A deep learning framework for causal shape transformation.

Area of Science:

• Engineering
• Computer Science
• Physics

Background:

• Recurrent Neural Networks (RNNs) and Long Short-term Memory (LSTM) networks excel at sequential data but are limited in visible domain dependencies.
• Predicting step-wise transformations dependent solely on previous visible states with a known end-state requires alternative architectures.
• Existing deep learning applications in inverse mapping for physical problems are limited, especially for high-dimensional challenges.

Purpose of the Study:

• To propose a hybrid Convolutional Neural Network (CNN) and Stacked Autoencoder (SAE) architecture for learning sequential causal actions.
• To enable non-linear transformations of input visual patterns to target patterns within the visible domain.
• To demonstrate the model's efficacy in solving a real-world, high-dimensional inverse problem in fluid physics.

Main Methods:

• Developed a hybrid CNN-SAE deep learning architecture.
• Trained the model to learn a sequence of causal actions for visual pattern transformation.
• Applied the model to a microfluidic flow sculpting problem, addressing a one-to-many inverse mapping challenge.

Main Results:

• Successfully learned a sequence of causal actions for non-linear transformation of visual patterns.
• Demonstrated the model's practicality and effectiveness in a real-world fluid physics engineering problem.
• Solved a high-dimensional inverse mapping problem where deep learning solutions are rarely explored.

Conclusions:

• The hybrid CNN-SAE model offers a viable solution for high-dimensional physical problems requiring step-wise transformations.
• This work highlights the potential of deep learning for inverse mapping in complex scientific domains.
• The findings may spur advancements in material sciences, medical biology, and other fields involving multistep topological transformations.