Extended Physics-Informed DeepONet (XPIDON) for Autoclave Composite Curing

Code for Extended Physics-Informed DeepONet (XPIDON) — accepted at the ICLR 2025 AI4Mat Workshop.

This work has been developed by Quantiphi’s PhiLabs.

This repository implements an Extended Physics-Informed DeepONet (XPIDON) model for simulating autoclave composite curing processes. It features a nonlinear decoder, subdomain-specific input normalization, and an adaptive and automated domain decomposition strategy.

Overview

The project addresses the challenge of long-duration transient simulations by decomposing the time domain into smaller subdomains. The XPIDON model is trained sequentially on these subdomains. An adaptive splitting mechanism refines subdomains if training losses exceed a tolerance, enabling automated refinement where the model struggles most.

Model Architecture

The XPIDON model is based on the DeepONet architecture, which learns operators mapping input functions to output functions. It comprises a Branch Network (for input functions), a Trunk Network (for spatio-temporal coordinates), and a Nonlinear Decoder. The use of a nonlinear decoder enhances the model's ability to capture complex relationships, while separate input normalization handle spectral bais.

a) Schematic representation of the composite-tool system inside an autoclave, including local coordinates x₁ and x₂. b) Architecture of the proposed sub-XPIDON model for predicting part temperature Gᵀᶜ. The same architecture is utilized for other output variables, including degree of cure Gᵅ and tool temperature Gᵀᵗ. c) Illustration of the proposed XPIDON framework with a nonlinear decoder and domain decomposition, designed for thermochemical analysis during the composite curing process.

Getting Started

1. Clone the repository:

   git clone https://github.com/Quantiphi-INC/XPIDON-Composites.git
   cd XPIDON-Composites

2. Install dependencies:

   pip install -r requirements.txt

   (Ensure JAX is installed correctly for your CPU/GPU setup as per JAX documentation.)
3. Prepare exp_params.json, phy_params.json and train_params.json in the root directory.
   • phy_params.json: Modify physical material properties (e.g., thermal conductivities, cure kinetics constants) here.
   • exp_params.json: Adjust experimental parameters such as the range of input functions (ramp rates, hold temperatures, durations, heat transfer coefficients, tool thickness), total process time, and initial conditions.
   • train_params.json: Configure training parameters, such as the number of collocation points, batch sizes, and weights assigned to each loss component.
4. Run the training:

   python main_run.py

   The main_run.py file orchestrates the training process. You can modify hyperparameters like network layer sizes (branch_layers, trunk_layers, nomad_layers_T), initial number of subdomains (init_sub_domain), and loss tolerance (Tolerance_level) directly within this file before running.

Notes

This codebase can also be adapted for our paper:
An advanced physics-informed neural operator for comprehensive design optimization of highly-nonlinear systems: An aerospace composites processing case study
(Requires removing the domain decomposition module and making a few modifications.)

Citation

If you find this work useful, please cite our paper:

@article{@article{PATEL2025112935,
title = {An extended physics-informed neural operator for accelerated design optimization in composites autoclave processing},
journal = {Composites Part B: Engineering},
pages = {112935},
year = {2025},
issn = {1359-8368},
doi = {https://doi.org/10.1016/j.compositesb.2025.112935},
url = {https://www.sciencedirect.com/science/article/pii/S1359836825008418},
author = {Janak M. Patel and Milad Ramezankhani and Anirudh Deodhar and Dagnachew Birru}}
}