This archive contains three sets of trained LSTM ensemble models for surrogate-based thermal history prediction during laser powder bed fusion (PBF-LB/M) of 42CrMo4 steel. Each ensemble consists of five independently seeded models trained with a six-stage curriculum that incrementally expands the training data selection. The three ensembles differ only in LSTM architecture depth and width, enabling a systematic comparison of model complexity. The trained weights are consumed by the companion framework via its inference and testing entry points.

Full-text publication: Code-Repository: https://doi.org/10.35097/dg39f4p0wxqdnfxy Training-Validation-Testing Dataset: https://doi.org/10.35097/pmem1cb9gu1ck8xz

Contains 3 .zip files with the trained model ensembles and the torch.nn.Module class: - 01Layers16Cells_Ensemble.zip - 02Layers32Cells_Ensemble.zip - 04Layers64Cells_Ensemble.zip - recurrent_neuralnetworks.py

Loading a checkpoint manually: python import torch from recurrent_neuralnetworks import LSTMModelWithTeacherForcing model = LSTMModelWithTeacherForcing( num_features=12, # 11 inputs + 1 temperature feedback num_hidden=32, # match the run (16 / 32 / 64) num_layers=2, # match the run (1 / 2 / 4) num_labels=1, ) state = torch.load("path/to/model_best_....pth", map_location="cpu") model.load_state_dict(state) model.eval()