ML in the Oil Industry: Can AI Kill a Well?

Maria Andryushina, Makar Ippolitov24 May 2025

Problem

Before a producing oil well can be repaired, it must be "killed": the well is filled with fluids that offset reservoir pressure and keep gas from escaping. In fractured-porous reservoirs, physics-based volume calculations work poorly, and the team has only about 1800 noisy, hand-entered operations. Too little fluid means gas breaks through, the kill is repeated and an emergency risk appears. Too much fluid is wasted, so the model must rarely under-predict without simply inflating every forecast.

Example from the article: a kill counts as successful when the predicted volume is at least 80% of the actual one. For problematic wells, the plain volume model reached 71% successful kills, against 75% in the historical baseline.

Solution in brief

A CatBoost model predicts fluid volumes. Failed attempts in a well's history are merged into one successful operation to produce valid targets. A custom loss function adds a penalty only for under-prediction (the K-method). On problematic wells this raised the share of successful kills from 75% to 86%, and on primary kills from 66% to 91%.

Questions this case answers
  • How to predict a required fluid volume when under-prediction causes a failed operation and over-prediction only wastes material
  • How to build a regression target from failed attempts when their recorded volumes are not valid labels
  • How to predict the required volume of a fluid when under-prediction causes a failed operation and over-prediction only wastes material
  • How to build a regression target when part of the historical operations failed and their volumes are not valid labels

Tools and methods

  • CatBoost — final regressor; chosen over linear regression, Random Forest and XGBoost
  • K-method — squared-error loss plus a linear penalty for under-prediction
  • Economic loss function — cost-sensitive loss from predictive maintenance literature; auxiliary only
  • Target reconstruction — merging a series of failed kills into one operation
  • SHAP — feature importance check
Continue reading — it's free

Google or a one-time email code. No password.