A new design exploring framework based on sensitivity analysis and Gaussian process regression in the early design stage

ABSTRACT With building energy codes getting strict, quantitative analysis is necessary in the early design stage of high-energy-performance buildings. To fully explore the design space, a highly efficient method is necessary. In the research, a new design exploring framework was proposed. Parameters are screened and separated into groups based on the sensitivity analysis, to reduce the dimension of inputs. Gaussian process regression, able to deal with the uncertainty in inputs, was used to make metamodels to reduce the calculation cost and fully explore the design space. Dashboards were made to visualize the data interactively, to help designers make decisions and make communications smooth. This framework was demonstrated with a case study and showed its efficiency.


Introduction
Buildings are responsible for more than 40% of global energy use and one-third of greenhouse gas emissions globally (Hirst 2013). In Japan, since the oil crisis in the 1970s, the energy consumption by building sector considerably increased by about 250%, which accounts for 34.5% of the total energy use in the country (Ministry of Economy, Trade and Industry, JP 2015). With. the progress of urbanization, the energy use of building sector will continuously increase by about 30% in the coming 20 years (U.S. Energy Information Administration 2017).
On the other hand, building sector has the largest potential for a reduction in energy consumption and greenhouse gas emission. It has been proved that the architecture design, especially in the early stage, has a significant impact on the energy performance (Naboni et al. 2015), while the cost of changing design will get higher and higher in the later stage (WBDG 2019). However, the most important decisions are made in the early design stage by architects (Schlueter and Thesseling 2009), usually with some rules of thumbs. With the energy codes getting more and more strict in recent years and Zero Energy House gradually becoming mandatory in Japan, it has been appointed out by some researchers that guild lines or rule of thumbs are not enough to ensure the energy performance (Attia 2018). Integrated Design Process is required to achieve ultra-low-energy design (Ferrara, Sirombo, and Fabrizio 2018). Quantitative energy analysis should be carried out in the early stage. However, limited by the low efficiency of the trial and error process, the possibilities of the design are often not fully explored (Attia 2018). Meanwhile the communication cost, within the design team and between designers and clients is considerable. Automatic optimization, with Genetic algorithm, Particle Swarming Optimization, etc., is another way to efficiently improve the energy performance of the design (Wang, Rivard, and Zmeureanu 2006). However, it is not very suitable in the early stage, as there are a lot of factors and criteria that cannot quantify be such as aesthetical or historical things. The optimized results might be undesirable to designers. It requires thousands of iterations of simulation to run a successful optimization, which is quite calculation costly and takes too much time. On the other hand, in the early stage, it is better to explore the design possibility as much as possible, rather than get accurate analysis or an optimized result.
The interactions between parameters are rarely analyzed in the design process. When the designers study a certain parameter, they usually assign constant values to other parameters based on some assumptions. The decisions made in earlier phases are often influenced by those made in the latter stages and become no more optimal, due to the interactions between the parameters. As a result, the designers have to move back and repeat the studies in previous phases.
This study focuses on four questions, (1) How to explore the design possibilities, (2) how to keep the calculation cost at a low level at the same time, (3) how to utilize the data to help the designers when making decisions and (4) how to avoid duplicated work. This paper has six sections, including this introduction.
Related works, including published tools and researches, are reviewed in Section 2. Section 3 describes the framework of design exploring, as well as the sensitivity analysis (SA), metamodeling and data visualization methods used in this study. Section 4 provides a case study as a demonstration of this framework. Conclusion and future work are presented in Section 5.

Dashboard tools
Dashboards, able to visualize the building parameters and the energy performance, are useable when designers make design decisions.
Insight 360 (Autodesk 2019) is a building energy performance assessment tool published by Autodesk. This designer-oriented dashboard tool is intuitive and easy to use. Good data visualization helps the designer understands the energy performances well. Parameters can be modified, and the feedback is realtime. However, there are still some limitations. The parameters and their range that can be studied are limited to the preset. Only the average value of energy consumption of all cases is displayed. Other criteria, such as thermal comfort, are not included or extendable.
Designer Explorer (Tomasetti 2019) is another tool under the concept of design dashboard. It consists of two parts, Grasshopper3D components and online data visualization. Designers can do the analysis in Grasshopper3D and output the results into a CSV file. The online data visualization tool visualizes the data and makes dashboards once this CSV file is uploaded. This process takes quite a long time to execute the simulations if the designer wants to explore the possibilities of design widely.

BIM-based tools
The authors have made several interviews with professional architects. They said that they prefer to stick to BIM tools from the very beginning, even though they know that Grasshopper3D might be more convenient in the early stage, so that they can push the work more smoothly to latter stage. Several Building Information Modeling (BIM)-based tools, aiming at the early design stage, have been developed by researchers. Schlueter and Thesseling (2009) developed a plug-in for Autodesk Revit called Design Performance Viewer, which could evaluate the energy performance of the building, using the information from BIM. Once the designers modified the model, they could get nearly real-time feedbacks. Though modeling and analyzing were integrated, designers still had to work in the trialand-error process. It was difficult to compare a lot of alternatives simultaneously. Rahmani Asl et al. (2015) integrated BIM and multi-objective optimization. They extracted the information from the BIM model to carry out energy, daylight and structure analysis. The design would be optimized by Optimp, an optimization tool integrated in Dynamo. The whole process was highly integrated, nevertheless, the calculation was still very costly. Jalaei and Jrade (2015) integrated BIM and LEED certification system for the conceptual stage. This integrated tool could generate LEED certification point using the information from BIM and help the designer selects the proper materials. However, LEED certification required a highly completed design. A lot of information necessary for LEED was lacked in the conceptual stage. As a result, it was still a question that whether LEED provides good criteria in the early stage.

Sensitivity analysis
SA has been adopted by many researchers to reduce the number of parameters and calculation cost. Heiselberg et al. (2009) introduced Morris method into the design of sustainable buildings. They screened the parameters by their importance and dropped less important parameters in the latter stage. They commented that SAshould be performed in the early stage when important factors were not decided. Garcia Sanchez et al. (2014) performed both firstand second -order analysis in building energy simulations. They commented that higher-order effects could help better understanding the results of SA. Østergård, Jensen, and Maagaard (2017) described an approach to explore the multi-dimensional design space. They used Morris method to screen parameters. The importance of each parameter helped designers to make decisions in the latter stage. Nguyen and Reiter (2015), Rivalin et al. (2018) and Gagnon, Gosselin, and Decker (2018) compared several SAtechnics applied to building energy performance assessment. Linear regression was reported not good enough by most researchers. Morris method showed both good efficiency and enough reliability. Variance-based methods, though had the best reliability, were very calculation costing. It was also pointed out that Latin hypercube outperformed the standard Monte Carlo in Morris method (Figure 1).

Metamodeling
Metamodeling is a very effective way to reduce calculation cost. Hygh et al. (2012) presented a Monte Carlo framework to develop a multivariate linear regression model based on 27 parameters. The coefficient of each parameter in the regressed model could be used as the sensitivity. Østergård, Jensen, and Maagaard (2017) also made metamodels with multivariate linear regression. They made a "what-if" dashboard, with the metamodels running in background, to help the designers making decisions. Gossard, Lartigue, and Thellier (2013) trained artificial neural network (ANN) with the simulation results of energy performance of the building. They used this trained ANN as a part of objective function in GA optimizations. Asadi et al. (2014) also combined ANN and GA in their optimization process and applied it in retrofit projects. Rivalin et al. (2018), Wei et al. (2015) and Østergård, Jensen, and Maagaard (2018) compared several techs of metamodeling applied to building energy performance assessment. Gaussian process was reported to be the most robust and easy-to use, neural network and multivariate adaptive regression splines also had good performance. In some cases, polynomial chaos showed better accuracy than Gaussian process did.

Targeted parameter and non-targeted parameter
In order to make this research more understandable, we defined two concepts, targeted parameters and non-targeted parameters. Target parameters, which is actually an alternative way to say parameters of interest, means the parameters that can be decided by the design team, such as window to wall ratio (WWR), insulation, etc. Other parameters influencing the energy performance but can hardly be decided or controlled by the design team, such as climate, occupancy, etc., are called non-targeted parameters in this research. We think that both targeted and nontargeted parameters should be considered, as well as the interactions between them.

Framework
In this research, we proposed a framework of design exploring based on SA ( Figure 2).
Architects will firstly work on the conceptual schema of the design, decide the rough shape, the volume or the topology of the functional spaces. Based on the conceptual design, the design team, including architects and experts from different disciplines, will list the targeted and non-targeted parameters and their ranges or distributions. The second step is to analyze the sensitivity of each parameter. Samples will be made with Latin Hyper-cube. An IDF file will be made for each sample parametrically and simulated with EnergyPlus. Morris method will be used  to get the sensitivity of each parameter based on the simulation results. Parameters will be screened based on their sensitivities, or importance in another word. The third step is to analyze the interactions between parameters with Expanded Morris method, by quantifying the higher-order effects. An interaction matrix will be made.
The design team can separate the targeted parameters into groups based on their interactions, so that one problem is divided into several problems and the dimension of each problem is reduced. The design team can study the problems phase by phase. In each phase, those non-targeted parameters, which interact with the targeted parameters in this phase, will also be studied together. The parameters will be sampled with Monte Carlo method and simulated with EnergyPlus. A meta-model will be made with Gaussian process. The targeted parameters and results from simulations will be used to train the model. Nontargeted parameters will be regarded as noise of the model.
With a trained meta-model, the design team is able to get real-time feedback when any parameter is modified, and test thousands of cases in minutes. A design  dashboard, visualizing the data, can help the design team understand the energy performance, make decisions and communicate in teamwork or with clients.
After decision made, the design team can move into the next phase, do simulations, train a metamodel, expand the design dashboard and make decisions on another group of targeted parameters.

Sensitivity and interaction analysis 3.3.1. Morris method.
The SA can be roughly divided into variance-based method and one-step-at-a-time method (OAT). As in the early stage of design, efficiency was more important than accuracy, we employed Morris method, a global OAT for global SA, to analyze the sensitivity of each parameter.
The first step is to generate the input matrix in the range of (0,1]. In this research, Latin hypercube sampling (LHS) was used to generate the input matrix Mo (Eq.1), as LHS can better fill the exploring space than random sampling. The shape of Mo is r in row and k in column. Column count k is equal to the number of parameters of interest. Row count r represents the number of times that OAT will be repeated.
One row in the input matrix can be regarded as a k-dimensional vector, X i . (Eq. 2) An element in the vector, x j i , will move a small distance of equal size Δ in its dimension (Eq. 3). The absolute value of the difference in two results, y X jÀ 1 i � � and y X j i � � , before and after moving, divided by the distance is used to represent the local sensitivity of the jth parameter (Eq.4), which can be regarded as an approximation of the absolute value of the partial derivative. The local sensitivity of each parameter will be calculated, and y X ð Þ will be calculated k þ 1 time for each row.
This procedure repeats r times for all the rows in the input matrix, so the calculation of y X ð Þ will be executed totally r k þ 1 ð Þ times. The average value of local sensitivity of jth parameter in all the rows will be used as its global sensitivity (Eq. 5).
Mo ¼  Figure 5. The movement of a sample, Expanded Morris method.

Extended Morris method.
In this study, we also studied the interactions between parameters. Interaction between two parameters means the influence of one parameter on the sensitivity of the other one. For example, if the interaction between x 1 and x 2 is strong, it means that the value of x 1 strongly influents the relationship between x 2 and y, also the relationship between x 1 and y is strongly influenced by x 2 . Extended Morris method was employed in this study to quantify the interactions between parameters. Like Morris method, an input matrix Mo (Eq. 1) is generated firstly. For each pair of parameters, jth and j'th, the vector X jÀ 1 i moves a small distance of equal size Δ along axis j (Eq.7), then along axis j' (Eq. 8), then along both axis j and j' (Eq. 9). The local interaction between jth and j'th parameters can be calculated with Eq.10, which quantifies the influence from a parameter on the local sensitivity of the other. For each row, y X ð Þ will be calculated � times in total. The average value of local interaction between jth and j'th parameters from all rows is used as the global interaction of them (Eq.11).

Gaussian process
In this research, we use Gaussian process regression (GPR) to build metamodels. Gaussian process, also known as Kriging method, is a kind of supervised learning method based on the Bayesian inference. A prior probability is defined based on the covariance matrix of the observations (Eq. 12). Predictions are made by interpolation governed by this probability (Eq. 13). So, GPR can be simply described as, the closer two points are located in the input space, the similar the predicted outputs will be. Usually, a kernel function is used to generate a prior covariance matrix based on the training data. Radial basis function kernel (RBF) (Eq. 16) is used in this research. The prediction given by a trained GPR is a normal distribution, rather than a single value, which makes GPR a very robust method to regress noisy data.

Interactive data visualization
We think the data visualization should be interactive.
Once the designers modify some parameters, the corresponding results should feedback in real-time. This allows designers, and even clients, quickly test their idea. In this research, Microsoft Power BI is used. Power BI is an interactive data visualization tool based on Power Query, a data retrieval technology. Power BI is able to retrieve data from a huge amount of data resources based on the filters set by users and redraw the graphs immediately. Once the data visualization has been published, it can be accessed from PC, iPad or smartphones, which is a big benefit during meetings.

Description of the case study
The case study in this research is a midrise office building located in Tokyo. There is a small lake to the south of the site. The main entrance will be located on the east, facing a road. The west side is adjacent to another building. The plan, consisting of two office rooms in south and north and one core, is quite widely used in energy performance study in Japan (Takizawa 1985). The design team decided to make big openings on the south façade to make full use of the nice view from south. However, based on the climate in Tokyo, big opening on south, even with overhangs, would result in higher energy consumption. How to balance the view and energy performance would be a key point in this design. Fins would be attached to east façade, to shade the building from the rising sun, which would also be important visual elements. How to decide to size of the fins, balance the appearance and performance is another question.

Parameters and criteria
In this study, we mainly focused on the envelop performance. Though orientation and aspect ratio had a huge impact on the energy performance, limited by the site, these two parameters were not included. The size of the opening was represented by WWR. As this study aimed at the early stage, we simply represented the performance of the window with Solar heat gain coefficient (SHGC) and U-factor of the glass. The ratio between overhang depth and window height was used to represent the shading on the south façade. The depth and width of the fins on the east façade were also studied. Thickness of insulation material was used to represent the insulation of opaque part of the envelop. The solar absorptance of the outside layer of the wall was also studied, as it was related to the color of the façade, which designers might be interested in. Non-targeted factors, heating and cooling setpoints, internal heat gain and air change rate, were also taken into consideration as they had a direct influence on the energy performance. In the GPR, however, these factors were regarded as noise, as they could not be precisely predicted in early stage. Tables 1 and 2 show the parameters and their ranges. Both energy performance and thermal comfort were studied. Hourly heating and cooling load were calculated and converted into annual energy density to evaluate the energy performance. The top 10 heating/cooling load from the 8760 results were picked up to get the heating/cooling peak load. Operative temperature was used to study the thermal comfort. The percentage of comfortable, hot and cold hours was calculated. We also picked up the top 10 and bottom 10 operative temperature from 8760 results to study the extreme situations. Table 3 shows the criteria.
EnergyPlus 8.8 was used to execute the simulations. The template of office building in climate zone 3a from ASHRAE 90.1 2010 was used. We extract one floor with the standard plan from the building to do the calculation. The constructions and schedules followed the template. The roof and floor were set to adiabatic. The surroundings were reproduced with shading surfaces.   and shadings were made and non-geometric properties were adjusted, such as SHGC, AC setpoints, etc. Figure 6 shows the flow that IDF files could automatically generated.

Sensitivity analysis
The sensitivity of each parameter on each criterion was analyzed with Morris methods. Twenty cases were generated with LHS. Each case was expended into 28 cases as there are 27 parameters. An EPW file of TMY from Meteonorm was used. As shown in Figure 7, the sensitivities of nontargeted factors were very high, which was disturbing. On the other hand, designers cannot really control them. For these reasons, we did not take them into consideration when determining the thresholds of sensitivities. For each criterion, the threshold was set to the half of the average value of the sensitivities of targeted parameters. A parameter would be skipped in further phases if its sensitivity was lower than the threshold. Appendix 1 shows the sensitivity of each parameter on each criterion.
It was found that in this case, as this was a coolingmain building, the influence of insulation and solar absorptance on the total annual electricity consumption was almost negligible, which meant that the designer could feel free to decide the color and the material of the external surface. The sensitivity of the insulation of the south façade on the operative temperature, however, passed the threshold. So, we ignored all the insulation and solar absorptance parameters besides the insulation on the south in latter phases. The fin depth just passed the threshold a little. Nineteen parameters left.
Twenty cases, though enough to screen parameters, were too rough to visualize the sensitivities in latter steps. We did additional 250 cases for the left 19 parameters. Each case was expanded into 20 cases. In this  step, we also took climate change into consideration. Each case was analyzed under four climates, TMY, 2020, 2050 and 2080 from Meteonorm. Twenty thousand simulations were executed in total. A sensitivity database was made.

Interaction analysis
The interactions of the 19 parameters with respect to each criterion were analyzed with Expanded Morris method. Twenty cases were generated using LHS, and each case was expanded into 191 cases. Appendix 2 to 6 show the interaction matrices. The data in the matrices have been divided by the average value of interactions between WWR and U-factor on four orientations. The threshold was set to 0.5. The parameters of each orientation have strong interactions with each other. The interactions between parameters belonging to different orientations are quite weak. The 19 parameters were separated into 4 groups (Table 4). When we study a parameter, it is better to study all the interactive parameters simultaneously. For example, when the designers try to decide the WWR on the south, it is recommended to take SHGC and U-value of the window on the south also into consideration. However, designers can study the south façade and the east façade independently.

Metamodels by Gaussian process regression
We firstly studied Group 1, the parameters related to the south façade. The values of the parameters in this group were decided randomly with LHS, other parameters kept constantly to the mid value of their range. 2000 cases were generated and simulated as the training set, other 500 cases as testing set. Each case was simulated with four climates, TMY, 2020, 2050 and 2080. In this phase, non-targeted parameters were regarded as noise and not included in the training. GPy (Sheffield ML 2019) was used to make the GPR models and train them.
Other 200 thousand cases were generated by LHS. The energy performance and thermal comfort of these cases are predicted with the trained GPR. As the output of GP for each case was a normal distribution, it was very difficult to analyze and visualize all the 200 thousand results. Symmetrically about the mean value, 60% Confidence interval was used (Figure 8). We extracted the bottom, mean and top value of the confidence interval to represent the possible energy performance and thermal comfort of a case. A database was made.
We repeatedly made the GPR for other three groups and add the predicted data into the database.

Data visualization and design dashboard
Power BI was used to visualize the data and make design dashboards.
For demonstration, we made three dashboards. Designers can adjust the parameters by moving the sliders of filters. The variation of criteria and sensitivities will be reflected in real-time. The database can be utilized continuously throughout the design process. New pages in dashboards can be easily made.
In the energy density and thermal comfort dashboard (Figure 9), designers could move the sliders on the left side to adjust the ranges of parameters. The EPW years could also be selected. Power BI would then filter all the cases that the parameters were within this ranges from the database and display their results on the right side. For each case, three indices of energy density, the bottom, mean and top of the 60% confidence interval, are displayed, named "Min prediction", "Mean prediction" and "Max prediction" in this dashboard. The average values of these three indices of all the filtered cases are displayed on the upper right in the form of a speedometer. Furthermore, we also displayed the distributions of the three indices of all the filtered cases, which can help designers better understand the possibilities and risks of their decisions.
Due to the interactions between parameters, the sensitivity of each parameter would change with the values of other parameters changed. Another dashboard was made to display the sensitivity of each parameter on each criterion ( Figure 10). By seeing the sensitivities, the designers can be informed that which  parameter is important, and they can make more efforts on those. This could be an important hint for designers. As a building would last for decades, climate change should also be considered in design stage. A climate change dashboard (Figure 11) was made. We thought that showing the influence from climate change on the sensitivities of parameters, rather than that on the energy consumption or thermal comfort, can better help designers understand the impact on the buildings. For example, we can see shadingrelated parameters getting more important but the insulation-related parameters getting less important. So, the designer should pay more attention to the shadings.

Additional predictions and simulations
When designers trying to explore the design space with the dashboard, it was found that, with the range of parameters getting small, the cases met the filters  got quite a few and the distribution of the results became unsmooth. We did additional 10 thousand predictions (for each climate and totally 40 thousand) with trained GPRs within the shrunk range and append the new data into the database. The distribution got smooth as shown in Figure 13. Designers could do more specific tests on each parameter.
From the sensitivity page, we found a blind spot. The sensitivity of SHGC of north opening was much higher than we expected. It is reasonable because this building is 15 degrees East of South, the solar radiation might get into the room in the summer afternoon. The design team decided to attach a fin to the north façade. We did other 1000 simulations additionally to study the opening on the north. WWR, SHGC, U-value of the north window and the newly added fin depth on north were adjusted with LHS, other parameters kept fixed. Another GPR was trained and new dashboard  ( Figure 12) was made based on the prediction. As the whole process was highly automated, it took just several hours to run additional simulations and make new dashboard.

Results
In the last section, we demonstrated the proposed process with a case study. The CPU of the computer we used in this demonstration was an 8-thread one, which was consumer-grade and widely used in architecture design studio. Eight simulations could be executed simultaneously. It took around 10 to 30 second the run one simulation.
All the IDF files were generated parametrically using Python codes. It took several minutes to generated more than thousands of IDF files.
There were 27 parameters at the beginning. To do the SA with Morris method, the simulations were executed 560 times, which took about 20 minutes. The number of parameters reduced from 27 to 19 based on their sensitivities. Then, the interactions between the 19 parameters were analyzed with Extended Morris method. The simulations were executed 3820 times, which took about 2 hours. Based on the interaction, the parameters were separated into four groups. In this way, the 27-dimensional problem has been simplified into one 8-dimensional, one 9-dimensional and two 7-dimensional problems, which helped the designers avoid the curse of dimensionality.
For each group, 2500 IDF files were generated and simulated with the weather data of TMY, 2020, 2050 and 2080. The simulations were executed 40,000 times, which took about 20 hours. The results were used to train and test Gaussian process models. With the trained GP models, the energy performance and thermal comfort of other 200 thousand cases were predicted within about 2 minutes.
Using the proposed process, a database containing 200 thousand cases could be made in two days. Comparatively, it takes more than 1 month to simulate so many cases even using a high-end workstation.
The non-targeted parameters, such as AC setpoints or internal heat gain, are actually noise, rather than parameters, in the early design stage. In this study, GPRshowed its robustness when dealing with noisy data. By outputting the boundaries of the confidence interval, designers can also notice the possibility and risk of each decision, rather than only watch on the mean value.
A blind point, the shadings on the north façade, was found by widely exploring the design possibility and showing the sensitivities. As the whole process is highly automated after the first run, even there was modification in the base IDF file, new database and dashboards were made with 1 day.
Interactive data visualization, with real-time feedback, allows designers to test their ideas more efficiently. It also benefits the communication, within the design team between different disciplines, or between the design team and clients.
The whole process ( Figure 14) can be concluded as, (1) Conceptual design (2) Sensitivity analysis and parameter screening (3) Interaction analysis and parameter grouping (4) Metamodeling and prediction (5) Data visualization and decision-making Designers can return to step 4 or step 1 when making decisions if necessary.
The proposed process was summed up into a template of Python codes, commented with details, and uploaded to the GitHub.

Conclusion and future works
A new design exploring process based on SA and GPR was proposed in this research and demonstrated with a case study. Sensitivity and interaction analysis can reduce the dimension of the problem. Informed of the sensitivities, designers can concentrate on the important parameters. Based on the interaction analysis, parameters separated into several groups, which can be studied independently phase by phase. What is more, it can also prevent earlier decisions from being influenced by latter ones, avoiding repeated works. Trained Gaussian process models can predict the energy performance of a huge number of cases in a moment. This process can help designers widely exploring the design possibilities.
However, we found some new problems in this research.
In the case study, it was found that the parameters belonging to different orientations were independent, which could also be inferred by common sense and was disappointing. Further studies about how to make better use of the interaction matrices are needed.
Though GPR showed its robustness when dealing with noisy data, it was also found in this study that if the variance of criteria caused by the non-targeted parameters was close to or even more than that by the targeted ones, GPR very likely failed to make proper predictions. Shrinking the ranges of non-targeted parameters can help to improve the prediction, but a very narrow range is against the original idea of this research.
As Power BI is a database-based tool and unable to utilize the metamodel, designers have to add data into the database when looking closer to a certain range of the design space. With the volume of data getting huger, it becomes very heavy and memory consuming to run the dashboard.
The further steps of this research would be, • Make the whole process more automatic • Make further use of interactions and study higherorder effects • Reduce the variance of the non-targeted parameters but keep their ranges reasonable • Introduce the variance of inputs into the GPR, for example, use Sparse GPR (Snelson and Ghahramani 2006) • Make cross-platform data visualization which can run metamodel in background