Side-by-side · baseline vs modern

Compare each predictor before you trust it.

Run the legacy baseline and the modern upgrade on the same input data, then judge by honest cross-validated metrics — not screenshots.

Cogging ENE predictor

MLP (legacy) vs Gradient Boosting (modern). Upload ONE Excel and both models train on it — fair side-by-side. Leave the file picker empty to use the bundled sample (48-row NSMLab dataset).

Sign in to upload your own files

File upload is part of the personal cabinet. In demo mode the platform runs every program on bundled sample data — sign up free to upload your own Excel / .dat / model files and save the results.

MLP baseline

MLP results will appear here

Single 80/20 train/test split with data-augmented training. Metrics live inside the generated image (Actual vs Predicted scatter, MAE curve).

Gradient Boosting (modern)

GB results will appear here

Processing map

Legacy: Prasad on the sparse 4×4 grid, interpolated by cubic griddata. Modern: a PINN learns the flow-stress surface from the same 16 curves and derives a smooth 50×50 η/ξ field via autodiff. Upload one Excel and both run on it — otherwise the bundled AISI 4340 sample is used.

Sign in to upload your own files

File upload is part of the personal cabinet. In demo mode the platform runs every program on bundled sample data — sign up free to upload your own Excel / .dat / model files and save the results.

Baseline · Prasad + griddata

Baseline plot will appear here

Discrete (T, ε̇) values interpolated globally. Reasonable inside the 4×4 grid but extrapolation/contours can be noisy.

PINN surrogate (modern)

PINN plot will appear here

3D Preform U-Net

Live inference comparison requires a trained Attention U-Net .h5 (offline, GPU). Until that's available, the comparison below is an architecture & capability summary.

Baseline U-Net

Trainable params~366 K

Input/Output128³ · single channel · sigmoid

Loss / metricweighted_folding_loss + dice

Standard encoder/decoder + skip concat
Loaded by giveStlModelBase64._load_or_cache_model
First inference 20–30 s, cached ~5 s

Attention U-Net (modern)

Trainable params~371 K

Input/Output128³ · single channel · sigmoid

Loss / metricweighted_folding_loss + dice

Attention gates on every skip connection (Oktay 2018)
Suppresses background voxels, focuses on preform shape
Drop-in compatible: same custom_objects, same I/O signature
Train offline (GPU/Colab) → drop the .h5 into sample_data/, no code change needed

How to train: on a GPU machine, run python -m threedlogic.attention_unet --inputs X/ --targets Y/ --out preform_attention.h5, then replace sample_data/unet_model.h5 with the new file. The web container will pick it up automatically.