Explore machine learning approaches for predicting seismic, acoustic, and atmospheric data in this 41-minute conference talk from the Fields Institute. Delve into enhanced K-Nearest Neighbors (KNN) techniques and their applications in atmospheric and material fracture studies. Learn about the importance of uncertainty quantification in machine learning and various performance metrics for evaluating predictions. Examine multivariate neural networks and heteroscedastic training loss functions for high-strain brittle fracture analysis. Gain insights from M. Giselle Fernández-Godino of Lawrence Livermore National Laboratory on cutting-edge methods for controlling error and improving efficiency in numerical models across diverse scientific domains.
Machine Learning Approaches for Predicting Seismic, Acoustic and Atmospheric Data
Overview
Syllabus
Machine Learning Approaches for Atmospheric and Material Fracture Applications and their Uncertainty Quantification
My background
Why Machine Learning (ML) Approaches?
Why Uncertainty Quantification (UQ) in ML?
Outline for this presentation
Problem of Interest and Motivation
Description of the Data
Overview of K-Nearest Neighbors Approach
KNN Graphical Examples
Our Enhanced KNN-based Approach
The KNN-based Prediction
Performance Metrics Options the number of neighbors used
Figure of Merit in the Space (FMS)
Normalized Root Mean Squared Error (NRMSE)
Fraction of Data (FAC2)
Fractional Bias
Coefficient of Determination, Slope and Intercept
Summary: KNN Approach
Summary: Performance Metrics
Uncertainty Bounds for Multivariate Machine Learning Predictions on High-Strain Brittle Fracture [1]
Multivariate Neural Networks Model
Heteroscedastic Training Loss Function
Multivariate Heteroscedastic Approach
Taught by
Fields Institute
