Starting a new job is always difficult. Everything is new — new workplace, new boss, new colleagues, new tasks and, sometimes, like it was in my case, a completely new industry. I have recently been moving between two different companies. It was a compelling experience and I want to share what I have learned.
I’ve transitioned from the position of Research Scientist at Pindrop, an American start-up which provides authentication and fraud detection for voice interactions, to being a Data Scientist at LYTT. LYTT (pronounced ‘light’) is a startup that delivers real-time downhole analytics for the oil and gas industry. The company is a pioneer in the field of using distributed fiber optic sensing for capturing temperature and sounds underground to detect subterranean events using machine learning-based models.
LYTT picked me for my background in audio processing, I received my PhD with a thesis on “Audio signal processing for context-aware applications” and then I have worked in acoustic and speech processing for fraud detection.
On my first day at LYTT I was given a clear task: they wanted me to apply my experience in audio processing to Distributed Acoustic Sensing (DAS) data. DAS is a technology that uses a fiber optic cable (the same cables used for delivering broadband internet connection) as a sensing device. I was asked to work on data collected from a new type of well and build an inflow profile of the well.
In short, building an inflow profile consists of answering 3 questions:
To me, this was unquestionably an impossible task because at that time I had no experience with fiber optics, no knowledge of the oil and gas industry and whilst LYTT had a lot of data from many different wells across the world this was their first job on a totally different type of well.
In addition, LYTT had already committed to the customer that we would have had the job done in 4 weeks. I thought they were crazy!
Prad, one of LYTT’s co-founders, believed that I could do it so I trusted him and started working on it, addressing one problem at the time.
I had no experience with fiber optics and before interviewing with LYTT I didn’t even know the existence of DAS. I had to quickly learn how DAS work. Basically, the propagation of the light through the fiber is marginally affected by the strain of the fiber, DAS exploits this marginal signal effectively transforming a 5km fiber into an array of 5,000 microphones spaced 1m apart from each other. Once I understood this similarity I immediately realized that this was something I could work with — I knew how microphone arrays work and I knew how to extract meaningful features from an acoustic signal.
I had no knowledge of the oil and gas industry and I didn’t know how wells are operated. I leveraged on my colleagues expertise: at LYTT we have many talented people with different expertise, I asked our Well Engineers how a well is drilled, I asked Petroleum Engineers how oil and gas are produced, I asked Petrophysicists how they differentiate between rocks, I asked our Geophysicists how sound propagates underground and underwater and I asked Data Scientists how fluid flowing underground affects the acoustic signal. I have learned that wells produce oil and gas for the majority of the time, however, they occasionally produce sand. Sand being solid makes a distinctive noise which is very different from the one generated by fluids. That rang a bell: I knew how to detect anomalies, it is what some of the models used for fraud detection do and so, for this task, sand was my fraudster.
LYTT had a lot of data from many oil wells but this was the first job on a different type of well, I had to improvise. The only DAS data related to the study we were looking at was from a lab environment that LYTT had acquired from a testing facility where different well and reservoir conditions i.e. different fluids (oil, water, gas), the rate they flow at, the mixtures of fluids and the mixtures of fluids and sand were simulated. The only problem was that the lab facility tested different rates to those observed in a real producing well. The solution here was to understand exactly how to calibrate my data: I needed to transform both the lab and the real-world data in such a way that they were independent of the rates.
I was then able to start building the inflow profile.
Going back to the beginning of my article, I was able to answer the first questions within my first week i.e. “Where?”
Fiber optics can provide temperature (Distributed Temperature Sensing or “DTS”) data along with acoustic data so I combined DAS with DTS and built a machine learning model able to predict whether a certain zone was producing fluid or not.
The following week I focused on the second question: “How much?”
This was the most difficult question because I had no reference point. I had to use the lab data and combine this with physics-based models to understand which depth was producing more and then quantify the inflow rate.
The third week I have tackled the third question: “What?”
I thought this was an easy question, I was expecting a hydrocarbon well to produce just one type of hydrocarbon. Well (pun intended), I’ve learned that downhole sometimes when people say oil well that does not mean the well produces oil alone. There are multiple fluid types that get produced. Hydrocarbon wells can produce water too. I had to build another machine learning based model to classify what fluid was flowing into the pipe.
There was one last question to answer: “Is the well producing sand?”
For answering this question I leveraged LYTT experience in sand management and I combined it my expertise in anomaly detection to build a sand log that showed where and when the well had produced sand.
Job done! Delivered to the customer in 4 weeks!
This month I went from knowing nothing about DAS and the oil industry to providing a full inflow profile for a new type of well. This was my first month at LYTT speed.
The business case for diversity, equality and inclusion is well documented, with hundreds of studies demonstrating th...