Predicting wind-driven spatial deposition through simulated color images using deep autoencoders

Area of Science:

• Geosciences
• Computational Physics
• Machine Learning

Background:

• Traditional scientific modeling relies on slow, iterative processes of observation and hypothesis testing.
• Machine learning offers a new paradigm for scientific discovery by learning directly from data, including visual information.
• Wind-driven spatial patterns, such as dune formation and pollution plumes, are complex phenomena often studied through physical models.

Purpose of the Study:

• To explore the use of deep convolutional neural network-based autoencoders for analyzing wind-driven spatial patterns.
• To reduce the dimensionality of image data representing these patterns for more efficient modeling.
• To develop a predictive model linking geographic and meteorological data to encoded spatial patterns.

Main Methods:

• Utilized computer simulations to generate RGB images of spatial deposition patterns.
• Employed deep convolutional neural network autoencoders to learn and compress the dimensionality of these image datasets.
• Trained fully connected neural networks to predict encoded spatial patterns from scalar input quantities.
• Reconstructed full spatial patterns using the decoder component of the autoencoder.

Main Results:

• Achieved significant data dimensionality reduction, compressing image data to 0.02% of its original size.
• The predictive model demonstrated high performance with a normalized root mean squared error of 8% on test data.
• Evaluated model accuracy using a figure of merit in space of 94% and a precision-recall area under the curve of 0.93.

Conclusions:

• Deep learning models, specifically autoencoders, can effectively learn and represent complex physics from image data.
• Dimensionality reduction using encoders facilitates the training of accurate predictive models for geoscientific phenomena.
• This image-based approach offers a powerful and efficient alternative to traditional methods for understanding wind-driven spatial patterns.