Research in the wild can reveal human behaviors, contexts, and needs around products that are difficult to observe in the lab. Telemetry data from the use of physical products can help facilitate in the wild research, in particular by suggesting hypotheses that can be explored through machine learning models. This paper explores ways for designers without strong data skills to engage with multimodal data to develop a contextual understanding of product use. This study is framed around a lightweight version of a data enhanced design research process where multimodal telemetry data was captured by a GoPro camera attached to a bicycle. This was combined with the video data and conversation with the rider to carry out an exploratory sensemaking process and generate design research questions that could potentially be addressed through data capture, annotation, and machine learning. We identify a range of ways that designers could make use of the data for ideation and developing context through annotating and exploring the data. Participants used data and annotation practices to connect the micro and macro, spot interesting moments, and frame questions around an unfamiliar problem. The work follows the designers’ questions, methods, and explorations, both immediate concerns and speculations about working at larger scales with machine learning models. This points to the possibility of tools that help designers to engage with machine learning, not just for optimization and refinement, but for creative ideation in the early stages of design processes.

@en

Recent observational studies (Jenkins et al. 2010 [1] and Dutrieux et al. 2013 [2]) have helped to constrain estimates of the melt behaviour underneath Pine Island Glacier (PIG) ice shelf, in western Antarctica. Generally however, observations are limited, due to the relatively inaccessible and inhospitable environment. A solid ice cover, up to many kilometres thick, bars access to the water column, so that observational data can only be obtained by inference from above, drilling holes through, or launching autonomous vehicles beneath the ice. This is further exacerbated by the fact that results of these recent studies have implied a significant proportion of the melting (~80%) occurs in networks of sub-kilometre scale basal channels close to the grounding line, some of the most inaccessible parts of sub-ice shelf ocean cavities. Accurately representing these small-scale processes in conventional ocean models is a huge challenge even in focused regional studies, and will not be possible in global coupled climate simulations in the near future. I will present the development of a new model of PIG that is capable of resolving the range of scales necessary to evaluate the melt distribution and forming processes that dominate. This is built on the Fluidity model (Piggott et al. 2008 [8]) that simulates non-hydrostatic dynamics on meshes that, like the FESOM model of Timmermann et al. 2012 [4], can be unstructured. In this case, the grid can be unstructured in all three dimensions and use an anisotropic adaptive-in-time resolution to optimise the mesh and calculation in response to evolving solution dynamics. The parameterisation of melting in this model has been validated in idealised cavity domains (Kimura et al. 2013 [5]) and a validation is underway for the dynamic treatment of the ice-ocean interface (Candy et al. 2016 [7]). Additionally, the model is not limited to a specific vertical coordinate system and can capture purely vertical features, which enables it to accurately represent ice fronts, and small shallow features. I will discuss the development of this model of PIG; including the cavity domain, conforming to appropriately filtered boundaries generated from data collected during the British Antarctic Survey Autosub 2009 expedition, and the simulation of non-hydrostatic dynamics to date (see Figure 1). This will include validation to observations and MITgcm model results in a Circumpolar Deep Water forcing scenario from measurements in 2012, as recently presented in Science (Dutrieux et al. 2014 [3]). The unstructured nature of the developed model (Candy et al. 2016 [6]) captures the high spatial variation seen in melt rates in the small-scale channels, that its difficult to resolve in other fixed-mesh state-of-the-art models @en