BAS & Equadratures Workshop | 1 Dec 2022

--- tags: workshop --- # BAS & Equadratures Workshop | 1 Dec 2022 Zoom: https://turing-uk.zoom.us/j/94968554438?pwd=TzJPczIvRnJHVjlFNUZ6TVIrTnlQUT09&from=addon [TOC] ### Useful links * [dobson-eq Github](https://github.com/alan-turing-institute/dobson-eq) * [28 March & 14 Oct meeting notes](https://hackmd.io/mCtUB5gYQAW9g4uKthbLdg) ### Agenda * Introductions * Progress update - (Peter: here are my [slides](https://www.dropbox.com/s/ep4ykgf7u1drt7r/Update_2.pdf?dl=0)) * Deployment Opportunities before March 2023 * Next Steps ### Attendees * Scott * James * Steve * Peter * Jen * Hari ### Meeting Notes * Presentation Notes * If we have a good interpolating model for direct sun, we can use this for zenith sun (relate zenith observables to actual ozone that could have been measured with direct sun) * Data processing/challenges * Minute resolution, missing data, noise * Springtime depletion (southern hemisphere) * Set up simplest model - binning observation by month (for 8 months of measurement) * Generally not a ton of spread besides month 11 * Introducing annual feature * Combo of GPs * Slide 5: Blue is the error (95% confidence interval) * SH: we can potentially find other info to lower error * SC: not sure because even satellites can't data because they rely on reflective sunlight * SH: maybe some balloon data? this is where AI can help to blend in datasets together * Discussion: * maybe a better fit to start the year in May? * Questions about tightening priors, bin to daily (does daily or weekly ozone make more sense?), transforms * SH comments * Setting of year boundaries starting Jan doesn't make sense re: atmospheric chemistry. Ozone season? * SC: No measurements from May - August because there is no sunlight and it can't work in the dark; stay clear of May and September * Perhaps 1 July or 1 August as starting point (June, July or August) * Minute binning feels very high; there is a potential cycle within the day - is daily averaging too coarse? * SC: direct sun measurements are manual measurement and there are 5-10 of those a day (Auto Dobson is every 5 min) * SC: maybe better to compare direct sun measurements with Auto Dobson at the same time (rather than a daily average) * Currently two equations (1. high sun 2. low sun) * **SC: what we're hoping to get from the AI is to get 10-20 zenith equations depending on the sun angle (and process using computer rather than by hand) * Next Steps** * SC: ozone layer moves around and changes throughout the day, so looking at individual observations * SC: there should be same moment between a few minutes (should be around the same amount of Ozone) * PY: instead of annual, focus on resolving on a space of an hour * **SC: Need to get down to specific observations and find the right equation based on mu value; we see a step change in the data right now because of where we change equations** * PY: Where are models for Zenith coming from? * SC: trial and error trying different models over time to get zenith/direct to match; 3rd, 4th, 5th order polynomials * Angle of sun has a big effect because of scattering * PY: are location important? should be fine to have a few km different; all measurements in this dataset taken from Halley station * Next steps: * Peter visits BAS to see Dobson - to schedule * Ideally Monday or Friday when Jonathan is there * SC: maybe one day before Christmas and then a few days in January (e.g. Mon, Tue, Wed)? * Start with GP; using equadratures * ASG end in March 2023 * SC: If we can come up with new Zenith equations to apply to auto-dobson data for past 3 winters, we could reprocess the data with new equations for more accurate measurements (this data is available online and labeled as provisional) * SH: Another potential option is to provide model online? Through a different website? * JB: BAS PR - if we are able to move towards demonstrating BAS automation efforts (not just progress on the science side) * SH: internal toy experiment: build a system - compare what's going online from Auto Dobson against model data * Potential pipeline: integrate Peter's model data to this? * Trustworthy AI * JB: Pipelines/automation - loop in James when it's becoming more important to create a robust implementation (data delivery/APIs) * SC: Tim Barnes may be the right person to speak to for automation (rather than someone coming in every week to process)