Abstract
It has been convincingly argued that computer simulation modeling differs from traditional science. If we understand simulation modeling as a new way of doing science, the manner in which scientists learn about the world through models must also be considered differently. This article examines how researchers learn about environmental processes through computer simulation modeling. Suggesting a conceptual framework anchored in a performative philosophical approach, we examine two modeling projects undertaken by research teams in England, both aiming to inform flood risk management. One of the modeling teams operated in the research wing of a consultancy firm, the other were university scientists taking part in an interdisciplinary project experimenting with public engagement. We found that in the first context the use of standardized software was critical to the process of improvisation, the obstacles emerging in the process concerned data and were resolved through exploiting affordances for generating, organizing, and combining scientific information in new ways. In the second context, an environmental competency group, obstacles were related to the computer program and affordances emerged in the combination of experience-based knowledge with the scientists’ skill enabling a reconfiguration of the mathematical structure of the model, allowing the group to learn about local flooding.
Introduction
This article examines how we can learn about the physical environment through computer simulation modeling. The question arises because of the distinct epistemological character of computer simulation models. Philosophers of science have argued convincingly that computer modeling is radically different from previous scientific approaches. For example, Stengers (2000, 135) explains that all models “say of themselves that they are fictions, and should be treated as such” because “several models can coexist without any problem, each defined by different variables, each having its zone of privileged validity or its specific advantages.” This ability to coexist distinguishes models from traditional theories in the natural sciences, where it is commonly thought that there can only be one correct explanation for each natural phenomenon. Winsberg (1999, 287-88) considers models to be so different that the “transformations involved in simulation cannot be judged solely by how their results compare to the world.” The impossibility of judging results by comparisons with physical phenomena arises, in part, because models may relate to entities that cannot be measured by other means. According to Knuuttila (2006, 43), modeling differs from traditional scientific knowledge creation because “that which is being represented when we build models is usually something that has already been theoretically rendered and conveyed.” Computer simulation modeling does not aim to produce new theoretical and general explanations of entities, as traditional scientific research is understood to do, but to show the potential consequences of interactions between theoretical principles. As objects involved in research, models have been conceptualized as doing a new type of work, providing “a partial representation that either abstracts from or translates into another form, the real nature of the system or theory, or one that is capable of embodying only a portion of a system” (Morrison and Morgan 1999, 27). This sets them apart from devices previously used in the natural sciences to generate knowledge about the physical world.
If we regard computer models as a different kind of scientific tool, which does a particular type of work, the ways in which models enable scientists to learn about the physical world must also be understood differently. In traditional philosophy of science, the creation of knowledge, accepted as true, and the process of learning about the world were assumed to coincide. Gooding’s (1990) historical analysis of the scientific discovery of electromagnetism is an illuminating example of intertwined learning and knowledge creation. Gooding maps Faraday’s research process in great detail, recording each attempt to make an experiment work and tracing the interpretation of failures and successes. The result is a story about how Faraday used the experimental apparatus to alter the material world in ways that revealed underlying regularities and principles. This simultaneous “learning whilst coming to know” in the research process is assumed to be the way science works. In contrast, computer simulation modeling does not physically interact with the phenomenon, rather it takes abstractions of the material world, in the form of mathematical formulations of physical laws, and sees how they behave in combination, through computer code, often sustained by other information (such as empirical rules). This changes the relationship between learning and creating new scientific knowledge: modelers learn about the environment by rendering physical processes and events virtual in their models, they do not manipulate physical objects to map their range of behaviors.
To address this issue, we introduce conceptual tools to allow for the separation of knowledge production from learning, based on understanding modeling as performative as proposed by Knuuttila (2006). In the following, we begin with an overview of the science studies discussion of computer simulation modeling; thereafter, we introduce a conceptual approach and outline our empirical methods. Then, we compare two cases of environmental computer simulation modeling of flood inundation in England. Finally, we reflect on what the newly introduced conceptual approach could contribute to the study of environmental computer simulation modeling.
Modeling and Learning
Computer simulation modeling has drawn the attention of science studies and so has learning, but not concurrently. 1 Studies of simulation modeling have characterized it in different ways, for example, as a new form of laboratory work (Merz 2006); as inherently conducive to policy needs (Bickerstaff and Simmons 2004); and as coproduced with social order within research communities (Sundberg 2009). Learning has been discussed, for example, in the history of science (Gooding 1990); as a feature in the development of expertise (Almklov and Hepsø 2011) and in medical training (Prentice 2007). We bring together the analysis of computer simulation modeling and learning because of the increasing reliance on models in environmental policy and decision making, often claimed to be a basis for “science-based” or “evidence-based” knowledge and policy making. It is important to clarify the unique characteristics of this type of learning, and its similarities with and differences from traditional scientific practices that aim to generate new general knowledge rather than to inform decision making.
As mentioned above, the philosophical perspective underpinning our analysis casts computer simulation modeling as performative rather than representational. We adopt this notion from Knuuttila (2006) who, in a discussion of modeling in linguistics, notes that what mattered to the modelers was the ability to mimic the phenomena under investigation. The value of a model depended upon its ability to bring about an outcome considered an adequate rendering of observed phenomena; whether the model exhaustively represented the causal mechanisms of language was not the key issue. This allows for a model to be “wrong” in the sense of being simplified, not providing realistic representations of detailed processes, but still producing results sufficiently correct to be useful. Knuuttila’s notion of modeling as performative focuses on the epistemic role of models in scientific research, which aims in principle to generate new theoretical knowledge. In contrast, the present article addresses how researchers learn about the environment through the process of modeling, regardless of whether the outcome produces new knowledge.
In order to study modeling in a way that addresses learning, as distinct from knowledge production, the notion of improvisation is useful. In the theater, improvisation means that the actors work without a script designed to convey a story and arrive at an ending. In this sense, improvisation resonates with the scientific research process as portrayed in ethnographic studies of research practices, which show that whether and how activities undertaken by researchers will succeed in generating new scientific knowledge is not known, or guaranteed, in advance. There are additional aspects of the notion that makes it particularly well suited for addressing learning, rather than the knowledge obtained. Of particular relevance is the emphasis on the practical process and the unplanned character of sequences of action. These aspects of improvisation have also been of interest to organization theorists, who have adopted the notion as a metaphor in order to redefine “the concept of structure to permit, creativity, innovation and continuous learning” (Kamoche and Pina e Cunha 2001, 756). 2 In organization studies, improvisation is commonly viewed as the opposite of science, because it “valorises subjectivity, emotion, the aesthetic” as well as “openness and uncertainty,” something that is believed to go “against the fundamental goals of prediction and control so highly valued by the traditional sciences” (Montuori 2003, 239). This demonstrates both the influence of positivist philosophy of science and a more widespread tendency to focus on the products of research, which contrast with science studies analyses that show creativity, openness, and uncertainty to be intrinsic to science-in-the-making. As portrayed in science studies, the research process shares with improvisation in the performing arts the quality of being able to generate “the unpredictable, the unusual, the unforeseen, within the pre-existing structures … navigating the edge between innovation and tradition” (Montuori 2003, 239).
The notion of improvisation developed in organization studies also highlights the role of obstacles that force actors to invent new ways of working. This accords with Gooding’s (1990) historical analysis of physical experiments as learning processes in which obstacles prompt invention. In effect, Gooding shows how Faraday learned through improvisation, by exploring the constraints that physical matter imposed on the experimental apparatus. In traditional scientific research, scientists learn by encountering nature as material constraints on experimental outcomes. In contrast, models are already known not to be true representations of the world and, as they exist in a realm not controlled by the same physical laws as the phenomenon modeled, constraints have to be of a different type. Edwards (2010) discusses obstacles encountered by climate scientists in terms of friction, which, he argues, has played a critical role in the evolution of climate modeling. He identifies three different types of friction—data, computation, and metadata—all of which prompt climate modelers to invent new ways of working. However, as compared with this study, Edwards focused on the development of better models, not on learning about particular geographically located environmental phenomena. In the following, we will develop the notion of obstacles further in relation to a learning process considered as improvisation.
Modes of Modeling
To understand the factors shaping the modeling processes in which learning takes place we compare two projects, undertaken in 2008-2009. Both aimed to inform flood risk management, thus they involved more than epistemic ambitions: there were risks to local communities, environmental management issues, and policy making on different levels for both projects. For the purpose of this article, we distinguish between the two wider contexts by using the notions of normal and postnormal flood management. Adopting this conceptual coupling from Funtowicz and Ravetz (1993), we deploy it as a methodological device that distinguishes flood modeling carried out within the established institutional order from flood modeling undertaken in the context of public controversy. 3 What we consider a normal context is defined by the established English regime of formal institutions and regulatory frameworks for managing flood risk in local communities, which involves expertise in computer simulation modeling (Landström et al. 2011). The vast majority of flood risk problems are solved within this framework. However, occasionally management objectives cannot be achieved with established means, and sometimes this triggers public controversy in which established expertise and institutions become part of the problem (Whatmore and Landström 2011). We understand attempts at addressing such situations as emerging in a postnormal context, in which knowledge is uncertain and expertise challenged locally. One of the modeling projects studied took place in the context of normal flood risk management, while the other, trialing a new experimental methodology for public engagement, was introduced in a postnormal situation. Below we draw on ethnography, 4 interviews, 5 and self-reporting 6 to analyze how the two modeling teams learned about local flooding.
Two Teams
One of the modeling teams operated in the research wing of a consultancy firm that provides expert knowledge about a range of water-related issues to a variety of clients within and outside of the United Kingdom, in a context of normal flood risk management. There are numerous businesses of this kind in the United Kingdom and internationally, “professional service providers” (PSFs) in the language of management studies, trading “mainly on the knowledge of its human capital … to develop and deliver intangible solutions to client problems” (Morris and Empson 1998, 610). The modeling undertaken by engineering consultancy firms ranges from run-of-the mill flood risk assessments using off-the-shelf software packages to simulate flood events in specific localities, to research-orientated work aiming to develop new methods in order to attract more business and new clients. The customers of this knowledge include public bodies such as the Environment Agency (EA) for England and Wales, private developers, the privatized water industry, and insurance companies. We understand this modeling practice as virtual engineering, undertaken by specialized teams, or individuals, within the companies (Landström et al. 2011). We followed one team that had been in existence for about eighteen months, working in an area of flood modeling that was new to the company. The leader wanted the team to grow and to develop unique skills and expertise; winning projects from clients was a driver and new projects translated into more team members. Over the year we followed the team, new members joined and they worked on several projects simultaneously. All projects were different, but they required modeling expertise to be successfully carried through to completion. Working with projects for different clients, the team intended to improvise, in the sense of “knowingly engage in an extemporaneous process and try to achieve an objective in a new way—new, at least, to them” (Vera and Crossan 2004, 733), a common feature of knowledge-producing work. For the purpose of this article, we focus on one specific project: the creation of a National Coastal Flood Map, an internal project funded by the company to create an object that could be used to develop client products. This is a common way to develop resources in PSFs, an aspect of investment in businesses selling knowledge and expertise (Morris and Empson 1998). Having a map of projected flood inundation along the coast of the United Kingdom with different return periods and flood defenses would enable the company to increase the number and variety of expert knowledge products they could offer to clients.
The other team we followed were university scientists taking part in an interdisciplinary experiment with public engagement with science in a postnormal context of public controversy over flood risk management. 7 Here, flood modeling was taking place in environmental competency groups (Whatmore 2009), staged to enable modelers, social scientists, and local people to work together to create an extended research collective (Callon, Lascoumes, and Barthe 2009). 8 The computer simulation modeling in this group aimed to generate knowledge relevant to the management of flood risk in the locality, in a situation where the standard approach of virtual engineering had failed. The university modeling team was brought together specifically for the larger research project that comprised two competency groups in which local residents and the project team collaborated for about one year to create new knowledge about flooding the rural localities. 9 The team was to exist only for the three years of project work and the duration of each local extended research collective was determined from the start. The modeling objective was to investigate the impact of possible alternative measures for reducing flood risk, a question of both scientific and local environmental management interest. In this article, we discuss the second competency group, in Uckfield, East Sussex, a locality with a long-term, ongoing controversy over flood risk management.
Starting Points
Organizational theorists have noted that “all improvisers draw on ready-mades (for example, short motifs or clichés) as they create their novel performance” (Vera and Crossan 2004, 741-42). This agrees with our findings, both flood modeling teams started with extant computer programs, stabilized codings of scientifically accepted mathematical representations of the general physical dynamics involved in local flooding. The virtual engineering team began their research with scientific knowledge encoded in a proprietary software package called TUFLOW. 10 This is one of the several such packages in use for production of flood expertise; the major difference between them is how the physical processes are represented, in one, two, or three spatial dimensions. They all start with physical rules based on the conservation of mass and momentum and are, as such, deterministic, even if they are eventually run probabilistically, that is, with very different boundary conditions to capture the uncertainty in input data. The dimensionality needed can be determined by the scale of application: 3-D over smaller scale and 1-D over the largest scale. In some cases, solutions of different dimensionality may be coupled: such as 1-D for flow in the river channel and 2-D to represent how water spreads out across a floodplain. TUFLOW is one of the five 2-D software packages commonly used in the United Kingdom. The choice of software is constrained by the requirements of the EA, which has established a process of benchmarking the different proprietary packages used by their consultants. The first such comparison was initiated by the National Rivers Authority (NRA) in 1993 and completed in 1997, when the NRA had been replaced by the EA. Later, the benchmarking of 1-D software has been used by the EA (and other operating authorities) to “identify and apply the best models [software packages] and practice” (EA 2005). The EA recently initiated a similar benchmarking of 2-D software packages (EA 2009). Hence, TUFLOW was important in the virtual engineering modeling project, not because it was the only available computational tool but because it was recognized in established practices and regulations. Using TUFLOW as the ready-made starting point for a process intended to extend established practice reaffirmed the social order and prepared for the possibility of forming future relationships with customers of the knowledge products generated from the new flood map. The creation of the map was a unique process, but the way in which it was generated followed industry standards and a methodology approved by the regulator, EA.
Working in a postnormal flood management context, in which the established institutional relationships were no longer operational and established modeling approaches were known not to be adequate for answering the questions of interest, the university team were free to choose whatever modeling tools they preferred. Aiming to generate new knowledge about the impact of land management on flood risk, they chose a physically based, distributed, hydrological model, CRUM2D v 3.1, with which they were familiar (Lane, Reaney, and Heathwaite 2009). This choice was directed by their objective of investigating the transformation of rain to overland flow; they needed a hydrological model. CRUM2D could also represent the impacts of land cover and soil infiltration, which potentially made it useful for modeling the impact of local land management changes on flood risk. The choice of this model as the ready-made template for the investigation of flooding in a particular locality confirmed the modelers’ position in the scientific debate over whether or not rural land management could be used to reduce downstream flood risk: they thought it might. To estimate changes in river flow, following the representation of land management interventions in CRUM2D, they intended to use one of the standardized 1-D hydraulic models approved by the EA.
From the start, both teams set out to produce new knowledge about flooding in specific localities, using the tools they were familiar with in new ways. Within the framing concept of improvisation, the extant computer programs amount to elements providing minimal structure in an open-ended process; as well as being props necessary to carry out the activity these also connected the teams to other actors. The choice of the virtual engineering team to use TUFLOW was determined by their ambition to provide new knowledge relevant to potential clients. The choice of the university modelers was based on their ambition to address new questions about catchment processes. Having used TUFLOW before, the virtual engineering team could envision a process of stepwise progression through a sequence of familiar tasks. In contrast, the university team was intending from the outset to use the prop to improvise. Working with local residents in the competency group was a new methodology, only tried once before by the same modeling team. They did not know if it would work in a new locality, if the objectives could be achieved, or how the process would unfold.
Both projects complicate the notion of model building as the coding of mathematical equations into computer programs, as discussed in philosophy and sociology of science (e.g., Winsberg 1999; Merz 2006). Coding was not where these two projects began. Instead, in these cases, model building meant rendering particular localities in digital form. Unlike the type of modeling usually discussed in the literature, these projects employed computer simulations to learn about the environment, rather than to learn about computer modeling.
Productive Obstacles
Obstacles are key features in improvisation, because when it becomes impossible to proceed as planned actors are forced to “explore the constraints created by the new conditions, but also the possibilities they offer” (Montuori 2003, 240). In technology studies, Bijker (1987) developed the notion of “obduracy” to capture the constraints resulting from the stabilization of sociotechnical systems. Hommels (2005, 334) clarifies this obduracy as a consequence of system integration where “the elements of a network are closely interrelated, the changing of one element requires the adaptation of other elements.” It follows that any process intending to do things in a new way will encounter obstacles, but it will not be possible to predict which exact aspects of existing systems may resist change. This notion of obstacles, as emerging when actors aim to do things differently because of obduracy in existing sociotechnical networks, works with the notion of improvisation, in which obstacles are connected with opportunity. Capturing such opportunities with regard to modeling, Knuuttila (2006, 52) identifies a dynamic in which the “epistemic functioning of both computer and theoretical models is due to the constraints and affordances they embody.” Further, she notes that “the constraints and the affordances of models are intertwined, and it is difficult to tell them apart” (Knuuttila 2006, 52). The notion of affordance has also been employed in technology studies to capture how a human becomes “an active organism that explores its environment, actively seeking (picking up) information” (Michael 2000, 61), always exists in relationships with nonhuman elements, comprising assemblages which make new forms of action possible. In the two modeling teams, we could see how the interplay of obstacles and affordances pushed improvisation in different directions depending on the assemblages found in specific contexts. This divergence began when the modelers started to render digital, in coded algorithms, the natural systems under investigation.
Modeling for Normal Flood Risk Management
The virtual engineering team encountered three major obstacles, all related to data. First, there was data scarcity. To produce a map of coastal inundation patterns under different circumstances, data are needed as boundary conditions to define the height of the extreme sea levels that drive inundation. Effective application of the TUFLOW modeling software therefore required data that reflected historical conditions, but the modeling team discovered that the data for sea levels in the last fifty years were scarce and patchy. A second obstacle concerned the model’s capacity to handle the amount of data defining the area that was to be mapped. It was not possible to represent the entire coast of England and Wales in one model; there are limits to how much data TUFLOW can handle at any one time. The third obstacle emerged as the modeling progressed, when visualizations of model runs looked strange. Approached as instances of obduracy, analysis of each obstacle illuminates how the team was forced to improvise in order to learn.
The data scarcity problem emerged in the encounter with the UK Tide Gauge Network, comprising about forty-five gauges along the coast. 11 Created after a severe flood event along the east coast of England in 1953, the network is owned and operated by the EA. The data collected are processed and banked to provide, with time, longtime series of sea levels. There are values recorded every fifteen minutes from January 1993 to present, and hourly values recorded for the years before 1993. There are also data for monthly mean surge and extreme values. 12 For the virtual engineering team, forty-five gauges providing continuous readings since 1993 were not nearly enough data to model the entire coastline of England and Wales with sufficient detail. In addition, the forty-five gauges have not all been continually operating since 1953, and earlier data are patchier. The Tide Gauge Network was created in response to one particular catastrophic flood event; computer simulation modeling is a new and different use of these data. Arguably, data collection is shaped by the context of use. Edwards (2010) discusses how climate data have changed since the 1950s, both with regard to the kind of data gathered and the amount. In addition to the problems with data scarcity and lack of long-term continuity of data sets, the spatial coverage of measurements was not sufficient for the modeling; the instruments had not been placed according to the data needs of 2-D software. This obstacle, expressing the obduracy of an existing sociotechnical network, forced the team to identify an affordance in order to enable the modeling to progress. Drawing on the knowledge and skills available, they were able to create a new computer modeling technique that could generate historical data from existing measurements. The team was able to construct “virtual gauges” to provide data points for the model, enabling them to do hindcasts of extreme sea levels in the past. 13 The complete set of sea level data that could be produced in this way provided the boundary values necessary to enable the TUFLOW model to calculate inundation patterns and they could progress.
The second difficulty, regarding the spatial extent of the area to be mapped, pushed the team to exploit another affordance. They figured out how to divide the representation of the coastline into sections that could be modeled one at the time, sequentially, along the coast. The separately modeled sections were reconnected using the output values from one model as input values for the next one in the sequence. This work required detailed checking of the digital terrain data to make sure that the elevations were right, particularly at the boundaries between model sections, but it did work.
The third obstacle in the virtual engineering team’s research emerged when they began to run the model to generate simulation output. Running the model sections along the coastline gave some strange results in terms of visible inundation patterns. Small bays, natural and artificial, filled up in a way that the team considered unrealistic. Closer analysis of the strange effect showed that the topographic data (light detection and ranging [LiDAR]) did not capture the parts of the landscape that were normally under water. LiDAR data are generated by airborne laser linked to Global Positioning System, a technology providing very high resolution (25 cm to 2 m) and precision (±0.15 to ±0.25 m). Data sets are collected by the Geomatics Group, a specialist business unit set up within the EA. 14 These data did not provide enough information about the coastal seabed for TUFLOW to calculate the inundation of bays correctly. The team was able to identify and exploit an affordance to combine LiDAR with bathymetry data in the affected sections. Bathymetry data are generated by sonar, which measures distances under water by emitting a short pulse of high-frequency sound and measuring the time until an echo is heard. There are different sonar techniques using, for example, ships, orbiting visible-light cameras or satellite radar. 15 The principal UK supplier of these data for governmental, commercial, and individual use is the British Geological Survey. 16 By combining the two types of data, held by different organizations, in locations with bays, the team could provide TUFLOW with input that made the inundation of the coastline appear more realistic in the simulations.
Paying close attention to the improvisation process of the virtual engineering team has illuminated the notion of affordance. We have used it to capture how the knowledge and skills of the modelers in relation to their digital tools enabled them to find ways to move forward, toward their specific aims in the particular circumstances. This clarifies how affordances link to obstacles—via an actor with appropriate expertise to improvise in the situation. We also note that the limitations of the LiDAR data only became visible when running the model, as the simulation was discovered to look strange when visualized. This complicates the sharp distinction between building and running a model, as assumed by philosophers and historians of science (Morrison and Morgan 1999). In hydraulic river modeling, it is possible to do repeated runs and modify the model set up, to get a rendering that can be considered to better capture the physical system, as understood by scientists. This distinguishes this mode of modeling from climate modeling with General Circulation Models, which are run for months by dedicated teams in facilities with extensive computer power and analyzed for years thereafter (Weart 2011). The virtual engineering team engaged in an iterative process of rendering flood inundation of the English and Welsh coastline virtual, in a way that resembled that which can be observed by the eye and measured by different scientific instruments as closely as possible. The process was carefully documented and the team developed document formats to standardize the accounting for the origin, characteristics, and quality of data. In addition, they wrote a manual for how to use the map and how to work with TUFLOW to interrogate the map in different ways. This research outcome was not intended to undergo the type of peer scrutiny that scientific knowledge claims are subjected to in journal publications. It was intended as a tool, used in the creation of science-based expert knowledge by the company, to sell on a market. Our empirical study did not generate information about the success of the team’s output; however, we do know that this type of knowledge is widely used in UK flood risk management, by the EA, local authorities, and private developers (Landström, Whatmore, and Lane 2011).
Modeling in a Postnormal Context
Data did not present a problem for the modeling in the extended research collective. The university modelers approached the issue differently from the outset, since data scarcity was one of the known issues for the modeling of local floods. One of the objectives of working with local people was to gather information about the catchment, the river and flood events that would not have been recorded by the scientific instruments in use. The modelers did initial work with existing hydrological and hydraulic data for the locality from established scientific sources and the local competency group generated more information. Already in the first meeting the scientists learned things about the river not recorded in any database, for example, that it was
… a canalised river, well, it is at the moment. Its original river bed is much to the north of where it is, and that is the lower end, topographically, it’s actually lower there than the river. So once the river is out of its water banks, it then flows to the left and that’s why it floods through the Somerfield supermarket, because that’s a lower element … (Uckfield Transcript 1: 32).
The research collective was also continuously extended, recruiting people who were known by local residents to have relevant knowledge about the catchment, for example, regarding soil porosity and depth.
The major obstacle encountered by the competency group related to the model structure. When the modelers had figured out how to represent different land management measures suggested by the group—riparian wooded corridors, buffer zones, ponds, and bunds—and had undertaken the first simulations with the model, they discovered that the interventions made no difference to the flood inundation pattern. This problem was traced to the structure of the CRUM2D model; it could not simulate overbank flow in the way required. When the water has reached the channel in this model, it is calculated with a Muskingum–Cunge routing equation that moves it through, approximating attenuation by energy losses derived from the known dimensions of the channel and the hydrograph input. This is a computational “black box” in which the exact morphology of the channel floodplain is represented through empirical parameters rather than being made explicit as morphology. Thus, when interventions were introduced that changed the morphology there was no obvious way of changing the empirical parameters to represent the change: the black box could not represent the phenomenon modeled. It would have been very hard to change the mathematical structure in ways that allowed modifications of the topography to represent the alterations the group wanted to explore. CRUM2D turned out to be a poor prop for the intended improvisation. It was not possible to use it to address the questions that the group was interested in and the performance failed. This resembles the type of obstacle Edwards (2010, 84) discusses in relation to climate modeling as “computational friction,” which “reduces the amount of information and knowledge that can be extracted from a given input.”
The impossibility of rendering the processes of interest to the competency group in CRUM2D led the modelers to abandon the computer program. Instead, while working with the local competency group, they developed a flow routing algorithm that could transform rainfall to overland flow and into river flow. This code provided a simple modeling tool with a graphical user interface (GUI) that could be used to explore the flood reduction options of interest to the group. The university team were able to develop this algorithm into a model that could simulate the processes that the extended research collective in Uckfield wanted to explore. They first brought it to the fourth competency group meeting to show the local collaborators. Using the model to simulate a flood that occurred in 2000, it was possible to see how the model worked and how it compared to experiences and measurements. In the following meetings, model runs facilitated discussion in the group about the representation of the system, for example, channel width and depth, as well as about the ways in which modeling requires generalizations and assumptions of details and processes. The GUI enabled all participants to see what the model did, making it transparent to people without a background in natural science or engineering, for example, the participating social scientists. Using different amounts of rainfall as input values the group could see where the model projected that water would go over the bank, onto the floodplain, and then back into the channel further downstream. The visual interface enabled the local participants to compare model simulations with actual experiences, to approve of, or to criticize the model, to propose interventions in the landscape, or the built environment.
I thought we said if we made it shallower, it will make it better.
So what shall we do—20%? How about that? And do you want to put in a buffer strip?
This is a hedge, is it?
Yes this is in your flood plain, so the … we are forcing the water over bank, and the idea is even though it has gone over bank, we still want to slow it down going over bank.
And what sort of buffer strip, is there any width or what? Or depth? Or height?
We can get a sense of how big the buffer strip is by looking here. These are 20 m cells, so not terribly wide at this point. It is a bit wider here, near the confluence, probably getting on for a hundred meters or more wide, there, but back it is maybe 60–80 meters wide, so it is not a terribly wide flood plain. But that is because it is quite steep sided, so it is flat, but it rises quite steeply up on the side.
How deep would it be, I mean how much buffer strip would it be? Would it be like a copse of trees, or just a single strand of hedge?
Well at the moment, the way the model is run, we simply buffer the whole reach. The extent of the flood plain, just in the flood plain. When we buffer the whole lot, later on if you want, we can get clever with it and say well ok, we will have some buffer there, and some buffer there.
OK.
106-7)
This extract from a transcribed audio recording shows how the visualization enabled interrogation of the data and the ways in which the physical system was rendered in the model in the extended research collective. Visualization, but also running the model repeatedly, was important for integrating the modeling with experience of floods in the locality. The extract shows how this computer program functioned as a prop with which it was possible to improvise model performances that responded to the queries by local participants.
In this case, we can see how the interest in exploring locally specific measures to reduce flooding prompted improvisation, pushing the modelers into less familiar territory. Obviously, modelers using CRUM2D previously had not tried to address this type of issue or to answer similar research questions. Here, the obduracy of CRUM2D showed that it was embedded in a different network, determining “its resistance to efforts aimed at changing it” (Hommels 2005, 334). Affordances emerge in the improvisations made possible by the relationships between objects and their users, depending on the situation and the experience of users. Drawing on their unique experiences of having already participated in a successful competency group, the modelers were able to let go of an element of structure that was scientifically established and to organize the modeling around an experimental algorithm instead. Fortunately, the routing algorithm worked and the modelers were able to progress with writing a model that could answer the extended research collective’s questions about the impacts of land management on the water levels in extreme weather conditions (Odoni and Lane 2010). The output of the competency group followed two distinct paths; one was the traditional scientific publication, enabling critical scrutiny of algorithms and codes in a community of peers (other scientific modelers). The other output was the local model indicating possible new flood mitigation options; this was presented to the agencies responsible for local flood risk management. The investigation underpinning the present article did not follow any of the outputs further along their respective paths; hence, we do not know whether the new modeling approach was accepted by the scientific community, or if the suggestions for local flood management were taken forward.
Concluding Discussion
This article has introduced conceptual tools for analyzing how researchers learn about the environment through computer simulation modeling. These tools are helpful because modelling differs from other scientific knowledge production methods. Instead of generating new general theoretical knowledge simultaneously with learning about particular phenomena, as traditional scientific research does, computer models and simulations are fictional renderings of the world based on contingent assumptions (Stengers 2000). Hence, we asked: how do researchers learn about nature from representations known to be “wrong”? A related issue concerns the constraints encountered in research, which force researchers to reconsider ideas and strategies—to learn. Traditionally, scientific learning has relied on material constraints mapped by using experimental apparatuses, but computer simulations are not constrained in physical encounters with matter. This prompts questions about which obstacles modelers encounter, and how they are overcome?
To address the questions, we first suggested a vocabulary based in a philosophical notion of modeling as a performative practice, in which the doing has value in itself (Knuuttila 2006). To examine, in depth, the learning processes in two flood inundation modeling projects, we adopted the notion of improvisation. This metaphor facilitated the examination of environmental computer simulation modeling shaped in specific local contexts. Tracing the research by the two modeling teams through a dynamics of obstacles and affordances, we were able to highlight the learning that occurred in the processes of rendering localities at risk of flooding virtual, in computer code, when aiming to answer questions about local inundation patterns under different circumstances. The comparative study was called for because previous studies show that there is significant variability in modeling across disciplines and subdisciplines. Hence, aiming for in-depth understanding of factors with impact on learning, we examined two different modes of flood modeling. While flooding was the concern in both cases, the social contexts differed significantly. To capture the differences, we borrowed the conceptual pairing of normal and postnormal (Funtowicz and Ravetz 1993). We adapted this well-known terminology into a methodological device distinguishing between a modeling context in which the established institutional order, procedures, and distribution of expertise were at hand, and one where controversy had erupted and modeling was part of a process redistributing agency.
Refiguring the distinction between normal and postnormal to capture the relationships of modeling to environmental management contexts made visible some important similarities in the two cases. First, we found that both modeling teams addressed research problems emerging in society that were specific to the localities. In both cases, model building comprised the rendering digital of a unique locality. This modeling differed from that undertaken within academic disciplines, the flood modeling we studied aimed to explicate specific, not general, mechanisms. A second similarity was that both teams started with existing computer programs in order to render particular environments virtual. We found that the ready-made elements affirmed the social relationships of both teams, as well as providing tried and tested modeling tools. A third similarity was the iterative process of model build and run. Making the models perform and evaluating their performance in a visual interface were critical for both teams. Model simulations were thus continuously related to other ways of knowing flooding—measurements and experience.
We could also discern differences between the learning processes in the two modeling teams. In the context of normal flood management, the use of standardized software was the key to generating outcomes that would become products on a market that provided knowledge to institutions and businesses. The obstacles that emerged in the process were expressed as “data friction” (Edwards 2010) and could be resolved by identifying affordances for generating, organizing, and combining standardized information in new ways, through improvising. The modelers were obliged to use particular software packages in order to maintain relationships with customers and regulatory authorities. This shaped their improvisations. When encountering obstacles, they exploited affordances in relation to data that they were expert at locating and working with. They were forced to identify affordances that would allow them to proceed with TUFLOW, and progress depended upon their ability to improvise around the obduracy of the data. In terms of learning through modeling, the improvisation that took place in the context of normal flood management prompted the modelers to learn new things about using their modeling tools in new circumstances, and they carefully documented their solutions to the problems encountered for use by other modelers in the firm. Their learning contributed to the value of the modeling software tools, the skill of the modeling team and the range of services the business could offer customers.
In the postnormal context of the competency group, obduracy manifested as computational friction and affordances that allowed the university modelers to reconsider the mathematical structure used in the model. The improvisation of the extended research collective entailed learning to model in a new way; instead of building trust in their model through multiple runs and other established scientific approaches, the scientists had to lay open the model assumptions for critical discussion in the competency group. When obstacles arose, the skills of the scientists were deployed to search for answers to the collectives’ questions, regardless of the lack of previous templates. The lack of formal social constraints imposed through relations with other actors, outside of the extended research collective, made it possible to let the nonscientific questions and understanding of the catchment lead the modeling process. This learning process generated a new constituency for modeling, moving the practice out from expert organizations, making it into a resource for a new set of local actors.
The concepts introduced in this article have facilitated an analysis of the learning process as independent of the knowledge products, which have not been discussed here. The interest in separating learning from knowledge outcomes originated in the intrigue that follows from the ways in which modeling is understood, by philosophers and social studies of science, to produce digital renderings rather than representational truth. To develop this understanding of modeling in the context of environmental modeling, we captured the context-specific dynamics of obstacles and affordances in which the two teams learned about flood inundation by improvising model performances, regardless of whether the modeling outcomes were later accepted as valid knowledge claims. The framing metaphor of improvisation enabled detailed analysis of the iterative processes involved in two different modes of flood modeling.
Footnotes
Declaration of Conflicting Interests
The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The research for this article was funded by grant RES-227-250-018 from the Rural Economy and Land Use program of three UK research councils (BBSRC, ESRC and NERC) and Defra.