Collaborative Machine Learning: using information from multiple mineral deposits to improve decision making

钻孔样本和遥感调查为矿产资源评估、采矿规划和项目评估提供了关于地下资源的信息。收集这些数据是昂贵的，但对于评估矿产资源和制定采矿计划是必要的。收集的额外数据减少了地下的不确定性，但收集成本很高。机器学习算法在地下建模工作流程中变得越来越常见，并被用作辅助决策的工具，但这些算法在访问更多训练数据时执行得更好。这个项目探索了将机器学习方法应用于地下建模问题的方法，使用来自不同公司持有的多个数据集的训练数据。矿业公司很少公布数据，因为数据收集费用和对竞争优势的渴望。一些算法培训环境，如联合学习，允许远程编译多个数据集，以协作的方式学习变量之间的基本关系；在此框架中，单独的数据集仍然是拥有数据库的每个公司的私有，但允许数据集/公司之间的学习。这可以与迁移学习相结合，后者侧重于将从先前问题(或数据集)中获得的知识应用于新问题(或数据集)。这些方法可以将具有密集数据的地质环境(即采出的矿藏)到具有稀疏数据的环境(即勘探)的知识结合起来，以改进地下建模和决策。这允许开发改进的机器学习算法，同时保持数据隐私和竞争优势。本文研究的地下建模问题包括岩心测井分析(无监督聚类)、岩石类型赋值(有监督聚类)、几何建模(人工神经网络)和空间不确定性建模(卷积神经网络)。

Drillhole samples and remote sensing surveys provide information on subsurface resources for mineral resource estimation, mine planning, and project evaluation. The collection of this data is expensive but necessary to evaluate mineral resources and build mine plans. Additional data collected reduces subsurface uncertainty but is expensive to collect. Machine learning algorithms are becoming common in subsurface modeling workflows and as tools to assist decision making, but these algorithms perform better with access to more training data. This project explores methodologies for applying machine learning methods to subsurface modeling problems using training data from multiple datasets held by different companies. Mining companies rarely publicize data because of the data collection expense and desire for competitive advantage. Some algorithm training environments, such as federated learning, allow for multiple datasets to be remotely compiled to learn underlying relations between variables in a collaborative way; in this framework, individual datasets remain private to each company that owns the databases but allows for learnings between datasets/companies. This can be combined with transfer learning, which focuses on using knowledge gained from a previous problem (or dataset) applied to a new problem (or dataset). These methods can incorporate knowledge from geological settings with dense data (i.e. mined out deposits) to settings with sparse data (i.e. exploration) to improve subsurface modeling and decision making. This allows for the development of improved machine learning algorithms while maintaining data privacy and competitive advantages. Subsurface modeling problems addressed in this work include core logging analysis (unsupervised clustering), rock type assignment (supervised clustering), geometallurgical modeling (artificial neural networks) and spatial uncertainty modeling (convolutional neural networks).