It's all about that Bayes: data-driven insights into energy devices without the black box

Large black-box models such as neural networks are exciting and powerful tools, but they're not the only way to learn from data! In this talk, I will showcase prior and ongoing work in using Bayesian parameter estimation to extract unprecedented materials-level insights from simple, automated electrical characterization of photovoltaic devices. By running a physical model many times to span the space of properties we wish to fit, and comparing its results to the physical measurements, we obtain a posterior distribution over the properties of interest. This approach has numerous advantages. First, with the democratization of computational power, the tradeoff between the extensive researcher time and labor to directly measure these properties and the computational effort to run a large number of simulations is increasingly favorable. Second, not only are the inferred values of comparable accuracy (and sometimes superior precision!) to direct physical probes, we can also be confident that they represent the most performance-relevant information about those properties, because we've measured them in the device context, rather than in a specially prepared sample that may have unrepresentative characteristics and measurement conditions. I will demonstrate the ability not only to extract these parameters (properties of both bulk materials and interfaces), but to easily observe relationships between them. Finally, I will also discuss ongoing work in utilizing this approach to directly inform/guide device engineering.