Philosophers of science study how science is practiced and how scientists develop new understanding of the systems they investigate. NCAR has recently had a number of philosophers of science as visiting scholars, who have collaborated with climate scientists and others in the organization to study how we work with numerical models and big data. In doing so, they have developed a number of interesting "big-picture" insights that will be useful for anyone who works on scientific modeling and analysis of big data to keep in mind, especially as the field begins to adopt AI and machine learning methodologies. In this talk, I will present a broad overview of the important ideas that have come out of our collaborations.

Work by Monica Ainhorn Morrison focuses on the fact that models are perspectival in nature: every model has some elements that it focuses on more closely than others, and tradeoffs made in model development affect how it represents the system it models. The representational perspective of a model is a reflection of its developers' aims and interests. Agreement between different models is a signal of robustness, but when models produce differing results, it can be difficult to determine which is more credible. By considering model perspective, we can leverage the disagreement to identify elements or processes in the system that are important and relevant. We can also examine representational priorities to determine whether or not a point
of difference is due an uncertainty that can be improved. These opportunities highlight the value of recording the choices made during the development of models, datasets, and analyses.

Work by Lisa Lloyd and company focuses on how climate scientists work with big data, and how that practice compares to other fields, such as biology. One important difference is that in climate science, data is handled in a more "hands-off" fashion than in biology, which relies heavily on ontologies that describe and categorize the objects of study and their properties. The recent history of biology has shown that the way that ontologies are used to organize big data can constrain downstream research in detrimental ways. Although this has not been a problem in climate science historically, there are risks that it may become one as we start to use tools from artificial intelligence machine learning, which depend heavily on ontology, classification, and interpretation in training datasets. By looking to the experience of other scientific disciplines, we can attempt to identify best practices to mitigate these problems before they occur.