Q: What are the filters the Petro.ai DSU Design Service initially applies to the wells used to train the model?
A: Kyle LaMotta explains the process, “Start big and then refine it over time. Once we do that, we see that the accuracy on the analogs increases because the model is using data that’s more similar to the area that we’re focused on. We try to feed the model with data that’s as representative of the geology, the completions type, and other critical features.
“If we can keep one variable constant then it doesn’t need to be a feature in the model. We keep that variable constant by applying filters on the data set.
“One of the places we start is with the time frame of the wells. When companies come with 2000 or 3000 wells, we need to apply filters in the model. There’s been a significant change in how wells are completed since the 2014-2016 timeframe. In most basins, there’s been a large shift in the way that these wells are completed. In most cases,anything completed before 2015 is no longer relevant.
“There are certain aspects of well design that are not reported publicly. Stage length, number of perforations, number of clusters per stage, etc. aren’t reported publicly so the model isn’t able to learn the relationship of stage length to overall production. We can filter out the wells that are older because we’re never going to complete a well like we did back in 2012. We don’t want the model to learn those relationships since the model is going to be underpredicting. We don’t want the model to learn from that. We’re only going to include more recent completions.
“Another common filter we apply is lateral length. We use lateral length as a feature in the model because a 3000 foot well is not going to perform as well per foot as a 2000 foot well. Some of the really short laterals might be good performers on a per foot basis, but it doesn’t scale linearly as you increase the well length. We typically filter out the really short wells and the really long wells. Most people are going to drill wells that are around 2 miles long, so that’s the data the model needs to learn from.
“Another filter that’s tied to the completions is completions intensity. In the Permian, anything smaller than 1000 lbs per foot is coming from older-vintage completions; we’ll filter those wells out.
“Our goal is to remove data that’s not representative of modern wells that are being drilled.
“The last item that I’d mention is that starting with 2000 wells in total, we like to shrink it down to an area that’s closer to the DSU that we’re designing. If we extend too broadly, there may be significant geological changes that aren’t obvious from the data we have available. Our customer’s geologists, however, will be able to look at a map and say anything west of this point is a totally different depositional environment. If the rock is totally different, we don’t want to include those wells in this model. Our goal is to only train the model on geology relevant to the target AOI.”