sections.
The residual errors for transverse cracking are show in figure 31. On two projects (1-2511BB and I-900AC), the residual error increases with age, and are the same sign, but have a relatively high error compared to the performance criteria (15 percent). Even so, adjustment of the calibration coefficients may improve the predicted performance. The third project (I-10CC) has a very low (or zero error), indicating the nationally calibrated model reasonably predicts transverse cracking on this pavement section.
The residual errors for faulting are show in figure 32. The residual errors increase with age, are the same sign, and are relatively high compared to the performance criteria (0.12 in [3.05 mm]). Therefore, adjustment of the calibration coefficients may improve the predicted performance.
As noted previously, the goal in the calibration process is to reduce the error between the measured and predicted distress. As with the HMA pavement sections, there are very few data points, the measured data are primarily zero quantities, and no data are available approaching the established failure criteria or near the end of the performance period, making model calibration challenging. The approach taken by the research team was to minimize the standard error with the available data points. Because there are no data available at later ages, it was assumed that the original pavement designs, on average, meet the selected limiting criteria and reliability. While this is a significant assumption, current project distress levels appear to support this assumption (i.e., actual field performance indicates no faulting and no cracking on the NCDOT pavement sections).
There are four calibration coefficients for transverse cracking: C1through C4. Adjustment of C1and C2had the greatest influence on reducing the differences in initial cracking estimates; however, how much of an adjustment is needed for characterizing the long-term performance of the NCDOT PCC pavements is unknown since long-term performance data are not available. With no data beyond 18 years, minimizing the error to only the years with data resulted in the prediction of early failure—an average of approximately 42 percent cracking at the selected reliability of 90 percent at the end of 30 years. If it is assumed that the original NCDOT designs will on average perform to the selected design criteria (15 percent slab cracking) at the specified reliability (90 percent), then the calibration coefficients shown in table 63 are obtained. Using the adjusted calibration coefficients, the resulting adjusted transverse cracking model for sections I-10CC, I-2511BB, and I-900AC are shown in figures 33, 34, and 35, respectively.
While the calibrated model appears to better characterize the observed performance of NCDOT PCC pavements (see also figure 36), the use of the adjusted calibration coefficients also comes with a caution. These coefficients adjust the fatigue damage calculations that, among other inputs, are dependent on the layer material properties and traffic characteristics. The layer properties for these design runs were selected primarily as default values, as were most of the traffic characteristics. While it may very well be that the adjusted calibration coefficients reflect local performance, efforts should be made to ensure that inputs are as accurate as possible for local conditions.
The faulting prediction model includes eight calibration coefficients, C1through C7. The adjusted faulting model coefficients are summarized in table 64 and figures 37 through 39 illustrated the nationally calibrated models, PMS data, and locally calibrated models using the derived coefficients.
The calibrated faulting model appears to better characterize the observed performance of NCDOT PCC pavements (see also figure 40); however, as noted with the transverse cracking model calibration, efforts should be made to ensure that inputs are as accurate as possible for local conditions.
For JCP, the IRI model is a function of initial IRI, percent of slabs with transverse cracks, percent of spalled joints, cumulative joint faulting, and a site factor (which considers pavement age, freezing index, and percent passing No sieve for the subgrade soil) (see also appendix D, table D–10 in volume 2). As with the HMA IRI model, the extensive number of pavement sections required for a valid calibration was not possible within the scope of this study.
With a focus on the implementation of the MEPDG over the next decade, many highway agencies will need to characterize existing pavement condition to aid in the local calibration process. One of the major challenges with calibration will be in correlating the pavement condition data collected as part of the LTPP program to that contained within each States pavement management system. There are a number of challenges in this process that include: LTPP sections are comprised of 500 ft (152.5 m) lengths and may not fully represent the project distress, the LTPP data definitions may not completely reflect the distress definitions of each SHA, and many highway agencies may have only limited pavement condition data; the latter is particularly critical because the calibration process requires numerous pavement sections with performance data that extends over the analysis period.
With these limitations, the research team demonstrated how existing pavement data for flexible and rigid pavement sections from the NCDOT could be used to calibrate the pavement distress models contained within the MEPDG. For the HMA pavement sections, a MS Excel® solver was used to iterate the calibration coefficients to result in a minimum error, while an iterative process was used for the JCP pavement sections. In both cases, revisions to the calibration coefficients produced a better fit between predicted and measured distress; however, caution was also noted that additional pavement sections and performance data were needed prior to NCDOT consideration for adoption of the recommended calibration coefficients.