Framing the Use of Climate Model Projections in Infrastructure Engineering: Practices, Uncertainties, and Recommendations

As presented in Part a of Fig. 1, the first step is to determine the required climate variables for particular engineering applications, including types of climate variables, temporal and spatial resolutions, and variable formats for the applications. While projections of temperature and precipitation are more commonly used and are also the focus of this study, some applications may require other variables, such as humidity (Meagher et al. 2012) and wind speed (Nguyen et al. 2013). Depending on the applications, required temporal resolutions can also be different, e.g., daily temperature and precipitation series are required as input for process models like rainfall-runoff and water-quality models to assess algal blooms in freshwater (Chapra et al. 2017), and subdaily precipitation series are required to examine changes in the frequency and intensity of subdaily rainfall events (Cook et al. 2017). With respect to the implementation of climate projections, a wide range of climate variable formats is used across applications, including climatic design values (Auld et al. 2010) or time series used for engineering process models (Appendix S2).

The included variables, temporal resolutions, and spatial resolutions of the projections used in previous studies are provided in Fig. 2. The required spatial and temporal resolutions of the applications can limit available climate projection products (especially for some earlier studies in Table 1), while some further processing techniques are available and can be used to tailor the temporal and spatial resolution of climate projection as desired (i.e., applying postprocessing techniques, discussed subsequently). As presented in Fig. 2, many of the 50 example studies utilized climate projections in daily resolution, although an hourly format may be required for some applications. Many of the applications used projections with spatial resolution of around $10–100\mathrm{km}$ (generally corresponding to the resolutions of available dynamically downscaled projections and some early developed statistically downscaled projections), whereas more recently developed statistically downscaled projections with finer resolutions ($<10\mathrm{km}$) were used in some recent studies.

The next step is to obtain climate model projections for engineering analyses (Part b of Fig. 1). Brief discussions on the background of GCMs (including the sources, scenarios, and downscaling methods) are provided in the first subsection (additional details can be found in Appendix S3), and discussions of the workflow for obtaining GCM projections as presented in Fig. 1 are provided in the second subsection. Note that the sequence of descriptions given in the first subsection on GCMs is slightly different compared to the workflow of Fig. 1 discussed in the second subsection because the availability of GCMs and scenarios differs among various downscaled projection products. For example, a particular downscaled projection product may include results for only one or two climate change scenarios, or it may only include results for a subset of GCMs. It is also worth noting that, though the use of GCM projections is the more common approach to considering future climate conditions in engineering, other forecasting techniques based on historical observations and statistical methods are available and can be utilized for engineering applications (e.g., Cheng et al. 2014; Hu and Ayyub 2018; Lai and Dzombak 2020).

The last step is to determine whether or which postprocessing techniques should be applied (Part c of Fig. 1). The use of postprocessing techniques is separated as an additional step because it requires an additional action (or decision) for engineers to process obtained climate model projections, although these techniques can also be categorized as downscaling, bias correction, or model output statistics techniques (e.g., in Maraun et al. 2010). The reasons for applying postprocessing techniques include increasing the alignment between projections and historical observations (Maraun 2016), performing temporal disaggregation (e.g., Coffel et al. 2017), and facilitating the production of particular files, such as weather files from weather generators (e.g., Shen 2017).

Common postprocessing techniques used in the identified engineering applications include change factor, quantile mapping, and GCM-modified weather files, as summarized in Fig. 1. For the change factor technique, historically observed climate variables are added (for temperature) or multiplied (for precipitation) by an estimated factor or ratio between GCM historical and future simulations (Maraun 2016). Quantile mapping, on the other hand, modifies the distributions of climate model projections based on a statistical relationship (or transfer function) (Maraun 2016) between distributions of historical observations and GCM historical simulations. Historical observations with a reference period are needed for change factor and quantile mapping techniques. Because weather generators are commonly used to provide weather files for engineering applications, such as in building energy use estimation (Shen 2017), the parameters of weather generators can be modified with the estimates from GCM projections to consider future climate change (i.e., to produce GCM-modified weather files), and these GCM-modified weather files have often been used in these engineering applications. Other postprocessing techniques are also available and have been used, for example, in Hu and Ayyub (2019). Additional descriptions of postprocessing techniques are provided in Appendix S4.

Postprocessing techniques have been applied in many studies, as presented in Fig. 2, and these techniques can be applied for different purposes. For example, quantile mapping has been used to further modify the distributions of downscaled projections to reduce bias (Mannshardt-Shamseldin et al. 2012) or to tailor projections to match required temporal resolutions (e.g., Coffel et al. 2017; Meagher et al. 2012).

Two example applications of using temperature and precipitation projections from GCMs in design-related engineering analyses were performed in a preliminary manner to assess and highlight the sensitivity and uncertainties with respect to the decisions involved. The high-temperature performance grade (PG) of asphalt binder for the design of asphalt concrete pavement in Los Angeles (LA) and the diameter of stormwater drainage pipe (for a design of a 1-h storm with 5-year return period) in New York City (NYC) were assessed. The selection of high-temperature PG and the pipe diameter follow calculation procedures similar to those presented in Underwood et al. (2017) and Cook et al. (2020). See Appendix S5 for more technical details on the calculation procedures.

Based on Part a of Fig. 1, the following decisions on the requirements of climate variables were made for the two case studies: temperature and precipitation of the two cities, daily resolution (1-h precipitation amount can be estimated using a relationship with 1-day precipitation amount), city-specific projections (i.e., projections at the level of the two downtown National Weather Service stations, LA Downtown and NYC Central Park stations), and climatic design values (i.e., maximum pavement design temperature and 1-day precipitation amount with 5-year return period) were determined. Based on these requirements, the daily projections of temperature and precipitation were obtained for the grid locations closest to the two weather stations. Because the required spatial resolution is at the station level, historical observations for the two weather stations (Lai and Dzombak 2019) were collected (starting from 1878 for LA and 1869 for NYC) and used for postprocessing climate model projections. These historical observations were considered accurate representations of local historical climate conditions and were also used to assess the performance of the projections.

The two case studies of the high-temperature PG and the stormwater drainage pipe diameter are presented as illustrative examples. A number of project-specific assumptions were made, and simplified calculation procedures were followed (including the estimation of 1-h precipitation intensity using a relationship with 1-day precipitation amount). The presented results are thus preliminary and subject to limitations. Considering the different observations and projections used, the presented results can be different from the results in Underwood et al. (2017) and Cook et al. (2020), although similar calculation procedures were applied.

Following Part b of Fig. 1, the first step of the analyses was to quantify the variations with respect to different sources of GCMs (i.e., CMIP phases). While many of the commonly used downscaled projection products are from CMIP5, as previously discussed, the downscaled GCM projections for the current phase (CMIP6) are expected to be used in future engineering applications. The assessment of projection results between CMIP5 and CMIP6 therefore provides a baseline for understanding the differences when using the projections of CMIP6. To evaluate the differences between the two phases, the GCM projections of CMIP5 and CMIP6 at the selected grid locations (without downscaling or postprocessing) were compared with the observations in terms of annual average temperature in LA and total precipitation in NYC. Two climate change scenarios (RCP4.5 and RCP8.5 from CMIP5 and the equivalent SSP2-4.5 and SSP5-8.5 from CMIP6) were assessed to provide some additional comparisons, although the selection of scenarios is a subsequent decision, as presented in Fig. 1.

Projection results from the GCMs of CMIP5 and their CMIP6 counterparts are provided in Fig. 3. As expected, the raw (low-resolution) GCM projections exhibit a wide range of model bias when directly compared to the regional temperature and precipitation, suggesting that the raw GCM projections should be further processed or downscaled for engineering applications. A visual comparison between CMIP5 and CMIP6 series does not reveal substantial differences in projected trend, with the exception that the temperature projections in CMIP6 seem to exhibit slightly greater increasing trends than those of CMIP5. Some recent studies, such as Meehl et al. (2020), suggest that the GCMs of CMIP6 exhibit greater sensitivity to increases in ${\mathrm{CO}}_2$ concentrations, i.e., with greater temperature changes, than CMIP5. Although further analyses of the derived variables—the pavement temperature and maximum 1-day precipitation amount—are needed, the results of annual average temperature and precipitation are similar between CMIP5 and CMIP6 for the two cities, as presented in Fig. 3.

Following Part b of Fig. 1, the performance of downscaling and the differences among the downscaled projection products were assessed. Aligning with the assessment of annual average temperature and precipitation in Fig. 3, the improvement from using a downscaled projection product (LOCA projections, with a resolution of around 6 km) was assessed in Figs. 4(a and b). The LOCA projections were used in this case because LOCA is a more recently developed downscaling method, recommended by Kilgore et al. (2019), and were used in the US National Climate Assessment (USGCRP 2018). To compare the different downscaled projection products, one GCM (CanESM2) was selected (based on its availability in the selected downscaled projection products) and assessed in Figs. 4(c–h) with the projections of RCP4.5 from the original CanESM2 and several downscaled projection products: the results of three RCMs from NA-CORDEX [CanRCM4 (approximately 25-km resolution), RCA4 (approximately 50 km), and CRCM5-UQAM (approximately 50 km)] and BCCA (approximately 12 km), LOCA (approximately 6 km), and MACA (approximately 4 km). The results of probability density functions (PDFs) of daily temperature or precipitation, the estimated design values (maximum pavement design temperature and 1-day precipitation amount with 5-year return period), and the selected increments for the high-temperature PG and pipe diameter are presented in Figs. 4(c–h).

The results in Figs. 4(a and b) suggest that downscaling can substantially improve projections at the regional scale, though some differences between the LOCA projections and historical observations can be observed in precipitation. Notably, for the temperature projections in Figs. 4(a and b), the LOCA downscaling substantially improved the GCM projections in terms of bringing the projections to the regional observation level. On the other hand, the LOCA projections of total precipitation in part (a2) exhibit a moderately large bias, suggesting that postprocessing is likely necessary for the precipitation results.

Similarly, in Figs. 4(c–h), the downscaling methods reduced the model bias, especially in temperature, though some noticeable differences among downscaled projection products are apparent. The PDFs of daily temperature and precipitation from downscaled projections in Figs. 4(c and f) are notably more comparable to historical observations than the raw GCM projections, with one exception in the BCCA results for precipitation. The results of pavement temperature suggested a substantial improvement with the downscaled projection products (i.e., aligning better with the historical estimates than was the case for the raw GCM projections) in Fig. 4(d), while some moderate improvement was obtained for the precipitation results in Fig. 4(g). However, the various downscaling products can yield quite different results, as was also observed by Lopez-Cantu et al. (2020). For the pavement temperature in Fig. 4(d), statistically downscaled projections generally preserved the long-term future trend exhibited in the GCM, while the results from the three RCMs suggested slightly different trends. For the precipitation in Fig. 4(g), the statistically downscaled projections exhibited slightly greater temporal changes in the future period than the original GCM projections, while the RCM results seemed to project temporal changes more consistent with the original GCM. The results of high-temperature PG and pipe diameter in Figs. 4(e and h) are similar to the results obtained with climatic design values in Figs. 4(d and g): the use of raw GCM projections leads to large model bias, the downscaled projections can reduce the model bias (more improvement in temperature), and the projection products can lead to different estimation results (especially for pipe diameters). Because RCMs include regional features that may not be represented in raw GCMs or in statistically downscaled GCM projections, the RCM-downscaled projections merit particular consideration, although these RCM projections may be less comparable to local historical observations than statistically downscaled projections, such as presented in Fig. 4(c) for the PDFs of daily temperature; in addition, the RCM projections are limited to the number of GCMs and scenarios available compared to the LOCA or other statistically downscaled projections.

The model bias from the downscaled projection products exhibited in Fig. 4 was likely caused by a combination of factors, including the limitation of GCMs in producing greater numbers of low-intensity precipitation events, also known as the drizzling effect (Gutowski et al. 2003; Pierce et al. 2014), the limitation of grid-level observations (used for model calibration and downscaling) in representing station-level observations especially for precipitation extremes (Huang et al. 2017), the bias caused by RCMs (Hall 2014), the large variation of future trends projected by RCMs (Karmalkar 2018), and a reduction of spatial and temporal variations in statistically downscaled GCM projections (Gutiérrez et al. 2013). Studies like that of Shepherd (2014) have shown that GCM projections of atmospheric circulations and related phenomena, such as precipitation, are subject to great uncertainty, which can be enlarged at regional scales, and in this case, the downscaled projections of precipitation generally exhibit larger model bias.

Based on the results in Fig. 4 (the mean absolute errors and the alignment with observations) and the recommendations from previous studies (e.g., Kilgore et al. 2019), the LOCA projections seem to be a moderately more appropriate option among statistically downscaled projection products and should be considered for use if resources are limited for considering other products. A careful evaluation of the obtained downscaled projections with historical observations is needed, however. Further analyses among different downscaled products and different locations are also needed for a more complete evaluation.

Following the workflow presented in Fig. 1, further analyses were conducted considering and using different scenarios and GCMs for the two case studies (the LOCA projections were used). The results are presented in Fig. 5, with the comparisons of results using RCP4.5 and RCP8.5 scenarios from one GCM (CanESM2, with LOCA downscaling) presented in Figs. 5(a–d); and the use of all available 32 GCMs (with LOCA downscaling) presented in Figs. 5(e–h). Additionally, the results from using the median projections from all 32 GCMs as a GCM ensemble are presented in Figs. 5(e–h).

The use of a multimodel ensemble provides a so-called consensus climate change trend (Taylor et al. 2012) and more insight into the range of model uncertainty than the use of a single GCM. As presented in Figs. 5(a and c), it is difficult to differentiate the long-term climate change trend from natural variability exhibited in the projection results of one GCM (CanESM2 in this case), especially for precipitation. In comparison, the use of a GCM ensemble, as presented in Figs. 5(b and d), provides climate change trend with less variability and more consistency between scenarios for the near-term period [up to 2050; different scenarios are expected to be similar in a near-term period (Hawkins and Sutton 2009)]. The use of different GCMs additionally provides an assessment of the possible range of model uncertainty, although such model uncertainty can be large especially for precipitation. The estimates of pavement temperature and 1-day precipitation amount using the GCM ensemble in Figs. 5(b and d) also exhibit some model bias compared to the historical estimates. Postprocessing of the LOCA projections may be needed.

Analyses were conducted on the postprocessed projections, corresponding to Part c of Fig. 1. A change factor method and a quantile mapping method (nonparametric, matching the empirical cumulative distributions with 20 quantiles) were used. Other postprocessing techniques, such as the integration with weather generators, have been investigated in studies like that of Maraun et al. (2010). The objective of these analyses was to assess the performance of postprocessing, possible variations when different postprocessing techniques are applied, and results when different stages of decisions are combined and cascading uncertainties ensue. The results of two case studies using the LOCA projections (32 GCMs) with the two postprocessing techniques are presented in Fig. 6. A separate analysis was also conducted comparing different downscaled projection products with postprocessing, which is further discussed in Appendix S5.

Based on the results of Fig. 6 (and additional analyses presented in Fig. S1 of Appendix S5), the postprocessing techniques can increase the alignment with historical observations, though these techniques are also subject to some limitations. In Fig. 6, both the projection uncertainty from different GCMs and projected future trends were modified following their application with postprocessing, especially for RCP8.5, suggesting the postprocessing can inflate future trends (Maraun 2013) and increase the uncertainty level. The selected high-temperature PG and pipe diameter are generally among the three increments, although the particular years for the increases in increment needed vary among the different results in Fig. 6, suggesting the sensitivity of the results with respect to particular postprocessing techniques (Cook et al. 2020) and historical reference periods (Hawkins and Sutton 2016). Additional analyses on the use of raw GCM projections with postprocessing (included in Appendix S5) also suggest that the raw GCM projections after postprocessing are comparable to the downscaled projection results, indicating that postprocessing may reduce so-called added values (Karmalkar 2018) from downscaling.

Considering that both statistical downscaling and postprocessing techniques modify the projections based on statistical methods and historical observations, one statistically downscaled projection product is likely sufficient when postprocessing is applied. Other postprocessing techniques that can better integrate the GCM projections with historical observations may be needed.

An overview of experience in applying climate model projections in engineering contexts is provided in this work, including summarized general procedures to obtain and use climate model projections, a review of commonly used projection results and processing techniques from a wide range of civil and environmental engineering topics, and two case studies to demonstrate the methods and evaluate the sensitivity and uncertainty with respect to different decisions and selections. Based on a review of 50 engineering studies in which GCM projections of future temperature and precipitation have been applied, a workflow for typical procedures used to incorporate climate model projections in engineering applications was developed (Fig. 1). The workflow includes three major steps: (a) determining requirements for climate variables, (b) obtaining climate projections, and (c) processing climate projections.

One main challenge of applying climate model projections is the underlying complexity of selecting and incorporating different downscaled projection products, climate change scenarios, GCMs, and postprocessing techniques. A summary of the findings based on the two case studies is provided in Table 2. For engineering applications that require future local climate information, one critical task is to obtain climate projections that are properly transformed to the desired local or regional scale. Downscaled projection products and postprocessing techniques are useful but can be subject to challenges. The cascading uncertainty from different GCMs, future climate change scenarios, downscaling methods, and postprocessing techniques can be impractically large, especially in providing climatic design values for a particular future period. Engineering decisions need to be made with caution when applying climate model projections because the results are sensitive to the particular selections made for data sets and processing techniques.

Advances from the climate modeling community are ongoing and could reduce the uncertainty in applying climate model projections in engineering applications, while the general procedures for applying climate model projections as presented in Fig. 1 are expected to remain applicable. Ongoing developments include the evaluation of downscaling methods (Jack and Katragkou 2019), improvement to the understanding of GCM performance (Eyring et al. 2019), and further collaboration between experts to improve state or regional climate change assessments (Moss et al. 2019). Further details are provided in Appendix S6.

Given that the use of climate model projections in engineering applications is rapidly increasing, a framework is urgently needed for the use of GCM projections in a broad range of civil and environmental engineering applications, in addition to the flowchart provided in Fig. 1. A common framework for engineers could reduce the workload involved in identifying and evaluating climate models and projection products and in updating climate change assessments when new model projections or downscaled projection products become available.

Based on the results from two engineering application case studies and generalizable findings from the existing literature, some observations and recommendations are offered. First, because engineers are often unfamiliar with the performances of individual GCMs for different climate variables, the use of an ensemble of GCMs with an average projected climate response is more appropriate, while guidance on individual GCM performance, such as that presented in Briley et al. (2020), can be highly informative. Second, based on the results of the two applied case studies and recommendations, such as those obtained by Kilgore et al. (2019), the LOCA projections seem to be a more appropriate option among commonly used statistically downscaled projection products, though a careful evaluation with historical observations is needed, especially when it comes to estimating precipitation extremes. Third, postprocessing techniques will likely be necessary for precipitation-related engineering applications because of the gaps between grid-level downscaled projections and station-level observations. Fourth, commonly used postprocessing techniques, like quantile mapping and change factors, are subject to several challenges, so other postprocessing or downscaling techniques (e.g., in McGinnis et al. 2015) should be used or developed to increase reliability, especially when long-term regional historical observations and raw GCM historical simulations are available. Finally, if local long-term historical observations are available (which facilitate the use of postprocessing), one statistical downscaling product (e.g., LOCA) could suffice in engineering applications, with the possible use of dynamically downscaled projections for additional evaluation.