Skip to content

Latest commit

 

History

History
26 lines (23 loc) · 2.53 KB

README.md

File metadata and controls

26 lines (23 loc) · 2.53 KB
Raw

Foundation Model for Chemical Process Modeling: Meta-Learning with Physics-Informed Adaptation

Zihao Wang, Zhe Wu
Paper: https://arxiv.org/abs/2405.11752

Requires: Python 3.11.3, Tensorflow Keras 2.13.0, Numpy, Sklearn, Pickle
File description:

  • Files in "Sine Wave Regression" folder are used to validate the performance of MAML and Reptile on modelling different sine waves.
  • Files in "Reactors" folder are used to model different types of reactors (i.e., including both normal data-driven method and physics-informed method), including continuous stirred-tank reactors (CSTRs), batch reactors (BRs), and plug flow reactors (PFRs).
  • Under "CSTR+Batch+PFR" folder:
    1. CSTR_Batch_PFR_Reptile.ipynb is used to generate the foundation model for all three types of reactors using Reptile, model_reptile_cstr_batch.sav is the corresponding trained foundation model.
    2. CSTR_Batch_PFR_Transfer.ipynb is used to generate the foundation model for all three types of reactors using transfer learning, model_transfer_cstr_batch.sav is the corresponding trained foundation model.
    3. CSTR_Fewshots.ipynb is used to fine-tune the foundation models using few-shot learning to adapt to the new chemical reactions in unseen CSTRs, while CSTR_FewShots_Collocation is used to investigate the relationship between few-shot performance in terms of testing MSE and the number of collocation points used in physics-informed-based modelling for CSTR-based reactions.
    4. Batch_Fewshots.ipynb is used to fine-tune the foundation models using few-shot learning to adapt to the new chemical reactions in unseen BRs, while Batch_FewShots_Collocation is used to investigate the relationship between few-shot performance in terms of testing MSE and the number of collocation points used in physics-informed-based modelling for BR-based reactions.
    5. PFR_Fewshots.ipynb is used to fine-tune the foundation models using few-shot learning to adapt to the new chemical reactions in unseen PFRs, while PFR_FewShots_Collocation is used to investigate the relationship between few-shot performance in terms of testing MSE and the number of collocation points used in physics-informed-based modelling for PFR-based reactions.

Citation

If you find our work relevant to your research, please cite:

@article{wang2024foundation,
  title={Foundation Model for Chemical Process Modeling: Meta-Learning with Physics-Informed Adaptation},
  author={Wang, Zihao and Wu, Zhe},
  journal={arXiv preprint arXiv:2405.11752},
  year={2024}
}