Stress Field Prediction Engine

Academic deep learning project for predicting scalar fields (stress/temperature) on arbitrary 2D meshes using Interpolated Multiresolution Convolutional Neural Networks. Uses a 6-layer multiresolution CNN with 20 filters.

Trained on 1600 combined Voronoi + Lattice geometries with 34,981 parameters. Achieves median R² of 0.925 (training) / 0.911 (testing) for stress, 0.99 for heat conduction. A fast alternative to finite element analysis.

Key Features

• 6-layer multiresolution CNN architecture
• Interpolation to arbitrary node positions
• Median R² of 0.925/0.911 (train/test) for stress
• R² of 0.99 for heat conduction prediction
• 34,981 trainable parameters
• 1600 training geometries (Voronoi + Lattice)
• Fast alternative to finite element analysis
• GPU acceleration (CUDA 12.1+)
• Trained model included (multi_model_6.pth)